[PATCH] mutex subsystem, semaphore to mutex: VFS, ->i_sem
[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/config.h>
9 #include <linux/mm.h>
10 #include <linux/hugetlb.h>
11 #include <linux/mman.h>
12 #include <linux/slab.h>
13 #include <linux/kernel_stat.h>
14 #include <linux/swap.h>
15 #include <linux/vmalloc.h>
16 #include <linux/pagemap.h>
17 #include <linux/namei.h>
18 #include <linux/shm.h>
19 #include <linux/blkdev.h>
20 #include <linux/writeback.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/init.h>
24 #include <linux/module.h>
25 #include <linux/rmap.h>
26 #include <linux/security.h>
27 #include <linux/backing-dev.h>
28 #include <linux/syscalls.h>
29
30 #include <asm/pgtable.h>
31 #include <asm/tlbflush.h>
32 #include <linux/swapops.h>
33
34 DEFINE_SPINLOCK(swap_lock);
35 unsigned int nr_swapfiles;
36 long total_swap_pages;
37 static int swap_overflow;
38
39 static const char Bad_file[] = "Bad swap file entry ";
40 static const char Unused_file[] = "Unused swap file entry ";
41 static const char Bad_offset[] = "Bad swap offset entry ";
42 static const char Unused_offset[] = "Unused swap offset entry ";
43
44 struct swap_list_t swap_list = {-1, -1};
45
46 struct swap_info_struct swap_info[MAX_SWAPFILES];
47
48 static DECLARE_MUTEX(swapon_sem);
49
50 /*
51  * We need this because the bdev->unplug_fn can sleep and we cannot
52  * hold swap_lock while calling the unplug_fn. And swap_lock
53  * cannot be turned into a semaphore.
54  */
55 static DECLARE_RWSEM(swap_unplug_sem);
56
57 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
58 {
59         swp_entry_t entry;
60
61         down_read(&swap_unplug_sem);
62         entry.val = page_private(page);
63         if (PageSwapCache(page)) {
64                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
65                 struct backing_dev_info *bdi;
66
67                 /*
68                  * If the page is removed from swapcache from under us (with a
69                  * racy try_to_unuse/swapoff) we need an additional reference
70                  * count to avoid reading garbage from page_private(page) above.
71                  * If the WARN_ON triggers during a swapoff it maybe the race
72                  * condition and it's harmless. However if it triggers without
73                  * swapoff it signals a problem.
74                  */
75                 WARN_ON(page_count(page) <= 1);
76
77                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
78                 blk_run_backing_dev(bdi, page);
79         }
80         up_read(&swap_unplug_sem);
81 }
82
83 #define SWAPFILE_CLUSTER        256
84 #define LATENCY_LIMIT           256
85
86 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
87 {
88         unsigned long offset, last_in_cluster;
89         int latency_ration = LATENCY_LIMIT;
90
91         /* 
92          * We try to cluster swap pages by allocating them sequentially
93          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
94          * way, however, we resort to first-free allocation, starting
95          * a new cluster.  This prevents us from scattering swap pages
96          * all over the entire swap partition, so that we reduce
97          * overall disk seek times between swap pages.  -- sct
98          * But we do now try to find an empty cluster.  -Andrea
99          */
100
101         si->flags += SWP_SCANNING;
102         if (unlikely(!si->cluster_nr)) {
103                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
104                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER)
105                         goto lowest;
106                 spin_unlock(&swap_lock);
107
108                 offset = si->lowest_bit;
109                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
110
111                 /* Locate the first empty (unaligned) cluster */
112                 for (; last_in_cluster <= si->highest_bit; offset++) {
113                         if (si->swap_map[offset])
114                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
115                         else if (offset == last_in_cluster) {
116                                 spin_lock(&swap_lock);
117                                 si->cluster_next = offset-SWAPFILE_CLUSTER-1;
118                                 goto cluster;
119                         }
120                         if (unlikely(--latency_ration < 0)) {
121                                 cond_resched();
122                                 latency_ration = LATENCY_LIMIT;
123                         }
124                 }
125                 spin_lock(&swap_lock);
126                 goto lowest;
127         }
128
129         si->cluster_nr--;
130 cluster:
131         offset = si->cluster_next;
132         if (offset > si->highest_bit)
133 lowest:         offset = si->lowest_bit;
134 checks: if (!(si->flags & SWP_WRITEOK))
135                 goto no_page;
136         if (!si->highest_bit)
137                 goto no_page;
138         if (!si->swap_map[offset]) {
139                 if (offset == si->lowest_bit)
140                         si->lowest_bit++;
141                 if (offset == si->highest_bit)
142                         si->highest_bit--;
143                 si->inuse_pages++;
144                 if (si->inuse_pages == si->pages) {
145                         si->lowest_bit = si->max;
146                         si->highest_bit = 0;
147                 }
148                 si->swap_map[offset] = 1;
149                 si->cluster_next = offset + 1;
150                 si->flags -= SWP_SCANNING;
151                 return offset;
152         }
153
154         spin_unlock(&swap_lock);
155         while (++offset <= si->highest_bit) {
156                 if (!si->swap_map[offset]) {
157                         spin_lock(&swap_lock);
158                         goto checks;
159                 }
160                 if (unlikely(--latency_ration < 0)) {
161                         cond_resched();
162                         latency_ration = LATENCY_LIMIT;
163                 }
164         }
165         spin_lock(&swap_lock);
166         goto lowest;
167
168 no_page:
169         si->flags -= SWP_SCANNING;
170         return 0;
171 }
172
173 swp_entry_t get_swap_page(void)
174 {
175         struct swap_info_struct *si;
176         pgoff_t offset;
177         int type, next;
178         int wrapped = 0;
179
180         spin_lock(&swap_lock);
181         if (nr_swap_pages <= 0)
182                 goto noswap;
183         nr_swap_pages--;
184
185         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
186                 si = swap_info + type;
187                 next = si->next;
188                 if (next < 0 ||
189                     (!wrapped && si->prio != swap_info[next].prio)) {
190                         next = swap_list.head;
191                         wrapped++;
192                 }
193
194                 if (!si->highest_bit)
195                         continue;
196                 if (!(si->flags & SWP_WRITEOK))
197                         continue;
198
199                 swap_list.next = next;
200                 offset = scan_swap_map(si);
201                 if (offset) {
202                         spin_unlock(&swap_lock);
203                         return swp_entry(type, offset);
204                 }
205                 next = swap_list.next;
206         }
207
208         nr_swap_pages++;
209 noswap:
210         spin_unlock(&swap_lock);
211         return (swp_entry_t) {0};
212 }
213
214 swp_entry_t get_swap_page_of_type(int type)
215 {
216         struct swap_info_struct *si;
217         pgoff_t offset;
218
219         spin_lock(&swap_lock);
220         si = swap_info + type;
221         if (si->flags & SWP_WRITEOK) {
222                 nr_swap_pages--;
223                 offset = scan_swap_map(si);
224                 if (offset) {
225                         spin_unlock(&swap_lock);
226                         return swp_entry(type, offset);
227                 }
228                 nr_swap_pages++;
229         }
230         spin_unlock(&swap_lock);
231         return (swp_entry_t) {0};
232 }
233
234 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
235 {
236         struct swap_info_struct * p;
237         unsigned long offset, type;
238
239         if (!entry.val)
240                 goto out;
241         type = swp_type(entry);
242         if (type >= nr_swapfiles)
243                 goto bad_nofile;
244         p = & swap_info[type];
245         if (!(p->flags & SWP_USED))
246                 goto bad_device;
247         offset = swp_offset(entry);
248         if (offset >= p->max)
249                 goto bad_offset;
250         if (!p->swap_map[offset])
251                 goto bad_free;
252         spin_lock(&swap_lock);
253         return p;
254
255 bad_free:
256         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
257         goto out;
258 bad_offset:
259         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
260         goto out;
261 bad_device:
262         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
263         goto out;
264 bad_nofile:
265         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
266 out:
267         return NULL;
268 }       
269
270 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
271 {
272         int count = p->swap_map[offset];
273
274         if (count < SWAP_MAP_MAX) {
275                 count--;
276                 p->swap_map[offset] = count;
277                 if (!count) {
278                         if (offset < p->lowest_bit)
279                                 p->lowest_bit = offset;
280                         if (offset > p->highest_bit)
281                                 p->highest_bit = offset;
282                         if (p->prio > swap_info[swap_list.next].prio)
283                                 swap_list.next = p - swap_info;
284                         nr_swap_pages++;
285                         p->inuse_pages--;
286                 }
287         }
288         return count;
289 }
290
291 /*
292  * Caller has made sure that the swapdevice corresponding to entry
293  * is still around or has not been recycled.
