[PATCH] Direct Migration V9: PageSwapCache checks
[linux-2.6.git] / mm / swapfile.c
1 /*
2  *  linux/mm/swapfile.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *  Swap reorganised 29.12.95, Stephen Tweedie
6  */
7
8 #include <linux/config.h>
9 #include <linux/mm.h>
10 #include <linux/hugetlb.h>
11 #include <linux/mman.h>
12 #include <linux/slab.h>
13 #include <linux/kernel_stat.h>
14 #include <linux/swap.h>
15 #include <linux/vmalloc.h>
16 #include <linux/pagemap.h>
17 #include <linux/namei.h>
18 #include <linux/shm.h>
19 #include <linux/blkdev.h>
20 #include <linux/writeback.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/init.h>
24 #include <linux/module.h>
25 #include <linux/rmap.h>
26 #include <linux/security.h>
27 #include <linux/backing-dev.h>
28 #include <linux/mutex.h>
29 #include <linux/capability.h>
30 #include <linux/syscalls.h>
31
32 #include <asm/pgtable.h>
33 #include <asm/tlbflush.h>
34 #include <linux/swapops.h>
35
36 DEFINE_SPINLOCK(swap_lock);
37 unsigned int nr_swapfiles;
38 long total_swap_pages;
39 static int swap_overflow;
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 DEFINE_MUTEX(swapon_mutex);
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 mutex.
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(page);
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(page) above.
73                  * If 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 swp_entry_t get_swap_page_of_type(int type)
217 {
218         struct swap_info_struct *si;
219         pgoff_t offset;
220
221         spin_lock(&swap_lock);
222         si = swap_info + type;
223         if (si->flags & SWP_WRITEOK) {
224                 nr_swap_pages--;
225                 offset = scan_swap_map(si);
226                 if (offset) {
227                         spin_unlock(&swap_lock);
228                         return swp_entry(type, offset);
229                 }
230                 nr_swap_pages++;
231         }
232         spin_unlock(&swap_lock);
233         return (swp_entry_t) {0};
234 }
235
236 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
237 {
238         struct swap_info_struct * p;
239         unsigned long offset, type;
240
241         if (!entry.val)
242                 goto out;
243         type = swp_type(entry);
244         if (type >= nr_swapfiles)
245                 goto bad_nofile;
246         p = & swap_info[type];
247         if (!(p->flags & SWP_USED))
248                 goto bad_device;
249         offset = swp_offset(entry);
250         if (offset >= p->max)
251                 goto bad_offset;
252         if (!p->swap_map[offset])
253                 goto bad_free;
254         spin_lock(&swap_lock);
255         return p;
256
257 bad_free:
258         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
259         goto out;
260 bad_offset:
261         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
262         goto out;
263 bad_device:
264         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
265         goto out;
266 bad_nofile:
267         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
268 out:
269         return NULL;
270 }       
271
272 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
273 {
274         int count = p->swap_map[offset];
275
276         if (count < SWAP_MAP_MAX) {
277                 count--;
278                 p->swap_map[offset] = count;
279                 if (!count) {
280                         if (offset < p->lowest_bit)
281                                 p->lowest_bit = offset;
282                         if (offset > p->highest_bit)
283                                 p->highest_bit = offset;
284                         if (p->prio > swap_info[swap_list.next].prio)
285                                 swap_list.next = p - swap_info;
286                         nr_swap_pages++;
287                         p->inuse_pages--;
288                 }
289         }
290         return count;
291 }
292
293 /*
294  * Caller has made sure that the swapdevice corresponding to entry
295  * is still around or has not been recycled.
296  */
297 void swap_free(swp_entry_t entry)
298 {
299         struct swap_info_struct * p;
300
301         p = swap_info_get(entry);
302         if (p) {
303                 swap_entry_free(p, swp_offset(entry));
304                 spin_unlock(&swap_lock);
305         }
306 }
307
308 /*
309  * How many references to page are currently swapped out?
