Merge master.kernel.org:/pub/scm/linux/kernel/git/davej/cpufreq
[linux-2.6] / mm / swapfile.c
1 /*
2  *  linux/mm/swapfile.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *  Swap reorganised 29.12.95, Stephen Tweedie
6  */
7
8 #include <linux/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 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_lock while calling the unplug_fn. And swap_lock
55  * cannot be turned into a semaphore.
56  */
57 static DECLARE_RWSEM(swap_unplug_sem);
58
59 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
60 {
61         swp_entry_t entry;
62
63         down_read(&swap_unplug_sem);
64         entry.val = page->private;
65         if (PageSwapCache(page)) {
66                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
67                 struct backing_dev_info *bdi;
68
69                 /*
70                  * If the page is removed from swapcache from under us (with a
71                  * racy try_to_unuse/swapoff) we need an additional reference
72                  * count to avoid reading garbage from page->private above. If
73                  * the WARN_ON triggers during a swapoff it maybe the race
74                  * condition and it's harmless. However if it triggers without
75                  * swapoff it signals a problem.
76                  */
77                 WARN_ON(page_count(page) <= 1);
78
79                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
80                 blk_run_backing_dev(bdi, page);
81         }
82         up_read(&swap_unplug_sem);
83 }
84
85 #define SWAPFILE_CLUSTER        256
86 #define LATENCY_LIMIT           256
87
88 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
89 {
90         unsigned long offset, last_in_cluster;
91         int latency_ration = LATENCY_LIMIT;
92
93         /* 
94          * We try to cluster swap pages by allocating them sequentially
95          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
96          * way, however, we resort to first-free allocation, starting
97          * a new cluster.  This prevents us from scattering swap pages
98          * all over the entire swap partition, so that we reduce
99          * overall disk seek times between swap pages.  -- sct
100          * But we do now try to find an empty cluster.  -Andrea
101          */
102
103         si->flags += SWP_SCANNING;
104         if (unlikely(!si->cluster_nr)) {
105                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
106                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER)
107                         goto lowest;
108                 spin_unlock(&swap_lock);
109
110                 offset = si->lowest_bit;
111                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
112
113                 /* Locate the first empty (unaligned) cluster */
114                 for (; last_in_cluster <= si->highest_bit; offset++) {
115                         if (si->swap_map[offset])
116                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
117                         else if (offset == last_in_cluster) {
118                                 spin_lock(&swap_lock);
119                                 si->cluster_next = offset-SWAPFILE_CLUSTER-1;
120                                 goto cluster;
121                         }
122                         if (unlikely(--latency_ration < 0)) {
123                                 cond_resched();
124                                 latency_ration = LATENCY_LIMIT;
125                         }
126                 }
127                 spin_lock(&swap_lock);
128                 goto lowest;
129         }
130
131         si->cluster_nr--;
132 cluster:
133         offset = si->cluster_next;
134         if (offset > si->highest_bit)
135 lowest:         offset = si->lowest_bit;
136 checks: if (!(si->flags & SWP_WRITEOK))
137                 goto no_page;
138         if (!si->highest_bit)
139                 goto no_page;
140         if (!si->swap_map[offset]) {
141                 if (offset == si->lowest_bit)
142                         si->lowest_bit++;
143                 if (offset == si->highest_bit)
144                         si->highest_bit--;
145                 si->inuse_pages++;
146                 if (si->inuse_pages == si->pages) {
147                         si->lowest_bit = si->max;
148                         si->highest_bit = 0;
149                 }
150                 si->swap_map[offset] = 1;
151                 si->cluster_next = offset + 1;
152                 si->flags -= SWP_SCANNING;
153                 return offset;
154         }
155
156         spin_unlock(&swap_lock);
157         while (++offset <= si->highest_bit) {
158                 if (!si->swap_map[offset]) {
159                         spin_lock(&swap_lock);
160                         goto checks;
161                 }
162                 if (unlikely(--latency_ration < 0)) {
163                         cond_resched();
164                         latency_ration = LATENCY_LIMIT;
165                 }
166         }
167         spin_lock(&swap_lock);
168         goto lowest;
169
170 no_page:
171         si->flags -= SWP_SCANNING;
172         return 0;
173 }
174
175 swp_entry_t get_swap_page(void)
176 {
177         struct swap_info_struct *si;
178         pgoff_t offset;
179         int type, next;
180         int wrapped = 0;
181
182         spin_lock(&swap_lock);
183         if (nr_swap_pages <= 0)
184                 goto noswap;
185         nr_swap_pages--;
186
187         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
188                 si = swap_info + type;
189                 next = si->next;
190                 if (next < 0 ||
191                     (!wrapped && si->prio != swap_info[next].prio)) {
192                         next = swap_list.head;
193                         wrapped++;
194                 }
195
196                 if (!si->highest_bit)
197                         continue;
198                 if (!(si->flags & SWP_WRITEOK))
199                         continue;
200
201                 swap_list.next = next;
202                 offset = scan_swap_map(si);
203                 if (offset) {
204                         spin_unlock(&swap_lock);
205                         return swp_entry(type, offset);
206                 }
207                 next = swap_list.next;
208         }
209
210         nr_swap_pages++;
211 noswap:
212         spin_unlock(&swap_lock);
213         return (swp_entry_t) {0};
214 }
215
216 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
217 {
218         struct swap_info_struct * p;
219         unsigned long offset, type;
220
221         if (!entry.val)
222                 goto out;
223         type = swp_type(entry);
224         if (type >= nr_swapfiles)
225                 goto bad_nofile;
226         p = & swap_info[type];
227         if (!(p->flags & SWP_USED))
228                 goto bad_device;
229         offset = swp_offset(entry);
230         if (offset >= p->max)
231                 goto bad_offset;
232         if (!p->swap_map[offset])
233                 goto bad_free;
234         spin_lock(&swap_lock);
235         return p;
236
237 bad_free:
238         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
239         goto out;
240 bad_offset:
241         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
242         goto out;
243 bad_device:
244         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
245         goto out;
246 bad_nofile:
247         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
248 out:
249         return NULL;
250 }       
251
252 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
253 {
254         int count = p->swap_map[offset];
255
256         if (count < SWAP_MAP_MAX) {
257                 count--;
258                 p->swap_map[offset] = count;
259                 if (!count) {
260                         if (offset < p->lowest_bit)
261                                 p->lowest_bit = offset;
262                         if (offset > p->highest_bit)
263                                 p->highest_bit = offset;
264                         if (p->prio > swap_info[swap_list.next].prio)
265                                 swap_list.next = p - swap_info;
266                         nr_swap_pages++;
267                         p->inuse_pages--;
268                 }
269         }
270         return count;
271 }
272
273 /*
274  * Caller has made sure that the swapdevice corresponding to entry
275  * is still around or has not been recycled.
