sh: propagate r_clk
[linux-2.6] / 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 total_swap_pages;
39 static int swap_overflow;
40 static int least_priority;
41
42 static const char Bad_file[] = "Bad swap file entry ";
43 static const char Unused_file[] = "Unused swap file entry ";
44 static const char Bad_offset[] = "Bad swap offset entry ";
45 static const char Unused_offset[] = "Unused swap offset entry ";
46
47 static struct swap_list_t swap_list = {-1, -1};
48
49 static struct swap_info_struct swap_info[MAX_SWAPFILES];
50
51 static DEFINE_MUTEX(swapon_mutex);
52
53 /*
54  * We need this because the bdev->unplug_fn can sleep and we cannot
55  * hold swap_lock while calling the unplug_fn. And swap_lock
56  * cannot be turned into a mutex.
57  */
58 static DECLARE_RWSEM(swap_unplug_sem);
59
60 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
61 {
62         swp_entry_t entry;
63
64         down_read(&swap_unplug_sem);
65         entry.val = page_private(page);
66         if (PageSwapCache(page)) {
67                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
68                 struct backing_dev_info *bdi;
69
70                 /*
71                  * If the page is removed from swapcache from under us (with a
72                  * racy try_to_unuse/swapoff) we need an additional reference
73                  * count to avoid reading garbage from page_private(page) above.
74                  * If the WARN_ON triggers during a swapoff it maybe the race
75                  * condition and it's harmless. However if it triggers without
76                  * swapoff it signals a problem.
77                  */
78                 WARN_ON(page_count(page) <= 1);
79
80                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
81                 blk_run_backing_dev(bdi, page);
82         }
83         up_read(&swap_unplug_sem);
84 }
85
86 #define SWAPFILE_CLUSTER        256
87 #define LATENCY_LIMIT           256
88
89 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
90 {
91         unsigned long offset, last_in_cluster;
92         int latency_ration = LATENCY_LIMIT;
93
94         /* 
95          * We try to cluster swap pages by allocating them sequentially
96          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
97          * way, however, we resort to first-free allocation, starting
98          * a new cluster.  This prevents us from scattering swap pages
99          * all over the entire swap partition, so that we reduce
100          * overall disk seek times between swap pages.  -- sct
101          * But we do now try to find an empty cluster.  -Andrea
102          */
103
104         si->flags += SWP_SCANNING;
105         if (unlikely(!si->cluster_nr)) {
106                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
107                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER)
108                         goto lowest;
109                 spin_unlock(&swap_lock);
110
111                 offset = si->lowest_bit;
112                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
113
114                 /* Locate the first empty (unaligned) cluster */
115                 for (; last_in_cluster <= si->highest_bit; offset++) {
116                         if (si->swap_map[offset])
117                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
118                         else if (offset == last_in_cluster) {
119                                 spin_lock(&swap_lock);
120                                 si->cluster_next = offset-SWAPFILE_CLUSTER+1;
121                                 goto cluster;
122                         }
123                         if (unlikely(--latency_ration < 0)) {
124                                 cond_resched();
125                                 latency_ration = LATENCY_LIMIT;
126                         }
127                 }
128                 spin_lock(&swap_lock);
129                 goto lowest;
130         }
131
132         si->cluster_nr--;
133 cluster:
134         offset = si->cluster_next;
135         if (offset > si->highest_bit)
136 lowest:         offset = si->lowest_bit;
137 checks: if (!(si->flags & SWP_WRITEOK))
138                 goto no_page;
139         if (!si->highest_bit)
140                 goto no_page;
141         if (!si->swap_map[offset]) {
142                 if (offset == si->lowest_bit)
143                         si->lowest_bit++;
144                 if (offset == si->highest_bit)
145                         si->highest_bit--;
146                 si->inuse_pages++;
147                 if (si->inuse_pages == si->pages) {
148                         si->lowest_bit = si->max;
149                         si->highest_bit = 0;
150                 }
151                 si->swap_map[offset] = 1;
152                 si->cluster_next = offset + 1;
153                 si->flags -= SWP_SCANNING;
154                 return offset;
155         }
156
157         spin_unlock(&swap_lock);
158         while (++offset <= si->highest_bit) {
159                 if (!si->swap_map[offset]) {
160                         spin_lock(&swap_lock);
161                         goto checks;
162                 }
163                 if (unlikely(--latency_ration < 0)) {
164                         cond_resched();
165                         latency_ration = LATENCY_LIMIT;
166                 }
167         }
168         spin_lock(&swap_lock);
169         goto lowest;
170
171 no_page:
172         si->flags -= SWP_SCANNING;
173         return 0;
174 }
175
176 swp_entry_t get_swap_page(void)
177 {
178         struct swap_info_struct *si;
179         pgoff_t offset;
180         int type, next;
181         int wrapped = 0;
182
183         spin_lock(&swap_lock);
184         if (nr_swap_pages <= 0)
185                 goto noswap;
186         nr_swap_pages--;
187
188         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
189                 si = swap_info + type;
190                 next = si->next;
191                 if (next < 0 ||
192                     (!wrapped && si->prio != swap_info[next].prio)) {
193                         next = swap_list.head;
194                         wrapped++;
195                 }
196
197                 if (!si->highest_bit)
198                         continue;
199                 if (!(si->flags & SWP_WRITEOK))
200                         continue;
201
202                 swap_list.next = next;
203                 offset = scan_swap_map(si);
204                 if (offset) {
205                         spin_unlock(&swap_lock);
206                         return swp_entry(type, offset);
207                 }
208                 next = swap_list.next;
209         }
210
211         nr_swap_pages++;
212 noswap:
213         spin_unlock(&swap_lock);
214         return (swp_entry_t) {0};
215 }
216
217 swp_entry_t get_swap_page_of_type(int type)
218 {
219         struct swap_info_struct *si;
220         pgoff_t offset;
221
222         spin_lock(&swap_lock);
223         si = swap_info + type;
224         if (si->flags & SWP_WRITEOK) {
225                 nr_swap_pages--;
226                 offset = scan_swap_map(si);
227                 if (offset) {
228                         spin_unlock(&swap_lock);
229                         return swp_entry(type, offset);
230                 }
231                 nr_swap_pages++;
232         }
233         spin_unlock(&swap_lock);
234         return (swp_entry_t) {0};
235 }
236
237 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
238 {
239         struct swap_info_struct * p;
240         unsigned long offset, type;
241
242         if (!entry.val)
243                 goto out;
244         type = swp_type(entry);
245         if (type >= nr_swapfiles)
246                 goto bad_nofile;
247         p = & swap_info[type];
248         if (!(p->flags & SWP_USED))
249                 goto bad_device;
250         offset = swp_offset(entry);
251         if (offset >= p->max)
252                 goto bad_offset;
253         if (!p->swap_map[offset])
254                 goto bad_free;
255         spin_lock(&swap_lock);
256         return p;
257
258 bad_free:
259         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
260         goto out;
261 bad_offset:
262         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
263         goto out;
264 bad_device:
265         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
266         goto out;
267 bad_nofile:
268         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
269 out:
270         return NULL;
271 }       
272
273 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
274 {
275         int count = p->swap_map[offset];
276
277         if (count < SWAP_MAP_MAX) {
278                 count--;
279                 p->swap_map[offset] = count;
280                 if (!count) {
281                         if (offset < p->lowest_bit)
282                                 p->lowest_bit = offset;
283                         if (offset > p->highest_bit)
284                                 p->highest_bit = offset;
285                         if (p->prio > swap_info[swap_list.next].prio)
286                                 swap_list.next = p - swap_info;
287                         nr_swap_pages++;
288                         p->inuse_pages--;
289                 }
290         }
291         return count;
292 }
293
294 /*
295  * Caller has made sure that the swapdevice corresponding to entry
296  * is still around or has not been recycled.
