Pull 5165 into release branch
[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 static const char Bad_file[] = "Bad swap file entry ";
40 static const char Unused_file[] = "Unused swap file entry ";
41 static const char Bad_offset[] = "Bad swap offset entry ";
42 static const char Unused_offset[] = "Unused swap offset entry ";
43
44 struct swap_list_t swap_list = {-1, -1};
45
46 struct swap_info_struct swap_info[MAX_SWAPFILES];
47
48 static DECLARE_MUTEX(swapon_sem);
49
50 /*
51  * We need this because the bdev->unplug_fn can sleep and we cannot
52  * hold swap_lock while calling the unplug_fn. And swap_lock
53  * cannot be turned into a semaphore.
54  */
55 static DECLARE_RWSEM(swap_unplug_sem);
56
57 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
58 {
59         swp_entry_t entry;
60
61         down_read(&swap_unplug_sem);
62         entry.val = page_private(page);
63         if (PageSwapCache(page)) {
64                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
65                 struct backing_dev_info *bdi;
66
67                 /*
68                  * If the page is removed from swapcache from under us (with a
69                  * racy try_to_unuse/swapoff) we need an additional reference
70                  * count to avoid reading garbage from page_private(page) above.
71                  * If the WARN_ON triggers during a swapoff it maybe the race
72                  * condition and it's harmless. However if it triggers without
73                  * swapoff it signals a problem.
74                  */
75                 WARN_ON(page_count(page) <= 1);
76
77                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
78                 blk_run_backing_dev(bdi, page);
79         }
80         up_read(&swap_unplug_sem);
81 }
82
83 #define SWAPFILE_CLUSTER        256
84 #define LATENCY_LIMIT           256
85
86 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
87 {
88         unsigned long offset, last_in_cluster;
89         int latency_ration = LATENCY_LIMIT;
90
91         /* 
92          * We try to cluster swap pages by allocating them sequentially
93          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
94          * way, however, we resort to first-free allocation, starting
95          * a new cluster.  This prevents us from scattering swap pages
96          * all over the entire swap partition, so that we reduce
97          * overall disk seek times between swap pages.  -- sct
98          * But we do now try to find an empty cluster.  -Andrea
99          */
100
101         si->flags += SWP_SCANNING;
102         if (unlikely(!si->cluster_nr)) {
103                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
104                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER)
105                         goto lowest;
106                 spin_unlock(&swap_lock);
107
108                 offset = si->lowest_bit;
109                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
110
111                 /* Locate the first empty (unaligned) cluster */
112                 for (; last_in_cluster <= si->highest_bit; offset++) {
113                         if (si->swap_map[offset])
114                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
115                         else if (offset == last_in_cluster) {
116                                 spin_lock(&swap_lock);
117                                 si->cluster_next = offset-SWAPFILE_CLUSTER-1;
118                                 goto cluster;
119                         }
120                         if (unlikely(--latency_ration < 0)) {
121                                 cond_resched();
122                                 latency_ration = LATENCY_LIMIT;
123                         }
124                 }
125                 spin_lock(&swap_lock);
126                 goto lowest;
127         }
128
129         si->cluster_nr--;
130 cluster:
131         offset = si->cluster_next;
132         if (offset > si->highest_bit)
133 lowest:         offset = si->lowest_bit;
134 checks: if (!(si->flags & SWP_WRITEOK))
135                 goto no_page;
136         if (!si->highest_bit)
137                 goto no_page;
138         if (!si->swap_map[offset]) {
139                 if (offset == si->lowest_bit)
140                         si->lowest_bit++;
141                 if (offset == si->highest_bit)
142                         si->highest_bit--;
143                 si->inuse_pages++;
144                 if (si->inuse_pages == si->pages) {
145                         si->lowest_bit = si->max;
146                         si->highest_bit = 0;
147                 }
148                 si->swap_map[offset] = 1;
149                 si->cluster_next = offset + 1;
150                 si->flags -= SWP_SCANNING;
151                 return offset;
152         }
153
154         spin_unlock(&swap_lock);
155         while (++offset <= si->highest_bit) {
156                 if (!si->swap_map[offset]) {
157                         spin_lock(&swap_lock);
158                         goto checks;
159                 }
160                 if (unlikely(--latency_ration < 0)) {
161                         cond_resched();
162                         latency_ration = LATENCY_LIMIT;
163                 }
164         }
165         spin_lock(&swap_lock);
166         goto lowest;
167
168 no_page:
169         si->flags -= SWP_SCANNING;
170         return 0;
171 }
172
173 swp_entry_t get_swap_page(void)
174 {
175         struct swap_info_struct *si;
176         pgoff_t offset;
177         int type, next;
178         int wrapped = 0;
179
180         spin_lock(&swap_lock);
181         if (nr_swap_pages <= 0)
182                 goto noswap;
183         nr_swap_pages--;
184
185         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
186                 si = swap_info + type;
187                 next = si->next;
188                 if (next < 0 ||
189                     (!wrapped && si->prio != swap_info[next].prio)) {
190                         next = swap_list.head;
191                         wrapped++;
192                 }
193
194                 if (!si->highest_bit)
195                         continue;
196                 if (!(si->flags & SWP_WRITEOK))
197                         continue;
198
199                 swap_list.next = next;
200                 offset = scan_swap_map(si);
201                 if (offset) {
202                         spin_unlock(&swap_lock);
203                         return swp_entry(type, offset);
204                 }
205                 next = swap_list.next;
206         }
207
208         nr_swap_pages++;
209 noswap:
210         spin_unlock(&swap_lock);
211         return (swp_entry_t) {0};
212 }
213
214 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
215 {
216         struct swap_info_struct * p;
217         unsigned long offset, type;
218
219         if (!entry.val)
220                 goto out;
221         type = swp_type(entry);
222         if (type >= nr_swapfiles)
223                 goto bad_nofile;
224         p = & swap_info[type];
225         if (!(p->flags & SWP_USED))
226                 goto bad_device;
227         offset = swp_offset(entry);
228         if (offset >= p->max)
229                 goto bad_offset;
230         if (!p->swap_map[offset])
231                 goto bad_free;
232         spin_lock(&swap_lock);
233         return p;
234
235 bad_free:
236         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
237         goto out;
238 bad_offset:
239         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
240         goto out;
241 bad_device:
242         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
243         goto out;
244 bad_nofile:
245         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
246 out:
247         return NULL;
248 }       
249
250 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
251 {
252         int count = p->swap_map[offset];
253
254         if (count < SWAP_MAP_MAX) {
255                 count--;
256                 p->swap_map[offset] = count;
257                 if (!count) {
258                         if (offset < p->lowest_bit)
259                                 p->lowest_bit = offset;
260                         if (offset > p->highest_bit)
261                                 p->highest_bit = offset;
262                         if (p->prio > swap_info[swap_list.next].prio)
263                                 swap_list.next = p - swap_info;
264                         nr_swap_pages++;
265                         p->inuse_pages--;
266                 }
267         }
268         return count;
269 }
270
271 /*
272  * Caller has made sure that the swapdevice corresponding to entry
273  * is still around or has not been recycled.