276  */
277 void swap_free(swp_entry_t entry)
278 {
279         struct swap_info_struct * p;
280
281         p = swap_info_get(entry);
282         if (p) {
283                 swap_entry_free(p, swp_offset(entry));
284                 spin_unlock(&swap_lock);
285         }
286 }
287
288 /*
289  * How many references to page are currently swapped out?
290  */
291 static inline int page_swapcount(struct page *page)
292 {
293         int count = 0;
294         struct swap_info_struct *p;
295         swp_entry_t entry;
296
297         entry.val = page->private;
298         p = swap_info_get(entry);
299         if (p) {
300                 /* Subtract the 1 for the swap cache itself */
301                 count = p->swap_map[swp_offset(entry)] - 1;
302                 spin_unlock(&swap_lock);
303         }
304         return count;
305 }
306
307 /*
308  * We can use this swap cache entry directly
309  * if there are no other references to it.
310  */
311 int can_share_swap_page(struct page *page)
312 {
313         int count;
314
315         BUG_ON(!PageLocked(page));
316         count = page_mapcount(page);
317         if (count <= 1 && PageSwapCache(page))
318                 count += page_swapcount(page);
319         return count == 1;
320 }
321
322 /*
323  * Work out if there are any other processes sharing this
324  * swap cache page. Free it if you can. Return success.
325  */
326 int remove_exclusive_swap_page(struct page *page)
327 {
328         int retval;
329         struct swap_info_struct * p;
330         swp_entry_t entry;
331
332         BUG_ON(PagePrivate(page));
333         BUG_ON(!PageLocked(page));
334
335         if (!PageSwapCache(page))
336                 return 0;
337         if (PageWriteback(page))
338                 return 0;
339         if (page_count(page) != 2) /* 2: us + cache */
340                 return 0;
341
342         entry.val = page->private;
343         p = swap_info_get(entry);
344         if (!p)
345                 return 0;
346
347         /* Is the only swap cache user the cache itself? */
348         retval = 0;
349         if (p->swap_map[swp_offset(entry)] == 1) {
350                 /* Recheck the page count with the swapcache lock held.. */
351                 write_lock_irq(&swapper_space.tree_lock);
352                 if ((page_count(page) == 2) && !PageWriteback(page)) {
353                         __delete_from_swap_cache(page);
354                         SetPageDirty(page);
355                         retval = 1;
356                 }
357                 write_unlock_irq(&swapper_space.tree_lock);
358         }
359         spin_unlock(&swap_lock);
360
361         if (retval) {
362                 swap_free(entry);
363                 page_cache_release(page);
364         }
365
366         return retval;
367 }
368
369 /*
370  * Free the swap entry like above, but also try to
371  * free the page cache entry if it is the last user.
372  */
373 void free_swap_and_cache(swp_entry_t entry)
374 {
375         struct swap_info_struct * p;
376         struct page *page = NULL;
377
378         p = swap_info_get(entry);
379         if (p) {
380                 if (swap_entry_free(p, swp_offset(entry)) == 1)
381                         page = find_trylock_page(&swapper_space, entry.val);
382                 spin_unlock(&swap_lock);
383         }
384         if (page) {
385                 int one_user;
386
387                 BUG_ON(PagePrivate(page));
388                 page_cache_get(page);
389                 one_user = (page_count(page) == 2);
390                 /* Only cache user (+us), or swap space full? Free it! */
391                 if (!PageWriteback(page) && (one_user || vm_swap_full())) {
392                         delete_from_swap_cache(page);
393                         SetPageDirty(page);
394                 }
395                 unlock_page(page);
396                 page_cache_release(page);
397         }
398 }
399
400 /*
401  * Always set the resulting pte to be nowrite (the same as COW pages
402  * after one process has exited).  We don't know just how many PTEs will
403  * share this swap entry, so be cautious and let do_wp_page work out
404  * what to do if a write is requested later.
405  *
406  * vma->vm_mm->page_table_lock is held.
407  */
408 static void unuse_pte(struct vm_area_struct *vma, pte_t *pte,
409                 unsigned long addr, swp_entry_t entry, struct page *page)
410 {
411         inc_mm_counter(vma->vm_mm, rss);
412         get_page(page);
413         set_pte_at(vma->vm_mm, addr, pte,
414                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
415         page_add_anon_rmap(page, vma, addr);
416         swap_free(entry);
417         /*
418          * Move the page to the active list so it is not
419          * immediately swapped out again after swapon.
420          */
421         activate_page(page);
422 }
423
424 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
425                                 unsigned long addr, unsigned long end,
426                                 swp_entry_t entry, struct page *page)
427 {
428         pte_t *pte;
429         pte_t swp_pte = swp_entry_to_pte(entry);
430
431         pte = pte_offset_map(pmd, addr);
432         do {
433                 /*
434                  * swapoff spends a _lot_ of time in this loop!
435                  * Test inline before going to call unuse_pte.
436                  */
437                 if (unlikely(pte_same(*pte, swp_pte))) {
438                         unuse_pte(vma, pte, addr, entry, page);
439                         pte_unmap(pte);
440                         return 1;
441                 }
442         } while (pte++, addr += PAGE_SIZE, addr != end);
443         pte_unmap(pte - 1);
444         return 0;
445 }
446
447 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
448                                 unsigned long addr, unsigned long end,
449                                 swp_entry_t entry, struct page *page)
450 {
451         pmd_t *pmd;
452         unsigned long next;
453
454         pmd = pmd_offset(pud, addr);
455         do {
456                 next = pmd_addr_end(addr, end);
457                 if (pmd_none_or_clear_bad(pmd))
458                         continue;
459                 if (unuse_pte_range(vma, pmd, addr, next, entry, page))
460                         return 1;
461         } while (pmd++, addr = next, addr != end);
462         return 0;
463 }
464
465 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
466                                 unsigned long addr, unsigned long end,
467                                 swp_entry_t entry, struct page *page)
468 {
469         pud_t *pud;
470         unsigned long next;
471
472         pud = pud_offset(pgd, addr);
473         do {
474                 next = pud_addr_end(addr, end);
475                 if (pud_none_or_clear_bad(pud))
476                         continue;
477                 if (unuse_pmd_range(vma, pud, addr, next, entry, page))
478                         return 1;
479         } while (pud++, addr = next, addr != end);
480         return 0;
481 }
482
483 static int unuse_vma(struct vm_area_struct *vma,
484                                 swp_entry_t entry, struct page *page)
485 {
486         pgd_t *pgd;
487         unsigned long addr, end, next;
488
489         if (page->mapping) {
490                 addr = page_address_in_vma(page, vma);
491                 if (addr == -EFAULT)
492                         return 0;
493                 else
494                         end = addr + PAGE_SIZE;
495         } else {
496                 addr = vma->vm_start;
497                 end = vma->vm_end;
498         }
499
500         pgd = pgd_offset(vma->vm_mm, addr);
501         do {
502                 next = pgd_addr_end(addr, end);
503                 if (pgd_none_or_clear_bad(pgd))
504                         continue;
505                 if (unuse_pud_range(vma, pgd, addr, next, entry, page))
506                         return 1;
507         } while (pgd++, addr = next, addr != end);
508         return 0;
509 }
510
511 static int unuse_mm(struct mm_struct *mm,
512                                 swp_entry_t entry, struct page *page)
513 {
514         struct vm_area_struct *vma;
515
516         if (!down_read_trylock(&mm->mmap_sem)) {
517                 /*
518                  * Activate page so shrink_cache is unlikely to unmap its
519                  * ptes while lock is dropped, so swapoff can make progress.