297  */
298 void swap_free(swp_entry_t entry)
299 {
300         struct swap_info_struct * p;
301
302         p = swap_info_get(entry);
303         if (p) {
304                 swap_entry_free(p, swp_offset(entry));
305                 spin_unlock(&swap_lock);
306         }
307 }
308
309 /*
310  * How many references to page are currently swapped out?
311  */
312 static inline int page_swapcount(struct page *page)
313 {
314         int count = 0;
315         struct swap_info_struct *p;
316         swp_entry_t entry;
317
318         entry.val = page_private(page);
319         p = swap_info_get(entry);
320         if (p) {
321                 /* Subtract the 1 for the swap cache itself */
322                 count = p->swap_map[swp_offset(entry)] - 1;
323                 spin_unlock(&swap_lock);
324         }
325         return count;
326 }
327
328 /*
329  * We can use this swap cache entry directly
330  * if there are no other references to it.
331  */
332 int can_share_swap_page(struct page *page)
333 {
334         int count;
335
336         BUG_ON(!PageLocked(page));
337         count = page_mapcount(page);
338         if (count <= 1 && PageSwapCache(page))
339                 count += page_swapcount(page);
340         return count == 1;
341 }
342
343 /*
344  * Work out if there are any other processes sharing this
345  * swap cache page. Free it if you can. Return success.
346  */
347 static int remove_exclusive_swap_page_count(struct page *page, int count)
348 {
349         int retval;
350         struct swap_info_struct * p;
351         swp_entry_t entry;
352
353         BUG_ON(PagePrivate(page));
354         BUG_ON(!PageLocked(page));
355
356         if (!PageSwapCache(page))
357                 return 0;
358         if (PageWriteback(page))
359                 return 0;
360         if (page_count(page) != count) /* us + cache + ptes */
361                 return 0;
362
363         entry.val = page_private(page);
364         p = swap_info_get(entry);
365         if (!p)
366                 return 0;
367
368         /* Is the only swap cache user the cache itself? */
369         retval = 0;
370         if (p->swap_map[swp_offset(entry)] == 1) {
371                 /* Recheck the page count with the swapcache lock held.. */
372                 spin_lock_irq(&swapper_space.tree_lock);
373                 if ((page_count(page) == count) && !PageWriteback(page)) {
374                         __delete_from_swap_cache(page);
375                         SetPageDirty(page);
376                         retval = 1;
377                 }
378                 spin_unlock_irq(&swapper_space.tree_lock);
379         }
380         spin_unlock(&swap_lock);
381
382         if (retval) {
383                 swap_free(entry);
384                 page_cache_release(page);
385         }
386
387         return retval;
388 }
389
390 /*
391  * Most of the time the page should have two references: one for the
392  * process and one for the swap cache.
393  */
394 int remove_exclusive_swap_page(struct page *page)
395 {
396         return remove_exclusive_swap_page_count(page, 2);
397 }
398
399 /*
400  * The pageout code holds an extra reference to the page.  That raises
401  * the reference count to test for to 2 for a page that is only in the
402  * swap cache plus 1 for each process that maps the page.
403  */
404 int remove_exclusive_swap_page_ref(struct page *page)
405 {
406         return remove_exclusive_swap_page_count(page, 2 + page_mapcount(page));
407 }
408
409 /*
410  * Free the swap entry like above, but also try to
411  * free the page cache entry if it is the last user.
412  */
413 void free_swap_and_cache(swp_entry_t entry)
414 {
415         struct swap_info_struct * p;
416         struct page *page = NULL;
417
418         if (is_migration_entry(entry))
419                 return;
420
421         p = swap_info_get(entry);
422         if (p) {
423                 if (swap_entry_free(p, swp_offset(entry)) == 1) {
424                         page = find_get_page(&swapper_space, entry.val);
425                         if (page && !trylock_page(page)) {
426                                 page_cache_release(page);
427                                 page = NULL;
428                         }
429                 }
430                 spin_unlock(&swap_lock);
431         }
432         if (page) {
433                 int one_user;
434
435                 BUG_ON(PagePrivate(page));
436                 one_user = (page_count(page) == 2);
437                 /* Only cache user (+us), or swap space full? Free it! */
438                 /* Also recheck PageSwapCache after page is locked (above) */
439                 if (PageSwapCache(page) && !PageWriteback(page) &&
440                                         (one_user || vm_swap_full())) {
441                         delete_from_swap_cache(page);
442                         SetPageDirty(page);
443                 }
444                 unlock_page(page);
445                 page_cache_release(page);
446         }
447 }
448
449 #ifdef CONFIG_HIBERNATION
450 /*
451  * Find the swap type that corresponds to given device (if any).
452  *
453  * @offset - number of the PAGE_SIZE-sized block of the device, starting
454  * from 0, in which the swap header is expected to be located.
455  *
456  * This is needed for the suspend to disk (aka swsusp).
457  */
458 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
459 {
460         struct block_device *bdev = NULL;
461         int i;
462
463         if (device)
464                 bdev = bdget(device);
465
466         spin_lock(&swap_lock);
467         for (i = 0; i < nr_swapfiles; i++) {
468                 struct swap_info_struct *sis = swap_info + i;
469
470                 if (!(sis->flags & SWP_WRITEOK))
471                         continue;
472
473                 if (!bdev) {
474                         if (bdev_p)
475                                 *bdev_p = sis->bdev;
476
477                         spin_unlock(&swap_lock);
478                         return i;
479                 }
480                 if (bdev == sis->bdev) {
481                         struct swap_extent *se;
482
483                         se = list_entry(sis->extent_list.next,
484                                         struct swap_extent, list);
485                         if (se->start_block == offset) {
486                                 if (bdev_p)
487                                         *bdev_p = sis->bdev;
488
489                                 spin_unlock(&swap_lock);
490                                 bdput(bdev);
491                                 return i;
492                         }
493                 }
494         }
495         spin_unlock(&swap_lock);
496         if (bdev)
497                 bdput(bdev);
498
499         return -ENODEV;
500 }
501
502 /*
503  * Return either the total number of swap pages of given type, or the number
504  * of free pages of that type (depending on @free)
505  *
506  * This is needed for software suspend
507  */
508 unsigned int count_swap_pages(int type, int free)
509 {
510         unsigned int n = 0;
511
512         if (type < nr_swapfiles) {
513                 spin_lock(&swap_lock);
514                 if (swap_info[type].flags & SWP_WRITEOK) {
515                         n = swap_info[type].pages;
516                         if (free)
517                                 n -= swap_info[type].inuse_pages;
518                 }
519                 spin_unlock(&swap_lock);
520         }
521         return n;
522 }
523 #endif
524
525 /*
526  * No need to decide whether this PTE shares the swap entry with others,
527  * just let do_wp_page work it out if a write is requested later - to
528  * force COW, vm_page_prot omits write permission from any private vma.
