Linux-2.6.12-rc2
[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(swaplock);
35 unsigned int nr_swapfiles;
36 long total_swap_pages;
37 static int swap_overflow;
38
39 EXPORT_SYMBOL(total_swap_pages);
40
41 static const char Bad_file[] = "Bad swap file entry ";
42 static const char Unused_file[] = "Unused swap file entry ";
43 static const char Bad_offset[] = "Bad swap offset entry ";
44 static const char Unused_offset[] = "Unused swap offset entry ";
45
46 struct swap_list_t swap_list = {-1, -1};
47
48 struct swap_info_struct swap_info[MAX_SWAPFILES];
49
50 static DECLARE_MUTEX(swapon_sem);
51
52 /*
53  * We need this because the bdev->unplug_fn can sleep and we cannot
54  * hold swap_list_lock while calling the unplug_fn. And swap_list_lock
55  * cannot be turned into a semaphore.
56  */
57 static DECLARE_RWSEM(swap_unplug_sem);
58
59 #define SWAPFILE_CLUSTER 256
60
61 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
62 {
63         swp_entry_t entry;
64
65         down_read(&swap_unplug_sem);
66         entry.val = page->private;
67         if (PageSwapCache(page)) {
68                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
69                 struct backing_dev_info *bdi;
70
71                 /*
72                  * If the page is removed from swapcache from under us (with a
73                  * racy try_to_unuse/swapoff) we need an additional reference
74                  * count to avoid reading garbage from page->private above. If
75                  * the WARN_ON triggers during a swapoff it maybe the race
76                  * condition and it's harmless. However if it triggers without
77                  * swapoff it signals a problem.
78                  */
79                 WARN_ON(page_count(page) <= 1);
80
81                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
82                 bdi->unplug_io_fn(bdi, page);
83         }
84         up_read(&swap_unplug_sem);
85 }
86
87 static inline int scan_swap_map(struct swap_info_struct *si)
88 {
89         unsigned long offset;
90         /* 
91          * We try to cluster swap pages by allocating them
92          * sequentially in swap.  Once we've allocated
93          * SWAPFILE_CLUSTER pages this way, however, we resort to
94          * first-free allocation, starting a new cluster.  This
95          * prevents us from scattering swap pages all over the entire
96          * swap partition, so that we reduce overall disk seek times
97          * between swap pages.  -- sct */
98         if (si->cluster_nr) {
99                 while (si->cluster_next <= si->highest_bit) {
100                         offset = si->cluster_next++;
101                         if (si->swap_map[offset])
102                                 continue;
103                         si->cluster_nr--;
104                         goto got_page;
105                 }
106         }
107         si->cluster_nr = SWAPFILE_CLUSTER;
108
109         /* try to find an empty (even not aligned) cluster. */
110         offset = si->lowest_bit;
111  check_next_cluster:
112         if (offset+SWAPFILE_CLUSTER-1 <= si->highest_bit)
113         {
114                 unsigned long nr;
115                 for (nr = offset; nr < offset+SWAPFILE_CLUSTER; nr++)
116                         if (si->swap_map[nr])
117                         {
118                                 offset = nr+1;
119                                 goto check_next_cluster;
120                         }
121                 /* We found a completly empty cluster, so start
122                  * using it.
123                  */
124                 goto got_page;
125         }
126         /* No luck, so now go finegrined as usual. -Andrea */
127         for (offset = si->lowest_bit; offset <= si->highest_bit ; offset++) {
128                 if (si->swap_map[offset])
129                         continue;
130                 si->lowest_bit = offset+1;
131         got_page:
132                 if (offset == si->lowest_bit)
133                         si->lowest_bit++;
134                 if (offset == si->highest_bit)
135                         si->highest_bit--;
136                 if (si->lowest_bit > si->highest_bit) {
137                         si->lowest_bit = si->max;
138                         si->highest_bit = 0;
139                 }
140                 si->swap_map[offset] = 1;
141                 si->inuse_pages++;
142                 nr_swap_pages--;
143                 si->cluster_next = offset+1;
144                 return offset;
145         }
146         si->lowest_bit = si->max;
147         si->highest_bit = 0;
148         return 0;
149 }
150
151 swp_entry_t get_swap_page(void)
152 {
153         struct swap_info_struct * p;
154         unsigned long offset;
155         swp_entry_t entry;
156         int type, wrapped = 0;
157
158         entry.val = 0;  /* Out of memory */
159         swap_list_lock();
160         type = swap_list.next;
161         if (type < 0)
162                 goto out;
163         if (nr_swap_pages <= 0)
164                 goto out;
165
166         while (1) {
167                 p = &swap_info[type];
168                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
169                         swap_device_lock(p);
170                         offset = scan_swap_map(p);
171                         swap_device_unlock(p);
172                         if (offset) {
173                                 entry = swp_entry(type,offset);
174                                 type = swap_info[type].next;
175                                 if (type < 0 ||
176                                         p->prio != swap_info[type].prio) {
177                                                 swap_list.next = swap_list.head;
178                                 } else {
179                                         swap_list.next = type;
180                                 }
181                                 goto out;
182                         }
183                 }
184                 type = p->next;
185                 if (!wrapped) {
186                         if (type < 0 || p->prio != swap_info[type].prio) {
187                                 type = swap_list.head;
188                                 wrapped = 1;
189                         }
190                 } else
191                         if (type < 0)
192                                 goto out;       /* out of swap space */
193         }
194 out:
195         swap_list_unlock();
196         return entry;
197 }
198
199 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
200 {
201         struct swap_info_struct * p;
202         unsigned long offset, type;
203
204         if (!entry.val)
205                 goto out;
206         type = swp_type(entry);
207         if (type >= nr_swapfiles)
208                 goto bad_nofile;
209         p = & swap_info[type];
210         if (!(p->flags & SWP_USED))
211                 goto bad_device;
212         offset = swp_offset(entry);
213         if (offset >= p->max)
214                 goto bad_offset;
215         if (!p->swap_map[offset])
216                 goto bad_free;
217         swap_list_lock();
218         if (p->prio > swap_info[swap_list.next].prio)
219                 swap_list.next = type;
220         swap_device_lock(p);
221         return p;
222
223 bad_free:
224         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
225         goto out;
226 bad_offset:
227         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
228         goto out;
229 bad_device:
230         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
231         goto out;
232 bad_nofile:
233         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
234 out:
235         return NULL;
236 }       
237
238 static void swap_info_put(struct swap_info_struct * p)
239 {
240         swap_device_unlock(p);
241         swap_list_unlock();
242 }
243
244 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
245 {
246         int count = p->swap_map[offset];
247
248         if (count < SWAP_MAP_MAX) {
249                 count--;
250                 p->swap_map[offset] = count;
251                 if (!count) {
252                         if (offset < p->lowest_bit)
253                                 p->lowest_bit = offset;
254                         if (offset > p->highest_bit)
255                                 p->highest_bit = offset;
256                         nr_swap_pages++;
257                         p->inuse_pages--;
258                 }
259         }
260         return count;
261 }
262
263 /*
264  * Caller has made sure that the swapdevice corresponding to entry
265  * is still around or has not been recycled.