520                  */
521                 activate_page(page);
522                 unlock_page(page);
523                 down_read(&mm->mmap_sem);
524                 lock_page(page);
525         }
526         spin_lock(&mm->page_table_lock);
527         for (vma = mm->mmap; vma; vma = vma->vm_next) {
528                 if (vma->anon_vma && unuse_vma(vma, entry, page))
529                         break;
530         }
531         spin_unlock(&mm->page_table_lock);
532         up_read(&mm->mmap_sem);
533         /*
534          * Currently unuse_mm cannot fail, but leave error handling
535          * at call sites for now, since we change it from time to time.
536          */
537         return 0;
538 }
539
540 /*
541  * Scan swap_map from current position to next entry still in use.
542  * Recycle to start on reaching the end, returning 0 when empty.
543  */
544 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
545                                         unsigned int prev)
546 {
547         unsigned int max = si->max;
548         unsigned int i = prev;
549         int count;
550
551         /*
552          * No need for swap_lock here: we're just looking
553          * for whether an entry is in use, not modifying it; false
554          * hits are okay, and sys_swapoff() has already prevented new
555          * allocations from this area (while holding swap_lock).
556          */
557         for (;;) {
558                 if (++i >= max) {
559                         if (!prev) {
560                                 i = 0;
561                                 break;
562                         }
563                         /*
564                          * No entries in use at top of swap_map,
565                          * loop back to start and recheck there.
566                          */
567                         max = prev + 1;
568                         prev = 0;
569                         i = 1;
570                 }
571                 count = si->swap_map[i];
572                 if (count && count != SWAP_MAP_BAD)
573                         break;
574         }
575         return i;
576 }
577
578 /*
579  * We completely avoid races by reading each swap page in advance,
580  * and then search for the process using it.  All the necessary
581  * page table adjustments can then be made atomically.
582  */
583 static int try_to_unuse(unsigned int type)
584 {
585         struct swap_info_struct * si = &swap_info[type];
586         struct mm_struct *start_mm;
587         unsigned short *swap_map;
588         unsigned short swcount;
589         struct page *page;
590         swp_entry_t entry;
591         unsigned int i = 0;
592         int retval = 0;
593         int reset_overflow = 0;
594         int shmem;
595
596         /*
597          * When searching mms for an entry, a good strategy is to
598          * start at the first mm we freed the previous entry from
599          * (though actually we don't notice whether we or coincidence
600          * freed the entry).  Initialize this start_mm with a hold.
601          *
602          * A simpler strategy would be to start at the last mm we
603          * freed the previous entry from; but that would take less
604          * advantage of mmlist ordering, which clusters forked mms
605          * together, child after parent.  If we race with dup_mmap(), we
606          * prefer to resolve parent before child, lest we miss entries
607          * duplicated after we scanned child: using last mm would invert
608          * that.  Though it's only a serious concern when an overflowed
609          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
610          */
611         start_mm = &init_mm;
612         atomic_inc(&init_mm.mm_users);
613
614         /*
615          * Keep on scanning until all entries have gone.  Usually,
616          * one pass through swap_map is enough, but not necessarily:
617          * there are races when an instance of an entry might be missed.
618          */
619         while ((i = find_next_to_unuse(si, i)) != 0) {
620                 if (signal_pending(current)) {
621                         retval = -EINTR;
622                         break;
623                 }
624
625                 /* 
626                  * Get a page for the entry, using the existing swap
627                  * cache page if there is one.  Otherwise, get a clean
628                  * page and read the swap into it. 
629                  */
630                 swap_map = &si->swap_map[i];
631                 entry = swp_entry(type, i);
632                 page = read_swap_cache_async(entry, NULL, 0);
633                 if (!page) {
634                         /*
635                          * Either swap_duplicate() failed because entry
636                          * has been freed independently, and will not be
637                          * reused since sys_swapoff() already disabled
638                          * allocation from here, or alloc_page() failed.
639                          */
640                         if (!*swap_map)
641                                 continue;
642                         retval = -ENOMEM;
643                         break;
644                 }
645
646                 /*
647                  * Don't hold on to start_mm if it looks like exiting.
648                  */
649                 if (atomic_read(&start_mm->mm_users) == 1) {
650                         mmput(start_mm);
651                         start_mm = &init_mm;
652                         atomic_inc(&init_mm.mm_users);
653                 }
654
655                 /*
656                  * Wait for and lock page.  When do_swap_page races with
657                  * try_to_unuse, do_swap_page can handle the fault much
658                  * faster than try_to_unuse can locate the entry.  This
659                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
660                  * defer to do_swap_page in such a case - in some tests,
661                  * do_swap_page and try_to_unuse repeatedly compete.
662                  */
663                 wait_on_page_locked(page);
664                 wait_on_page_writeback(page);
665                 lock_page(page);
666                 wait_on_page_writeback(page);
667
668                 /*
669                  * Remove all references to entry.
670                  * Whenever we reach init_mm, there's no address space
671                  * to search, but use it as a reminder to search shmem.