274  */
275 void swap_free(swp_entry_t entry)
276 {
277         struct swap_info_struct * p;
278
279         p = swap_info_get(entry);
280         if (p) {
281                 swap_entry_free(p, swp_offset(entry));
282                 spin_unlock(&swap_lock);
283         }
284 }
285
286 /*
287  * How many references to page are currently swapped out?
288  */
289 static inline int page_swapcount(struct page *page)
290 {
291         int count = 0;
292         struct swap_info_struct *p;
293         swp_entry_t entry;
294
295         entry.val = page_private(page);
296         p = swap_info_get(entry);
297         if (p) {
298                 /* Subtract the 1 for the swap cache itself */
299                 count = p->swap_map[swp_offset(entry)] - 1;
300                 spin_unlock(&swap_lock);
301         }
302         return count;
303 }
304
305 /*
306  * We can use this swap cache entry directly
307  * if there are no other references to it.
308  */
309 int can_share_swap_page(struct page *page)
310 {
311         int count;
312
313         BUG_ON(!PageLocked(page));
314         count = page_mapcount(page);
315         if (count <= 1 && PageSwapCache(page))
316                 count += page_swapcount(page);
317         return count == 1;
318 }
319
320 /*
321  * Work out if there are any other processes sharing this
322  * swap cache page. Free it if you can. Return success.
323  */
324 int remove_exclusive_swap_page(struct page *page)
325 {
326         int retval;
327         struct swap_info_struct * p;
328         swp_entry_t entry;
329
330         BUG_ON(PagePrivate(page));
331         BUG_ON(!PageLocked(page));
332
333         if (!PageSwapCache(page))
334                 return 0;
335         if (PageWriteback(page))
336                 return 0;
337         if (page_count(page) != 2) /* 2: us + cache */
338                 return 0;
339
340         entry.val = page_private(page);
341         p = swap_info_get(entry);
342         if (!p)
343                 return 0;
344
345         /* Is the only swap cache user the cache itself? */
346         retval = 0;
347         if (p->swap_map[swp_offset(entry)] == 1) {
348                 /* Recheck the page count with the swapcache lock held.. */
349                 write_lock_irq(&swapper_space.tree_lock);
350                 if ((page_count(page) == 2) && !PageWriteback(page)) {
351                         __delete_from_swap_cache(page);
352                         SetPageDirty(page);
353                         retval = 1;
354                 }
355                 write_unlock_irq(&swapper_space.tree_lock);
356         }
357         spin_unlock(&swap_lock);
358
359         if (retval) {
360                 swap_free(entry);
361                 page_cache_release(page);
362         }
363
364         return retval;
365 }
366
367 /*
368  * Free the swap entry like above, but also try to
369  * free the page cache entry if it is the last user.
370  */
371 void free_swap_and_cache(swp_entry_t entry)
372 {
373         struct swap_info_struct * p;
374         struct page *page = NULL;
375
376         p = swap_info_get(entry);
377         if (p) {
378                 if (swap_entry_free(p, swp_offset(entry)) == 1)
379                         page = find_trylock_page(&swapper_space, entry.val);
380                 spin_unlock(&swap_lock);
381         }
382         if (page) {
383                 int one_user;
384
385                 BUG_ON(PagePrivate(page));
386                 page_cache_get(page);
387                 one_user = (page_count(page) == 2);
388                 /* Only cache user (+us), or swap space full? Free it! */
389                 if (!PageWriteback(page) && (one_user || vm_swap_full())) {
390                         delete_from_swap_cache(page);
391                         SetPageDirty(page);
392                 }
393                 unlock_page(page);
394                 page_cache_release(page);
395         }
396 }
397
398 /*
399  * No need to decide whether this PTE shares the swap entry with others,
400  * just let do_wp_page work it out if a write is requested later - to
401  * force COW, vm_page_prot omits write permission from any private vma.
402  */
403 static void unuse_pte(struct vm_area_struct *vma, pte_t *pte,
404                 unsigned long addr, swp_entry_t entry, struct page *page)
405 {
406         inc_mm_counter(vma->vm_mm, anon_rss);
407         get_page(page);
408         set_pte_at(vma->vm_mm, addr, pte,
409                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
410         page_add_anon_rmap(page, vma, addr);
411         swap_free(entry);
412         /*
413          * Move the page to the active list so it is not
414          * immediately swapped out again after swapon.
415          */
416         activate_page(page);
417 }
418
419 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
420                                 unsigned long addr, unsigned long end,
421                                 swp_entry_t entry, struct page *page)
422 {
423         pte_t swp_pte = swp_entry_to_pte(entry);
424         pte_t *pte;
425         spinlock_t *ptl;
426         int found = 0;
427
428         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
429         do {
430                 /*
431                  * swapoff spends a _lot_ of time in this loop!
432                  * Test inline before going to call unuse_pte.