266  */
267 void swap_free(swp_entry_t entry)
268 {
269         struct swap_info_struct * p;
270
271         p = swap_info_get(entry);
272         if (p) {
273                 swap_entry_free(p, swp_offset(entry));
274                 swap_info_put(p);
275         }
276 }
277
278 /*
279  * Check if we're the only user of a swap page,
280  * when the page is locked.
281  */
282 static int exclusive_swap_page(struct page *page)
283 {
284         int retval = 0;
285         struct swap_info_struct * p;
286         swp_entry_t entry;
287
288         entry.val = page->private;
289         p = swap_info_get(entry);
290         if (p) {
291                 /* Is the only swap cache user the cache itself? */
292                 if (p->swap_map[swp_offset(entry)] == 1) {
293                         /* Recheck the page count with the swapcache lock held.. */
294                         write_lock_irq(&swapper_space.tree_lock);
295                         if (page_count(page) == 2)
296                                 retval = 1;
297                         write_unlock_irq(&swapper_space.tree_lock);
298                 }
299                 swap_info_put(p);
300         }
301         return retval;
302 }
303
304 /*
305  * We can use this swap cache entry directly
306  * if there are no other references to it.
307  *
308  * Here "exclusive_swap_page()" does the real
309  * work, but we opportunistically check whether
310  * we need to get all the locks first..
311  */
312 int can_share_swap_page(struct page *page)
313 {
314         int retval = 0;
315
316         if (!PageLocked(page))
317                 BUG();
318         switch (page_count(page)) {
319         case 3:
320                 if (!PagePrivate(page))
321                         break;
322                 /* Fallthrough */
323         case 2:
324                 if (!PageSwapCache(page))
325                         break;
326                 retval = exclusive_swap_page(page);
327                 break;
328         case 1:
329                 if (PageReserved(page))
330                         break;
331                 retval = 1;
332         }
333         return retval;
334 }
335
336 /*
337  * Work out if there are any other processes sharing this
338  * swap cache page. Free it if you can. Return success.
339  */
340 int remove_exclusive_swap_page(struct page *page)
341 {
342         int retval;
343         struct swap_info_struct * p;
344         swp_entry_t entry;
345
346         BUG_ON(PagePrivate(page));
347         BUG_ON(!PageLocked(page));
348
349         if (!PageSwapCache(page))
350                 return 0;
351         if (PageWriteback(page))
352                 return 0;
353         if (page_count(page) != 2) /* 2: us + cache */
354                 return 0;
355
356         entry.val = page->private;
357         p = swap_info_get(entry);
358         if (!p)
359                 return 0;
360
361         /* Is the only swap cache user the cache itself? */
362         retval = 0;
363         if (p->swap_map[swp_offset(entry)] == 1) {
364                 /* Recheck the page count with the swapcache lock held.. */
365                 write_lock_irq(&swapper_space.tree_lock);
366                 if ((page_count(page) == 2) && !PageWriteback(page)) {
367                         __delete_from_swap_cache(page);
368                         SetPageDirty(page);
369                         retval = 1;
370                 }
371                 write_unlock_irq(&swapper_space.tree_lock);
372         }
373         swap_info_put(p);
374
375         if (retval) {
376                 swap_free(entry);
377                 page_cache_release(page);
378         }
379
380         return retval;
381 }
382
383 /*
384  * Free the swap entry like above, but also try to
385  * free the page cache entry if it is the last user.
386  */
387 void free_swap_and_cache(swp_entry_t entry)
388 {
389         struct swap_info_struct * p;
390         struct page *page = NULL;
391
392         p = swap_info_get(entry);
393         if (p) {
394                 if (swap_entry_free(p, swp_offset(entry)) == 1)
395                         page = find_trylock_page(&swapper_space, entry.val);
396                 swap_info_put(p);
397         }
398         if (page) {
399                 int one_user;
400
401                 BUG_ON(PagePrivate(page));
402                 page_cache_get(page);
403                 one_user = (page_count(page) == 2);
404                 /* Only cache user (+us), or swap space full? Free it! */
405                 if (!PageWriteback(page) && (one_user || vm_swap_full())) {
406                         delete_from_swap_cache(page);
407                         SetPageDirty(page);
408                 }
409                 unlock_page(page);
410                 page_cache_release(page);
411         }
412 }
413
414 /*
415  * Always set the resulting pte to be nowrite (the same as COW pages
416  * after one process has exited).  We don't know just how many PTEs will
417  * share this swap entry, so be cautious and let do_wp_page work out
418  * what to do if a write is requested later.
419  *
420  * vma->vm_mm->page_table_lock is held.
421  */
422 static void unuse_pte(struct vm_area_struct *vma, pte_t *pte,
423                 unsigned long addr, swp_entry_t entry, struct page *page)
424 {
425         inc_mm_counter(vma->vm_mm, rss);
426         get_page(page);
427         set_pte_at(vma->vm_mm, addr, pte,
428                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
429         page_add_anon_rmap(page, vma, addr);
430         swap_free(entry);
431         /*
432          * Move the page to the active list so it is not
433          * immediately swapped out again after swapon.
434          */
435         activate_page(page);
436 }
437
438 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
439                                 unsigned long addr, unsigned long end,
440                                 swp_entry_t entry, struct page *page)
441 {
442         pte_t *pte;
443         pte_t swp_pte = swp_entry_to_pte(entry);
444
445         pte = pte_offset_map(pmd, addr);
446         do {
447                 /*
448                  * swapoff spends a _lot_ of time in this loop!
449                  * Test inline before going to call unuse_pte.