529  */
530 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
531                 unsigned long addr, swp_entry_t entry, struct page *page)
532 {
533         spinlock_t *ptl;
534         pte_t *pte;
535         int ret = 1;
536
537         if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
538                 ret = -ENOMEM;
539
540         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
541         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
542                 if (ret > 0)
543                         mem_cgroup_uncharge_page(page);
544                 ret = 0;
545                 goto out;
546         }
547
548         inc_mm_counter(vma->vm_mm, anon_rss);
549         get_page(page);
550         set_pte_at(vma->vm_mm, addr, pte,
551                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
552         page_add_anon_rmap(page, vma, addr);
553         swap_free(entry);
554         /*
555          * Move the page to the active list so it is not
556          * immediately swapped out again after swapon.
557          */
558         activate_page(page);
559 out:
560         pte_unmap_unlock(pte, ptl);
561         return ret;
562 }
563
564 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
565                                 unsigned long addr, unsigned long end,
566                                 swp_entry_t entry, struct page *page)
567 {
568         pte_t swp_pte = swp_entry_to_pte(entry);
569         pte_t *pte;
570         int ret = 0;
571
572         /*
573          * We don't actually need pte lock while scanning for swp_pte: since
574          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
575          * page table while we're scanning; though it could get zapped, and on
576          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
577          * of unmatched parts which look like swp_pte, so unuse_pte must
578          * recheck under pte lock.  Scanning without pte lock lets it be
579          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
580          */
581         pte = pte_offset_map(pmd, addr);
582         do {
583                 /*
584                  * swapoff spends a _lot_ of time in this loop!
585                  * Test inline before going to call unuse_pte.
586                  */
587                 if (unlikely(pte_same(*pte, swp_pte))) {
588                         pte_unmap(pte);
589                         ret = unuse_pte(vma, pmd, addr, entry, page);
590                         if (ret)
591                                 goto out;
592                         pte = pte_offset_map(pmd, addr);
593                 }
594         } while (pte++, addr += PAGE_SIZE, addr != end);
595         pte_unmap(pte - 1);
596 out:
597         return ret;
598 }
599
600 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
601                                 unsigned long addr, unsigned long end,
602                                 swp_entry_t entry, struct page *page)
603 {
604         pmd_t *pmd;
605         unsigned long next;
606         int ret;
607
608         pmd = pmd_offset(pud, addr);
609         do {
610                 next = pmd_addr_end(addr, end);
611                 if (pmd_none_or_clear_bad(pmd))
612                         continue;
613                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
614                 if (ret)
615                         return ret;
616         } while (pmd++, addr = next, addr != end);
617         return 0;
618 }
619
620 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
621                                 unsigned long addr, unsigned long end,
622                                 swp_entry_t entry, struct page *page)
623 {
624         pud_t *pud;
625         unsigned long next;
626         int ret;
627
628         pud = pud_offset(pgd, addr);
629         do {
630                 next = pud_addr_end(addr, end);
631                 if (pud_none_or_clear_bad(pud))
632                         continue;
633                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
634                 if (ret)
635                         return ret;
636         } while (pud++, addr = next, addr != end);
637         return 0;
638 }
639
640 static int unuse_vma(struct vm_area_struct *vma,
641                                 swp_entry_t entry, struct page *page)
642 {
643         pgd_t *pgd;
644         unsigned long addr, end, next;
645         int ret;
646
647         if (page->mapping) {
648                 addr = page_address_in_vma(page, vma);
649                 if (addr == -EFAULT)
650                         return 0;
651                 else
652                         end = addr + PAGE_SIZE;
653         } else {
654                 addr = vma->vm_start;
655                 end = vma->vm_end;
656         }
657
658         pgd = pgd_offset(vma->vm_mm, addr);
659         do {
660                 next = pgd_addr_end(addr, end);
661                 if (pgd_none_or_clear_bad(pgd))
662                         continue;
663                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
664                 if (ret)
665                         return ret;
666         } while (pgd++, addr = next, addr != end);
667         return 0;
668 }
669
670 static int unuse_mm(struct mm_struct *mm,
671                                 swp_entry_t entry, struct page *page)
672 {
673         struct vm_area_struct *vma;
674         int ret = 0;
675
676         if (!down_read_trylock(&mm->mmap_sem)) {
677                 /*
678                  * Activate page so shrink_inactive_list is unlikely to unmap
679                  * its ptes while lock is dropped, so swapoff can make progress.
680                  */
681                 activate_page(page);
682                 unlock_page(page);
683                 down_read(&mm->mmap_sem);
684                 lock_page(page);
685         }
686         for (vma = mm->mmap; vma; vma = vma->vm_next) {
687                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
688                         break;
689         }
690         up_read(&mm->mmap_sem);
691         return (ret < 0)? ret: 0;
692 }
693
694 /*
695  * Scan swap_map from current position to next entry still in use.
696  * Recycle to start on reaching the end, returning 0 when empty.
697  */
698 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
699                                         unsigned int prev)
700 {
701         unsigned int max = si->max;
702         unsigned int i = prev;
703         int count;
704
705         /*
706          * No need for swap_lock here: we're just looking
707          * for whether an entry is in use, not modifying it; false
708          * hits are okay, and sys_swapoff() has already prevented new
709          * allocations from this area (while holding swap_lock).
710          */
711         for (;;) {
712                 if (++i >= max) {
713                         if (!prev) {
714                                 i = 0;
715                                 break;
716                         }
717                         /*
718                          * No entries in use at top of swap_map,
719                          * loop back to start and recheck there.
720                          */
721                         max = prev + 1;
722                         prev = 0;
723                         i = 1;
724                 }
725                 count = si->swap_map[i];
726                 if (count && count != SWAP_MAP_BAD)
727                         break;
728         }
729         return i;
730 }
731
732 /*
733  * We completely avoid races by reading each swap page in advance,
734  * and then search for the process using it.  All the necessary
735  * page table adjustments can then be made atomically.
736  */
737 static int try_to_unuse(unsigned int type)
738 {
739         struct swap_info_struct * si = &swap_info[type];
740         struct mm_struct *start_mm;
741         unsigned short *swap_map;
742         unsigned short swcount;
743         struct page *page;
744         swp_entry_t entry;
745         unsigned int i = 0;
746         int retval = 0;
747         int reset_overflow = 0;
748         int shmem;
749
750         /*
751          * When searching mms for an entry, a good strategy is to
752          * start at the first mm we freed the previous entry from
753          * (though actually we don't notice whether we or coincidence
754          * freed the entry).  Initialize this start_mm with a hold.
755          *
756          * A simpler strategy would be to start at the last mm we
757          * freed the previous entry from; but that would take less
758          * advantage of mmlist ordering, which clusters forked mms
759          * together, child after parent.  If we race with dup_mmap(), we
760          * prefer to resolve parent before child, lest we miss entries
761          * duplicated after we scanned child: using last mm would invert
762          * that.  Though it's only a serious concern when an overflowed
763          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
764          */
765         start_mm = &init_mm;
766         atomic_inc(&init_mm.mm_users);
767
768         /*
769          * Keep on scanning until all entries have gone.  Usually,
770          * one pass through swap_map is enough, but not necessarily:
771          * there are races when an instance of an entry might be missed.