433                  */
434                 if (unlikely(pte_same(*pte, swp_pte))) {
435                         unuse_pte(vma, pte++, addr, entry, page);
436                         found = 1;
437                         break;
438                 }
439         } while (pte++, addr += PAGE_SIZE, addr != end);
440         pte_unmap_unlock(pte - 1, ptl);
441         return found;
442 }
443
444 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
445                                 unsigned long addr, unsigned long end,
446                                 swp_entry_t entry, struct page *page)
447 {
448         pmd_t *pmd;
449         unsigned long next;
450
451         pmd = pmd_offset(pud, addr);
452         do {
453                 next = pmd_addr_end(addr, end);
454                 if (pmd_none_or_clear_bad(pmd))
455                         continue;
456                 if (unuse_pte_range(vma, pmd, addr, next, entry, page))
457                         return 1;
458         } while (pmd++, addr = next, addr != end);
459         return 0;
460 }
461
462 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
463                                 unsigned long addr, unsigned long end,
464                                 swp_entry_t entry, struct page *page)
465 {
466         pud_t *pud;
467         unsigned long next;
468
469         pud = pud_offset(pgd, addr);
470         do {
471                 next = pud_addr_end(addr, end);
472                 if (pud_none_or_clear_bad(pud))
473                         continue;
474                 if (unuse_pmd_range(vma, pud, addr, next, entry, page))
475                         return 1;
476         } while (pud++, addr = next, addr != end);
477         return 0;
478 }
479
480 static int unuse_vma(struct vm_area_struct *vma,
481                                 swp_entry_t entry, struct page *page)
482 {
483         pgd_t *pgd;
484         unsigned long addr, end, next;
485
486         if (page->mapping) {
487                 addr = page_address_in_vma(page, vma);
488                 if (addr == -EFAULT)
489                         return 0;
490                 else
491                         end = addr + PAGE_SIZE;
492         } else {
493                 addr = vma->vm_start;
494                 end = vma->vm_end;
495         }
496
497         pgd = pgd_offset(vma->vm_mm, addr);
498         do {
499                 next = pgd_addr_end(addr, end);
500                 if (pgd_none_or_clear_bad(pgd))
501                         continue;
502                 if (unuse_pud_range(vma, pgd, addr, next, entry, page))
503                         return 1;
504         } while (pgd++, addr = next, addr != end);
505         return 0;
506 }
507
508 static int unuse_mm(struct mm_struct *mm,
509                                 swp_entry_t entry, struct page *page)
510 {
511         struct vm_area_struct *vma;
512
513         if (!down_read_trylock(&mm->mmap_sem)) {
514                 /*
515                  * Activate page so shrink_cache is unlikely to unmap its
516                  * ptes while lock is dropped, so swapoff can make progress.
517                  */
518                 activate_page(page);
519                 unlock_page(page);
520                 down_read(&mm->mmap_sem);
521                 lock_page(page);
522         }
523         for (vma = mm->mmap; vma; vma = vma->vm_next) {
524                 if (vma->anon_vma && unuse_vma(vma, entry, page))
525                         break;
526         }
527         up_read(&mm->mmap_sem);
528         /*
529          * Currently unuse_mm cannot fail, but leave error handling
530          * at call sites for now, since we change it from time to time.
531          */
532         return 0;
533 }
534
535 /*
536  * Scan swap_map from current position to next entry still in use.
537  * Recycle to start on reaching the end, returning 0 when empty.
538  */
539 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
540                                         unsigned int prev)
541 {
542         unsigned int max = si->max;
543         unsigned int i = prev;
544         int count;
545
546         /*
547          * No need for swap_lock here: we're just looking
548          * for whether an entry is in use, not modifying it; false
549          * hits are okay, and sys_swapoff() has already prevented new
550          * allocations from this area (while holding swap_lock).
551          */
552         for (;;) {
553                 if (++i >= max) {
554                         if (!prev) {
555                                 i = 0;
556                                 break;
557                         }
558                         /*
559                          * No entries in use at top of swap_map,
560                          * loop back to start and recheck there.
561                          */
562                         max = prev + 1;
563                         prev = 0;
564                         i = 1;
565                 }
566                 count = si->swap_map[i];
567                 if (count && count != SWAP_MAP_BAD)
568                         break;
569         }
570         return i;
571 }
572
573 /*
574  * We completely avoid races by reading each swap page in advance,
575  * and then search for the process using it.  All the necessary
576  * page table adjustments can then be made atomically.
577  */
578 static int try_to_unuse(unsigned int type)
579 {
580         struct swap_info_struct * si = &swap_info[type];
581         struct mm_struct *start_mm;
582         unsigned short *swap_map;
583         unsigned short swcount;
584         struct page *page;
585         swp_entry_t entry;
586         unsigned int i = 0;
587         int retval = 0;
588         int reset_overflow = 0;
589         int shmem;
590
591         /*
592          * When searching mms for an entry, a good strategy is to
593          * start at the first mm we freed the previous entry from
594          * (though actually we don't notice whether we or coincidence
595          * freed the entry).  Initialize this start_mm with a hold.
596          *
597          * A simpler strategy would be to start at the last mm we
598          * freed the previous entry from; but that would take less
599          * advantage of mmlist ordering, which clusters forked mms
600          * together, child after parent.  If we race with dup_mmap(), we
601          * prefer to resolve parent before child, lest we miss entries
602          * duplicated after we scanned child: using last mm would invert
603          * that.  Though it's only a serious concern when an overflowed
604          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
605          */
606         start_mm = &init_mm;
607         atomic_inc(&init_mm.mm_users);
608
609         /*
610          * Keep on scanning until all entries have gone.  Usually,
611          * one pass through swap_map is enough, but not necessarily:
612          * there are races when an instance of an entry might be missed.
613          */
614         while ((i = find_next_to_unuse(si, i)) != 0) {
615                 if (signal_pending(current)) {
616                         retval = -EINTR;
617                         break;
618                 }
619
620                 /* 
621                  * Get a page for the entry, using the existing swap
622                  * cache page if there is one.  Otherwise, get a clean
623                  * page and read the swap into it. 
624                  */
625                 swap_map = &si->swap_map[i];
626                 entry = swp_entry(type, i);
627                 page = read_swap_cache_async(entry, NULL, 0);
628                 if (!page) {
629                         /*
630                          * Either swap_duplicate() failed because entry
631                          * has been freed independently, and will not be
632                          * reused since sys_swapoff() already disabled
633                          * allocation from here, or alloc_page() failed.
634                          */
635                         if (!*swap_map)
636                                 continue;
637                         retval = -ENOMEM;
638                         break;
639                 }
640
641                 /*
642                  * Don't hold on to start_mm if it looks like exiting.
643                  */
644                 if (atomic_read(&start_mm->mm_users) == 1) {
645                         mmput(start_mm);
646                         start_mm = &init_mm;
647                         atomic_inc(&init_mm.mm_users);
648                 }
649
650                 /*
651                  * Wait for and lock page.  When do_swap_page races with
652                  * try_to_unuse, do_swap_page can handle the fault much
653                  * faster than try_to_unuse can locate the entry.  This
654                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
655                  * defer to do_swap_page in such a case - in some tests,
656                  * do_swap_page and try_to_unuse repeatedly compete.
657                  */
658                 wait_on_page_locked(page);
659                 wait_on_page_writeback(page);
660                 lock_page(page);
661                 wait_on_page_writeback(page);
662
663                 /*
664                  * Remove all references to entry.
665                  * Whenever we reach init_mm, there's no address space
666                  * to search, but use it as a reminder to search shmem.