450                  */
451                 if (unlikely(pte_same(*pte, swp_pte))) {
452                         unuse_pte(vma, pte, addr, entry, page);
453                         pte_unmap(pte);
454                         return 1;
455                 }
456         } while (pte++, addr += PAGE_SIZE, addr != end);
457         pte_unmap(pte - 1);
458         return 0;
459 }
460
461 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
462                                 unsigned long addr, unsigned long end,
463                                 swp_entry_t entry, struct page *page)
464 {
465         pmd_t *pmd;
466         unsigned long next;
467
468         pmd = pmd_offset(pud, addr);
469         do {
470                 next = pmd_addr_end(addr, end);
471                 if (pmd_none_or_clear_bad(pmd))
472                         continue;
473                 if (unuse_pte_range(vma, pmd, addr, next, entry, page))
474                         return 1;
475         } while (pmd++, addr = next, addr != end);
476         return 0;
477 }
478
479 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
480                                 unsigned long addr, unsigned long end,
481                                 swp_entry_t entry, struct page *page)
482 {
483         pud_t *pud;
484         unsigned long next;
485
486         pud = pud_offset(pgd, addr);
487         do {
488                 next = pud_addr_end(addr, end);
489                 if (pud_none_or_clear_bad(pud))
490                         continue;
491                 if (unuse_pmd_range(vma, pud, addr, next, entry, page))
492                         return 1;
493         } while (pud++, addr = next, addr != end);
494         return 0;
495 }
496
497 static int unuse_vma(struct vm_area_struct *vma,
498                                 swp_entry_t entry, struct page *page)
499 {
500         pgd_t *pgd;
501         unsigned long addr, end, next;
502
503         if (page->mapping) {
504                 addr = page_address_in_vma(page, vma);
505                 if (addr == -EFAULT)
506                         return 0;
507                 else
508                         end = addr + PAGE_SIZE;
509         } else {
510                 addr = vma->vm_start;
511                 end = vma->vm_end;
512         }
513
514         pgd = pgd_offset(vma->vm_mm, addr);
515         do {
516                 next = pgd_addr_end(addr, end);
517                 if (pgd_none_or_clear_bad(pgd))
518                         continue;
519                 if (unuse_pud_range(vma, pgd, addr, next, entry, page))
520                         return 1;
521         } while (pgd++, addr = next, addr != end);
522         return 0;
523 }
524
525 static int unuse_mm(struct mm_struct *mm,
526                                 swp_entry_t entry, struct page *page)
527 {
528         struct vm_area_struct *vma;
529
530         if (!down_read_trylock(&mm->mmap_sem)) {
531                 /*
532                  * Our reference to the page stops try_to_unmap_one from
533                  * unmapping its ptes, so swapoff can make progress.
534                  */
535                 unlock_page(page);
536                 down_read(&mm->mmap_sem);
537                 lock_page(page);
538         }
539         spin_lock(&mm->page_table_lock);
540         for (vma = mm->mmap; vma; vma = vma->vm_next) {
541                 if (vma->anon_vma && unuse_vma(vma, entry, page))
542                         break;
543         }
544         spin_unlock(&mm->page_table_lock);
545         up_read(&mm->mmap_sem);
546         /*
547          * Currently unuse_mm cannot fail, but leave error handling
548          * at call sites for now, since we change it from time to time.
549          */
550         return 0;
551 }
552
553 /*
554  * Scan swap_map from current position to next entry still in use.
555  * Recycle to start on reaching the end, returning 0 when empty.
556  */
557 static int find_next_to_unuse(struct swap_info_struct *si, int prev)
558 {
559         int max = si->max;
560         int i = prev;
561         int count;
562
563         /*
564          * No need for swap_device_lock(si) here: we're just looking
565          * for whether an entry is in use, not modifying it; false
566          * hits are okay, and sys_swapoff() has already prevented new
567          * allocations from this area (while holding swap_list_lock()).
568          */
569         for (;;) {
570                 if (++i >= max) {
571                         if (!prev) {
572                                 i = 0;
573                                 break;
574                         }
575                         /*
576                          * No entries in use at top of swap_map,
577                          * loop back to start and recheck there.
578                          */
579                         max = prev + 1;
580                         prev = 0;
581                         i = 1;
582                 }
583                 count = si->swap_map[i];
584                 if (count && count != SWAP_MAP_BAD)
585                         break;
586         }
587         return i;
588 }
589
590 /*
591  * We completely avoid races by reading each swap page in advance,
592  * and then search for the process using it.  All the necessary
593  * page table adjustments can then be made atomically.
594  */
595 static int try_to_unuse(unsigned int type)
596 {
597         struct swap_info_struct * si = &swap_info[type];
598         struct mm_struct *start_mm;
599         unsigned short *swap_map;
600         unsigned short swcount;
601         struct page *page;
602         swp_entry_t entry;
603         int i = 0;
604         int retval = 0;
605         int reset_overflow = 0;
606         int shmem;
607
608         /*
609          * When searching mms for an entry, a good strategy is to
610          * start at the first mm we freed the previous entry from
611          * (though actually we don't notice whether we or coincidence
612          * freed the entry).  Initialize this start_mm with a hold.
613          *
614          * A simpler strategy would be to start at the last mm we
615          * freed the previous entry from; but that would take less
616          * advantage of mmlist ordering, which clusters forked mms
617          * together, child after parent.  If we race with dup_mmap(), we
618          * prefer to resolve parent before child, lest we miss entries
619          * duplicated after we scanned child: using last mm would invert
620          * that.  Though it's only a serious concern when an overflowed
621          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
622          */
623         start_mm = &init_mm;
624         atomic_inc(&init_mm.mm_users);
625
626         /*
627          * Keep on scanning until all entries have gone.  Usually,
628          * one pass through swap_map is enough, but not necessarily:
629          * there are races when an instance of an entry might be missed.
630          */
631         while ((i = find_next_to_unuse(si, i)) != 0) {
632                 if (signal_pending(current)) {
633                         retval = -EINTR;
634                         break;
635                 }
636
637                 /* 
638                  * Get a page for the entry, using the existing swap
639                  * cache page if there is one.  Otherwise, get a clean
640                  * page and read the swap into it. 
641                  */
642                 swap_map = &si->swap_map[i];
643                 entry = swp_entry(type, i);
644                 page = read_swap_cache_async(entry, NULL, 0);
645                 if (!page) {
646                         /*
647                          * Either swap_duplicate() failed because entry
648                          * has been freed independently, and will not be
649                          * reused since sys_swapoff() already disabled
650                          * allocation from here, or alloc_page() failed.
651                          */
652                         if (!*swap_map)
653                                 continue;
654                         retval = -ENOMEM;
655                         break;
656                 }
657
658                 /*
659                  * Don't hold on to start_mm if it looks like exiting.
660                  */
661                 if (atomic_read(&start_mm->mm_users) == 1) {
662                         mmput(start_mm);
663                         start_mm = &init_mm;
664                         atomic_inc(&init_mm.mm_users);
665                 }
666
667                 /*
668                  * Wait for and lock page.  When do_swap_page races with
669                  * try_to_unuse, do_swap_page can handle the fault much
670                  * faster than try_to_unuse can locate the entry.  This
671                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
672                  * defer to do_swap_page in such a case - in some tests,
673                  * do_swap_page and try_to_unuse repeatedly compete.