672                  */
673                 shmem = 0;
674                 swcount = *swap_map;
675                 if (swcount > 1) {
676                         if (start_mm == &init_mm)
677                                 shmem = shmem_unuse(entry, page);
678                         else
679                                 retval = unuse_mm(start_mm, entry, page);
680                 }
681                 if (*swap_map > 1) {
682                         int set_start_mm = (*swap_map >= swcount);
683                         struct list_head *p = &start_mm->mmlist;
684                         struct mm_struct *new_start_mm = start_mm;
685                         struct mm_struct *prev_mm = start_mm;
686                         struct mm_struct *mm;
687
688                         atomic_inc(&new_start_mm->mm_users);
689                         atomic_inc(&prev_mm->mm_users);
690                         spin_lock(&mmlist_lock);
691                         while (*swap_map > 1 && !retval &&
692                                         (p = p->next) != &start_mm->mmlist) {
693                                 mm = list_entry(p, struct mm_struct, mmlist);
694                                 if (atomic_inc_return(&mm->mm_users) == 1) {
695                                         atomic_dec(&mm->mm_users);
696                                         continue;
697                                 }
698                                 spin_unlock(&mmlist_lock);
699                                 mmput(prev_mm);
700                                 prev_mm = mm;
701
702                                 cond_resched();
703
704                                 swcount = *swap_map;
705                                 if (swcount <= 1)
706                                         ;
707                                 else if (mm == &init_mm) {
708                                         set_start_mm = 1;
709                                         shmem = shmem_unuse(entry, page);
710                                 } else
711                                         retval = unuse_mm(mm, entry, page);
712                                 if (set_start_mm && *swap_map < swcount) {
713                                         mmput(new_start_mm);
714                                         atomic_inc(&mm->mm_users);
715                                         new_start_mm = mm;
716                                         set_start_mm = 0;
717                                 }
718                                 spin_lock(&mmlist_lock);
719                         }
720                         spin_unlock(&mmlist_lock);
721                         mmput(prev_mm);
722                         mmput(start_mm);
723                         start_mm = new_start_mm;
724                 }
725                 if (retval) {
726                         unlock_page(page);
727                         page_cache_release(page);
728                         break;
729                 }
730
731                 /*
732                  * How could swap count reach 0x7fff when the maximum
733                  * pid is 0x7fff, and there's no way to repeat a swap
734                  * page within an mm (except in shmem, where it's the
735                  * shared object which takes the reference count)?
736                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
737                  *
738                  * If that's wrong, then we should worry more about
739                  * exit_mmap() and do_munmap() cases described above:
740                  * we might be resetting SWAP_MAP_MAX too early here.
741                  * We know "Undead"s can happen, they're okay, so don't
742                  * report them; but do report if we reset SWAP_MAP_MAX.
743                  */
744                 if (*swap_map == SWAP_MAP_MAX) {
745                         spin_lock(&swap_lock);
746                         *swap_map = 1;
747                         spin_unlock(&swap_lock);
748                         reset_overflow = 1;
749                 }
750
751                 /*
752                  * If a reference remains (rare), we would like to leave
753                  * the page in the swap cache; but try_to_unmap could
754                  * then re-duplicate the entry once we drop page lock,
755                  * so we might loop indefinitely; also, that page could
756                  * not be swapped out to other storage meanwhile.  So:
757                  * delete from cache even if there's another reference,
758                  * after ensuring that the data has been saved to disk -
759                  * since if the reference remains (rarer), it will be
760                  * read from disk into another page.  Splitting into two
761                  * pages would be incorrect if swap supported "shared
762                  * private" pages, but they are handled by tmpfs files.
763                  *
764                  * Note shmem_unuse already deleted a swappage from
765                  * the swap cache, unless the move to filepage failed:
766                  * in which case it left swappage in cache, lowered its
767                  * swap count to pass quickly through the loops above,
768                  * and now we must reincrement count to try again later.
769                  */
770                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
771                         struct writeback_control wbc = {
772                                 .sync_mode = WB_SYNC_NONE,
773                         };
774
775                         swap_writepage(page, &wbc);
776                         lock_page(page);
777                         wait_on_page_writeback(page);
778                 }
779                 if (PageSwapCache(page)) {
780                         if (shmem)
781                                 swap_duplicate(entry);
782                         else
783                                 delete_from_swap_cache(page);
784                 }
785
786                 /*
787                  * So we could skip searching mms once swap count went
788                  * to 1, we did not mark any present ptes as dirty: must
789                  * mark page dirty so shrink_list will preserve it.
790                  */
791                 SetPageDirty(page);
792                 unlock_page(page);
793                 page_cache_release(page);
794
795                 /*
796                  * Make sure that we aren't completely killing
797                  * interactive performance.
798                  */
799                 cond_resched();
800         }
801
802         mmput(start_mm);
803         if (reset_overflow) {
804                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
805                 swap_overflow = 0;
806         }
807         return retval;
808 }
809
810 /*
811  * After a successful try_to_unuse, if no swap is now in use, we know
812  * we can empty the mmlist.  swap_lock must be held on entry and exit.
813  * Note that mmlist_lock nests inside swap_lock, and an mm must be
814  * added to the mmlist just after page_duplicate - before would be racy.
815  */
816 static void drain_mmlist(void)
817 {
818         struct list_head *p, *next;
819         unsigned int i;
820
821         for (i = 0; i < nr_swapfiles; i++)
822                 if (swap_info[i].inuse_pages)
823                         return;
824         spin_lock(&mmlist_lock);
825         list_for_each_safe(p, next, &init_mm.mmlist)
826                 list_del_init(p);
827         spin_unlock(&mmlist_lock);
828 }
829
830 /*
831  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
832  * corresponds to page offset `offset'.
833  */
834 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
835 {
836         struct swap_extent *se = sis->curr_swap_extent;
837         struct swap_extent *start_se = se;
838
839         for ( ; ; ) {
840                 struct list_head *lh;
841
842                 if (se->start_page <= offset &&
843                                 offset < (se->start_page + se->nr_pages)) {
844                         return se->start_block + (offset - se->start_page);
845                 }
846                 lh = se->list.next;
847                 if (lh == &sis->extent_list)
848                         lh = lh->next;
849                 se = list_entry(lh, struct swap_extent, list);
850                 sis->curr_swap_extent = se;
851                 BUG_ON(se == start_se);         /* It *must* be present */
852         }
853 }
854
855 /*
856  * Free all of a swapdev's extent information
857  */
858 static void destroy_swap_extents(struct swap_info_struct *sis)
859 {
860         while (!list_empty(&sis->extent_list)) {
861                 struct swap_extent *se;
862
863                 se = list_entry(sis->extent_list.next,
864                                 struct swap_extent, list);
865                 list_del(&se->list);
866                 kfree(se);
867         }
868 }
869
870 /*
871  * Add a block range (and the corresponding page range) into this swapdev's
872  * extent list.  The extent list is kept sorted in page order.
873  *
874  * This function rather assumes that it is called in ascending page order.