667                  */
668                 shmem = 0;
669                 swcount = *swap_map;
670                 if (swcount > 1) {
671                         if (start_mm == &init_mm)
672                                 shmem = shmem_unuse(entry, page);
673                         else
674                                 retval = unuse_mm(start_mm, entry, page);
675                 }
676                 if (*swap_map > 1) {
677                         int set_start_mm = (*swap_map >= swcount);
678                         struct list_head *p = &start_mm->mmlist;
679                         struct mm_struct *new_start_mm = start_mm;
680                         struct mm_struct *prev_mm = start_mm;
681                         struct mm_struct *mm;
682
683                         atomic_inc(&new_start_mm->mm_users);
684                         atomic_inc(&prev_mm->mm_users);
685                         spin_lock(&mmlist_lock);
686                         while (*swap_map > 1 && !retval &&
687                                         (p = p->next) != &start_mm->mmlist) {
688                                 mm = list_entry(p, struct mm_struct, mmlist);
689                                 if (atomic_inc_return(&mm->mm_users) == 1) {
690                                         atomic_dec(&mm->mm_users);
691                                         continue;
692                                 }
693                                 spin_unlock(&mmlist_lock);
694                                 mmput(prev_mm);
695                                 prev_mm = mm;
696
697                                 cond_resched();
698
699                                 swcount = *swap_map;
700                                 if (swcount <= 1)
701                                         ;
702                                 else if (mm == &init_mm) {
703                                         set_start_mm = 1;
704                                         shmem = shmem_unuse(entry, page);
705                                 } else
706                                         retval = unuse_mm(mm, entry, page);
707                                 if (set_start_mm && *swap_map < swcount) {
708                                         mmput(new_start_mm);
709                                         atomic_inc(&mm->mm_users);
710                                         new_start_mm = mm;
711                                         set_start_mm = 0;
712                                 }
713                                 spin_lock(&mmlist_lock);
714                         }
715                         spin_unlock(&mmlist_lock);
716                         mmput(prev_mm);
717                         mmput(start_mm);
718                         start_mm = new_start_mm;
719                 }
720                 if (retval) {
721                         unlock_page(page);
722                         page_cache_release(page);
723                         break;
724                 }
725
726                 /*
727                  * How could swap count reach 0x7fff when the maximum
728                  * pid is 0x7fff, and there's no way to repeat a swap
729                  * page within an mm (except in shmem, where it's the
730                  * shared object which takes the reference count)?
731                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
732                  *
733                  * If that's wrong, then we should worry more about
734                  * exit_mmap() and do_munmap() cases described above:
735                  * we might be resetting SWAP_MAP_MAX too early here.
736                  * We know "Undead"s can happen, they're okay, so don't
737                  * report them; but do report if we reset SWAP_MAP_MAX.
738                  */
739                 if (*swap_map == SWAP_MAP_MAX) {
740                         spin_lock(&swap_lock);
741                         *swap_map = 1;
742                         spin_unlock(&swap_lock);
743                         reset_overflow = 1;
744                 }
745
746                 /*
747                  * If a reference remains (rare), we would like to leave
748                  * the page in the swap cache; but try_to_unmap could
749                  * then re-duplicate the entry once we drop page lock,
750                  * so we might loop indefinitely; also, that page could
751                  * not be swapped out to other storage meanwhile.  So:
752                  * delete from cache even if there's another reference,
753                  * after ensuring that the data has been saved to disk -
754                  * since if the reference remains (rarer), it will be
755                  * read from disk into another page.  Splitting into two
756                  * pages would be incorrect if swap supported "shared
757                  * private" pages, but they are handled by tmpfs files.
758                  *
759                  * Note shmem_unuse already deleted a swappage from
760                  * the swap cache, unless the move to filepage failed:
761                  * in which case it left swappage in cache, lowered its
762                  * swap count to pass quickly through the loops above,
763                  * and now we must reincrement count to try again later.
764                  */
765                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
766                         struct writeback_control wbc = {
767                                 .sync_mode = WB_SYNC_NONE,
768                         };
769
770                         swap_writepage(page, &wbc);
771                         lock_page(page);
772                         wait_on_page_writeback(page);
773                 }
774                 if (PageSwapCache(page)) {
775                         if (shmem)
776                                 swap_duplicate(entry);
777                         else
778                                 delete_from_swap_cache(page);
779                 }
780
781                 /*
782                  * So we could skip searching mms once swap count went
783                  * to 1, we did not mark any present ptes as dirty: must
784                  * mark page dirty so shrink_list will preserve it.
785                  */
786                 SetPageDirty(page);
787                 unlock_page(page);
788                 page_cache_release(page);
789
790                 /*
791                  * Make sure that we aren't completely killing
792                  * interactive performance.
793                  */
794                 cond_resched();
795         }
796
797         mmput(start_mm);
798         if (reset_overflow) {
799                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
800                 swap_overflow = 0;
801         }
802         return retval;
803 }
804
805 /*
806  * After a successful try_to_unuse, if no swap is now in use, we know
807  * we can empty the mmlist.  swap_lock must be held on entry and exit.
808  * Note that mmlist_lock nests inside swap_lock, and an mm must be
809  * added to the mmlist just after page_duplicate - before would be racy.
810  */
811 static void drain_mmlist(void)
812 {
813         struct list_head *p, *next;
814         unsigned int i;
815
816         for (i = 0; i < nr_swapfiles; i++)
817                 if (swap_info[i].inuse_pages)
818                         return;
819         spin_lock(&mmlist_lock);
820         list_for_each_safe(p, next, &init_mm.mmlist)
821                 list_del_init(p);
822         spin_unlock(&mmlist_lock);
823 }
824
825 /*
826  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
827  * corresponds to page offset `offset'.
828  */
829 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
830 {
831         struct swap_extent *se = sis->curr_swap_extent;
832         struct swap_extent *start_se = se;
833
834         for ( ; ; ) {
835                 struct list_head *lh;
836
837                 if (se->start_page <= offset &&
838                                 offset < (se->start_page + se->nr_pages)) {
839                         return se->start_block + (offset - se->start_page);
840                 }
841                 lh = se->list.next;
842                 if (lh == &sis->extent_list)
843                         lh = lh->next;
844                 se = list_entry(lh, struct swap_extent, list);
845                 sis->curr_swap_extent = se;
846                 BUG_ON(se == start_se);         /* It *must* be present */
847         }
848 }
849
850 /*
851  * Free all of a swapdev's extent information
852  */
853 static void destroy_swap_extents(struct swap_info_struct *sis)
854 {
855         while (!list_empty(&sis->extent_list)) {
856                 struct swap_extent *se;
857
858                 se = list_entry(sis->extent_list.next,
859                                 struct swap_extent, list);
860                 list_del(&se->list);
861                 kfree(se);
862         }
863 }
864
865 /*
866  * Add a block range (and the corresponding page range) into this swapdev's
867  * extent list.  The extent list is kept sorted in page order.
868  *
869  * This function rather assumes that it is called in ascending page order.