674                  */
675                 wait_on_page_locked(page);
676                 wait_on_page_writeback(page);
677                 lock_page(page);
678                 wait_on_page_writeback(page);
679
680                 /*
681                  * Remove all references to entry.
682                  * Whenever we reach init_mm, there's no address space
683                  * to search, but use it as a reminder to search shmem.
684                  */
685                 shmem = 0;
686                 swcount = *swap_map;
687                 if (swcount > 1) {
688                         if (start_mm == &init_mm)
689                                 shmem = shmem_unuse(entry, page);
690                         else
691                                 retval = unuse_mm(start_mm, entry, page);
692                 }
693                 if (*swap_map > 1) {
694                         int set_start_mm = (*swap_map >= swcount);
695                         struct list_head *p = &start_mm->mmlist;
696                         struct mm_struct *new_start_mm = start_mm;
697                         struct mm_struct *prev_mm = start_mm;
698                         struct mm_struct *mm;
699
700                         atomic_inc(&new_start_mm->mm_users);
701                         atomic_inc(&prev_mm->mm_users);
702                         spin_lock(&mmlist_lock);
703                         while (*swap_map > 1 && !retval &&
704                                         (p = p->next) != &start_mm->mmlist) {
705                                 mm = list_entry(p, struct mm_struct, mmlist);
706                                 if (atomic_inc_return(&mm->mm_users) == 1) {
707                                         atomic_dec(&mm->mm_users);
708                                         continue;
709                                 }
710                                 spin_unlock(&mmlist_lock);
711                                 mmput(prev_mm);
712                                 prev_mm = mm;
713
714                                 cond_resched();
715
716                                 swcount = *swap_map;
717                                 if (swcount <= 1)
718                                         ;
719                                 else if (mm == &init_mm) {
720                                         set_start_mm = 1;
721                                         shmem = shmem_unuse(entry, page);
722                                 } else
723                                         retval = unuse_mm(mm, entry, page);
724                                 if (set_start_mm && *swap_map < swcount) {
725                                         mmput(new_start_mm);
726                                         atomic_inc(&mm->mm_users);
727                                         new_start_mm = mm;
728                                         set_start_mm = 0;
729                                 }
730                                 spin_lock(&mmlist_lock);
731                         }
732                         spin_unlock(&mmlist_lock);
733                         mmput(prev_mm);
734                         mmput(start_mm);
735                         start_mm = new_start_mm;
736                 }
737                 if (retval) {
738                         unlock_page(page);
739                         page_cache_release(page);
740                         break;
741                 }
742
743                 /*
744                  * How could swap count reach 0x7fff when the maximum
745                  * pid is 0x7fff, and there's no way to repeat a swap
746                  * page within an mm (except in shmem, where it's the
747                  * shared object which takes the reference count)?
748                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
749                  *
750                  * If that's wrong, then we should worry more about
751                  * exit_mmap() and do_munmap() cases described above:
752                  * we might be resetting SWAP_MAP_MAX too early here.
753                  * We know "Undead"s can happen, they're okay, so don't
754                  * report them; but do report if we reset SWAP_MAP_MAX.
755                  */
756                 if (*swap_map == SWAP_MAP_MAX) {
757                         swap_device_lock(si);
758                         *swap_map = 1;
759                         swap_device_unlock(si);
760                         reset_overflow = 1;
761                 }
762
763                 /*
764                  * If a reference remains (rare), we would like to leave
765                  * the page in the swap cache; but try_to_unmap could
766                  * then re-duplicate the entry once we drop page lock,
767                  * so we might loop indefinitely; also, that page could
768                  * not be swapped out to other storage meanwhile.  So:
769                  * delete from cache even if there's another reference,
770                  * after ensuring that the data has been saved to disk -
771                  * since if the reference remains (rarer), it will be
772                  * read from disk into another page.  Splitting into two
773                  * pages would be incorrect if swap supported "shared
774                  * private" pages, but they are handled by tmpfs files.
775                  *
776                  * Note shmem_unuse already deleted a swappage from
777                  * the swap cache, unless the move to filepage failed:
778                  * in which case it left swappage in cache, lowered its
779                  * swap count to pass quickly through the loops above,
780                  * and now we must reincrement count to try again later.
781                  */
782                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
783                         struct writeback_control wbc = {
784                                 .sync_mode = WB_SYNC_NONE,
785                         };
786
787                         swap_writepage(page, &wbc);
788                         lock_page(page);
789                         wait_on_page_writeback(page);
790                 }
791                 if (PageSwapCache(page)) {
792                         if (shmem)
793                                 swap_duplicate(entry);
794                         else
795                                 delete_from_swap_cache(page);
796                 }
797
798                 /*
799                  * So we could skip searching mms once swap count went
800                  * to 1, we did not mark any present ptes as dirty: must
801                  * mark page dirty so shrink_list will preserve it.
802                  */
803                 SetPageDirty(page);
804                 unlock_page(page);
805                 page_cache_release(page);
806
807                 /*
808                  * Make sure that we aren't completely killing
809                  * interactive performance.
810                  */
811                 cond_resched();
812         }
813
814         mmput(start_mm);
815         if (reset_overflow) {
816                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
817                 swap_overflow = 0;
818         }
819         return retval;
820 }
821
822 /*
823  * After a successful try_to_unuse, if no swap is now in use, we know we
824  * can empty the mmlist.  swap_list_lock must be held on entry and exit.
825  * Note that mmlist_lock nests inside swap_list_lock, and an mm must be
826  * added to the mmlist just after page_duplicate - before would be racy.
827  */
828 static void drain_mmlist(void)
829 {
830         struct list_head *p, *next;
831         unsigned int i;
832
833         for (i = 0; i < nr_swapfiles; i++)
834                 if (swap_info[i].inuse_pages)
835                         return;
836         spin_lock(&mmlist_lock);
837         list_for_each_safe(p, next, &init_mm.mmlist)
838                 list_del_init(p);
839         spin_unlock(&mmlist_lock);
840 }
841
842 /*
843  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
844  * corresponds to page offset `offset'.