772          */
773         while ((i = find_next_to_unuse(si, i)) != 0) {
774                 if (signal_pending(current)) {
775                         retval = -EINTR;
776                         break;
777                 }
778
779                 /* 
780                  * Get a page for the entry, using the existing swap
781                  * cache page if there is one.  Otherwise, get a clean
782                  * page and read the swap into it. 
783                  */
784                 swap_map = &si->swap_map[i];
785                 entry = swp_entry(type, i);
786                 page = read_swap_cache_async(entry,
787                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
788                 if (!page) {
789                         /*
790                          * Either swap_duplicate() failed because entry
791                          * has been freed independently, and will not be
792                          * reused since sys_swapoff() already disabled
793                          * allocation from here, or alloc_page() failed.
794                          */
795                         if (!*swap_map)
796                                 continue;
797                         retval = -ENOMEM;
798                         break;
799                 }
800
801                 /*
802                  * Don't hold on to start_mm if it looks like exiting.
803                  */
804                 if (atomic_read(&start_mm->mm_users) == 1) {
805                         mmput(start_mm);
806                         start_mm = &init_mm;
807                         atomic_inc(&init_mm.mm_users);
808                 }
809
810                 /*
811                  * Wait for and lock page.  When do_swap_page races with
812                  * try_to_unuse, do_swap_page can handle the fault much
813                  * faster than try_to_unuse can locate the entry.  This
814                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
815                  * defer to do_swap_page in such a case - in some tests,
816                  * do_swap_page and try_to_unuse repeatedly compete.
817                  */
818                 wait_on_page_locked(page);
819                 wait_on_page_writeback(page);
820                 lock_page(page);
821                 wait_on_page_writeback(page);
822
823                 /*
824                  * Remove all references to entry.
825                  * Whenever we reach init_mm, there's no address space
826                  * to search, but use it as a reminder to search shmem.
827                  */
828                 shmem = 0;
829                 swcount = *swap_map;
830                 if (swcount > 1) {
831                         if (start_mm == &init_mm)
832                                 shmem = shmem_unuse(entry, page);
833                         else
834                                 retval = unuse_mm(start_mm, entry, page);
835                 }
836                 if (*swap_map > 1) {
837                         int set_start_mm = (*swap_map >= swcount);
838                         struct list_head *p = &start_mm->mmlist;
839                         struct mm_struct *new_start_mm = start_mm;
840                         struct mm_struct *prev_mm = start_mm;
841                         struct mm_struct *mm;
842
843                         atomic_inc(&new_start_mm->mm_users);
844                         atomic_inc(&prev_mm->mm_users);
845                         spin_lock(&mmlist_lock);
846                         while (*swap_map > 1 && !retval && !shmem &&
847                                         (p = p->next) != &start_mm->mmlist) {
848                                 mm = list_entry(p, struct mm_struct, mmlist);
849                                 if (!atomic_inc_not_zero(&mm->mm_users))
850                                         continue;
851                                 spin_unlock(&mmlist_lock);
852                                 mmput(prev_mm);
853                                 prev_mm = mm;
854
855                                 cond_resched();
856
857                                 swcount = *swap_map;
858                                 if (swcount <= 1)
859                                         ;
860                                 else if (mm == &init_mm) {
861                                         set_start_mm = 1;
862                                         shmem = shmem_unuse(entry, page);
863                                 } else
864                                         retval = unuse_mm(mm, entry, page);
865                                 if (set_start_mm && *swap_map < swcount) {
866                                         mmput(new_start_mm);
867                                         atomic_inc(&mm->mm_users);
868                                         new_start_mm = mm;
869                                         set_start_mm = 0;
870                                 }
871                                 spin_lock(&mmlist_lock);
872                         }
873                         spin_unlock(&mmlist_lock);
874                         mmput(prev_mm);
875                         mmput(start_mm);
876                         start_mm = new_start_mm;
877                 }
878                 if (shmem) {
879                         /* page has already been unlocked and released */
880                         if (shmem > 0)
881                                 continue;
882                         retval = shmem;
883                         break;
884                 }
885                 if (retval) {
886                         unlock_page(page);
887                         page_cache_release(page);
888                         break;
889                 }
890
891                 /*
892                  * How could swap count reach 0x7fff when the maximum
893                  * pid is 0x7fff, and there's no way to repeat a swap
894                  * page within an mm (except in shmem, where it's the
895                  * shared object which takes the reference count)?
896                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
897                  *
898                  * If that's wrong, then we should worry more about
899                  * exit_mmap() and do_munmap() cases described above:
900                  * we might be resetting SWAP_MAP_MAX too early here.
901                  * We know "Undead"s can happen, they're okay, so don't
902                  * report them; but do report if we reset SWAP_MAP_MAX.
903                  */
904                 if (*swap_map == SWAP_MAP_MAX) {
905                         spin_lock(&swap_lock);
906                         *swap_map = 1;
907                         spin_unlock(&swap_lock);
908                         reset_overflow = 1;
909                 }
910
911                 /*
912                  * If a reference remains (rare), we would like to leave
913                  * the page in the swap cache; but try_to_unmap could
914                  * then re-duplicate the entry once we drop page lock,
915                  * so we might loop indefinitely; also, that page could
916                  * not be swapped out to other storage meanwhile.  So:
917                  * delete from cache even if there's another reference,
918                  * after ensuring that the data has been saved to disk -
919                  * since if the reference remains (rarer), it will be
920                  * read from disk into another page.  Splitting into two
921                  * pages would be incorrect if swap supported "shared
922                  * private" pages, but they are handled by tmpfs files.
923                  */
924                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
925                         struct writeback_control wbc = {
926                                 .sync_mode = WB_SYNC_NONE,
927                         };
928
929                         swap_writepage(page, &wbc);
930                         lock_page(page);
931                         wait_on_page_writeback(page);
932                 }
933                 if (PageSwapCache(page))
934                         delete_from_swap_cache(page);
935
936                 /*
937                  * So we could skip searching mms once swap count went
938                  * to 1, we did not mark any present ptes as dirty: must
939                  * mark page dirty so shrink_page_list will preserve it.
940                  */
941                 SetPageDirty(page);
942                 unlock_page(page);
943                 page_cache_release(page);
944
945                 /*
946                  * Make sure that we aren't completely killing
947                  * interactive performance.
948                  */
949                 cond_resched();
950         }
951
952         mmput(start_mm);
953         if (reset_overflow) {
954                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
955                 swap_overflow = 0;
956         }
957         return retval;
958 }
959
960 /*
961  * After a successful try_to_unuse, if no swap is now in use, we know
962  * we can empty the mmlist.  swap_lock must be held on entry and exit.
963  * Note that mmlist_lock nests inside swap_lock, and an mm must be
964  * added to the mmlist just after page_duplicate - before would be racy.
965  */
966 static void drain_mmlist(void)
967 {
968         struct list_head *p, *next;
969         unsigned int i;
970
971         for (i = 0; i < nr_swapfiles; i++)
972                 if (swap_info[i].inuse_pages)
973                         return;
974         spin_lock(&mmlist_lock);
975         list_for_each_safe(p, next, &init_mm.mmlist)
976                 list_del_init(p);
977         spin_unlock(&mmlist_lock);
978 }
979
980 /*
981  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
982  * corresponds to page offset `offset'.