870  */
871 static int
872 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
873                 unsigned long nr_pages, sector_t start_block)
874 {
875         struct swap_extent *se;
876         struct swap_extent *new_se;
877         struct list_head *lh;
878
879         lh = sis->extent_list.prev;     /* The highest page extent */
880         if (lh != &sis->extent_list) {
881                 se = list_entry(lh, struct swap_extent, list);
882                 BUG_ON(se->start_page + se->nr_pages != start_page);
883                 if (se->start_block + se->nr_pages == start_block) {
884                         /* Merge it */
885                         se->nr_pages += nr_pages;
886                         return 0;
887                 }
888         }
889
890         /*
891          * No merge.  Insert a new extent, preserving ordering.
892          */
893         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
894         if (new_se == NULL)
895                 return -ENOMEM;
896         new_se->start_page = start_page;
897         new_se->nr_pages = nr_pages;
898         new_se->start_block = start_block;
899
900         list_add_tail(&new_se->list, &sis->extent_list);
901         return 1;
902 }
903
904 /*
905  * A `swap extent' is a simple thing which maps a contiguous range of pages
906  * onto a contiguous range of disk blocks.  An ordered list of swap extents
907  * is built at swapon time and is then used at swap_writepage/swap_readpage
908  * time for locating where on disk a page belongs.
909  *
910  * If the swapfile is an S_ISBLK block device, a single extent is installed.
911  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
912  * swap files identically.
913  *
914  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
915  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
916  * swapfiles are handled *identically* after swapon time.
917  *
918  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
919  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
920  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
921  * requirements, they are simply tossed out - we will never use those blocks
922  * for swapping.
923  *
924  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
925  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
926  * which will scribble on the fs.
927  *
928  * The amount of disk space which a single swap extent represents varies.
929  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
930  * extents in the list.  To avoid much list walking, we cache the previous
931  * search location in `curr_swap_extent', and start new searches from there.
932  * This is extremely effective.  The average number of iterations in
933  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
934  */
935 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
936 {
937         struct inode *inode;
938         unsigned blocks_per_page;
939         unsigned long page_no;
940         unsigned blkbits;
941         sector_t probe_block;
942         sector_t last_block;
943         sector_t lowest_block = -1;
944         sector_t highest_block = 0;
945         int nr_extents = 0;
946         int ret;
947
948         inode = sis->swap_file->f_mapping->host;
949         if (S_ISBLK(inode->i_mode)) {
950                 ret = add_swap_extent(sis, 0, sis->max, 0);
951                 *span = sis->pages;
952                 goto done;
953         }
954
955         blkbits = inode->i_blkbits;
956         blocks_per_page = PAGE_SIZE >> blkbits;
957
958         /*
959          * Map all the blocks into the extent list.  This code doesn't try
960          * to be very smart.
961          */
962         probe_block = 0;
963         page_no = 0;
964         last_block = i_size_read(inode) >> blkbits;
965         while ((probe_block + blocks_per_page) <= last_block &&
966                         page_no < sis->max) {
967                 unsigned block_in_page;
968                 sector_t first_block;
969
970                 first_block = bmap(inode, probe_block);
971                 if (first_block == 0)
972                         goto bad_bmap;
973
974                 /*
975                  * It must be PAGE_SIZE aligned on-disk
976                  */
977                 if (first_block & (blocks_per_page - 1)) {
978                         probe_block++;
979                         goto reprobe;
980                 }
981
982                 for (block_in_page = 1; block_in_page < blocks_per_page;
983                                         block_in_page++) {
984                         sector_t block;
985
986                         block = bmap(inode, probe_block + block_in_page);
987                         if (block == 0)
988                                 goto bad_bmap;
989                         if (block != first_block + block_in_page) {
990                                 /* Discontiguity */
991                                 probe_block++;
992                                 goto reprobe;
993                         }
994                 }
995
996                 first_block >>= (PAGE_SHIFT - blkbits);
997                 if (page_no) {  /* exclude the header page */
998                         if (first_block < lowest_block)
999                                 lowest_block = first_block;
1000                         if (first_block > highest_block)
1001                                 highest_block = first_block;
1002                 }
1003
1004                 /*
1005                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1006                  */
1007                 ret = add_swap_extent(sis, page_no, 1, first_block);
1008                 if (ret < 0)
1009                         goto out;
1010                 nr_extents += ret;
1011                 page_no++;
1012                 probe_block += blocks_per_page;
1013 reprobe:
1014                 continue;
1015         }
1016         ret = nr_extents;
1017         *span = 1 + highest_block - lowest_block;
1018         if (page_no == 0)
1019                 page_no = 1;    /* force Empty message */
1020         sis->max = page_no;
1021         sis->pages = page_no - 1;
1022         sis->highest_bit = page_no - 1;
1023 done:
1024         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1025                                         struct swap_extent, list);
1026         goto out;
1027 bad_bmap:
1028         printk(KERN_ERR "swapon: swapfile has holes\n");
1029         ret = -EINVAL;
1030 out:
1031         return ret;
1032 }
1033
1034 #if 0   /* We don't need this yet */
1035 #include <linux/backing-dev.h>
1036 int page_queue_congested(struct page *page)
1037 {
1038         struct backing_dev_info *bdi;
1039
1040         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1041
1042         if (PageSwapCache(page)) {
1043                 swp_entry_t entry = { .val = page_private(page) };
1044                 struct swap_info_struct *sis;
1045
1046                 sis = get_swap_info_struct(swp_type(entry));
1047                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1048         } else
1049                 bdi = page->mapping->backing_dev_info;
1050         return bdi_write_congested(bdi);
1051 }
1052 #endif
1053
1054 asmlinkage long sys_swapoff(const char __user * specialfile)
1055 {
1056         struct swap_info_struct * p = NULL;
1057         unsigned short *swap_map;
1058         struct file *swap_file, *victim;
1059         struct address_space *mapping;
1060         struct inode *inode;
1061         char * pathname;
1062         int i, type, prev;
1063         int err;
1064         
1065         if (!capable(CAP_SYS_ADMIN))
1066                 return -EPERM;
1067
1068         pathname = getname(specialfile);
1069         err = PTR_ERR(pathname);
1070         if (IS_ERR(pathname))
1071                 goto out;
1072
1073         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1074         putname(pathname);
1075         err = PTR_ERR(victim);
1076         if (IS_ERR(victim))
1077                 goto out;
1078
1079         mapping = victim->f_mapping;
1080         prev = -1;
1081         spin_lock(&swap_lock);
1082         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1083                 p = swap_info + type;
1084                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1085                         if (p->swap_file->f_mapping == mapping)