845  */
846 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
847 {
848         struct swap_extent *se = sis->curr_swap_extent;
849         struct swap_extent *start_se = se;
850
851         for ( ; ; ) {
852                 struct list_head *lh;
853
854                 if (se->start_page <= offset &&
855                                 offset < (se->start_page + se->nr_pages)) {
856                         return se->start_block + (offset - se->start_page);
857                 }
858                 lh = se->list.prev;
859                 if (lh == &sis->extent_list)
860                         lh = lh->prev;
861                 se = list_entry(lh, struct swap_extent, list);
862                 sis->curr_swap_extent = se;
863                 BUG_ON(se == start_se);         /* It *must* be present */
864         }
865 }
866
867 /*
868  * Free all of a swapdev's extent information
869  */
870 static void destroy_swap_extents(struct swap_info_struct *sis)
871 {
872         while (!list_empty(&sis->extent_list)) {
873                 struct swap_extent *se;
874
875                 se = list_entry(sis->extent_list.next,
876                                 struct swap_extent, list);
877                 list_del(&se->list);
878                 kfree(se);
879         }
880         sis->nr_extents = 0;
881 }
882
883 /*
884  * Add a block range (and the corresponding page range) into this swapdev's
885  * extent list.  The extent list is kept sorted in block order.
886  *
887  * This function rather assumes that it is called in ascending sector_t order.
888  * It doesn't look for extent coalescing opportunities.
889  */
890 static int
891 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
892                 unsigned long nr_pages, sector_t start_block)
893 {
894         struct swap_extent *se;
895         struct swap_extent *new_se;
896         struct list_head *lh;
897
898         lh = sis->extent_list.next;     /* The highest-addressed block */
899         while (lh != &sis->extent_list) {
900                 se = list_entry(lh, struct swap_extent, list);
901                 if (se->start_block + se->nr_pages == start_block &&
902                     se->start_page  + se->nr_pages == start_page) {
903                         /* Merge it */
904                         se->nr_pages += nr_pages;
905                         return 0;
906                 }
907                 lh = lh->next;
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         lh = sis->extent_list.prev;     /* The lowest block */
921         while (lh != &sis->extent_list) {
922                 se = list_entry(lh, struct swap_extent, list);
923                 if (se->start_block > start_block)
924                         break;
925                 lh = lh->prev;
926         }
927         list_add_tail(&new_se->list, lh);
928         sis->nr_extents++;
929         return 0;
930 }
931
932 /*
933  * A `swap extent' is a simple thing which maps a contiguous range of pages
934  * onto a contiguous range of disk blocks.  An ordered list of swap extents
935  * is built at swapon time and is then used at swap_writepage/swap_readpage
936  * time for locating where on disk a page belongs.
937  *
938  * If the swapfile is an S_ISBLK block device, a single extent is installed.
939  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
940  * swap files identically.
941  *
942  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
943  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
944  * swapfiles are handled *identically* after swapon time.
945  *
946  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
947  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
948  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
949  * requirements, they are simply tossed out - we will never use those blocks
950  * for swapping.
951  *
952  * For S_ISREG swapfiles we hold i_sem across the life of the swapon.  This
953  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
954  * which will scribble on the fs.
955  *
956  * The amount of disk space which a single swap extent represents varies.
957  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
958  * extents in the list.  To avoid much list walking, we cache the previous
959  * search location in `curr_swap_extent', and start new searches from there.
960  * This is extremely effective.  The average number of iterations in
961  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
962  */
963 static int setup_swap_extents(struct swap_info_struct *sis)
964 {
965         struct inode *inode;
966         unsigned blocks_per_page;
967         unsigned long page_no;
968         unsigned blkbits;
969         sector_t probe_block;
970         sector_t last_block;
971         int ret;
972
973         inode = sis->swap_file->f_mapping->host;
974         if (S_ISBLK(inode->i_mode)) {
975                 ret = add_swap_extent(sis, 0, sis->max, 0);
976                 goto done;
977         }
978
979         blkbits = inode->i_blkbits;
980         blocks_per_page = PAGE_SIZE >> blkbits;
981
982         /*
983          * Map all the blocks into the extent list.  This code doesn't try
984          * to be very smart.
985          */
986         probe_block = 0;
987         page_no = 0;
988         last_block = i_size_read(inode) >> blkbits;
989         while ((probe_block + blocks_per_page) <= last_block &&
990                         page_no < sis->max) {
991                 unsigned block_in_page;
992                 sector_t first_block;
993
994                 first_block = bmap(inode, probe_block);
995                 if (first_block == 0)
996                         goto bad_bmap;
997
998                 /*
999                  * It must be PAGE_SIZE aligned on-disk
1000                  */
1001                 if (first_block & (blocks_per_page - 1)) {
1002                         probe_block++;
1003                         goto reprobe;
1004                 }
1005
1006                 for (block_in_page = 1; block_in_page < blocks_per_page;
1007                                         block_in_page++) {
1008                         sector_t block;
1009
1010                         block = bmap(inode, probe_block + block_in_page);
1011                         if (block == 0)
1012                                 goto bad_bmap;
1013                         if (block != first_block + block_in_page) {
1014                                 /* Discontiguity */
1015                                 probe_block++;
1016                                 goto reprobe;
1017                         }
1018                 }
1019
1020                 /*
1021                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1022                  */
1023                 ret = add_swap_extent(sis, page_no, 1,
1024                                 first_block >> (PAGE_SHIFT - blkbits));
1025                 if (ret)
1026                         goto out;
1027                 page_no++;
1028                 probe_block += blocks_per_page;
1029 reprobe:
1030                 continue;
1031         }
1032         ret = 0;
1033         if (page_no == 0)
1034                 ret = -EINVAL;
1035         sis->max = page_no;
1036         sis->highest_bit = page_no - 1;
1037 done:
1038         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1039                                         struct swap_extent, list);
1040         goto out;
1041 bad_bmap:
1042         printk(KERN_ERR "swapon: swapfile has holes\n");
1043         ret = -EINVAL;
1044 out:
1045         return ret;
1046 }
1047
1048 #if 0   /* We don't need this yet */
1049 #include <linux/backing-dev.h>
1050 int page_queue_congested(struct page *page)
1051 {
1052         struct backing_dev_info *bdi;
1053
1054         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1055
1056         if (PageSwapCache(page)) {
1057                 swp_entry_t entry = { .val = page->private };
1058                 struct swap_info_struct *sis;
1059
1060                 sis = get_swap_info_struct(swp_type(entry));
1061                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1062         } else
1063                 bdi = page->mapping->backing_dev_info;
1064         return bdi_write_congested(bdi);
1065 }
1066 #endif
1067
1068 asmlinkage long sys_swapoff(const char __user * specialfile)
1069 {
1070         struct swap_info_struct * p = NULL;
1071         unsigned short *swap_map;
1072         struct file *swap_file, *victim;
1073         struct address_space *mapping;
1074         struct inode *inode;
1075         char * pathname;
1076         int i, type, prev;
1077         int err;
1078         
1079         if (!capable(CAP_SYS_ADMIN))