983  */
984 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
985 {
986         struct swap_extent *se = sis->curr_swap_extent;
987         struct swap_extent *start_se = se;
988
989         for ( ; ; ) {
990                 struct list_head *lh;
991
992                 if (se->start_page <= offset &&
993                                 offset < (se->start_page + se->nr_pages)) {
994                         return se->start_block + (offset - se->start_page);
995                 }
996                 lh = se->list.next;
997                 if (lh == &sis->extent_list)
998                         lh = lh->next;
999                 se = list_entry(lh, struct swap_extent, list);
1000                 sis->curr_swap_extent = se;
1001                 BUG_ON(se == start_se);         /* It *must* be present */
1002         }
1003 }
1004
1005 #ifdef CONFIG_HIBERNATION
1006 /*
1007  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1008  * corresponding to given index in swap_info (swap type).
1009  */
1010 sector_t swapdev_block(int swap_type, pgoff_t offset)
1011 {
1012         struct swap_info_struct *sis;
1013
1014         if (swap_type >= nr_swapfiles)
1015                 return 0;
1016
1017         sis = swap_info + swap_type;
1018         return (sis->flags & SWP_WRITEOK) ? map_swap_page(sis, offset) : 0;
1019 }
1020 #endif /* CONFIG_HIBERNATION */
1021
1022 /*
1023  * Free all of a swapdev's extent information
1024  */
1025 static void destroy_swap_extents(struct swap_info_struct *sis)
1026 {
1027         while (!list_empty(&sis->extent_list)) {
1028                 struct swap_extent *se;
1029
1030                 se = list_entry(sis->extent_list.next,
1031                                 struct swap_extent, list);
1032                 list_del(&se->list);
1033                 kfree(se);
1034         }
1035 }
1036
1037 /*
1038  * Add a block range (and the corresponding page range) into this swapdev's
1039  * extent list.  The extent list is kept sorted in page order.
1040  *
1041  * This function rather assumes that it is called in ascending page order.
1042  */
1043 static int
1044 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1045                 unsigned long nr_pages, sector_t start_block)
1046 {
1047         struct swap_extent *se;
1048         struct swap_extent *new_se;
1049         struct list_head *lh;
1050
1051         lh = sis->extent_list.prev;     /* The highest page extent */
1052         if (lh != &sis->extent_list) {
1053                 se = list_entry(lh, struct swap_extent, list);
1054                 BUG_ON(se->start_page + se->nr_pages != start_page);
1055                 if (se->start_block + se->nr_pages == start_block) {
1056                         /* Merge it */
1057                         se->nr_pages += nr_pages;
1058                         return 0;
1059                 }
1060         }
1061
1062         /*
1063          * No merge.  Insert a new extent, preserving ordering.
1064          */
1065         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1066         if (new_se == NULL)
1067                 return -ENOMEM;
1068         new_se->start_page = start_page;
1069         new_se->nr_pages = nr_pages;
1070         new_se->start_block = start_block;
1071
1072         list_add_tail(&new_se->list, &sis->extent_list);
1073         return 1;
1074 }
1075
1076 /*
1077  * A `swap extent' is a simple thing which maps a contiguous range of pages
1078  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1079  * is built at swapon time and is then used at swap_writepage/swap_readpage
1080  * time for locating where on disk a page belongs.
1081  *
1082  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1083  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1084  * swap files identically.
1085  *
1086  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1087  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1088  * swapfiles are handled *identically* after swapon time.
1089  *
1090  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1091  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1092  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1093  * requirements, they are simply tossed out - we will never use those blocks
1094  * for swapping.
1095  *
1096  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1097  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1098  * which will scribble on the fs.
1099  *
1100  * The amount of disk space which a single swap extent represents varies.
1101  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1102  * extents in the list.  To avoid much list walking, we cache the previous
1103  * search location in `curr_swap_extent', and start new searches from there.
1104  * This is extremely effective.  The average number of iterations in
1105  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1106  */
1107 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1108 {
1109         struct inode *inode;
1110         unsigned blocks_per_page;
1111         unsigned long page_no;
1112         unsigned blkbits;
1113         sector_t probe_block;
1114         sector_t last_block;
1115         sector_t lowest_block = -1;
1116         sector_t highest_block = 0;
1117         int nr_extents = 0;
1118         int ret;
1119
1120         inode = sis->swap_file->f_mapping->host;
1121         if (S_ISBLK(inode->i_mode)) {
1122                 ret = add_swap_extent(sis, 0, sis->max, 0);
1123                 *span = sis->pages;
1124                 goto done;
1125         }
1126
1127         blkbits = inode->i_blkbits;
1128         blocks_per_page = PAGE_SIZE >> blkbits;
1129
1130         /*
1131          * Map all the blocks into the extent list.  This code doesn't try
1132          * to be very smart.
1133          */
1134         probe_block = 0;
1135         page_no = 0;
1136         last_block = i_size_read(inode) >> blkbits;
1137         while ((probe_block + blocks_per_page) <= last_block &&
1138                         page_no < sis->max) {
1139                 unsigned block_in_page;
1140                 sector_t first_block;
1141
1142                 first_block = bmap(inode, probe_block);
1143                 if (first_block == 0)
1144                         goto bad_bmap;
1145
1146                 /*
1147                  * It must be PAGE_SIZE aligned on-disk
1148                  */
1149                 if (first_block & (blocks_per_page - 1)) {
1150                         probe_block++;
1151                         goto reprobe;
1152                 }
1153
1154                 for (block_in_page = 1; block_in_page < blocks_per_page;
1155                                         block_in_page++) {
1156                         sector_t block;
1157
1158                         block = bmap(inode, probe_block + block_in_page);
1159                         if (block == 0)
1160                                 goto bad_bmap;
1161                         if (block != first_block + block_in_page) {
1162                                 /* Discontiguity */
1163                                 probe_block++;
1164                                 goto reprobe;
1165                         }
1166                 }
1167
1168                 first_block >>= (PAGE_SHIFT - blkbits);
1169                 if (page_no) {  /* exclude the header page */
1170                         if (first_block < lowest_block)
1171                                 lowest_block = first_block;
1172                         if (first_block > highest_block)
1173                                 highest_block = first_block;
1174                 }
1175
1176                 /*
1177                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1178                  */
1179                 ret = add_swap_extent(sis, page_no, 1, first_block);
1180                 if (ret < 0)
1181                         goto out;
1182                 nr_extents += ret;
1183                 page_no++;
1184                 probe_block += blocks_per_page;
1185 reprobe:
1186                 continue;
1187         }
1188         ret = nr_extents;
1189         *span = 1 + highest_block - lowest_block;