1086                                 break;
1087                 }
1088                 prev = type;
1089         }
1090         if (type < 0) {
1091                 err = -EINVAL;
1092                 spin_unlock(&swap_lock);
1093                 goto out_dput;
1094         }
1095         if (!security_vm_enough_memory(p->pages))
1096                 vm_unacct_memory(p->pages);
1097         else {
1098                 err = -ENOMEM;
1099                 spin_unlock(&swap_lock);
1100                 goto out_dput;
1101         }
1102         if (prev < 0) {
1103                 swap_list.head = p->next;
1104         } else {
1105                 swap_info[prev].next = p->next;
1106         }
1107         if (type == swap_list.next) {
1108                 /* just pick something that's safe... */
1109                 swap_list.next = swap_list.head;
1110         }
1111         nr_swap_pages -= p->pages;
1112         total_swap_pages -= p->pages;
1113         p->flags &= ~SWP_WRITEOK;
1114         spin_unlock(&swap_lock);
1115
1116         current->flags |= PF_SWAPOFF;
1117         err = try_to_unuse(type);
1118         current->flags &= ~PF_SWAPOFF;
1119
1120         if (err) {
1121                 /* re-insert swap space back into swap_list */
1122                 spin_lock(&swap_lock);
1123                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1124                         if (p->prio >= swap_info[i].prio)
1125                                 break;
1126                 p->next = i;
1127                 if (prev < 0)
1128                         swap_list.head = swap_list.next = p - swap_info;
1129                 else
1130                         swap_info[prev].next = p - swap_info;
1131                 nr_swap_pages += p->pages;
1132                 total_swap_pages += p->pages;
1133                 p->flags |= SWP_WRITEOK;
1134                 spin_unlock(&swap_lock);
1135                 goto out_dput;
1136         }
1137
1138         /* wait for any unplug function to finish */
1139         down_write(&swap_unplug_sem);
1140         up_write(&swap_unplug_sem);
1141
1142         destroy_swap_extents(p);
1143         down(&swapon_sem);
1144         spin_lock(&swap_lock);
1145         drain_mmlist();
1146
1147         /* wait for anyone still in scan_swap_map */
1148         p->highest_bit = 0;             /* cuts scans short */
1149         while (p->flags >= SWP_SCANNING) {
1150                 spin_unlock(&swap_lock);
1151                 schedule_timeout_uninterruptible(1);
1152                 spin_lock(&swap_lock);
1153         }
1154
1155         swap_file = p->swap_file;
1156         p->swap_file = NULL;
1157         p->max = 0;
1158         swap_map = p->swap_map;
1159         p->swap_map = NULL;
1160         p->flags = 0;
1161         spin_unlock(&swap_lock);
1162         up(&swapon_sem);
1163         vfree(swap_map);
1164         inode = mapping->host;
1165         if (S_ISBLK(inode->i_mode)) {
1166                 struct block_device *bdev = I_BDEV(inode);
1167                 set_blocksize(bdev, p->old_block_size);
1168                 bd_release(bdev);
1169         } else {
1170                 down(&inode->i_sem);
1171                 inode->i_flags &= ~S_SWAPFILE;
1172                 up(&inode->i_sem);
1173         }
1174         filp_close(swap_file, NULL);
1175         err = 0;
1176
1177 out_dput:
1178         filp_close(victim, NULL);
1179 out:
1180         return err;
1181 }
1182
1183 #ifdef CONFIG_PROC_FS
1184 /* iterator */
1185 static void *swap_start(struct seq_file *swap, loff_t *pos)
1186 {
1187         struct swap_info_struct *ptr = swap_info;
1188         int i;
1189         loff_t l = *pos;
1190
1191         down(&swapon_sem);
1192
1193         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1194                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1195                         continue;
1196                 if (!l--)
1197                         return ptr;
1198         }
1199
1200         return NULL;
1201 }
1202
1203 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1204 {
1205         struct swap_info_struct *ptr = v;
1206         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1207
1208         for (++ptr; ptr < endptr; ptr++) {
1209                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1210                         continue;
1211                 ++*pos;
1212                 return ptr;
1213         }
1214
1215         return NULL;
1216 }
1217
1218 static void swap_stop(struct seq_file *swap, void *v)
1219 {
1220         up(&swapon_sem);
1221 }
1222
1223 static int swap_show(struct seq_file *swap, void *v)
1224 {
1225         struct swap_info_struct *ptr = v;
1226         struct file *file;
1227         int len;
1228
1229         if (v == swap_info)
1230                 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1231
1232         file = ptr->swap_file;
1233         len = seq_path(swap, file->f_vfsmnt, file->f_dentry, " \t\n\\");
1234         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1235                        len < 40 ? 40 - len : 1, " ",
1236                        S_ISBLK(file->f_dentry->d_inode->i_mode) ?
1237                                 "partition" : "file\t",
1238                        ptr->pages << (PAGE_SHIFT - 10),
1239                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1240                        ptr->prio);
1241         return 0;
1242 }
1243
1244 static struct seq_operations swaps_op = {
1245         .start =        swap_start,
1246         .next =         swap_next,
1247         .stop =         swap_stop,
1248         .show =         swap_show
1249 };
1250
1251 static int swaps_open(struct inode *inode, struct file *file)
1252 {
1253         return seq_open(file, &swaps_op);
1254 }
1255
1256 static struct file_operations proc_swaps_operations = {
1257         .open           = swaps_open,
1258         .read           = seq_read,
1259         .llseek         = seq_lseek,
1260         .release        = seq_release,
1261 };
1262
1263 static int __init procswaps_init(void)
1264 {
1265         struct proc_dir_entry *entry;
1266
1267         entry = create_proc_entry("swaps", 0, NULL);
1268         if (entry)
1269                 entry->proc_fops = &proc_swaps_operations;
1270         return 0;
1271 }
1272 __initcall(procswaps_init);
1273 #endif /* CONFIG_PROC_FS */
1274
1275 /*
1276  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1277  *
1278  * The swapon system call
1279  */
1280 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1281 {
1282         struct swap_info_struct * p;
1283         char *name = NULL;
1284         struct block_device *bdev = NULL;
1285         struct file *swap_file = NULL;
1286         struct address_space *mapping;
1287         unsigned int type;
1288         int i, prev;
1289         int error;
1290         static int least_priority;
1291         union swap_header *swap_header = NULL;
1292         int swap_header_version;
1293         unsigned int nr_good_pages = 0;
1294         int nr_extents = 0;
1295         sector_t span;
1296         unsigned long maxpages = 1;
1297         int swapfilesize;
1298         unsigned short *swap_map;
1299         struct page *page = NULL;
1300         struct inode *inode = NULL;
1301         int did_down = 0;
1302
1303         if (!capable(CAP_SYS_ADMIN))
1304                 return -EPERM;
1305         spin_lock(&swap_lock);
1306         p = swap_info;
1307         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1308                 if (!(p->flags & SWP_USED))
1309                         break;
1310         error = -EPERM;
1311         /*
1312          * Test if adding another swap device is possible. There are
1313          * two limiting factors: 1) the number of bits for the swap
1314          * type swp_entry_t definition and 2) the number of bits for
1315          * the swap type in the swap ptes as defined by the different
1316          * architectures. To honor both limitations a swap entry
1317          * with swap offset 0 and swap type ~0UL is created, encoded
1318          * to a swap pte, decoded to a swp_entry_t again and finally
1319          * the swap type part is extracted. This will mask all bits
1320          * from the initial ~0UL that can't be encoded in either the
1321          * swp_entry_t or the architecture definition of a swap pte.