1080                 return -EPERM;
1081
1082         pathname = getname(specialfile);
1083         err = PTR_ERR(pathname);
1084         if (IS_ERR(pathname))
1085                 goto out;
1086
1087         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1088         putname(pathname);
1089         err = PTR_ERR(victim);
1090         if (IS_ERR(victim))
1091                 goto out;
1092
1093         mapping = victim->f_mapping;
1094         prev = -1;
1095         swap_list_lock();
1096         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1097                 p = swap_info + type;
1098                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1099                         if (p->swap_file->f_mapping == mapping)
1100                                 break;
1101                 }
1102                 prev = type;
1103         }
1104         if (type < 0) {
1105                 err = -EINVAL;
1106                 swap_list_unlock();
1107                 goto out_dput;
1108         }
1109         if (!security_vm_enough_memory(p->pages))
1110                 vm_unacct_memory(p->pages);
1111         else {
1112                 err = -ENOMEM;
1113                 swap_list_unlock();
1114                 goto out_dput;
1115         }
1116         if (prev < 0) {
1117                 swap_list.head = p->next;
1118         } else {
1119                 swap_info[prev].next = p->next;
1120         }
1121         if (type == swap_list.next) {
1122                 /* just pick something that's safe... */
1123                 swap_list.next = swap_list.head;
1124         }
1125         nr_swap_pages -= p->pages;
1126         total_swap_pages -= p->pages;
1127         p->flags &= ~SWP_WRITEOK;
1128         swap_list_unlock();
1129         current->flags |= PF_SWAPOFF;
1130         err = try_to_unuse(type);
1131         current->flags &= ~PF_SWAPOFF;
1132
1133         /* wait for any unplug function to finish */
1134         down_write(&swap_unplug_sem);
1135         up_write(&swap_unplug_sem);
1136
1137         if (err) {
1138                 /* re-insert swap space back into swap_list */
1139                 swap_list_lock();
1140                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1141                         if (p->prio >= swap_info[i].prio)
1142                                 break;
1143                 p->next = i;
1144                 if (prev < 0)
1145                         swap_list.head = swap_list.next = p - swap_info;
1146                 else
1147                         swap_info[prev].next = p - swap_info;
1148                 nr_swap_pages += p->pages;
1149                 total_swap_pages += p->pages;
1150                 p->flags |= SWP_WRITEOK;
1151                 swap_list_unlock();
1152                 goto out_dput;
1153         }
1154         down(&swapon_sem);
1155         swap_list_lock();
1156         drain_mmlist();
1157         swap_device_lock(p);
1158         swap_file = p->swap_file;
1159         p->swap_file = NULL;
1160         p->max = 0;
1161         swap_map = p->swap_map;
1162         p->swap_map = NULL;
1163         p->flags = 0;
1164         destroy_swap_extents(p);
1165         swap_device_unlock(p);
1166         swap_list_unlock();
1167         up(&swapon_sem);
1168         vfree(swap_map);
1169         inode = mapping->host;
1170         if (S_ISBLK(inode->i_mode)) {
1171                 struct block_device *bdev = I_BDEV(inode);
1172                 set_blocksize(bdev, p->old_block_size);
1173                 bd_release(bdev);
1174         } else {
1175                 down(&inode->i_sem);
1176                 inode->i_flags &= ~S_SWAPFILE;
1177                 up(&inode->i_sem);
1178         }
1179         filp_close(swap_file, NULL);
1180         err = 0;
1181
1182 out_dput:
1183         filp_close(victim, NULL);
1184 out:
1185         return err;
1186 }
1187
1188 #ifdef CONFIG_PROC_FS
1189 /* iterator */
1190 static void *swap_start(struct seq_file *swap, loff_t *pos)
1191 {
1192         struct swap_info_struct *ptr = swap_info;
1193         int i;
1194         loff_t l = *pos;
1195
1196         down(&swapon_sem);
1197
1198         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1199                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1200                         continue;
1201                 if (!l--)
1202                         return ptr;
1203         }
1204
1205         return NULL;
1206 }
1207
1208 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1209 {
1210         struct swap_info_struct *ptr = v;
1211         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1212
1213         for (++ptr; ptr < endptr; ptr++) {
1214                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1215                         continue;
1216                 ++*pos;
1217                 return ptr;
1218         }
1219
1220         return NULL;
1221 }
1222
1223 static void swap_stop(struct seq_file *swap, void *v)
1224 {
1225         up(&swapon_sem);
1226 }
1227
1228 static int swap_show(struct seq_file *swap, void *v)
1229 {
1230         struct swap_info_struct *ptr = v;
1231         struct file *file;
1232         int len;
1233
1234         if (v == swap_info)
1235                 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1236
1237         file = ptr->swap_file;
1238         len = seq_path(swap, file->f_vfsmnt, file->f_dentry, " \t\n\\");
1239         seq_printf(swap, "%*s%s\t%d\t%ld\t%d\n",
1240                        len < 40 ? 40 - len : 1, " ",
1241                        S_ISBLK(file->f_dentry->d_inode->i_mode) ?
1242                                 "partition" : "file\t",
1243                        ptr->pages << (PAGE_SHIFT - 10),
1244                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1245                        ptr->prio);
1246         return 0;
1247 }
1248
1249 static struct seq_operations swaps_op = {
1250         .start =        swap_start,
1251         .next =         swap_next,
1252         .stop =         swap_stop,
1253         .show =         swap_show
1254 };
1255
1256 static int swaps_open(struct inode *inode, struct file *file)
1257 {
1258         return seq_open(file, &swaps_op);
1259 }
1260
1261 static struct file_operations proc_swaps_operations = {
1262         .open           = swaps_open,
1263         .read           = seq_read,
1264         .llseek         = seq_lseek,
1265         .release        = seq_release,
1266 };
1267
1268 static int __init procswaps_init(void)
1269 {
1270         struct proc_dir_entry *entry;
1271
1272         entry = create_proc_entry("swaps", 0, NULL);
1273         if (entry)
1274                 entry->proc_fops = &proc_swaps_operations;
1275         return 0;
1276 }
1277 __initcall(procswaps_init);
1278 #endif /* CONFIG_PROC_FS */
1279
1280 /*
1281  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1282  *
1283  * The swapon system call
1284  */
1285 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1286 {
1287         struct swap_info_struct * p;
1288         char *name = NULL;
1289         struct block_device *bdev = NULL;
1290         struct file *swap_file = NULL;
1291         struct address_space *mapping;
1292         unsigned int type;
1293         int i, prev;
1294         int error;
1295         static int least_priority;
1296         union swap_header *swap_header = NULL;
1297         int swap_header_version;
1298         int nr_good_pages = 0;
1299         unsigned long maxpages = 1;
1300         int swapfilesize;
1301         unsigned short *swap_map;
1302         struct page *page = NULL;
1303         struct inode *inode = NULL;
1304         int did_down = 0;
1305
1306         if (!capable(CAP_SYS_ADMIN))
1307                 return -EPERM;
1308         swap_list_lock();
1309         p = swap_info;
1310         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1311                 if (!(p->flags & SWP_USED))
1312                         break;
1313         error = -EPERM;
1314         /*
1315          * Test if adding another swap device is possible. There are
1316          * two limiting factors: 1) the number of bits for the swap
1317          * type swp_entry_t definition and 2) the number of bits for
1318          * the swap type in the swap ptes as defined by the different
1319          * architectures. To honor both limitations a swap entry
1320          * with swap offset 0 and swap type ~0UL is created, encoded
1321          * to a swap pte, decoded to a swp_entry_t again and finally
1322          * the swap type part is extracted. This will mask all bits
1323          * from the initial ~0UL that can't be encoded in either the
1324          * swp_entry_t or the architecture definition of a swap pte.