310  */
311 static inline int page_swapcount(struct page *page)
312 {
313         int count = 0;
314         struct swap_info_struct *p;
315         swp_entry_t entry;
316
317         entry.val = page_private(page);
318         p = swap_info_get(entry);
319         if (p) {
320                 /* Subtract the 1 for the swap cache itself */
321                 count = p->swap_map[swp_offset(entry)] - 1;
322                 spin_unlock(&swap_lock);
323         }
324         return count;
325 }
326
327 /*
328  * We can use this swap cache entry directly
329  * if there are no other references to it.
330  */
331 int can_share_swap_page(struct page *page)
332 {
333         int count;
334
335         BUG_ON(!PageLocked(page));
336         count = page_mapcount(page);
337         if (count <= 1 && PageSwapCache(page))
338                 count += page_swapcount(page);
339         return count == 1;
340 }
341
342 /*
343  * Work out if there are any other processes sharing this
344  * swap cache page. Free it if you can. Return success.
345  */
346 int remove_exclusive_swap_page(struct page *page)
347 {
348         int retval;
349         struct swap_info_struct * p;
350         swp_entry_t entry;
351
352         BUG_ON(PagePrivate(page));
353         BUG_ON(!PageLocked(page));
354
355         if (!PageSwapCache(page))
356                 return 0;
357         if (PageWriteback(page))
358                 return 0;
359         if (page_count(page) != 2) /* 2: us + cache */
360                 return 0;
361
362         entry.val = page_private(page);
363         p = swap_info_get(entry);
364         if (!p)
365                 return 0;
366
367         /* Is the only swap cache user the cache itself? */
368         retval = 0;
369         if (p->swap_map[swp_offset(entry)] == 1) {
370                 /* Recheck the page count with the swapcache lock held.. */
371                 write_lock_irq(&swapper_space.tree_lock);
372                 if ((page_count(page) == 2) && !PageWriteback(page)) {
373                         __delete_from_swap_cache(page);
374                         SetPageDirty(page);
375                         retval = 1;
376                 }
377                 write_unlock_irq(&swapper_space.tree_lock);
378         }
379         spin_unlock(&swap_lock);
380
381         if (retval) {
382                 swap_free(entry);
383                 page_cache_release(page);
384         }
385
386         return retval;
387 }
388
389 /*
390  * Free the swap entry like above, but also try to
391  * free the page cache entry if it is the last user.
392  */
393 void free_swap_and_cache(swp_entry_t entry)
394 {
395         struct swap_info_struct * p;
396         struct page *page = NULL;
397
398         p = swap_info_get(entry);
399         if (p) {
400                 if (swap_entry_free(p, swp_offset(entry)) == 1)
401                         page = find_trylock_page(&swapper_space, entry.val);
402                 spin_unlock(&swap_lock);
403         }
404         if (page) {
405                 int one_user;
406
407                 BUG_ON(PagePrivate(page));
408                 page_cache_get(page);
409                 one_user = (page_count(page) == 2);
410                 /* Only cache user (+us), or swap space full? Free it! */
411                 if (!PageWriteback(page) && (one_user || vm_swap_full())) {
412                         delete_from_swap_cache(page);
413                         SetPageDirty(page);
414                 }
415                 unlock_page(page);
416                 page_cache_release(page);
417         }
418 }
419
420 /*
421  * No need to decide whether this PTE shares the swap entry with others,
422  * just let do_wp_page work it out if a write is requested later - to
423  * force COW, vm_page_prot omits write permission from any private vma.
424  */
425 static void unuse_pte(struct vm_area_struct *vma, pte_t *pte,
426                 unsigned long addr, swp_entry_t entry, struct page *page)
427 {
428         inc_mm_counter(vma->vm_mm, anon_rss);
429         get_page(page);
430         set_pte_at(vma->vm_mm, addr, pte,
431                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
432         page_add_anon_rmap(page, vma, addr);
433         swap_free(entry);
434         /*
435          * Move the page to the active list so it is not
436          * immediately swapped out again after swapon.