1190         if (page_no == 0)
1191                 page_no = 1;    /* force Empty message */
1192         sis->max = page_no;
1193         sis->pages = page_no - 1;
1194         sis->highest_bit = page_no - 1;
1195 done:
1196         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1197                                         struct swap_extent, list);
1198         goto out;
1199 bad_bmap:
1200         printk(KERN_ERR "swapon: swapfile has holes\n");
1201         ret = -EINVAL;
1202 out:
1203         return ret;
1204 }
1205
1206 #if 0   /* We don't need this yet */
1207 #include <linux/backing-dev.h>
1208 int page_queue_congested(struct page *page)
1209 {
1210         struct backing_dev_info *bdi;
1211
1212         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1213
1214         if (PageSwapCache(page)) {
1215                 swp_entry_t entry = { .val = page_private(page) };
1216                 struct swap_info_struct *sis;
1217
1218                 sis = get_swap_info_struct(swp_type(entry));
1219                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1220         } else
1221                 bdi = page->mapping->backing_dev_info;
1222         return bdi_write_congested(bdi);
1223 }
1224 #endif
1225
1226 asmlinkage long sys_swapoff(const char __user * specialfile)
1227 {
1228         struct swap_info_struct * p = NULL;
1229         unsigned short *swap_map;
1230         struct file *swap_file, *victim;
1231         struct address_space *mapping;
1232         struct inode *inode;
1233         char * pathname;
1234         int i, type, prev;
1235         int err;
1236         
1237         if (!capable(CAP_SYS_ADMIN))
1238                 return -EPERM;
1239
1240         pathname = getname(specialfile);
1241         err = PTR_ERR(pathname);
1242         if (IS_ERR(pathname))
1243                 goto out;
1244
1245         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1246         putname(pathname);
1247         err = PTR_ERR(victim);
1248         if (IS_ERR(victim))
1249                 goto out;
1250
1251         mapping = victim->f_mapping;
1252         prev = -1;
1253         spin_lock(&swap_lock);
1254         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1255                 p = swap_info + type;
1256                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1257                         if (p->swap_file->f_mapping == mapping)
1258                                 break;
1259                 }
1260                 prev = type;
1261         }
1262         if (type < 0) {
1263                 err = -EINVAL;
1264                 spin_unlock(&swap_lock);
1265                 goto out_dput;
1266         }
1267         if (!security_vm_enough_memory(p->pages))
1268                 vm_unacct_memory(p->pages);
1269         else {
1270                 err = -ENOMEM;
1271                 spin_unlock(&swap_lock);
1272                 goto out_dput;
1273         }
1274         if (prev < 0) {
1275                 swap_list.head = p->next;
1276         } else {
1277                 swap_info[prev].next = p->next;
1278         }
1279         if (type == swap_list.next) {
1280                 /* just pick something that's safe... */
1281                 swap_list.next = swap_list.head;
1282         }
1283         if (p->prio < 0) {
1284                 for (i = p->next; i >= 0; i = swap_info[i].next)
1285                         swap_info[i].prio = p->prio--;
1286                 least_priority++;
1287         }
1288         nr_swap_pages -= p->pages;
1289         total_swap_pages -= p->pages;
1290         p->flags &= ~SWP_WRITEOK;
1291         spin_unlock(&swap_lock);
1292
1293         current->flags |= PF_SWAPOFF;
1294         err = try_to_unuse(type);
1295         current->flags &= ~PF_SWAPOFF;
1296
1297         if (err) {
1298                 /* re-insert swap space back into swap_list */
1299                 spin_lock(&swap_lock);
1300                 if (p->prio < 0)
1301                         p->prio = --least_priority;
1302                 prev = -1;
1303                 for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1304                         if (p->prio >= swap_info[i].prio)
1305                                 break;
1306                         prev = i;
1307                 }
1308                 p->next = i;
1309                 if (prev < 0)
1310                         swap_list.head = swap_list.next = p - swap_info;
1311                 else
1312                         swap_info[prev].next = p - swap_info;
1313                 nr_swap_pages += p->pages;
1314                 total_swap_pages += p->pages;
1315                 p->flags |= SWP_WRITEOK;
1316                 spin_unlock(&swap_lock);
1317                 goto out_dput;
1318         }
1319
1320         /* wait for any unplug function to finish */
1321         down_write(&swap_unplug_sem);
1322         up_write(&swap_unplug_sem);
1323
1324         destroy_swap_extents(p);
1325         mutex_lock(&swapon_mutex);
1326         spin_lock(&swap_lock);
1327         drain_mmlist();
1328
1329         /* wait for anyone still in scan_swap_map */
1330         p->highest_bit = 0;             /* cuts scans short */
1331         while (p->flags >= SWP_SCANNING) {
1332                 spin_unlock(&swap_lock);
1333                 schedule_timeout_uninterruptible(1);
1334                 spin_lock(&swap_lock);
1335         }
1336
1337         swap_file = p->swap_file;
1338         p->swap_file = NULL;
1339         p->max = 0;
1340         swap_map = p->swap_map;
1341         p->swap_map = NULL;
1342         p->flags = 0;
1343         spin_unlock(&swap_lock);
1344         mutex_unlock(&swapon_mutex);
1345         vfree(swap_map);
1346         inode = mapping->host;
1347         if (S_ISBLK(inode->i_mode)) {
1348                 struct block_device *bdev = I_BDEV(inode);
1349                 set_blocksize(bdev, p->old_block_size);
1350                 bd_release(bdev);
1351         } else {
1352                 mutex_lock(&inode->i_mutex);
1353                 inode->i_flags &= ~S_SWAPFILE;
1354                 mutex_unlock(&inode->i_mutex);
1355         }
1356         filp_close(swap_file, NULL);
1357         err = 0;
1358
1359 out_dput:
1360         filp_close(victim, NULL);
1361 out:
1362         return err;
1363 }
1364
1365 #ifdef CONFIG_PROC_FS
1366 /* iterator */
1367 static void *swap_start(struct seq_file *swap, loff_t *pos)
1368 {
1369         struct swap_info_struct *ptr = swap_info;
1370         int i;
1371         loff_t l = *pos;
1372
1373         mutex_lock(&swapon_mutex);
1374
1375         if (!l)
1376                 return SEQ_START_TOKEN;
1377
1378         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1379                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1380                         continue;
1381                 if (!--l)
1382                         return ptr;
1383         }
1384
1385         return NULL;
1386 }
1387
1388 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1389 {
1390         struct swap_info_struct *ptr;
1391         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1392
1393         if (v == SEQ_START_TOKEN)
1394                 ptr = swap_info;
1395         else {
1396                 ptr = v;
1397                 ptr++;
1398         }
1399
1400         for (; ptr < endptr; ptr++) {
1401                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1402                         continue;
1403                 ++*pos;
1404                 return ptr;
1405         }
1406
1407         return NULL;
1408 }
1409
1410 static void swap_stop(struct seq_file *swap, void *v)
1411 {
1412         mutex_unlock(&swapon_mutex);
1413 }
1414
1415 static int swap_show(struct seq_file *swap, void *v)
1416 {
1417         struct swap_info_struct *ptr = v;
1418         struct file *file;
1419         int len;
1420
1421         if (ptr == SEQ_START_TOKEN) {
1422                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1423                 return 0;
1424         }
1425
1426         file = ptr->swap_file;
1427         len = seq_path(swap, &file->f_path, " \t\n\\");