875  */
876 static int
877 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
878                 unsigned long nr_pages, sector_t start_block)
879 {
880         struct swap_extent *se;
881         struct swap_extent *new_se;
882         struct list_head *lh;
883
884         lh = sis->extent_list.prev;     /* The highest page extent */
885         if (lh != &sis->extent_list) {
886                 se = list_entry(lh, struct swap_extent, list);
887                 BUG_ON(se->start_page + se->nr_pages != start_page);
888                 if (se->start_block + se->nr_pages == start_block) {
889                         /* Merge it */
890                         se->nr_pages += nr_pages;
891                         return 0;
892                 }
893         }
894
895         /*
896          * No merge.  Insert a new extent, preserving ordering.
897          */
898         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
899         if (new_se == NULL)
900                 return -ENOMEM;
901         new_se->start_page = start_page;
902         new_se->nr_pages = nr_pages;
903         new_se->start_block = start_block;
904
905         list_add_tail(&new_se->list, &sis->extent_list);
906         return 1;
907 }
908
909 /*
910  * A `swap extent' is a simple thing which maps a contiguous range of pages
911  * onto a contiguous range of disk blocks.  An ordered list of swap extents
912  * is built at swapon time and is then used at swap_writepage/swap_readpage
913  * time for locating where on disk a page belongs.
914  *
915  * If the swapfile is an S_ISBLK block device, a single extent is installed.
916  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
917  * swap files identically.
918  *
919  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
920  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
921  * swapfiles are handled *identically* after swapon time.
922  *
923  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
924  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
925  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
926  * requirements, they are simply tossed out - we will never use those blocks
927  * for swapping.
928  *
929  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
930  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
931  * which will scribble on the fs.
932  *
933  * The amount of disk space which a single swap extent represents varies.
934  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
935  * extents in the list.  To avoid much list walking, we cache the previous
936  * search location in `curr_swap_extent', and start new searches from there.
937  * This is extremely effective.  The average number of iterations in
938  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
939  */
940 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
941 {
942         struct inode *inode;
943         unsigned blocks_per_page;
944         unsigned long page_no;
945         unsigned blkbits;
946         sector_t probe_block;
947         sector_t last_block;
948         sector_t lowest_block = -1;
949         sector_t highest_block = 0;
950         int nr_extents = 0;
951         int ret;
952
953         inode = sis->swap_file->f_mapping->host;
954         if (S_ISBLK(inode->i_mode)) {
955                 ret = add_swap_extent(sis, 0, sis->max, 0);
956                 *span = sis->pages;
957                 goto done;
958         }
959
960         blkbits = inode->i_blkbits;
961         blocks_per_page = PAGE_SIZE >> blkbits;
962
963         /*
964          * Map all the blocks into the extent list.  This code doesn't try
965          * to be very smart.
966          */
967         probe_block = 0;
968         page_no = 0;
969         last_block = i_size_read(inode) >> blkbits;
970         while ((probe_block + blocks_per_page) <= last_block &&
971                         page_no < sis->max) {
972                 unsigned block_in_page;
973                 sector_t first_block;
974
975                 first_block = bmap(inode, probe_block);
976                 if (first_block == 0)
977                         goto bad_bmap;
978
979                 /*
980                  * It must be PAGE_SIZE aligned on-disk
981                  */
982                 if (first_block & (blocks_per_page - 1)) {
983                         probe_block++;
984                         goto reprobe;
985                 }
986
987                 for (block_in_page = 1; block_in_page < blocks_per_page;
988                                         block_in_page++) {
989                         sector_t block;
990
991                         block = bmap(inode, probe_block + block_in_page);
992                         if (block == 0)
993                                 goto bad_bmap;
994                         if (block != first_block + block_in_page) {
995                                 /* Discontiguity */
996                                 probe_block++;
997                                 goto reprobe;
998                         }
999                 }
1000
1001                 first_block >>= (PAGE_SHIFT - blkbits);
1002                 if (page_no) {  /* exclude the header page */
1003                         if (first_block < lowest_block)
1004                                 lowest_block = first_block;
1005                         if (first_block > highest_block)
1006                                 highest_block = first_block;
1007                 }
1008
1009                 /*
1010                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1011                  */
1012                 ret = add_swap_extent(sis, page_no, 1, first_block);
1013                 if (ret < 0)
1014                         goto out;
1015                 nr_extents += ret;
1016                 page_no++;
1017                 probe_block += blocks_per_page;
1018 reprobe:
1019                 continue;
1020         }
1021         ret = nr_extents;
1022         *span = 1 + highest_block - lowest_block;
1023         if (page_no == 0)
1024                 page_no = 1;    /* force Empty message */
1025         sis->max = page_no;
1026         sis->pages = page_no - 1;
1027         sis->highest_bit = page_no - 1;
1028 done:
1029         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1030                                         struct swap_extent, list);
1031         goto out;
1032 bad_bmap:
1033         printk(KERN_ERR "swapon: swapfile has holes\n");
1034         ret = -EINVAL;
1035 out:
1036         return ret;
1037 }
1038
1039 #if 0   /* We don't need this yet */
1040 #include <linux/backing-dev.h>
1041 int page_queue_congested(struct page *page)
1042 {
1043         struct backing_dev_info *bdi;
1044
1045         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1046
1047         if (PageSwapCache(page)) {
1048                 swp_entry_t entry = { .val = page->private };
1049                 struct swap_info_struct *sis;
1050
1051                 sis = get_swap_info_struct(swp_type(entry));
1052                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1053         } else
1054                 bdi = page->mapping->backing_dev_info;
1055         return bdi_write_congested(bdi);
1056 }
1057 #endif
1058
1059 asmlinkage long sys_swapoff(const char __user * specialfile)
1060 {
1061         struct swap_info_struct * p = NULL;
1062         unsigned short *swap_map;
1063         struct file *swap_file, *victim;
1064         struct address_space *mapping;
1065         struct inode *inode;
1066         char * pathname;
1067         int i, type, prev;
1068         int err;
1069         
1070         if (!capable(CAP_SYS_ADMIN))
1071                 return -EPERM;
1072
1073         pathname = getname(specialfile);
1074         err = PTR_ERR(pathname);
1075         if (IS_ERR(pathname))
1076                 goto out;
1077
1078         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1079         putname(pathname);
1080         err = PTR_ERR(victim);
1081         if (IS_ERR(victim))
1082                 goto out;
1083
1084         mapping = victim->f_mapping;
1085         prev = -1;
1086         spin_lock(&swap_lock);
1087         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1088                 p = swap_info + type;
1089                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1090                         if (p->swap_file->f_mapping == mapping)