437          */
438         activate_page(page);
439 }
440
441 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
442                                 unsigned long addr, unsigned long end,
443                                 swp_entry_t entry, struct page *page)
444 {
445         pte_t swp_pte = swp_entry_to_pte(entry);
446         pte_t *pte;
447         spinlock_t *ptl;
448         int found = 0;
449
450         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
451         do {
452                 /*
453                  * swapoff spends a _lot_ of time in this loop!
454                  * Test inline before going to call unuse_pte.
455                  */
456                 if (unlikely(pte_same(*pte, swp_pte))) {
457                         unuse_pte(vma, pte++, addr, entry, page);
458                         found = 1;
459                         break;
460                 }
461         } while (pte++, addr += PAGE_SIZE, addr != end);
462         pte_unmap_unlock(pte - 1, ptl);
463         return found;
464 }
465
466 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
467                                 unsigned long addr, unsigned long end,
468                                 swp_entry_t entry, struct page *page)
469 {
470         pmd_t *pmd;
471         unsigned long next;
472
473         pmd = pmd_offset(pud, addr);
474         do {
475                 next = pmd_addr_end(addr, end);
476                 if (pmd_none_or_clear_bad(pmd))
477                         continue;
478                 if (unuse_pte_range(vma, pmd, addr, next, entry, page))
479                         return 1;
480         } while (pmd++, addr = next, addr != end);
481         return 0;
482 }
483
484 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
485                                 unsigned long addr, unsigned long end,
486                                 swp_entry_t entry, struct page *page)
487 {
488         pud_t *pud;
489         unsigned long next;
490
491         pud = pud_offset(pgd, addr);
492         do {
493                 next = pud_addr_end(addr, end);
494                 if (pud_none_or_clear_bad(pud))
495                         continue;
496                 if (unuse_pmd_range(vma, pud, addr, next, entry, page))
497                         return 1;
498         } while (pud++, addr = next, addr != end);
499         return 0;
500 }
501
502 static int unuse_vma(struct vm_area_struct *vma,
503                                 swp_entry_t entry, struct page *page)
504 {
505         pgd_t *pgd;
506         unsigned long addr, end, next;
507
508         if (page->mapping) {
509                 addr = page_address_in_vma(page, vma);
510                 if (addr == -EFAULT)
511                         return 0;
512                 else
513                         end = addr + PAGE_SIZE;
514         } else {
515                 addr = vma->vm_start;
516                 end = vma->vm_end;
517         }
518
519         pgd = pgd_offset(vma->vm_mm, addr);
520         do {
521                 next = pgd_addr_end(addr, end);
522                 if (pgd_none_or_clear_bad(pgd))
523                         continue;
524                 if (unuse_pud_range(vma, pgd, addr, next, entry, page))
525                         return 1;
526         } while (pgd++, addr = next, addr != end);
527         return 0;
528 }
529
530 static int unuse_mm(struct mm_struct *mm,
531                                 swp_entry_t entry, struct page *page)
532 {
533         struct vm_area_struct *vma;
534
535         if (!down_read_trylock(&mm->mmap_sem)) {
536                 /*
537                  * Activate page so shrink_cache is unlikely to unmap its
538                  * ptes while lock is dropped, so swapoff can make progress.
539                  */
540                 activate_page(page);
541                 unlock_page(page);
542                 down_read(&mm->mmap_sem);
543                 lock_page(page);
544         }
545         for (vma = mm->mmap; vma; vma = vma->vm_next) {
546                 if (vma->anon_vma && unuse_vma(vma, entry, page))
547                         break;
548         }
549         up_read(&mm->mmap_sem);
550         /*
551          * Currently unuse_mm cannot fail, but leave error handling
552          * at call sites for now, since we change it from time to time.
553          */
554         return 0;
555 }
556
557 /*
558  * Scan swap_map from current position to next entry still in use.
559  * Recycle to start on reaching the end, returning 0 when empty.