1325          */
1326         if (type > swp_type(pte_to_swp_entry(swp_entry_to_pte(swp_entry(~0UL,0))))) {
1327                 swap_list_unlock();
1328                 goto out;
1329         }
1330         if (type >= nr_swapfiles)
1331                 nr_swapfiles = type+1;
1332         INIT_LIST_HEAD(&p->extent_list);
1333         p->flags = SWP_USED;
1334         p->nr_extents = 0;
1335         p->swap_file = NULL;
1336         p->old_block_size = 0;
1337         p->swap_map = NULL;
1338         p->lowest_bit = 0;
1339         p->highest_bit = 0;
1340         p->cluster_nr = 0;
1341         p->inuse_pages = 0;
1342         spin_lock_init(&p->sdev_lock);
1343         p->next = -1;
1344         if (swap_flags & SWAP_FLAG_PREFER) {
1345                 p->prio =
1346                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1347         } else {
1348                 p->prio = --least_priority;
1349         }
1350         swap_list_unlock();
1351         name = getname(specialfile);
1352         error = PTR_ERR(name);
1353         if (IS_ERR(name)) {
1354                 name = NULL;
1355                 goto bad_swap_2;
1356         }
1357         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1358         error = PTR_ERR(swap_file);
1359         if (IS_ERR(swap_file)) {
1360                 swap_file = NULL;
1361                 goto bad_swap_2;
1362         }
1363
1364         p->swap_file = swap_file;
1365         mapping = swap_file->f_mapping;
1366         inode = mapping->host;
1367
1368         error = -EBUSY;
1369         for (i = 0; i < nr_swapfiles; i++) {
1370                 struct swap_info_struct *q = &swap_info[i];
1371
1372                 if (i == type || !q->swap_file)
1373                         continue;
1374                 if (mapping == q->swap_file->f_mapping)
1375                         goto bad_swap;
1376         }
1377
1378         error = -EINVAL;
1379         if (S_ISBLK(inode->i_mode)) {
1380                 bdev = I_BDEV(inode);
1381                 error = bd_claim(bdev, sys_swapon);
1382                 if (error < 0) {
1383                         bdev = NULL;
1384                         goto bad_swap;
1385                 }
1386                 p->old_block_size = block_size(bdev);
1387                 error = set_blocksize(bdev, PAGE_SIZE);
1388                 if (error < 0)
1389                         goto bad_swap;
1390                 p->bdev = bdev;
1391         } else if (S_ISREG(inode->i_mode)) {
1392                 p->bdev = inode->i_sb->s_bdev;
1393                 down(&inode->i_sem);
1394                 did_down = 1;
1395                 if (IS_SWAPFILE(inode)) {
1396                         error = -EBUSY;
1397                         goto bad_swap;
1398                 }
1399         } else {
1400                 goto bad_swap;
1401         }
1402
1403         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1404
1405         /*
1406          * Read the swap header.
1407          */
1408         if (!mapping->a_ops->readpage) {
1409                 error = -EINVAL;
1410                 goto bad_swap;
1411         }
1412         page = read_cache_page(mapping, 0,
1413                         (filler_t *)mapping->a_ops->readpage, swap_file);
1414         if (IS_ERR(page)) {
1415                 error = PTR_ERR(page);
1416                 goto bad_swap;
1417         }
1418         wait_on_page_locked(page);
1419         if (!PageUptodate(page))
1420                 goto bad_swap;
1421         kmap(page);
1422         swap_header = page_address(page);
1423
1424         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1425                 swap_header_version = 1;
1426         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1427                 swap_header_version = 2;
1428         else {
1429                 printk("Unable to find swap-space signature\n");
1430                 error = -EINVAL;
1431                 goto bad_swap;
1432         }
1433         
1434         switch (swap_header_version) {
1435         case 1:
1436                 printk(KERN_ERR "version 0 swap is no longer supported. "
1437                         "Use mkswap -v1 %s\n", name);
1438                 error = -EINVAL;
1439                 goto bad_swap;
1440         case 2:
1441                 /* Check the swap header's sub-version and the size of
1442                    the swap file and bad block lists */
1443                 if (swap_header->info.version != 1) {
1444                         printk(KERN_WARNING
1445                                "Unable to handle swap header version %d\n",
1446                                swap_header->info.version);
1447                         error = -EINVAL;
1448                         goto bad_swap;
1449                 }
1450
1451                 p->lowest_bit  = 1;
1452                 /*
1453                  * Find out how many pages are allowed for a single swap
1454                  * device. There are two limiting factors: 1) the number of
1455                  * bits for the swap offset in the swp_entry_t type and
1456                  * 2) the number of bits in the a swap pte as defined by
1457                  * the different architectures. In order to find the
1458                  * largest possible bit mask a swap entry with swap type 0
1459                  * and swap offset ~0UL is created, encoded to a swap pte,
1460                  * decoded to a swp_entry_t again and finally the swap
1461                  * offset is extracted. This will mask all the bits from
1462                  * the initial ~0UL mask that can't be encoded in either
1463                  * the swp_entry_t or the architecture definition of a
1464                  * swap pte.