294  */
295 void swap_free(swp_entry_t entry)
296 {
297         struct swap_info_struct * p;
298
299         p = swap_info_get(entry);
300         if (p) {
301                 swap_entry_free(p, swp_offset(entry));
302                 spin_unlock(&swap_lock);
303         }
304 }
305
306 /*
307  * How many references to page are currently swapped out?
308  */
309 static inline int page_swapcount(struct page *page)
310 {
311         int count = 0;
312         struct swap_info_struct *p;
313         swp_entry_t entry;
314
315         entry.val = page_private(page);
316         p = swap_info_get(entry);
317         if (p) {
318                 /* Subtract the 1 for the swap cache itself */
319                 count = p->swap_map[swp_offset(entry)] - 1;
320                 spin_unlock(&swap_lock);
321         }
322         return count;
323 }
324
325 /*
326  * We can use this swap cache entry directly
327  * if there are no other references to it.
328  */
329 int can_share_swap_page(struct page *page)
330 {
331         int count;
332
333         BUG_ON(!PageLocked(page));
334         count = page_mapcount(page);
335         if (count <= 1 && PageSwapCache(page))
336                 count += page_swapcount(page);
337         return count == 1;
338 }
339
340 /*
341  * Work out if there are any other processes sharing this
342  * swap cache page. Free it if you can. Return success.
343  */
344 int remove_exclusive_swap_page(struct page *page)
345 {
346         int retval;
347         struct swap_info_struct * p;
348         swp_entry_t entry;
349
350         BUG_ON(PagePrivate(page));
351         BUG_ON(!PageLocked(page));
352
353         if (!PageSwapCache(page))
354                 return 0;
355         if (PageWriteback(page))
356                 return 0;
357         if (page_count(page) != 2) /* 2: us + cache */
358                 return 0;
359
360         entry.val = page_private(page);
361         p = swap_info_get(entry);
362         if (!p)
363                 return 0;
364
365         /* Is the only swap cache user the cache itself? */
366         retval = 0;
367         if (p->swap_map[swp_offset(entry)] == 1) {
368                 /* Recheck the page count with the swapcache lock held.. */
369                 write_lock_irq(&swapper_space.tree_lock);
370                 if ((page_count(page) == 2) && !PageWriteback(page)) {
371                         __delete_from_swap_cache(page);
372                         SetPageDirty(page);
373                         retval = 1;
374                 }
375                 write_unlock_irq(&swapper_space.tree_lock);
376         }
377         spin_unlock(&swap_lock);
378
379         if (retval) {
380                 swap_free(entry);
381                 page_cache_release(page);
382         }
383
384         return retval;
385 }
386
387 /*
388  * Free the swap entry like above, but also try to
389  * free the page cache entry if it is the last user.
390  */
391 void free_swap_and_cache(swp_entry_t entry)
392 {
393         struct swap_info_struct * p;
394         struct page *page = NULL;
395
396         p = swap_info_get(entry);
397         if (p) {
398                 if (swap_entry_free(p, swp_offset(entry)) == 1)
399                         page = find_trylock_page(&swapper_space, entry.val);
400                 spin_unlock(&swap_lock);
401         }
402         if (page) {
403                 int one_user;
404
405                 BUG_ON(PagePrivate(page));
406                 page_cache_get(page);
407                 one_user = (page_count(page) == 2);
408                 /* Only cache user (+us), or swap space full? Free it! */
409                 if (!PageWriteback(page) && (one_user || vm_swap_full())) {
410                         delete_from_swap_cache(page);
411                         SetPageDirty(page);
412                 }
413                 unlock_page(page);
414                 page_cache_release(page);
415         }
416 }
417
418 /*
419  * No need to decide whether this PTE shares the swap entry with others,
420  * just let do_wp_page work it out if a write is requested later - to
421  * force COW, vm_page_prot omits write permission from any private vma.
422  */
423 static void unuse_pte(struct vm_area_struct *vma, pte_t *pte,
424                 unsigned long addr, swp_entry_t entry, struct page *page)
425 {
426         inc_mm_counter(vma->vm_mm, anon_rss);
427         get_page(page);
428         set_pte_at(vma->vm_mm, addr, pte,
429                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
430         page_add_anon_rmap(page, vma, addr);
431         swap_free(entry);
432         /*
433          * Move the page to the active list so it is not
434          * immediately swapped out again after swapon.
435          */
436         activate_page(page);
437 }
438
439 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
440                                 unsigned long addr, unsigned long end,
441                                 swp_entry_t entry, struct page *page)
442 {
443         pte_t swp_pte = swp_entry_to_pte(entry);
444         pte_t *pte;
445         spinlock_t *ptl;
446         int found = 0;
447
448         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
449         do {
450                 /*
451                  * swapoff spends a _lot_ of time in this loop!
452                  * Test inline before going to call unuse_pte.