560  */
561 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
562                                         unsigned int prev)
563 {
564         unsigned int max = si->max;
565         unsigned int i = prev;
566         int count;
567
568         /*
569          * No need for swap_lock here: we're just looking
570          * for whether an entry is in use, not modifying it; false
571          * hits are okay, and sys_swapoff() has already prevented new
572          * allocations from this area (while holding swap_lock).
573          */
574         for (;;) {
575                 if (++i >= max) {
576                         if (!prev) {
577                                 i = 0;
578                                 break;
579                         }
580                         /*
581                          * No entries in use at top of swap_map,
582                          * loop back to start and recheck there.
583                          */
584                         max = prev + 1;
585                         prev = 0;
586                         i = 1;
587                 }
588                 count = si->swap_map[i];
589                 if (count && count != SWAP_MAP_BAD)
590                         break;
591         }
592         return i;
593 }
594
595 /*
596  * We completely avoid races by reading each swap page in advance,
597  * and then search for the process using it.  All the necessary
598  * page table adjustments can then be made atomically.
599  */
600 static int try_to_unuse(unsigned int type)
601 {
602         struct swap_info_struct * si = &swap_info[type];
603         struct mm_struct *start_mm;
604         unsigned short *swap_map;
605         unsigned short swcount;
606         struct page *page;
607         swp_entry_t entry;
608         unsigned int i = 0;
609         int retval = 0;
610         int reset_overflow = 0;
611         int shmem;
612
613         /*
614          * When searching mms for an entry, a good strategy is to
615          * start at the first mm we freed the previous entry from
616          * (though actually we don't notice whether we or coincidence
617          * freed the entry).  Initialize this start_mm with a hold.
618          *
619          * A simpler strategy would be to start at the last mm we
620          * freed the previous entry from; but that would take less
621          * advantage of mmlist ordering, which clusters forked mms
622          * together, child after parent.  If we race with dup_mmap(), we
623          * prefer to resolve parent before child, lest we miss entries
624          * duplicated after we scanned child: using last mm would invert
625          * that.  Though it's only a serious concern when an overflowed
626          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
627          */
628         start_mm = &init_mm;
629         atomic_inc(&init_mm.mm_users);
630
631         /*
632          * Keep on scanning until all entries have gone.  Usually,
633          * one pass through swap_map is enough, but not necessarily:
634          * there are races when an instance of an entry might be missed.
635          */
636         while ((i = find_next_to_unuse(si, i)) != 0) {
637                 if (signal_pending(current)) {
638                         retval = -EINTR;
639                         break;
640                 }
641
642                 /* 
643                  * Get a page for the entry, using the existing swap
644                  * cache page if there is one.  Otherwise, get a clean
645                  * page and read the swap into it. 
646                  */
647                 swap_map = &si->swap_map[i];
648                 entry = swp_entry(type, i);
649 again:
650                 page = read_swap_cache_async(entry, NULL, 0);
651                 if (!page) {
652                         /*
653                          * Either swap_duplicate() failed because entry
654                          * has been freed independently, and will not be
655                          * reused since sys_swapoff() already disabled
656                          * allocation from here, or alloc_page() failed.
657                          */
658                         if (!*swap_map)
659                                 continue;
660                         retval = -ENOMEM;
661                         break;
662                 }
663
664                 /*
665                  * Don't hold on to start_mm if it looks like exiting.
666                  */
667                 if (atomic_read(&start_mm->mm_users) == 1) {
668                         mmput(start_mm);
669                         start_mm = &init_mm;
670                         atomic_inc(&init_mm.mm_users);
671                 }
672
673                 /*
674                  * Wait for and lock page.  When do_swap_page races with
675                  * try_to_unuse, do_swap_page can handle the fault much
676                  * faster than try_to_unuse can locate the entry.  This
677                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
678                  * defer to do_swap_page in such a case - in some tests,
679                  * do_swap_page and try_to_unuse repeatedly compete.
680                  */
681                 wait_on_page_locked(page);
682                 wait_on_page_writeback(page);
683                 lock_page(page);
684                 if (!PageSwapCache(page)) {
685                         /* Page migration has occured */
686                         unlock_page(page);
687                         page_cache_release(page);
688                         goto again;
689                 }
690                 wait_on_page_writeback(page);
691
692                 /*
693                  * Remove all references to entry.