1428         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1429                        len < 40 ? 40 - len : 1, " ",
1430                        S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1431                                 "partition" : "file\t",
1432                        ptr->pages << (PAGE_SHIFT - 10),
1433                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1434                        ptr->prio);
1435         return 0;
1436 }
1437
1438 static const struct seq_operations swaps_op = {
1439         .start =        swap_start,
1440         .next =         swap_next,
1441         .stop =         swap_stop,
1442         .show =         swap_show
1443 };
1444
1445 static int swaps_open(struct inode *inode, struct file *file)
1446 {
1447         return seq_open(file, &swaps_op);
1448 }
1449
1450 static const struct file_operations proc_swaps_operations = {
1451         .open           = swaps_open,
1452         .read           = seq_read,
1453         .llseek         = seq_lseek,
1454         .release        = seq_release,
1455 };
1456
1457 static int __init procswaps_init(void)
1458 {
1459         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1460         return 0;
1461 }
1462 __initcall(procswaps_init);
1463 #endif /* CONFIG_PROC_FS */
1464
1465 /*
1466  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1467  *
1468  * The swapon system call
1469  */
1470 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1471 {
1472         struct swap_info_struct * p;
1473         char *name = NULL;
1474         struct block_device *bdev = NULL;
1475         struct file *swap_file = NULL;
1476         struct address_space *mapping;
1477         unsigned int type;
1478         int i, prev;
1479         int error;
1480         union swap_header *swap_header = NULL;
1481         int swap_header_version;
1482         unsigned int nr_good_pages = 0;
1483         int nr_extents = 0;
1484         sector_t span;
1485         unsigned long maxpages = 1;
1486         int swapfilesize;
1487         unsigned short *swap_map = NULL;
1488         struct page *page = NULL;
1489         struct inode *inode = NULL;
1490         int did_down = 0;
1491
1492         if (!capable(CAP_SYS_ADMIN))
1493                 return -EPERM;
1494         spin_lock(&swap_lock);
1495         p = swap_info;
1496         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1497                 if (!(p->flags & SWP_USED))
1498                         break;
1499         error = -EPERM;
1500         if (type >= MAX_SWAPFILES) {
1501                 spin_unlock(&swap_lock);
1502                 goto out;
1503         }
1504         if (type >= nr_swapfiles)
1505                 nr_swapfiles = type+1;
1506         memset(p, 0, sizeof(*p));
1507         INIT_LIST_HEAD(&p->extent_list);
1508         p->flags = SWP_USED;
1509         p->next = -1;
1510         spin_unlock(&swap_lock);
1511         name = getname(specialfile);
1512         error = PTR_ERR(name);
1513         if (IS_ERR(name)) {
1514                 name = NULL;
1515                 goto bad_swap_2;
1516         }
1517         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1518         error = PTR_ERR(swap_file);
1519         if (IS_ERR(swap_file)) {
1520                 swap_file = NULL;
1521                 goto bad_swap_2;
1522         }
1523
1524         p->swap_file = swap_file;
1525         mapping = swap_file->f_mapping;
1526         inode = mapping->host;
1527
1528         error = -EBUSY;
1529         for (i = 0; i < nr_swapfiles; i++) {
1530                 struct swap_info_struct *q = &swap_info[i];
1531
1532                 if (i == type || !q->swap_file)
1533                         continue;
1534                 if (mapping == q->swap_file->f_mapping)
1535                         goto bad_swap;
1536         }
1537
1538         error = -EINVAL;
1539         if (S_ISBLK(inode->i_mode)) {
1540                 bdev = I_BDEV(inode);
1541                 error = bd_claim(bdev, sys_swapon);
1542                 if (error < 0) {
1543                         bdev = NULL;
1544                         error = -EINVAL;
1545                         goto bad_swap;
1546                 }
1547                 p->old_block_size = block_size(bdev);
1548                 error = set_blocksize(bdev, PAGE_SIZE);
1549                 if (error < 0)
1550                         goto bad_swap;
1551                 p->bdev = bdev;
1552         } else if (S_ISREG(inode->i_mode)) {
1553                 p->bdev = inode->i_sb->s_bdev;
1554                 mutex_lock(&inode->i_mutex);
1555                 did_down = 1;
1556                 if (IS_SWAPFILE(inode)) {
1557                         error = -EBUSY;
1558                         goto bad_swap;
1559                 }
1560         } else {
1561                 goto bad_swap;
1562         }
1563
1564         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1565
1566         /*
1567          * Read the swap header.
1568          */
1569         if (!mapping->a_ops->readpage) {
1570                 error = -EINVAL;
1571                 goto bad_swap;
1572         }
1573         page = read_mapping_page(mapping, 0, swap_file);
1574         if (IS_ERR(page)) {
1575                 error = PTR_ERR(page);
1576                 goto bad_swap;
1577         }
1578         kmap(page);
1579         swap_header = page_address(page);
1580
1581         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1582                 swap_header_version = 1;
1583         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1584                 swap_header_version = 2;
1585         else {
1586                 printk(KERN_ERR "Unable to find swap-space signature\n");
1587                 error = -EINVAL;
1588                 goto bad_swap;
1589         }
1590         
1591         switch (swap_header_version) {
1592         case 1:
1593                 printk(KERN_ERR "version 0 swap is no longer supported. "
1594                         "Use mkswap -v1 %s\n", name);
1595                 error = -EINVAL;
1596                 goto bad_swap;
1597         case 2:
1598                 /* swap partition endianess hack... */
1599                 if (swab32(swap_header->info.version) == 1) {
1600                         swab32s(&swap_header->info.version);
1601                         swab32s(&swap_header->info.last_page);
1602                         swab32s(&swap_header->info.nr_badpages);
1603                         for (i = 0; i < swap_header->info.nr_badpages; i++)
1604                                 swab32s(&swap_header->info.badpages[i]);
1605                 }
1606                 /* Check the swap header's sub-version and the size of
1607                    the swap file and bad block lists */
1608                 if (swap_header->info.version != 1) {
1609                         printk(KERN_WARNING
1610                                "Unable to handle swap header version %d\n",
1611                                swap_header->info.version);
1612                         error = -EINVAL;
1613                         goto bad_swap;
1614                 }
1615
1616                 p->lowest_bit  = 1;
1617                 p->cluster_next = 1;
1618
1619                 /*
1620                  * Find out how many pages are allowed for a single swap
1621                  * device. There are two limiting factors: 1) the number of
1622                  * bits for the swap offset in the swp_entry_t type and
1623                  * 2) the number of bits in the a swap pte as defined by
1624                  * the different architectures. In order to find the
1625                  * largest possible bit mask a swap entry with swap type 0
1626                  * and swap offset ~0UL is created, encoded to a swap pte,
1627                  * decoded to a swp_entry_t again and finally the swap
1628                  * offset is extracted. This will mask all the bits from
1629                  * the initial ~0UL mask that can't be encoded in either
1630                  * the swp_entry_t or the architecture definition of a
1631                  * swap pte.