453                  */
454                 if (unlikely(pte_same(*pte, swp_pte))) {
455                         unuse_pte(vma, pte++, addr, entry, page);
456                         found = 1;
457                         break;
458                 }
459         } while (pte++, addr += PAGE_SIZE, addr != end);
460         pte_unmap_unlock(pte - 1, ptl);
461         return found;
462 }
463
464 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
465                                 unsigned long addr, unsigned long end,
466                                 swp_entry_t entry, struct page *page)
467 {
468         pmd_t *pmd;
469         unsigned long next;
470
471         pmd = pmd_offset(pud, addr);
472         do {
473                 next = pmd_addr_end(addr, end);
474                 if (pmd_none_or_clear_bad(pmd))
475                         continue;
476                 if (unuse_pte_range(vma, pmd, addr, next, entry, page))
477                         return 1;
478         } while (pmd++, addr = next, addr != end);
479         return 0;
480 }
481
482 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
483                                 unsigned long addr, unsigned long end,
484                                 swp_entry_t entry, struct page *page)
485 {
486         pud_t *pud;
487         unsigned long next;
488
489         pud = pud_offset(pgd, addr);
490         do {
491                 next = pud_addr_end(addr, end);
492                 if (pud_none_or_clear_bad(pud))
493                         continue;
494                 if (unuse_pmd_range(vma, pud, addr, next, entry, page))
495                         return 1;
496         } while (pud++, addr = next, addr != end);
497         return 0;
498 }
499
500 static int unuse_vma(struct vm_area_struct *vma,
501                                 swp_entry_t entry, struct page *page)
502 {
503         pgd_t *pgd;
504         unsigned long addr, end, next;
505
506         if (page->mapping) {
507                 addr = page_address_in_vma(page, vma);
508                 if (addr == -EFAULT)
509                         return 0;
510                 else
511                         end = addr + PAGE_SIZE;
512         } else {
513                 addr = vma->vm_start;
514                 end = vma->vm_end;
515         }
516
517         pgd = pgd_offset(vma->vm_mm, addr);
518         do {
519                 next = pgd_addr_end(addr, end);
520                 if (pgd_none_or_clear_bad(pgd))
521                         continue;
522                 if (unuse_pud_range(vma, pgd, addr, next, entry, page))
523                         return 1;
524         } while (pgd++, addr = next, addr != end);
525         return 0;
526 }
527
528 static int unuse_mm(struct mm_struct *mm,
529                                 swp_entry_t entry, struct page *page)
530 {
531         struct vm_area_struct *vma;
532
533         if (!down_read_trylock(&mm->mmap_sem)) {
534                 /*
535                  * Activate page so shrink_cache is unlikely to unmap its
536                  * ptes while lock is dropped, so swapoff can make progress.
537                  */
538                 activate_page(page);
539                 unlock_page(page);
540                 down_read(&mm->mmap_sem);
541                 lock_page(page);
542         }
543         for (vma = mm->mmap; vma; vma = vma->vm_next) {
544                 if (vma->anon_vma && unuse_vma(vma, entry, page))
545                         break;
546         }
547         up_read(&mm->mmap_sem);
548         /*
549          * Currently unuse_mm cannot fail, but leave error handling
550          * at call sites for now, since we change it from time to time.
551          */
552         return 0;
553 }
554
555 /*
556  * Scan swap_map from current position to next entry still in use.
557  * Recycle to start on reaching the end, returning 0 when empty.
558  */
559 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
560                                         unsigned int prev)
561 {
562         unsigned int max = si->max;
563         unsigned int i = prev;
564         int count;
565
566         /*
567          * No need for swap_lock here: we're just looking
568          * for whether an entry is in use, not modifying it; false
569          * hits are okay, and sys_swapoff() has already prevented new
570          * allocations from this area (while holding swap_lock).
571          */
572         for (;;) {
573                 if (++i >= max) {
574                         if (!prev) {
575                                 i = 0;
576                                 break;
577                         }
578                         /*
579                          * No entries in use at top of swap_map,
580                          * loop back to start and recheck there.
581                          */
582                         max = prev + 1;
583                         prev = 0;
584                         i = 1;
585                 }
586                 count = si->swap_map[i];
587                 if (count && count != SWAP_MAP_BAD)
588                         break;
589         }
590         return i;
591 }
592
593 /*
594  * We completely avoid races by reading each swap page in advance,
595  * and then search for the process using it.  All the necessary
596  * page table adjustments can then be made atomically.
597  */
598 static int try_to_unuse(unsigned int type)
599 {
600         struct swap_info_struct * si = &swap_info[type];
601         struct mm_struct *start_mm;
602         unsigned short *swap_map;
603         unsigned short swcount;
604         struct page *page;
605         swp_entry_t entry;
606         unsigned int i = 0;
607         int retval = 0;
608         int reset_overflow = 0;
609         int shmem;
610
611         /*
612          * When searching mms for an entry, a good strategy is to
613          * start at the first mm we freed the previous entry from
614          * (though actually we don't notice whether we or coincidence
615          * freed the entry).  Initialize this start_mm with a hold.
616          *
617          * A simpler strategy would be to start at the last mm we
618          * freed the previous entry from; but that would take less
619          * advantage of mmlist ordering, which clusters forked mms
620          * together, child after parent.  If we race with dup_mmap(), we
621          * prefer to resolve parent before child, lest we miss entries
622          * duplicated after we scanned child: using last mm would invert
623          * that.  Though it's only a serious concern when an overflowed
624          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
625          */
626         start_mm = &init_mm;
627         atomic_inc(&init_mm.mm_users);
628
629         /*
630          * Keep on scanning until all entries have gone.  Usually,
631          * one pass through swap_map is enough, but not necessarily:
632          * there are races when an instance of an entry might be missed.
633          */
634         while ((i = find_next_to_unuse(si, i)) != 0) {
635                 if (signal_pending(current)) {
636                         retval = -EINTR;
637                         break;
638                 }
639
640                 /* 
641                  * Get a page for the entry, using the existing swap
642                  * cache page if there is one.  Otherwise, get a clean
643                  * page and read the swap into it. 
644                  */
645                 swap_map = &si->swap_map[i];
646                 entry = swp_entry(type, i);
647                 page = read_swap_cache_async(entry, NULL, 0);
648                 if (!page) {
649                         /*
650                          * Either swap_duplicate() failed because entry
651                          * has been freed independently, and will not be
652                          * reused since sys_swapoff() already disabled
653                          * allocation from here, or alloc_page() failed.
654                          */
655                         if (!*swap_map)
656                                 continue;
657                         retval = -ENOMEM;
658                         break;
659                 }
660
661                 /*
662                  * Don't hold on to start_mm if it looks like exiting.