1322          */
1323         if (type > swp_type(pte_to_swp_entry(swp_entry_to_pte(swp_entry(~0UL,0))))) {
1324                 spin_unlock(&swap_lock);
1325                 goto out;
1326         }
1327         if (type >= nr_swapfiles)
1328                 nr_swapfiles = type+1;
1329         INIT_LIST_HEAD(&p->extent_list);
1330         p->flags = SWP_USED;
1331         p->swap_file = NULL;
1332         p->old_block_size = 0;
1333         p->swap_map = NULL;
1334         p->lowest_bit = 0;
1335         p->highest_bit = 0;
1336         p->cluster_nr = 0;
1337         p->inuse_pages = 0;
1338         p->next = -1;
1339         if (swap_flags & SWAP_FLAG_PREFER) {
1340                 p->prio =
1341                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1342         } else {
1343                 p->prio = --least_priority;
1344         }
1345         spin_unlock(&swap_lock);
1346         name = getname(specialfile);
1347         error = PTR_ERR(name);
1348         if (IS_ERR(name)) {
1349                 name = NULL;
1350                 goto bad_swap_2;
1351         }
1352         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1353         error = PTR_ERR(swap_file);
1354         if (IS_ERR(swap_file)) {
1355                 swap_file = NULL;
1356                 goto bad_swap_2;
1357         }
1358
1359         p->swap_file = swap_file;
1360         mapping = swap_file->f_mapping;
1361         inode = mapping->host;
1362
1363         error = -EBUSY;
1364         for (i = 0; i < nr_swapfiles; i++) {
1365                 struct swap_info_struct *q = &swap_info[i];
1366
1367                 if (i == type || !q->swap_file)
1368                         continue;
1369                 if (mapping == q->swap_file->f_mapping)
1370                         goto bad_swap;
1371         }
1372
1373         error = -EINVAL;
1374         if (S_ISBLK(inode->i_mode)) {
1375                 bdev = I_BDEV(inode);
1376                 error = bd_claim(bdev, sys_swapon);
1377                 if (error < 0) {
1378                         bdev = NULL;
1379                         error = -EINVAL;
1380                         goto bad_swap;
1381                 }
1382                 p->old_block_size = block_size(bdev);
1383                 error = set_blocksize(bdev, PAGE_SIZE);
1384                 if (error < 0)
1385                         goto bad_swap;
1386                 p->bdev = bdev;
1387         } else if (S_ISREG(inode->i_mode)) {
1388                 p->bdev = inode->i_sb->s_bdev;
1389                 down(&inode->i_sem);
1390                 did_down = 1;
1391                 if (IS_SWAPFILE(inode)) {
1392                         error = -EBUSY;
1393                         goto bad_swap;
1394                 }
1395         } else {
1396                 goto bad_swap;
1397         }
1398
1399         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1400
1401         /*
1402          * Read the swap header.
1403          */
1404         if (!mapping->a_ops->readpage) {
1405                 error = -EINVAL;
1406                 goto bad_swap;
1407         }
1408         page = read_cache_page(mapping, 0,
1409                         (filler_t *)mapping->a_ops->readpage, swap_file);
1410         if (IS_ERR(page)) {
1411                 error = PTR_ERR(page);
1412                 goto bad_swap;
1413         }
1414         wait_on_page_locked(page);
1415         if (!PageUptodate(page))
1416                 goto bad_swap;
1417         kmap(page);
1418         swap_header = page_address(page);
1419
1420         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1421                 swap_header_version = 1;
1422         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1423                 swap_header_version = 2;
1424         else {
1425                 printk("Unable to find swap-space signature\n");
1426                 error = -EINVAL;
1427                 goto bad_swap;
1428         }
1429         
1430         switch (swap_header_version) {
1431         case 1:
1432                 printk(KERN_ERR "version 0 swap is no longer supported. "
1433                         "Use mkswap -v1 %s\n", name);
1434                 error = -EINVAL;
1435                 goto bad_swap;
1436         case 2:
1437                 /* Check the swap header's sub-version and the size of
1438                    the swap file and bad block lists */
1439                 if (swap_header->info.version != 1) {
1440                         printk(KERN_WARNING
1441                                "Unable to handle swap header version %d\n",
1442                                swap_header->info.version);
1443                         error = -EINVAL;
1444                         goto bad_swap;
1445                 }
1446
1447                 p->lowest_bit  = 1;
1448                 p->cluster_next = 1;
1449
1450                 /*
1451                  * Find out how many pages are allowed for a single swap
1452                  * device. There are two limiting factors: 1) the number of
1453                  * bits for the swap offset in the swp_entry_t type and
1454                  * 2) the number of bits in the a swap pte as defined by
1455                  * the different architectures. In order to find the
1456                  * largest possible bit mask a swap entry with swap type 0
1457                  * and swap offset ~0UL is created, encoded to a swap pte,
1458                  * decoded to a swp_entry_t again and finally the swap
1459                  * offset is extracted. This will mask all the bits from
1460                  * the initial ~0UL mask that can't be encoded in either
1461                  * the swp_entry_t or the architecture definition of a
1462                  * swap pte.