1632                  */
1633                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1634                 if (maxpages > swap_header->info.last_page)
1635                         maxpages = swap_header->info.last_page;
1636                 p->highest_bit = maxpages - 1;
1637
1638                 error = -EINVAL;
1639                 if (!maxpages)
1640                         goto bad_swap;
1641                 if (swapfilesize && maxpages > swapfilesize) {
1642                         printk(KERN_WARNING
1643                                "Swap area shorter than signature indicates\n");
1644                         goto bad_swap;
1645                 }
1646                 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1647                         goto bad_swap;
1648                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1649                         goto bad_swap;
1650
1651                 /* OK, set up the swap map and apply the bad block list */
1652                 swap_map = vmalloc(maxpages * sizeof(short));
1653                 if (!swap_map) {
1654                         error = -ENOMEM;
1655                         goto bad_swap;
1656                 }
1657
1658                 error = 0;
1659                 memset(swap_map, 0, maxpages * sizeof(short));
1660                 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1661                         int page_nr = swap_header->info.badpages[i];
1662                         if (page_nr <= 0 || page_nr >= swap_header->info.last_page)
1663                                 error = -EINVAL;
1664                         else
1665                                 swap_map[page_nr] = SWAP_MAP_BAD;
1666                 }
1667                 nr_good_pages = swap_header->info.last_page -
1668                                 swap_header->info.nr_badpages -
1669                                 1 /* header page */;
1670                 if (error)
1671                         goto bad_swap;
1672         }
1673
1674         if (nr_good_pages) {
1675                 swap_map[0] = SWAP_MAP_BAD;
1676                 p->max = maxpages;
1677                 p->pages = nr_good_pages;
1678                 nr_extents = setup_swap_extents(p, &span);
1679                 if (nr_extents < 0) {
1680                         error = nr_extents;
1681                         goto bad_swap;
1682                 }
1683                 nr_good_pages = p->pages;
1684         }
1685         if (!nr_good_pages) {
1686                 printk(KERN_WARNING "Empty swap-file\n");
1687                 error = -EINVAL;
1688                 goto bad_swap;
1689         }
1690
1691         mutex_lock(&swapon_mutex);
1692         spin_lock(&swap_lock);
1693         if (swap_flags & SWAP_FLAG_PREFER)
1694                 p->prio =
1695                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
1696         else
1697                 p->prio = --least_priority;
1698         p->swap_map = swap_map;
1699         p->flags = SWP_ACTIVE;
1700         nr_swap_pages += nr_good_pages;
1701         total_swap_pages += nr_good_pages;
1702
1703         printk(KERN_INFO "Adding %uk swap on %s.  "
1704                         "Priority:%d extents:%d across:%lluk\n",
1705                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1706                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1707
1708         /* insert swap space into swap_list: */
1709         prev = -1;
1710         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1711                 if (p->prio >= swap_info[i].prio) {
1712                         break;
1713                 }
1714                 prev = i;
1715         }
1716         p->next = i;
1717         if (prev < 0) {
1718                 swap_list.head = swap_list.next = p - swap_info;
1719         } else {
1720                 swap_info[prev].next = p - swap_info;
1721         }
1722         spin_unlock(&swap_lock);
1723         mutex_unlock(&swapon_mutex);
1724         error = 0;
1725         goto out;
1726 bad_swap:
1727         if (bdev) {
1728                 set_blocksize(bdev, p->old_block_size);
1729                 bd_release(bdev);
1730         }
1731         destroy_swap_extents(p);
1732 bad_swap_2:
1733         spin_lock(&swap_lock);
1734         p->swap_file = NULL;
1735         p->flags = 0;
1736         spin_unlock(&swap_lock);
1737         vfree(swap_map);
1738         if (swap_file)
1739                 filp_close(swap_file, NULL);
1740 out:
1741         if (page && !IS_ERR(page)) {
1742                 kunmap(page);
1743                 page_cache_release(page);
1744         }
1745         if (name)
1746                 putname(name);
1747         if (did_down) {
1748                 if (!error)
1749                         inode->i_flags |= S_SWAPFILE;
1750                 mutex_unlock(&inode->i_mutex);
1751         }
1752         return error;
1753 }
1754
1755 void si_swapinfo(struct sysinfo *val)
1756 {
1757         unsigned int i;
1758         unsigned long nr_to_be_unused = 0;
1759
1760         spin_lock(&swap_lock);
1761         for (i = 0; i < nr_swapfiles; i++) {
1762                 if (!(swap_info[i].flags & SWP_USED) ||
1763                      (swap_info[i].flags & SWP_WRITEOK))
1764                         continue;
1765                 nr_to_be_unused += swap_info[i].inuse_pages;
1766         }
1767         val->freeswap = nr_swap_pages + nr_to_be_unused;
1768         val->totalswap = total_swap_pages + nr_to_be_unused;
1769         spin_unlock(&swap_lock);
1770 }
1771
1772 /*
1773  * Verify that a swap entry is valid and increment its swap map count.
1774  *
1775  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1776  * "permanent", but will be reclaimed by the next swapoff.
1777  */
1778 int swap_duplicate(swp_entry_t entry)
1779 {
1780         struct swap_info_struct * p;
1781         unsigned long offset, type;
1782         int result = 0;
1783
1784         if (is_migration_entry(entry))
1785                 return 1;
1786
1787         type = swp_type(entry);
1788         if (type >= nr_swapfiles)
1789                 goto bad_file;
1790         p = type + swap_info;
1791         offset = swp_offset(entry);
1792
1793         spin_lock(&swap_lock);
1794         if (offset < p->max && p->swap_map[offset]) {
1795                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1796                         p->swap_map[offset]++;
1797                         result = 1;
1798                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1799                         if (swap_overflow++ < 5)
1800                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1801                         p->swap_map[offset] = SWAP_MAP_MAX;
1802                         result = 1;
1803                 }
1804         }
1805         spin_unlock(&swap_lock);
1806 out:
1807         return result;
1808
1809 bad_file:
1810         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1811         goto out;
1812 }
1813
1814 struct swap_info_struct *
1815 get_swap_info_struct(unsigned type)
1816 {
1817         return &swap_info[type];
1818 }
1819
1820 /*
1821  * swap_lock prevents swap_map being freed. Don't grab an extra
1822  * reference on the swaphandle, it doesn't matter if it becomes unused.
1823  */
1824 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1825 {
1826         struct swap_info_struct *si;
1827         int our_page_cluster = page_cluster;
1828         pgoff_t target, toff;
1829         pgoff_t base, end;
1830         int nr_pages = 0;
1831
1832         if (!our_page_cluster)  /* no readahead */
1833                 return 0;
1834
1835         si = &swap_info[swp_type(entry)];
1836         target = swp_offset(entry);
1837         base = (target >> our_page_cluster) << our_page_cluster;
1838         end = base + (1 << our_page_cluster);
1839         if (!base)              /* first page is swap header */
1840                 base++;
1841
1842         spin_lock(&swap_lock);
1843         if (end > si->max)      /* don't go beyond end of map */
1844                 end = si->max;
1845
1846         /* Count contiguous allocated slots above our target */
1847         for (toff = target; ++toff < end; nr_pages++) {
1848                 /* Don't read in free or bad pages */
1849                 if (!si->swap_map[toff])
1850                         break;
1851                 if (si->swap_map[toff] == SWAP_MAP_BAD)
1852                         break;
1853         }
1854         /* Count contiguous allocated slots below our target */
1855         for (toff = target; --toff >= base; nr_pages++) {
1856                 /* Don't read in free or bad pages */
1857                 if (!si->swap_map[toff])
1858                         break;
1859                 if (si->swap_map[toff] == SWAP_MAP_BAD)
1860                         break;
1861         }
1862         spin_unlock(&swap_lock);
1863
1864         /*
1865          * Indicate starting offset, and return number of pages to get:
1866          * if only 1, say 0, since there's then no readahead to be done.
1867          */
1868         *offset = ++toff;
1869         return nr_pages? ++nr_pages: 0;
1870 }