663                  */
664                 if (atomic_read(&start_mm->mm_users) == 1) {
665                         mmput(start_mm);
666                         start_mm = &init_mm;
667                         atomic_inc(&init_mm.mm_users);
668                 }
669
670                 /*
671                  * Wait for and lock page.  When do_swap_page races with
672                  * try_to_unuse, do_swap_page can handle the fault much
673                  * faster than try_to_unuse can locate the entry.  This
674                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
675                  * defer to do_swap_page in such a case - in some tests,
676                  * do_swap_page and try_to_unuse repeatedly compete.
677                  */
678                 wait_on_page_locked(page);
679                 wait_on_page_writeback(page);
680                 lock_page(page);
681                 wait_on_page_writeback(page);
682
683                 /*
684                  * Remove all references to entry.
685                  * Whenever we reach init_mm, there's no address space
686                  * to search, but use it as a reminder to search shmem.
687                  */
688                 shmem = 0;
689                 swcount = *swap_map;
690                 if (swcount > 1) {
691                         if (start_mm == &init_mm)
692                                 shmem = shmem_unuse(entry, page);
693                         else
694                                 retval = unuse_mm(start_mm, entry, page);
695                 }
696                 if (*swap_map > 1) {
697                         int set_start_mm = (*swap_map >= swcount);
698                         struct list_head *p = &start_mm->mmlist;
699                         struct mm_struct *new_start_mm = start_mm;
700                         struct mm_struct *prev_mm = start_mm;
701                         struct mm_struct *mm;
702
703                         atomic_inc(&new_start_mm->mm_users);
704                         atomic_inc(&prev_mm->mm_users);
705                         spin_lock(&mmlist_lock);
706                         while (*swap_map > 1 && !retval &&
707                                         (p = p->next) != &start_mm->mmlist) {
708                                 mm = list_entry(p, struct mm_struct, mmlist);
709                                 if (atomic_inc_return(&mm->mm_users) == 1) {
710                                         atomic_dec(&mm->mm_users);
711                                         continue;
712                                 }
713                                 spin_unlock(&mmlist_lock);
714                                 mmput(prev_mm);
715                                 prev_mm = mm;
716
717                                 cond_resched();
718
719                                 swcount = *swap_map;
720                                 if (swcount <= 1)
721                                         ;
722                                 else if (mm == &init_mm) {
723                                         set_start_mm = 1;
724                                         shmem = shmem_unuse(entry, page);
725                                 } else
726                                         retval = unuse_mm(mm, entry, page);
727                                 if (set_start_mm && *swap_map < swcount) {
728                                         mmput(new_start_mm);
729                                         atomic_inc(&mm->mm_users);
730                                         new_start_mm = mm;
731                                         set_start_mm = 0;
732                                 }
733                                 spin_lock(&mmlist_lock);
734                         }
735                         spin_unlock(&mmlist_lock);
736                         mmput(prev_mm);
737                         mmput(start_mm);
738                         start_mm = new_start_mm;
739                 }
740                 if (retval) {
741                         unlock_page(page);
742                         page_cache_release(page);
743                         break;
744                 }
745
746                 /*
747                  * How could swap count reach 0x7fff when the maximum
748                  * pid is 0x7fff, and there's no way to repeat a swap
749                  * page within an mm (except in shmem, where it's the
750                  * shared object which takes the reference count)?
751                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
752                  *
753                  * If that's wrong, then we should worry more about
754                  * exit_mmap() and do_munmap() cases described above:
755                  * we might be resetting SWAP_MAP_MAX too early here.
756                  * We know "Undead"s can happen, they're okay, so don't
757                  * report them; but do report if we reset SWAP_MAP_MAX.
758                  */
759                 if (*swap_map == SWAP_MAP_MAX) {
760                         spin_lock(&swap_lock);
761                         *swap_map = 1;
762                         spin_unlock(&swap_lock);
763                         reset_overflow = 1;
764                 }
765
766                 /*
767                  * If a reference remains (rare), we would like to leave
768                  * the page in the swap cache; but try_to_unmap could
769                  * then re-duplicate the entry once we drop page lock,
770                  * so we might loop indefinitely; also, that page could
771                  * not be swapped out to other storage meanwhile.  So:
772                  * delete from cache even if there's another reference,
773                  * after ensuring that the data has been saved to disk -
774                  * since if the reference remains (rarer), it will be
775                  * read from disk into another page.  Splitting into two
776                  * pages would be incorrect if swap supported "shared
777                  * private" pages, but they are handled by tmpfs files.
778                  *
779                  * Note shmem_unuse already deleted a swappage from
780                  * the swap cache, unless the move to filepage failed:
781                  * in which case it left swappage in cache, lowered its
782                  * swap count to pass quickly through the loops above,
783                  * and now we must reincrement count to try again later.
784                  */
785                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
786                         struct writeback_control wbc = {
787                                 .sync_mode = WB_SYNC_NONE,
788                         };
789
790                         swap_writepage(page, &wbc);
791                         lock_page(page);
792                         wait_on_page_writeback(page);
793                 }
794                 if (PageSwapCache(page)) {
795                         if (shmem)
796                                 swap_duplicate(entry);
797                         else
798                                 delete_from_swap_cache(page);
799                 }
800
801                 /*
802                  * So we could skip searching mms once swap count went
803                  * to 1, we did not mark any present ptes as dirty: must
804                  * mark page dirty so shrink_list will preserve it.
805                  */
806                 SetPageDirty(page);
807                 unlock_page(page);
808                 page_cache_release(page);
809
810                 /*
811                  * Make sure that we aren't completely killing
812                  * interactive performance.
813                  */
814                 cond_resched();
815         }
816
817         mmput(start_mm);
818         if (reset_overflow) {
819                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
820                 swap_overflow = 0;
821         }
822         return retval;
823 }
824
825 /*
826  * After a successful try_to_unuse, if no swap is now in use, we know
827  * we can empty the mmlist.  swap_lock must be held on entry and exit.
828  * Note that mmlist_lock nests inside swap_lock, and an mm must be
829  * added to the mmlist just after page_duplicate - before would be racy.
830  */
831 static void drain_mmlist(void)
832 {
833         struct list_head *p, *next;
834         unsigned int i;
835
836         for (i = 0; i < nr_swapfiles; i++)
837                 if (swap_info[i].inuse_pages)
838                         return;
839         spin_lock(&mmlist_lock);
840         list_for_each_safe(p, next, &init_mm.mmlist)
841                 list_del_init(p);
842         spin_unlock(&mmlist_lock);
843 }
844
845 /*
846  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
847  * corresponds to page offset `offset'.