1091                                 break;
1092                 }
1093                 prev = type;
1094         }
1095         if (type < 0) {
1096                 err = -EINVAL;
1097                 spin_unlock(&swap_lock);
1098                 goto out_dput;
1099         }
1100         if (!security_vm_enough_memory(p->pages))
1101                 vm_unacct_memory(p->pages);
1102         else {
1103                 err = -ENOMEM;
1104                 spin_unlock(&swap_lock);
1105                 goto out_dput;
1106         }
1107         if (prev < 0) {
1108                 swap_list.head = p->next;
1109         } else {
1110                 swap_info[prev].next = p->next;
1111         }
1112         if (type == swap_list.next) {
1113                 /* just pick something that's safe... */
1114                 swap_list.next = swap_list.head;
1115         }
1116         nr_swap_pages -= p->pages;
1117         total_swap_pages -= p->pages;
1118         p->flags &= ~SWP_WRITEOK;
1119         spin_unlock(&swap_lock);
1120
1121         current->flags |= PF_SWAPOFF;
1122         err = try_to_unuse(type);
1123         current->flags &= ~PF_SWAPOFF;
1124
1125         if (err) {
1126                 /* re-insert swap space back into swap_list */
1127                 spin_lock(&swap_lock);
1128                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1129                         if (p->prio >= swap_info[i].prio)
1130                                 break;
1131                 p->next = i;
1132                 if (prev < 0)
1133                         swap_list.head = swap_list.next = p - swap_info;
1134                 else
1135                         swap_info[prev].next = p - swap_info;
1136                 nr_swap_pages += p->pages;
1137                 total_swap_pages += p->pages;
1138                 p->flags |= SWP_WRITEOK;
1139                 spin_unlock(&swap_lock);
1140                 goto out_dput;
1141         }
1142
1143         /* wait for any unplug function to finish */
1144         down_write(&swap_unplug_sem);
1145         up_write(&swap_unplug_sem);
1146
1147         destroy_swap_extents(p);
1148         down(&swapon_sem);
1149         spin_lock(&swap_lock);
1150         drain_mmlist();
1151
1152         /* wait for anyone still in scan_swap_map */
1153         p->highest_bit = 0;             /* cuts scans short */
1154         while (p->flags >= SWP_SCANNING) {
1155                 spin_unlock(&swap_lock);
1156                 set_current_state(TASK_UNINTERRUPTIBLE);
1157                 schedule_timeout(1);
1158                 spin_lock(&swap_lock);
1159         }
1160
1161         swap_file = p->swap_file;
1162         p->swap_file = NULL;
1163         p->max = 0;
1164         swap_map = p->swap_map;
1165         p->swap_map = NULL;
1166         p->flags = 0;
1167         spin_unlock(&swap_lock);
1168         up(&swapon_sem);
1169         vfree(swap_map);
1170         inode = mapping->host;
1171         if (S_ISBLK(inode->i_mode)) {
1172                 struct block_device *bdev = I_BDEV(inode);
1173                 set_blocksize(bdev, p->old_block_size);
1174                 bd_release(bdev);
1175         } else {
1176                 down(&inode->i_sem);
1177                 inode->i_flags &= ~S_SWAPFILE;
1178                 up(&inode->i_sem);
1179         }
1180         filp_close(swap_file, NULL);
1181         err = 0;
1182
1183 out_dput:
1184         filp_close(victim, NULL);
1185 out:
1186         return err;
1187 }
1188
1189 #ifdef CONFIG_PROC_FS
1190 /* iterator */
1191 static void *swap_start(struct seq_file *swap, loff_t *pos)
1192 {
1193         struct swap_info_struct *ptr = swap_info;
1194         int i;
1195         loff_t l = *pos;
1196
1197         down(&swapon_sem);
1198
1199         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1200                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1201                         continue;
1202                 if (!l--)
1203                         return ptr;
1204         }
1205
1206         return NULL;
1207 }
1208
1209 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1210 {
1211         struct swap_info_struct *ptr = v;
1212         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1213
1214         for (++ptr; ptr < endptr; ptr++) {
1215                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1216                         continue;
1217                 ++*pos;
1218                 return ptr;
1219         }
1220
1221         return NULL;
1222 }
1223
1224 static void swap_stop(struct seq_file *swap, void *v)
1225 {
1226         up(&swapon_sem);
1227 }
1228
1229 static int swap_show(struct seq_file *swap, void *v)
1230 {
1231         struct swap_info_struct *ptr = v;
1232         struct file *file;
1233         int len;
1234
1235         if (v == swap_info)
1236                 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1237
1238         file = ptr->swap_file;
1239         len = seq_path(swap, file->f_vfsmnt, file->f_dentry, " \t\n\\");
1240         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1241                        len < 40 ? 40 - len : 1, " ",
1242                        S_ISBLK(file->f_dentry->d_inode->i_mode) ?
1243                                 "partition" : "file\t",
1244                        ptr->pages << (PAGE_SHIFT - 10),
1245                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1246                        ptr->prio);
1247         return 0;
1248 }
1249
1250 static struct seq_operations swaps_op = {
1251         .start =        swap_start,
1252         .next =         swap_next,
1253         .stop =         swap_stop,
1254         .show =         swap_show
1255 };
1256
1257 static int swaps_open(struct inode *inode, struct file *file)
1258 {
1259         return seq_open(file, &swaps_op);
1260 }
1261
1262 static struct file_operations proc_swaps_operations = {
1263         .open           = swaps_open,
1264         .read           = seq_read,
1265         .llseek         = seq_lseek,
1266         .release        = seq_release,
1267 };
1268
1269 static int __init procswaps_init(void)
1270 {
1271         struct proc_dir_entry *entry;
1272
1273         entry = create_proc_entry("swaps", 0, NULL);
1274         if (entry)
1275                 entry->proc_fops = &proc_swaps_operations;
1276         return 0;
1277 }
1278 __initcall(procswaps_init);
1279 #endif /* CONFIG_PROC_FS */
1280
1281 /*
1282  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1283  *
1284  * The swapon system call
1285  */
1286 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1287 {
1288         struct swap_info_struct * p;
1289         char *name = NULL;
1290         struct block_device *bdev = NULL;
1291         struct file *swap_file = NULL;
1292         struct address_space *mapping;
1293         unsigned int type;
1294         int i, prev;
1295         int error;
1296         static int least_priority;
1297         union swap_header *swap_header = NULL;
1298         int swap_header_version;
1299         unsigned int nr_good_pages = 0;
1300         int nr_extents = 0;
1301         sector_t span;
1302         unsigned long maxpages = 1;
1303         int swapfilesize;
1304         unsigned short *swap_map;
1305         struct page *page = NULL;
1306         struct inode *inode = NULL;
1307         int did_down = 0;
1308
1309         if (!capable(CAP_SYS_ADMIN))
1310                 return -EPERM;
1311         spin_lock(&swap_lock);
1312         p = swap_info;
1313         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1314                 if (!(p->flags & SWP_USED))
1315                         break;
1316         error = -EPERM;
1317         /*
1318          * Test if adding another swap device is possible. There are
1319          * two limiting factors: 1) the number of bits for the swap
1320          * type swp_entry_t definition and 2) the number of bits for
1321          * the swap type in the swap ptes as defined by the different
1322          * architectures. To honor both limitations a swap entry
1323          * with swap offset 0 and swap type ~0UL is created, encoded
1324          * to a swap pte, decoded to a swp_entry_t again and finally
1325          * the swap type part is extracted. This will mask all bits
1326          * from the initial ~0UL that can't be encoded in either the
1327          * swp_entry_t or the architecture definition of a swap pte.