694                  * Whenever we reach init_mm, there's no address space
695                  * to search, but use it as a reminder to search shmem.
696                  */
697                 shmem = 0;
698                 swcount = *swap_map;
699                 if (swcount > 1) {
700                         if (start_mm == &init_mm)
701                                 shmem = shmem_unuse(entry, page);
702                         else
703                                 retval = unuse_mm(start_mm, entry, page);
704                 }
705                 if (*swap_map > 1) {
706                         int set_start_mm = (*swap_map >= swcount);
707                         struct list_head *p = &start_mm->mmlist;
708                         struct mm_struct *new_start_mm = start_mm;
709                         struct mm_struct *prev_mm = start_mm;
710                         struct mm_struct *mm;
711
712                         atomic_inc(&new_start_mm->mm_users);
713                         atomic_inc(&prev_mm->mm_users);
714                         spin_lock(&mmlist_lock);
715                         while (*swap_map > 1 && !retval &&
716                                         (p = p->next) != &start_mm->mmlist) {
717                                 mm = list_entry(p, struct mm_struct, mmlist);
718                                 if (atomic_inc_return(&mm->mm_users) == 1) {
719                                         atomic_dec(&mm->mm_users);
720                                         continue;
721                                 }
722                                 spin_unlock(&mmlist_lock);
723                                 mmput(prev_mm);
724                                 prev_mm = mm;
725
726                                 cond_resched();
727
728                                 swcount = *swap_map;
729                                 if (swcount <= 1)
730                                         ;
731                                 else if (mm == &init_mm) {
732                                         set_start_mm = 1;
733                                         shmem = shmem_unuse(entry, page);
734                                 } else
735                                         retval = unuse_mm(mm, entry, page);
736                                 if (set_start_mm && *swap_map < swcount) {
737                                         mmput(new_start_mm);
738                                         atomic_inc(&mm->mm_users);
739                                         new_start_mm = mm;
740                                         set_start_mm = 0;
741                                 }
742                                 spin_lock(&mmlist_lock);
743                         }
744                         spin_unlock(&mmlist_lock);
745                         mmput(prev_mm);
746                         mmput(start_mm);
747                         start_mm = new_start_mm;
748                 }
749                 if (retval) {
750                         unlock_page(page);
751                         page_cache_release(page);
752                         break;
753                 }
754
755                 /*
756                  * How could swap count reach 0x7fff when the maximum
757                  * pid is 0x7fff, and there's no way to repeat a swap
758                  * page within an mm (except in shmem, where it's the
759                  * shared object which takes the reference count)?
760                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
761                  *
762                  * If that's wrong, then we should worry more about
763                  * exit_mmap() and do_munmap() cases described above:
764                  * we might be resetting SWAP_MAP_MAX too early here.
765                  * We know "Undead"s can happen, they're okay, so don't
766                  * report them; but do report if we reset SWAP_MAP_MAX.
767                  */
768                 if (*swap_map == SWAP_MAP_MAX) {
769                         spin_lock(&swap_lock);
770                         *swap_map = 1;
771                         spin_unlock(&swap_lock);
772                         reset_overflow = 1;
773                 }
774
775                 /*
776                  * If a reference remains (rare), we would like to leave
777                  * the page in the swap cache; but try_to_unmap could
778                  * then re-duplicate the entry once we drop page lock,
779                  * so we might loop indefinitely; also, that page could
780                  * not be swapped out to other storage meanwhile.  So:
781                  * delete from cache even if there's another reference,
782                  * after ensuring that the data has been saved to disk -
783                  * since if the reference remains (rarer), it will be
784                  * read from disk into another page.  Splitting into two
785                  * pages would be incorrect if swap supported "shared
786                  * private" pages, but they are handled by tmpfs files.