848  */
849 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
850 {
851         struct swap_extent *se = sis->curr_swap_extent;
852         struct swap_extent *start_se = se;
853
854         for ( ; ; ) {
855                 struct list_head *lh;
856
857                 if (se->start_page <= offset &&
858                                 offset < (se->start_page + se->nr_pages)) {
859                         return se->start_block + (offset - se->start_page);
860                 }
861                 lh = se->list.next;
862                 if (lh == &sis->extent_list)
863                         lh = lh->next;
864                 se = list_entry(lh, struct swap_extent, list);
865                 sis->curr_swap_extent = se;
866                 BUG_ON(se == start_se);         /* It *must* be present */
867         }
868 }
869
870 /*
871  * Free all of a swapdev's extent information
872  */
873 static void destroy_swap_extents(struct swap_info_struct *sis)
874 {
875         while (!list_empty(&sis->extent_list)) {
876                 struct swap_extent *se;
877
878                 se = list_entry(sis->extent_list.next,
879                                 struct swap_extent, list);
880                 list_del(&se->list);
881                 kfree(se);
882         }
883 }
884
885 /*
886  * Add a block range (and the corresponding page range) into this swapdev's
887  * extent list.  The extent list is kept sorted in page order.
888  *
889  * This function rather assumes that it is called in ascending page order.
890  */
891 static int
892 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
893                 unsigned long nr_pages, sector_t start_block)
894 {
895         struct swap_extent *se;
896         struct swap_extent *new_se;
897         struct list_head *lh;
898
899         lh = sis->extent_list.prev;     /* The highest page extent */
900         if (lh != &sis->extent_list) {
901                 se = list_entry(lh, struct swap_extent, list);
902                 BUG_ON(se->start_page + se->nr_pages != start_page);
903                 if (se->start_block + se->nr_pages == start_block) {
904                         /* Merge it */
905                         se->nr_pages += nr_pages;
906                         return 0;
907                 }
908         }
909
910         /*
911          * No merge.  Insert a new extent, preserving ordering.
912          */
913         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
914         if (new_se == NULL)
915                 return -ENOMEM;
916         new_se->start_page = start_page;
917         new_se->nr_pages = nr_pages;
918         new_se->start_block = start_block;
919
920         list_add_tail(&new_se->list, &sis->extent_list);
921         return 1;
922 }
923
924 /*
925  * A `swap extent' is a simple thing which maps a contiguous range of pages
926  * onto a contiguous range of disk blocks.  An ordered list of swap extents
927  * is built at swapon time and is then used at swap_writepage/swap_readpage
928  * time for locating where on disk a page belongs.
929  *
930  * If the swapfile is an S_ISBLK block device, a single extent is installed.
931  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
932  * swap files identically.
933  *
934  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
935  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
936  * swapfiles are handled *identically* after swapon time.
937  *
938  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
939  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
940  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
941  * requirements, they are simply tossed out - we will never use those blocks
942  * for swapping.
943  *
944  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
945  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
946  * which will scribble on the fs.
947  *
948  * The amount of disk space which a single swap extent represents varies.
949  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
950  * extents in the list.  To avoid much list walking, we cache the previous
951  * search location in `curr_swap_extent', and start new searches from there.
952  * This is extremely effective.  The average number of iterations in
953  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
954  */
955 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
956 {
957         struct inode *inode;
958         unsigned blocks_per_page;
959         unsigned long page_no;
960         unsigned blkbits;
961         sector_t probe_block;
962         sector_t last_block;
963         sector_t lowest_block = -1;
964         sector_t highest_block = 0;
965         int nr_extents = 0;
966         int ret;
967
968         inode = sis->swap_file->f_mapping->host;
969         if (S_ISBLK(inode->i_mode)) {
970                 ret = add_swap_extent(sis, 0, sis->max, 0);
971                 *span = sis->pages;
972                 goto done;
973         }
974
975         blkbits = inode->i_blkbits;
976         blocks_per_page = PAGE_SIZE >> blkbits;
977
978         /*
979          * Map all the blocks into the extent list.  This code doesn't try
980          * to be very smart.
981          */
982         probe_block = 0;
983         page_no = 0;
984         last_block = i_size_read(inode) >> blkbits;
985         while ((probe_block + blocks_per_page) <= last_block &&
986                         page_no < sis->max) {
987                 unsigned block_in_page;
988                 sector_t first_block;
989
990                 first_block = bmap(inode, probe_block);
991                 if (first_block == 0)
992                         goto bad_bmap;
993
994                 /*
995                  * It must be PAGE_SIZE aligned on-disk
996                  */
997                 if (first_block & (blocks_per_page - 1)) {
998                         probe_block++;
999                         goto reprobe;
1000                 }
1001
1002                 for (block_in_page = 1; block_in_page < blocks_per_page;
1003                                         block_in_page++) {
1004                         sector_t block;
1005
1006                         block = bmap(inode, probe_block + block_in_page);
1007                         if (block == 0)
1008                                 goto bad_bmap;
1009                         if (block != first_block + block_in_page) {
1010                                 /* Discontiguity */
1011                                 probe_block++;
1012                                 goto reprobe;
1013                         }
1014                 }
1015
1016                 first_block >>= (PAGE_SHIFT - blkbits);
1017                 if (page_no) {  /* exclude the header page */
1018                         if (first_block < lowest_block)
1019                                 lowest_block = first_block;
1020                         if (first_block > highest_block)
1021                                 highest_block = first_block;
1022                 }
1023
1024                 /*
1025                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1026                  */
1027                 ret = add_swap_extent(sis, page_no, 1, first_block);
1028                 if (ret < 0)
1029                         goto out;
1030                 nr_extents += ret;
1031                 page_no++;
1032                 probe_block += blocks_per_page;
1033 reprobe:
1034                 continue;
1035         }
1036         ret = nr_extents;
1037         *span = 1 + highest_block - lowest_block;
1038         if (page_no == 0)
1039                 page_no = 1;    /* force Empty message */
1040         sis->max = page_no;
1041         sis->pages = page_no - 1;
1042         sis->highest_bit = page_no - 1;
1043 done:
1044         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1045                                         struct swap_extent, list);
1046         goto out;
1047 bad_bmap:
1048         printk(KERN_ERR "swapon: swapfile has holes\n");
1049         ret = -EINVAL;
1050 out:
1051         return ret;
1052 }
1053
1054 #if 0   /* We don't need this yet */
1055 #include <linux/backing-dev.h>
1056 int page_queue_congested(struct page *page)
1057 {
1058         struct backing_dev_info *bdi;
1059
1060         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1061
1062         if (PageSwapCache(page)) {
1063                 swp_entry_t entry = { .val = page_private(page) };
1064                 struct swap_info_struct *sis;
1065
1066                 sis = get_swap_info_struct(swp_type(entry));
1067                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1068         } else
1069                 bdi = page->mapping->backing_dev_info;
1070         return bdi_write_congested(bdi);
1071 }
1072 #endif
1073
1074 asmlinkage long sys_swapoff(const char __user * specialfile)
1075 {
1076         struct swap_info_struct * p = NULL;
1077         unsigned short *swap_map;
1078         struct file *swap_file, *victim;
1079         struct address_space *mapping;
1080         struct inode *inode;
1081         char * pathname;
1082         int i, type, prev;
1083         int err;
1084         