1328          */
1329         if (type > swp_type(pte_to_swp_entry(swp_entry_to_pte(swp_entry(~0UL,0))))) {
1330                 spin_unlock(&swap_lock);
1331                 goto out;
1332         }
1333         if (type >= nr_swapfiles)
1334                 nr_swapfiles = type+1;
1335         INIT_LIST_HEAD(&p->extent_list);
1336         p->flags = SWP_USED;
1337         p->swap_file = NULL;
1338         p->old_block_size = 0;
1339         p->swap_map = NULL;
1340         p->lowest_bit = 0;
1341         p->highest_bit = 0;
1342         p->cluster_nr = 0;
1343         p->inuse_pages = 0;
1344         p->next = -1;
1345         if (swap_flags & SWAP_FLAG_PREFER) {
1346                 p->prio =
1347                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1348         } else {
1349                 p->prio = --least_priority;
1350         }
1351         spin_unlock(&swap_lock);
1352         name = getname(specialfile);
1353         error = PTR_ERR(name);
1354         if (IS_ERR(name)) {
1355                 name = NULL;
1356                 goto bad_swap_2;
1357         }
1358         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1359         error = PTR_ERR(swap_file);
1360         if (IS_ERR(swap_file)) {
1361                 swap_file = NULL;
1362                 goto bad_swap_2;
1363         }
1364
1365         p->swap_file = swap_file;
1366         mapping = swap_file->f_mapping;
1367         inode = mapping->host;
1368
1369         error = -EBUSY;
1370         for (i = 0; i < nr_swapfiles; i++) {
1371                 struct swap_info_struct *q = &swap_info[i];
1372
1373                 if (i == type || !q->swap_file)
1374                         continue;
1375                 if (mapping == q->swap_file->f_mapping)
1376                         goto bad_swap;
1377         }
1378
1379         error = -EINVAL;
1380         if (S_ISBLK(inode->i_mode)) {
1381                 bdev = I_BDEV(inode);
1382                 error = bd_claim(bdev, sys_swapon);
1383                 if (error < 0) {
1384                         bdev = NULL;
1385                         goto bad_swap;
1386                 }
1387                 p->old_block_size = block_size(bdev);
1388                 error = set_blocksize(bdev, PAGE_SIZE);
1389                 if (error < 0)
1390                         goto bad_swap;
1391                 p->bdev = bdev;
1392         } else if (S_ISREG(inode->i_mode)) {
1393                 p->bdev = inode->i_sb->s_bdev;
1394                 down(&inode->i_sem);
1395                 did_down = 1;
1396                 if (IS_SWAPFILE(inode)) {
1397                         error = -EBUSY;
1398                         goto bad_swap;
1399                 }
1400         } else {
1401                 goto bad_swap;
1402         }
1403
1404         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1405
1406         /*
1407          * Read the swap header.
1408          */
1409         if (!mapping->a_ops->readpage) {
1410                 error = -EINVAL;
1411                 goto bad_swap;
1412         }
1413         page = read_cache_page(mapping, 0,
1414                         (filler_t *)mapping->a_ops->readpage, swap_file);
1415         if (IS_ERR(page)) {
1416                 error = PTR_ERR(page);
1417                 goto bad_swap;
1418         }
1419         wait_on_page_locked(page);
1420         if (!PageUptodate(page))
1421                 goto bad_swap;
1422         kmap(page);
1423         swap_header = page_address(page);
1424
1425         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1426                 swap_header_version = 1;
1427         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1428                 swap_header_version = 2;
1429         else {
1430                 printk("Unable to find swap-space signature\n");
1431                 error = -EINVAL;
1432                 goto bad_swap;
1433         }
1434         
1435         switch (swap_header_version) {
1436         case 1:
1437                 printk(KERN_ERR "version 0 swap is no longer supported. "
1438                         "Use mkswap -v1 %s\n", name);
1439                 error = -EINVAL;
1440                 goto bad_swap;
1441         case 2:
1442                 /* Check the swap header's sub-version and the size of
1443                    the swap file and bad block lists */
1444                 if (swap_header->info.version != 1) {
1445                         printk(KERN_WARNING
1446                                "Unable to handle swap header version %d\n",
1447                                swap_header->info.version);
1448                         error = -EINVAL;
1449                         goto bad_swap;
1450                 }
1451
1452                 p->lowest_bit  = 1;
1453                 p->cluster_next = 1;
1454
1455                 /*
1456                  * Find out how many pages are allowed for a single swap
1457                  * device. There are two limiting factors: 1) the number of
1458                  * bits for the swap offset in the swp_entry_t type and
1459                  * 2) the number of bits in the a swap pte as defined by
1460                  * the different architectures. In order to find the
1461                  * largest possible bit mask a swap entry with swap type 0
1462                  * and swap offset ~0UL is created, encoded to a swap pte,
1463                  * decoded to a swp_entry_t again and finally the swap
1464                  * offset is extracted. This will mask all the bits from
1465                  * the initial ~0UL mask that can't be encoded in either
1466                  * the swp_entry_t or the architecture definition of a
1467                  * swap pte.