1465                  */
1466                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1467                 if (maxpages > swap_header->info.last_page)
1468                         maxpages = swap_header->info.last_page;
1469                 p->highest_bit = maxpages - 1;
1470
1471                 error = -EINVAL;
1472                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1473                         goto bad_swap;
1474                 
1475                 /* OK, set up the swap map and apply the bad block list */
1476                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1477                         error = -ENOMEM;
1478                         goto bad_swap;
1479                 }
1480
1481                 error = 0;
1482                 memset(p->swap_map, 0, maxpages * sizeof(short));
1483                 for (i=0; i<swap_header->info.nr_badpages; i++) {
1484                         int page = swap_header->info.badpages[i];
1485                         if (page <= 0 || page >= swap_header->info.last_page)
1486                                 error = -EINVAL;
1487                         else
1488                                 p->swap_map[page] = SWAP_MAP_BAD;
1489                 }
1490                 nr_good_pages = swap_header->info.last_page -
1491                                 swap_header->info.nr_badpages -
1492                                 1 /* header page */;
1493                 if (error) 
1494                         goto bad_swap;
1495         }
1496         
1497         if (swapfilesize && maxpages > swapfilesize) {
1498                 printk(KERN_WARNING
1499                        "Swap area shorter than signature indicates\n");
1500                 error = -EINVAL;
1501                 goto bad_swap;
1502         }
1503         if (!nr_good_pages) {
1504                 printk(KERN_WARNING "Empty swap-file\n");
1505                 error = -EINVAL;
1506                 goto bad_swap;
1507         }
1508         p->swap_map[0] = SWAP_MAP_BAD;
1509         p->max = maxpages;
1510         p->pages = nr_good_pages;
1511
1512         error = setup_swap_extents(p);
1513         if (error)
1514                 goto bad_swap;
1515
1516         down(&swapon_sem);
1517         swap_list_lock();
1518         swap_device_lock(p);
1519         p->flags = SWP_ACTIVE;
1520         nr_swap_pages += nr_good_pages;
1521         total_swap_pages += nr_good_pages;
1522         printk(KERN_INFO "Adding %dk swap on %s.  Priority:%d extents:%d\n",
1523                 nr_good_pages<<(PAGE_SHIFT-10), name,
1524                 p->prio, p->nr_extents);
1525
1526         /* insert swap space into swap_list: */
1527         prev = -1;
1528         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1529                 if (p->prio >= swap_info[i].prio) {
1530                         break;
1531                 }
1532                 prev = i;
1533         }
1534         p->next = i;
1535         if (prev < 0) {
1536                 swap_list.head = swap_list.next = p - swap_info;
1537         } else {
1538                 swap_info[prev].next = p - swap_info;
1539         }
1540         swap_device_unlock(p);
1541         swap_list_unlock();
1542         up(&swapon_sem);
1543         error = 0;
1544         goto out;
1545 bad_swap:
1546         if (bdev) {
1547                 set_blocksize(bdev, p->old_block_size);
1548                 bd_release(bdev);
1549         }
1550 bad_swap_2:
1551         swap_list_lock();
1552         swap_map = p->swap_map;
1553         p->swap_file = NULL;
1554         p->swap_map = NULL;
1555         p->flags = 0;
1556         if (!(swap_flags & SWAP_FLAG_PREFER))
1557                 ++least_priority;
1558         swap_list_unlock();
1559         destroy_swap_extents(p);
1560         vfree(swap_map);
1561         if (swap_file)
1562                 filp_close(swap_file, NULL);
1563 out:
1564         if (page && !IS_ERR(page)) {
1565                 kunmap(page);
1566                 page_cache_release(page);
1567         }
1568         if (name)
1569                 putname(name);
1570         if (did_down) {
1571                 if (!error)
1572                         inode->i_flags |= S_SWAPFILE;
1573                 up(&inode->i_sem);
1574         }
1575         return error;
1576 }
1577
1578 void si_swapinfo(struct sysinfo *val)
1579 {
1580         unsigned int i;
1581         unsigned long nr_to_be_unused = 0;
1582
1583         swap_list_lock();
1584         for (i = 0; i < nr_swapfiles; i++) {
1585                 if (!(swap_info[i].flags & SWP_USED) ||
1586                      (swap_info[i].flags & SWP_WRITEOK))
1587                         continue;
1588                 nr_to_be_unused += swap_info[i].inuse_pages;
1589         }
1590         val->freeswap = nr_swap_pages + nr_to_be_unused;
1591         val->totalswap = total_swap_pages + nr_to_be_unused;
1592         swap_list_unlock();
1593 }
1594
1595 /*
1596  * Verify that a swap entry is valid and increment its swap map count.
1597  *
1598  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1599  * "permanent", but will be reclaimed by the next swapoff.
1600  */
1601 int swap_duplicate(swp_entry_t entry)
1602 {
1603         struct swap_info_struct * p;
1604         unsigned long offset, type;
1605         int result = 0;
1606
1607         type = swp_type(entry);
1608         if (type >= nr_swapfiles)
1609                 goto bad_file;
1610         p = type + swap_info;
1611         offset = swp_offset(entry);
1612
1613         swap_device_lock(p);
1614         if (offset < p->max && p->swap_map[offset]) {
1615                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1616                         p->swap_map[offset]++;
1617                         result = 1;
1618                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1619                         if (swap_overflow++ < 5)
1620                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1621                         p->swap_map[offset] = SWAP_MAP_MAX;
1622                         result = 1;
1623                 }
1624         }
1625         swap_device_unlock(p);
1626 out:
1627         return result;
1628
1629 bad_file:
1630         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1631         goto out;
1632 }
1633
1634 struct swap_info_struct *
1635 get_swap_info_struct(unsigned type)
1636 {
1637         return &swap_info[type];
1638 }
1639
1640 /*
1641  * swap_device_lock prevents swap_map being freed. Don't grab an extra
1642  * reference on the swaphandle, it doesn't matter if it becomes unused.
1643  */
1644 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1645 {
1646         int ret = 0, i = 1 << page_cluster;
1647         unsigned long toff;
1648         struct swap_info_struct *swapdev = swp_type(entry) + swap_info;
1649
1650         if (!page_cluster)      /* no readahead */
1651                 return 0;
1652         toff = (swp_offset(entry) >> page_cluster) << page_cluster;
1653         if (!toff)              /* first page is swap header */
1654                 toff++, i--;
1655         *offset = toff;
1656
1657         swap_device_lock(swapdev);
1658         do {
1659                 /* Don't read-ahead past the end of the swap area */
1660                 if (toff >= swapdev->max)
1661                         break;
1662                 /* Don't read in free or bad pages */
1663                 if (!swapdev->swap_map[toff])
1664                         break;
1665                 if (swapdev->swap_map[toff] == SWAP_MAP_BAD)
1666                         break;
1667                 toff++;
1668                 ret++;
1669         } while (--i);
1670         swap_device_unlock(swapdev);
1671         return ret;
1672 }