1085         if (!capable(CAP_SYS_ADMIN))
1086                 return -EPERM;
1087
1088         pathname = getname(specialfile);
1089         err = PTR_ERR(pathname);
1090         if (IS_ERR(pathname))
1091                 goto out;
1092
1093         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1094         putname(pathname);
1095         err = PTR_ERR(victim);
1096         if (IS_ERR(victim))
1097                 goto out;
1098
1099         mapping = victim->f_mapping;
1100         prev = -1;
1101         spin_lock(&swap_lock);
1102         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1103                 p = swap_info + type;
1104                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1105                         if (p->swap_file->f_mapping == mapping)
1106                                 break;
1107                 }
1108                 prev = type;
1109         }
1110         if (type < 0) {
1111                 err = -EINVAL;
1112                 spin_unlock(&swap_lock);
1113                 goto out_dput;
1114         }
1115         if (!security_vm_enough_memory(p->pages))
1116                 vm_unacct_memory(p->pages);
1117         else {
1118                 err = -ENOMEM;
1119                 spin_unlock(&swap_lock);
1120                 goto out_dput;
1121         }
1122         if (prev < 0) {
1123                 swap_list.head = p->next;
1124         } else {
1125                 swap_info[prev].next = p->next;
1126         }
1127         if (type == swap_list.next) {
1128                 /* just pick something that's safe... */
1129                 swap_list.next = swap_list.head;
1130         }
1131         nr_swap_pages -= p->pages;
1132         total_swap_pages -= p->pages;
1133         p->flags &= ~SWP_WRITEOK;
1134         spin_unlock(&swap_lock);
1135
1136         current->flags |= PF_SWAPOFF;
1137         err = try_to_unuse(type);
1138         current->flags &= ~PF_SWAPOFF;
1139
1140         if (err) {
1141                 /* re-insert swap space back into swap_list */
1142                 spin_lock(&swap_lock);
1143                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1144                         if (p->prio >= swap_info[i].prio)
1145                                 break;
1146                 p->next = i;
1147                 if (prev < 0)
1148                         swap_list.head = swap_list.next = p - swap_info;
1149                 else
1150                         swap_info[prev].next = p - swap_info;
1151                 nr_swap_pages += p->pages;
1152                 total_swap_pages += p->pages;
1153                 p->flags |= SWP_WRITEOK;
1154                 spin_unlock(&swap_lock);
1155                 goto out_dput;
1156         }
1157
1158         /* wait for any unplug function to finish */
1159         down_write(&swap_unplug_sem);
1160         up_write(&swap_unplug_sem);
1161
1162         destroy_swap_extents(p);
1163         down(&swapon_sem);
1164         spin_lock(&swap_lock);
1165         drain_mmlist();
1166
1167         /* wait for anyone still in scan_swap_map */
1168         p->highest_bit = 0;             /* cuts scans short */
1169         while (p->flags >= SWP_SCANNING) {
1170                 spin_unlock(&swap_lock);
1171                 schedule_timeout_uninterruptible(1);
1172                 spin_lock(&swap_lock);
1173         }
1174
1175         swap_file = p->swap_file;
1176         p->swap_file = NULL;
1177         p->max = 0;
1178         swap_map = p->swap_map;
1179         p->swap_map = NULL;
1180         p->flags = 0;
1181         spin_unlock(&swap_lock);
1182         up(&swapon_sem);
1183         vfree(swap_map);
1184         inode = mapping->host;
1185         if (S_ISBLK(inode->i_mode)) {
1186                 struct block_device *bdev = I_BDEV(inode);
1187                 set_blocksize(bdev, p->old_block_size);
1188                 bd_release(bdev);
1189         } else {
1190                 mutex_lock(&inode->i_mutex);
1191                 inode->i_flags &= ~S_SWAPFILE;
1192                 mutex_unlock(&inode->i_mutex);
1193         }
1194         filp_close(swap_file, NULL);
1195         err = 0;
1196
1197 out_dput:
1198         filp_close(victim, NULL);
1199 out:
1200         return err;
1201 }
1202
1203 #ifdef CONFIG_PROC_FS
1204 /* iterator */
1205 static void *swap_start(struct seq_file *swap, loff_t *pos)
1206 {
1207         struct swap_info_struct *ptr = swap_info;
1208         int i;
1209         loff_t l = *pos;
1210
1211         down(&swapon_sem);
1212
1213         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1214                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1215                         continue;
1216                 if (!l--)
1217                         return ptr;
1218         }
1219
1220         return NULL;
1221 }
1222
1223 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1224 {
1225         struct swap_info_struct *ptr = v;
1226         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1227
1228         for (++ptr; ptr < endptr; ptr++) {
1229                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1230                         continue;
1231                 ++*pos;
1232                 return ptr;
1233         }
1234
1235         return NULL;
1236 }
1237
1238 static void swap_stop(struct seq_file *swap, void *v)
1239 {
1240         up(&swapon_sem);
1241 }
1242
1243 static int swap_show(struct seq_file *swap, void *v)
1244 {
1245         struct swap_info_struct *ptr = v;
1246         struct file *file;
1247         int len;
1248
1249         if (v == swap_info)
1250                 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1251
1252         file = ptr->swap_file;
1253         len = seq_path(swap, file->f_vfsmnt, file->f_dentry, " \t\n\\");
1254         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1255                        len < 40 ? 40 - len : 1, " ",
1256                        S_ISBLK(file->f_dentry->d_inode->i_mode) ?
1257                                 "partition" : "file\t",
1258                        ptr->pages << (PAGE_SHIFT - 10),
1259                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1260                        ptr->prio);
1261         return 0;
1262 }
1263
1264 static struct seq_operations swaps_op = {
1265         .start =        swap_start,
1266         .next =         swap_next,
1267         .stop =         swap_stop,
1268         .show =         swap_show
1269 };
1270
1271 static int swaps_open(struct inode *inode, struct file *file)
1272 {
1273         return seq_open(file, &swaps_op);
1274 }
1275
1276 static struct file_operations proc_swaps_operations = {
1277         .open           = swaps_open,
1278         .read           = seq_read,
1279         .llseek         = seq_lseek,
1280         .release        = seq_release,
1281 };
1282
1283 static int __init procswaps_init(void)
1284 {
1285         struct proc_dir_entry *entry;
1286
1287         entry = create_proc_entry("swaps", 0, NULL);
1288         if (entry)
1289                 entry->proc_fops = &proc_swaps_operations;
1290         return 0;
1291 }
1292 __initcall(procswaps_init);
1293 #endif /* CONFIG_PROC_FS */
1294
1295 /*
1296  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1297  *
1298  * The swapon system call
1299  */
1300 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1301 {
1302         struct swap_info_struct * p;
1303         char *name = NULL;
1304         struct block_device *bdev = NULL;
1305         struct file *swap_file = NULL;
1306         struct address_space *mapping;
1307         unsigned int type;
1308         int i, prev;
1309         int error;
1310         static int least_priority;
1311         union swap_header *swap_header = NULL;
1312         int swap_header_version;
1313         unsigned int nr_good_pages = 0;
1314         int nr_extents = 0;
1315         sector_t span;
1316         unsigned long maxpages = 1;
1317         int swapfilesize;
1318         unsigned short *swap_map;
1319         struct page *page = NULL;
1320         struct inode *inode = NULL;
1321         int did_down = 0;
1322
1323         if (!capable(CAP_SYS_ADMIN))
1324                 return -EPERM;
1325         spin_lock(&swap_lock);
1326         p = swap_info;
1327         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1328                 if (!(p->flags & SWP_USED))
1329                         break;
1330         error = -EPERM;
1331         /*
1332          * Test if adding another swap device is possible. There are
1333          * two limiting factors: 1) the number of bits for the swap
1334          * type swp_entry_t definition and 2) the number of bits for
1335          * the swap type in the swap ptes as defined by the different
1336          * architectures. To honor both limitations a swap entry
1337          * with swap offset 0 and swap type ~0UL is created, encoded
1338          * to a swap pte, decoded to a swp_entry_t again and finally
1339          * the swap type part is extracted. This will mask all bits
1340          * from the initial ~0UL that can't be encoded in either the
1341          * swp_entry_t or the architecture definition of a swap pte.