1468                  */
1469                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1470                 if (maxpages > swap_header->info.last_page)
1471                         maxpages = swap_header->info.last_page;
1472                 p->highest_bit = maxpages - 1;
1473
1474                 error = -EINVAL;
1475                 if (!maxpages)
1476                         goto bad_swap;
1477                 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1478                         goto bad_swap;
1479                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1480                         goto bad_swap;
1481                 
1482                 /* OK, set up the swap map and apply the bad block list */
1483                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1484                         error = -ENOMEM;
1485                         goto bad_swap;
1486                 }
1487
1488                 error = 0;
1489                 memset(p->swap_map, 0, maxpages * sizeof(short));
1490                 for (i=0; i<swap_header->info.nr_badpages; i++) {
1491                         int page = swap_header->info.badpages[i];
1492                         if (page <= 0 || page >= swap_header->info.last_page)
1493                                 error = -EINVAL;
1494                         else
1495                                 p->swap_map[page] = SWAP_MAP_BAD;
1496                 }
1497                 nr_good_pages = swap_header->info.last_page -
1498                                 swap_header->info.nr_badpages -
1499                                 1 /* header page */;
1500                 if (error) 
1501                         goto bad_swap;
1502         }
1503
1504         if (swapfilesize && maxpages > swapfilesize) {
1505                 printk(KERN_WARNING
1506                        "Swap area shorter than signature indicates\n");
1507                 error = -EINVAL;
1508                 goto bad_swap;
1509         }
1510         if (nr_good_pages) {
1511                 p->swap_map[0] = SWAP_MAP_BAD;
1512                 p->max = maxpages;
1513                 p->pages = nr_good_pages;
1514                 nr_extents = setup_swap_extents(p, &span);
1515                 if (nr_extents < 0) {
1516                         error = nr_extents;
1517                         goto bad_swap;
1518                 }
1519                 nr_good_pages = p->pages;
1520         }
1521         if (!nr_good_pages) {
1522                 printk(KERN_WARNING "Empty swap-file\n");
1523                 error = -EINVAL;
1524                 goto bad_swap;
1525         }
1526
1527         down(&swapon_sem);
1528         spin_lock(&swap_lock);
1529         p->flags = SWP_ACTIVE;
1530         nr_swap_pages += nr_good_pages;
1531         total_swap_pages += nr_good_pages;
1532
1533         printk(KERN_INFO "Adding %uk swap on %s.  "
1534                         "Priority:%d extents:%d across:%lluk\n",
1535                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1536                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1537
1538         /* insert swap space into swap_list: */
1539         prev = -1;
1540         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1541                 if (p->prio >= swap_info[i].prio) {
1542                         break;
1543                 }
1544                 prev = i;
1545         }
1546         p->next = i;
1547         if (prev < 0) {
1548                 swap_list.head = swap_list.next = p - swap_info;
1549         } else {
1550                 swap_info[prev].next = p - swap_info;
1551         }
1552         spin_unlock(&swap_lock);
1553         up(&swapon_sem);
1554         error = 0;
1555         goto out;
1556 bad_swap:
1557         if (bdev) {
1558                 set_blocksize(bdev, p->old_block_size);
1559                 bd_release(bdev);
1560         }
1561         destroy_swap_extents(p);
1562 bad_swap_2:
1563         spin_lock(&swap_lock);
1564         swap_map = p->swap_map;
1565         p->swap_file = NULL;
1566         p->swap_map = NULL;
1567         p->flags = 0;
1568         if (!(swap_flags & SWAP_FLAG_PREFER))
1569                 ++least_priority;
1570         spin_unlock(&swap_lock);
1571         vfree(swap_map);
1572         if (swap_file)
1573                 filp_close(swap_file, NULL);
1574 out:
1575         if (page && !IS_ERR(page)) {
1576                 kunmap(page);
1577                 page_cache_release(page);
1578         }
1579         if (name)
1580                 putname(name);
1581         if (did_down) {
1582                 if (!error)
1583                         inode->i_flags |= S_SWAPFILE;
1584                 up(&inode->i_sem);
1585         }
1586         return error;
1587 }
1588
1589 void si_swapinfo(struct sysinfo *val)
1590 {
1591         unsigned int i;
1592         unsigned long nr_to_be_unused = 0;
1593
1594         spin_lock(&swap_lock);
1595         for (i = 0; i < nr_swapfiles; i++) {
1596                 if (!(swap_info[i].flags & SWP_USED) ||
1597                      (swap_info[i].flags & SWP_WRITEOK))
1598                         continue;
1599                 nr_to_be_unused += swap_info[i].inuse_pages;
1600         }
1601         val->freeswap = nr_swap_pages + nr_to_be_unused;
1602         val->totalswap = total_swap_pages + nr_to_be_unused;
1603         spin_unlock(&swap_lock);
1604 }
1605
1606 /*
1607  * Verify that a swap entry is valid and increment its swap map count.
1608  *
1609  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1610  * "permanent", but will be reclaimed by the next swapoff.
1611  */
1612 int swap_duplicate(swp_entry_t entry)
1613 {
1614         struct swap_info_struct * p;
1615         unsigned long offset, type;
1616         int result = 0;
1617
1618         type = swp_type(entry);
1619         if (type >= nr_swapfiles)
1620                 goto bad_file;
1621         p = type + swap_info;
1622         offset = swp_offset(entry);
1623
1624         spin_lock(&swap_lock);
1625         if (offset < p->max && p->swap_map[offset]) {
1626                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1627                         p->swap_map[offset]++;
1628                         result = 1;
1629                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1630                         if (swap_overflow++ < 5)
1631                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1632                         p->swap_map[offset] = SWAP_MAP_MAX;
1633                         result = 1;
1634                 }
1635         }
1636         spin_unlock(&swap_lock);
1637 out:
1638         return result;
1639
1640 bad_file:
1641         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1642         goto out;
1643 }
1644
1645 struct swap_info_struct *
1646 get_swap_info_struct(unsigned type)
1647 {
1648         return &swap_info[type];
1649 }
1650
1651 /*
1652  * swap_lock prevents swap_map being freed. Don't grab an extra
1653  * reference on the swaphandle, it doesn't matter if it becomes unused.
1654  */
1655 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1656 {
1657         int ret = 0, i = 1 << page_cluster;
1658         unsigned long toff;
1659         struct swap_info_struct *swapdev = swp_type(entry) + swap_info;
1660
1661         if (!page_cluster)      /* no readahead */
1662                 return 0;
1663         toff = (swp_offset(entry) >> page_cluster) << page_cluster;
1664         if (!toff)              /* first page is swap header */
1665                 toff++, i--;
1666         *offset = toff;
1667
1668         spin_lock(&swap_lock);
1669         do {
1670                 /* Don't read-ahead past the end of the swap area */
1671                 if (toff >= swapdev->max)
1672                         break;
1673                 /* Don't read in free or bad pages */
1674                 if (!swapdev->swap_map[toff])
1675                         break;
1676                 if (swapdev->swap_map[toff] == SWAP_MAP_BAD)
1677                         break;
1678                 toff++;
1679                 ret++;
1680         } while (--i);
1681         spin_unlock(&swap_lock);
1682         return ret;
1683 }