1463                  */
1464                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1465                 if (maxpages > swap_header->info.last_page)
1466                         maxpages = swap_header->info.last_page;
1467                 p->highest_bit = maxpages - 1;
1468
1469                 error = -EINVAL;
1470                 if (!maxpages)
1471                         goto bad_swap;
1472                 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1473                         goto bad_swap;
1474                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1475                         goto bad_swap;
1476                 
1477                 /* OK, set up the swap map and apply the bad block list */
1478                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1479                         error = -ENOMEM;
1480                         goto bad_swap;
1481                 }
1482
1483                 error = 0;
1484                 memset(p->swap_map, 0, maxpages * sizeof(short));
1485                 for (i=0; i<swap_header->info.nr_badpages; i++) {
1486                         int page = swap_header->info.badpages[i];
1487                         if (page <= 0 || page >= swap_header->info.last_page)
1488                                 error = -EINVAL;
1489                         else
1490                                 p->swap_map[page] = SWAP_MAP_BAD;
1491                 }
1492                 nr_good_pages = swap_header->info.last_page -
1493                                 swap_header->info.nr_badpages -
1494                                 1 /* header page */;
1495                 if (error) 
1496                         goto bad_swap;
1497         }
1498
1499         if (swapfilesize && maxpages > swapfilesize) {
1500                 printk(KERN_WARNING
1501                        "Swap area shorter than signature indicates\n");
1502                 error = -EINVAL;
1503                 goto bad_swap;
1504         }
1505         if (nr_good_pages) {
1506                 p->swap_map[0] = SWAP_MAP_BAD;
1507                 p->max = maxpages;
1508                 p->pages = nr_good_pages;
1509                 nr_extents = setup_swap_extents(p, &span);
1510                 if (nr_extents < 0) {
1511                         error = nr_extents;
1512                         goto bad_swap;
1513                 }
1514                 nr_good_pages = p->pages;
1515         }
1516         if (!nr_good_pages) {
1517                 printk(KERN_WARNING "Empty swap-file\n");
1518                 error = -EINVAL;
1519                 goto bad_swap;
1520         }
1521
1522         down(&swapon_sem);
1523         spin_lock(&swap_lock);
1524         p->flags = SWP_ACTIVE;
1525         nr_swap_pages += nr_good_pages;
1526         total_swap_pages += nr_good_pages;
1527
1528         printk(KERN_INFO "Adding %uk swap on %s.  "
1529                         "Priority:%d extents:%d across:%lluk\n",
1530                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1531                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1532
1533         /* insert swap space into swap_list: */
1534         prev = -1;
1535         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1536                 if (p->prio >= swap_info[i].prio) {
1537                         break;
1538                 }
1539                 prev = i;
1540         }
1541         p->next = i;
1542         if (prev < 0) {
1543                 swap_list.head = swap_list.next = p - swap_info;
1544         } else {
1545                 swap_info[prev].next = p - swap_info;
1546         }
1547         spin_unlock(&swap_lock);
1548         up(&swapon_sem);
1549         error = 0;
1550         goto out;
1551 bad_swap:
1552         if (bdev) {
1553                 set_blocksize(bdev, p->old_block_size);
1554                 bd_release(bdev);
1555         }
1556         destroy_swap_extents(p);
1557 bad_swap_2:
1558         spin_lock(&swap_lock);
1559         swap_map = p->swap_map;
1560         p->swap_file = NULL;
1561         p->swap_map = NULL;
1562         p->flags = 0;
1563         if (!(swap_flags & SWAP_FLAG_PREFER))
1564                 ++least_priority;
1565         spin_unlock(&swap_lock);
1566         vfree(swap_map);
1567         if (swap_file)
1568                 filp_close(swap_file, NULL);
1569 out:
1570         if (page && !IS_ERR(page)) {
1571                 kunmap(page);
1572                 page_cache_release(page);
1573         }
1574         if (name)
1575                 putname(name);
1576         if (did_down) {
1577                 if (!error)
1578                         inode->i_flags |= S_SWAPFILE;
1579                 up(&inode->i_sem);
1580         }
1581         return error;
1582 }
1583
1584 void si_swapinfo(struct sysinfo *val)
1585 {
1586         unsigned int i;
1587         unsigned long nr_to_be_unused = 0;
1588
1589         spin_lock(&swap_lock);
1590         for (i = 0; i < nr_swapfiles; i++) {
1591                 if (!(swap_info[i].flags & SWP_USED) ||
1592                      (swap_info[i].flags & SWP_WRITEOK))
1593                         continue;
1594                 nr_to_be_unused += swap_info[i].inuse_pages;
1595         }
1596         val->freeswap = nr_swap_pages + nr_to_be_unused;
1597         val->totalswap = total_swap_pages + nr_to_be_unused;
1598         spin_unlock(&swap_lock);
1599 }
1600
1601 /*
1602  * Verify that a swap entry is valid and increment its swap map count.
1603  *
1604  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1605  * "permanent", but will be reclaimed by the next swapoff.
1606  */
1607 int swap_duplicate(swp_entry_t entry)
1608 {
1609         struct swap_info_struct * p;
1610         unsigned long offset, type;
1611         int result = 0;
1612
1613         type = swp_type(entry);
1614         if (type >= nr_swapfiles)
1615                 goto bad_file;
1616         p = type + swap_info;
1617         offset = swp_offset(entry);
1618
1619         spin_lock(&swap_lock);
1620         if (offset < p->max && p->swap_map[offset]) {
1621                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1622                         p->swap_map[offset]++;
1623                         result = 1;
1624                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1625                         if (swap_overflow++ < 5)
1626                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1627                         p->swap_map[offset] = SWAP_MAP_MAX;
1628                         result = 1;
1629                 }
1630         }
1631         spin_unlock(&swap_lock);
1632 out:
1633         return result;
1634
1635 bad_file:
1636         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1637         goto out;
1638 }
1639
1640 struct swap_info_struct *
1641 get_swap_info_struct(unsigned type)
1642 {
1643         return &swap_info[type];
1644 }
1645
1646 /*
1647  * swap_lock prevents swap_map being freed. Don't grab an extra
1648  * reference on the swaphandle, it doesn't matter if it becomes unused.
1649  */
1650 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1651 {
1652         int ret = 0, i = 1 << page_cluster;
1653         unsigned long toff;
1654         struct swap_info_struct *swapdev = swp_type(entry) + swap_info;
1655
1656         if (!page_cluster)      /* no readahead */
1657                 return 0;
1658         toff = (swp_offset(entry) >> page_cluster) << page_cluster;
1659         if (!toff)              /* first page is swap header */
1660                 toff++, i--;
1661         *offset = toff;
1662
1663         spin_lock(&swap_lock);
1664         do {
1665                 /* Don't read-ahead past the end of the swap area */
1666                 if (toff >= swapdev->max)
1667                         break;
1668                 /* Don't read in free or bad pages */
1669                 if (!swapdev->swap_map[toff])
1670                         break;
1671                 if (swapdev->swap_map[toff] == SWAP_MAP_BAD)
1672                         break;
1673                 toff++;
1674                 ret++;
1675         } while (--i);
1676         spin_unlock(&swap_lock);
1677         return ret;
1678 }