1342          */
1343         if (type > swp_type(pte_to_swp_entry(swp_entry_to_pte(swp_entry(~0UL,0))))) {
1344                 spin_unlock(&swap_lock);
1345                 goto out;
1346         }
1347         if (type >= nr_swapfiles)
1348                 nr_swapfiles = type+1;
1349         INIT_LIST_HEAD(&p->extent_list);
1350         p->flags = SWP_USED;
1351         p->swap_file = NULL;
1352         p->old_block_size = 0;
1353         p->swap_map = NULL;
1354         p->lowest_bit = 0;
1355         p->highest_bit = 0;
1356         p->cluster_nr = 0;
1357         p->inuse_pages = 0;
1358         p->next = -1;
1359         if (swap_flags & SWAP_FLAG_PREFER) {
1360                 p->prio =
1361                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1362         } else {
1363                 p->prio = --least_priority;
1364         }
1365         spin_unlock(&swap_lock);
1366         name = getname(specialfile);
1367         error = PTR_ERR(name);
1368         if (IS_ERR(name)) {
1369                 name = NULL;
1370                 goto bad_swap_2;
1371         }
1372         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1373         error = PTR_ERR(swap_file);
1374         if (IS_ERR(swap_file)) {
1375                 swap_file = NULL;
1376                 goto bad_swap_2;
1377         }
1378
1379         p->swap_file = swap_file;
1380         mapping = swap_file->f_mapping;
1381         inode = mapping->host;
1382
1383         error = -EBUSY;
1384         for (i = 0; i < nr_swapfiles; i++) {
1385                 struct swap_info_struct *q = &swap_info[i];
1386
1387                 if (i == type || !q->swap_file)
1388                         continue;
1389                 if (mapping == q->swap_file->f_mapping)
1390                         goto bad_swap;
1391         }
1392
1393         error = -EINVAL;
1394         if (S_ISBLK(inode->i_mode)) {
1395                 bdev = I_BDEV(inode);
1396                 error = bd_claim(bdev, sys_swapon);
1397                 if (error < 0) {
1398                         bdev = NULL;
1399                         error = -EINVAL;
1400                         goto bad_swap;
1401                 }
1402                 p->old_block_size = block_size(bdev);
1403                 error = set_blocksize(bdev, PAGE_SIZE);
1404                 if (error < 0)
1405                         goto bad_swap;
1406                 p->bdev = bdev;
1407         } else if (S_ISREG(inode->i_mode)) {
1408                 p->bdev = inode->i_sb->s_bdev;
1409                 mutex_lock(&inode->i_mutex);
1410                 did_down = 1;
1411                 if (IS_SWAPFILE(inode)) {
1412                         error = -EBUSY;
1413                         goto bad_swap;
1414                 }
1415         } else {
1416                 goto bad_swap;
1417         }
1418
1419         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1420
1421         /*
1422          * Read the swap header.
1423          */
1424         if (!mapping->a_ops->readpage) {
1425                 error = -EINVAL;
1426                 goto bad_swap;
1427         }
1428         page = read_cache_page(mapping, 0,
1429                         (filler_t *)mapping->a_ops->readpage, swap_file);
1430         if (IS_ERR(page)) {
1431                 error = PTR_ERR(page);
1432                 goto bad_swap;
1433         }
1434         wait_on_page_locked(page);
1435         if (!PageUptodate(page))
1436                 goto bad_swap;
1437         kmap(page);
1438         swap_header = page_address(page);
1439
1440         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1441                 swap_header_version = 1;
1442         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1443                 swap_header_version = 2;
1444         else {
1445                 printk("Unable to find swap-space signature\n");
1446                 error = -EINVAL;
1447                 goto bad_swap;
1448         }
1449         
1450         switch (swap_header_version) {
1451         case 1:
1452                 printk(KERN_ERR "version 0 swap is no longer supported. "
1453                         "Use mkswap -v1 %s\n", name);
1454                 error = -EINVAL;
1455                 goto bad_swap;
1456         case 2:
1457                 /* Check the swap header's sub-version and the size of
1458                    the swap file and bad block lists */
1459                 if (swap_header->info.version != 1) {
1460                         printk(KERN_WARNING
1461                                "Unable to handle swap header version %d\n",
1462                                swap_header->info.version);
1463                         error = -EINVAL;
1464                         goto bad_swap;
1465                 }
1466
1467                 p->lowest_bit  = 1;
1468                 p->cluster_next = 1;
1469
1470                 /*
1471                  * Find out how many pages are allowed for a single swap
1472                  * device. There are two limiting factors: 1) the number of
1473                  * bits for the swap offset in the swp_entry_t type and
1474                  * 2) the number of bits in the a swap pte as defined by
1475                  * the different architectures. In order to find the
1476                  * largest possible bit mask a swap entry with swap type 0
1477                  * and swap offset ~0UL is created, encoded to a swap pte,
1478                  * decoded to a swp_entry_t again and finally the swap
1479                  * offset is extracted. This will mask all the bits from
1480                  * the initial ~0UL mask that can't be encoded in either
1481                  * the swp_entry_t or the architecture definition of a
1482                  * swap pte.
1483                  */
1484                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1485                 if (maxpages > swap_header->info.last_page)
1486                         maxpages = swap_header->info.last_page;
1487                 p->highest_bit = maxpages - 1;
1488
1489                 error = -EINVAL;
1490                 if (!maxpages)
1491                         goto bad_swap;
1492                 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1493                         goto bad_swap;
1494                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1495                         goto bad_swap;
1496
1497                 /* OK, set up the swap map and apply the bad block list */
1498                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1499                         error = -ENOMEM;
1500                         goto bad_swap;
1501                 }
1502
1503                 error = 0;
1504                 memset(p->swap_map, 0, maxpages * sizeof(short));
1505                 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1506                         int page_nr = swap_header->info.badpages[i];
1507                         if (page_nr <= 0 || page_nr >= swap_header->info.last_page)
1508                                 error = -EINVAL;
1509                         else
1510                                 p->swap_map[page_nr] = SWAP_MAP_BAD;
1511                 }
1512                 nr_good_pages = swap_header->info.last_page -
1513                                 swap_header->info.nr_badpages -
1514                                 1 /* header page */;
1515                 if (error)
1516                         goto bad_swap;
1517         }
1518
1519         if (swapfilesize && maxpages > swapfilesize) {
1520                 printk(KERN_WARNING
1521                        "Swap area shorter than signature indicates\n");
1522                 error = -EINVAL;
1523                 goto bad_swap;
1524         }
1525         if (nr_good_pages) {
1526                 p->swap_map[0] = SWAP_MAP_BAD;
1527                 p->max = maxpages;
1528                 p->pages = nr_good_pages;
1529                 nr_extents = setup_swap_extents(p, &span);
1530                 if (nr_extents < 0) {
1531                         error = nr_extents;
1532                         goto bad_swap;
1533                 }
1534                 nr_good_pages = p->pages;
1535         }
1536         if (!nr_good_pages) {
1537                 printk(KERN_WARNING "Empty swap-file\n");
1538                 error = -EINVAL;
1539                 goto bad_swap;
1540         }
1541
1542         down(&swapon_sem);
1543         spin_lock(&swap_lock);
1544         p->flags = SWP_ACTIVE;
1545         nr_swap_pages += nr_good_pages;
1546         total_swap_pages += nr_good_pages;
1547
1548         printk(KERN_INFO "Adding %uk swap on %s.  "
1549                         "Priority:%d extents:%d across:%lluk\n",
1550                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1551                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1552
1553         /* insert swap space into swap_list: */
1554         prev = -1;
1555         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1556                 if (p->prio >= swap_info[i].prio) {
1557                         break;
1558                 }
1559                 prev = i;
1560         }
1561         p->next = i;
1562         if (prev < 0) {
1563                 swap_list.head = swap_list.next = p - swap_info;
1564         } else {
1565                 swap_info[prev].next = p - swap_info;
1566         }
1567         spin_unlock(&swap_lock);
1568         up(&swapon_sem);
1569         error = 0;
1570         goto out;
1571 bad_swap:
1572         if (bdev) {
1573                 set_blocksize(bdev, p->old_block_size);
1574                 bd_release(bdev);
1575         }
1576         destroy_swap_extents(p);
1577 bad_swap_2:
1578         spin_lock(&swap_lock);
1579         swap_map = p->swap_map;
1580         p->swap_file = NULL;
1581         p->swap_map = NULL;
1582         p->flags = 0;
1583         if (!(swap_flags & SWAP_FLAG_PREFER))
1584                 ++least_priority;
1585         spin_unlock(&swap_lock);
1586         vfree(swap_map);
1587         if (swap_file)
1588                 filp_close(swap_file, NULL);
1589 out:
1590         if (page && !IS_ERR(page)) {
1591                 kunmap(page);
1592                 page_cache_release(page);
1593         }
1594         if (name)
1595                 putname(name);
1596         if (did_down) {
1597                 if (!error)
1598                         inode->i_flags |= S_SWAPFILE;
1599                 mutex_unlock(&inode->i_mutex);
1600         }
1601         return error;
1602 }
1603
1604 void si_swapinfo(struct sysinfo *val)
1605 {
1606         unsigned int i;
1607         unsigned long nr_to_be_unused = 0;
1608
1609         spin_lock(&swap_lock);
1610         for (i = 0; i < nr_swapfiles; i++) {
1611                 if (!(swap_info[i].flags & SWP_USED) ||
1612                      (swap_info[i].flags & SWP_WRITEOK))
1613                         continue;
1614                 nr_to_be_unused += swap_info[i].inuse_pages;
1615         }
1616         val->freeswap = nr_swap_pages + nr_to_be_unused;
1617         val->totalswap = total_swap_pages + nr_to_be_unused;
1618         spin_unlock(&swap_lock);
1619 }
1620
1621 /*
1622  * Verify that a swap entry is valid and increment its swap map count.
1623  *
1624  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1625  * "permanent", but will be reclaimed by the next swapoff.
1626  */
1627 int swap_duplicate(swp_entry_t entry)
1628 {
1629         struct swap_info_struct * p;
1630         unsigned long offset, type;
1631         int result = 0;
1632
1633         type = swp_type(entry);
1634         if (type >= nr_swapfiles)
1635                 goto bad_file;
1636         p = type + swap_info;
1637         offset = swp_offset(entry);
1638
1639         spin_lock(&swap_lock);
1640         if (offset < p->max && p->swap_map[offset]) {
1641                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1642                         p->swap_map[offset]++;
1643                         result = 1;
1644                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1645                         if (swap_overflow++ < 5)
1646                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1647                         p->swap_map[offset] = SWAP_MAP_MAX;
1648                         result = 1;
1649                 }
1650         }
1651         spin_unlock(&swap_lock);
1652 out:
1653         return result;
1654
1655 bad_file:
1656         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1657         goto out;
1658 }
1659
1660 struct swap_info_struct *
1661 get_swap_info_struct(unsigned type)
1662 {
1663         return &swap_info[type];
1664 }
1665
1666 /*
1667  * swap_lock prevents swap_map being freed. Don't grab an extra
1668  * reference on the swaphandle, it doesn't matter if it becomes unused.
1669  */
1670 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1671 {
1672         int ret = 0, i = 1 << page_cluster;
1673         unsigned long toff;
1674         struct swap_info_struct *swapdev = swp_type(entry) + swap_info;
1675
1676         if (!page_cluster)      /* no readahead */
1677                 return 0;
1678         toff = (swp_offset(entry) >> page_cluster) << page_cluster;
1679         if (!toff)              /* first page is swap header */
1680                 toff++, i--;
1681         *offset = toff;
1682
1683         spin_lock(&swap_lock);
1684         do {
1685                 /* Don't read-ahead past the end of the swap area */
1686                 if (toff >= swapdev->max)
1687                         break;
1688                 /* Don't read in free or bad pages */
1689                 if (!swapdev->swap_map[toff])
1690                         break;
1691                 if (swapdev->swap_map[toff] == SWAP_MAP_BAD)
1692                         break;
1693                 toff++;
1694                 ret++;
1695         } while (--i);
1696         spin_unlock(&swap_lock);
1697         return ret;
1698 }