787                  *
788                  * Note shmem_unuse already deleted a swappage from
789                  * the swap cache, unless the move to filepage failed:
790                  * in which case it left swappage in cache, lowered its
791                  * swap count to pass quickly through the loops above,
792                  * and now we must reincrement count to try again later.
793                  */
794                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
795                         struct writeback_control wbc = {
796                                 .sync_mode = WB_SYNC_NONE,
797                         };
798
799                         swap_writepage(page, &wbc);
800                         lock_page(page);
801                         wait_on_page_writeback(page);
802                 }
803                 if (PageSwapCache(page)) {
804                         if (shmem)
805                                 swap_duplicate(entry);
806                         else
807                                 delete_from_swap_cache(page);
808                 }
809
810                 /*
811                  * So we could skip searching mms once swap count went
812                  * to 1, we did not mark any present ptes as dirty: must
813                  * mark page dirty so shrink_list will preserve it.
814                  */
815                 SetPageDirty(page);
816                 unlock_page(page);
817                 page_cache_release(page);
818
819                 /*
820                  * Make sure that we aren't completely killing
821                  * interactive performance.
822                  */
823                 cond_resched();
824         }
825
826         mmput(start_mm);
827         if (reset_overflow) {
828                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
829                 swap_overflow = 0;
830         }
831         return retval;
832 }
833
834 /*
835  * After a successful try_to_unuse, if no swap is now in use, we know
836  * we can empty the mmlist.  swap_lock must be held on entry and exit.
837  * Note that mmlist_lock nests inside swap_lock, and an mm must be
838  * added to the mmlist just after page_duplicate - before would be racy.
839  */
840 static void drain_mmlist(void)
841 {
842         struct list_head *p, *next;
843         unsigned int i;
844
845         for (i = 0; i < nr_swapfiles; i++)
846                 if (swap_info[i].inuse_pages)
847                         return;
848         spin_lock(&mmlist_lock);
849         list_for_each_safe(p, next, &init_mm.mmlist)
850                 list_del_init(p);
851         spin_unlock(&mmlist_lock);
852 }
853
854 /*
855  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
856  * corresponds to page offset `offset'.
857  */
858 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
859 {
860         struct swap_extent *se = sis->curr_swap_extent;
861         struct swap_extent *start_se = se;
862
863         for ( ; ; ) {
864                 struct list_head *lh;
865
866                 if (se->start_page <= offset &&
867                                 offset < (se->start_page + se->nr_pages)) {
868                         return se->start_block + (offset - se->start_page);
869                 }
870                 lh = se->list.next;
871                 if (lh == &sis->extent_list)
872                         lh = lh->next;
873                 se = list_entry(lh, struct swap_extent, list);
874                 sis->curr_swap_extent = se;
875                 BUG_ON(se == start_se);         /* It *must* be present */
876         }
877 }
878
879 /*
880  * Free all of a swapdev's extent information
881  */
882 static void destroy_swap_extents(struct swap_info_struct *sis)
883 {
884         while (!list_empty(&sis->extent_list)) {
885                 struct swap_extent *se;
886
887                 se = list_entry(sis->extent_list.next,
888                                 struct swap_extent, list);
889                 list_del(&se->list);
890                 kfree(se);
891         }
892 }
893
894 /*
895  * Add a block range (and the corresponding page range) into this swapdev's
896  * extent list.  The extent list is kept sorted in page order.
897  *
898  * This function rather assumes that it is called in ascending page order.
899  */
900 static int
901 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
902                 unsigned long nr_pages, sector_t start_block)
903 {
904         struct swap_extent *se;
905         struct swap_extent *new_se;
906         struct list_head *lh;
907
908         lh = sis->extent_list.prev;     /* The highest page extent */
909         if (lh != &sis->extent_list) {
910                 se = list_entry(lh, struct swap_extent, list);
911                 BUG_ON(se->start_page + se->nr_pages != start_page);
912                 if (se->start_block + se->nr_pages == start_block) {
913                         /* Merge it */
914                         se->nr_pages += nr_pages;
915                         return 0;
916                 }
917         }
918
919         /*
920          * No merge.  Insert a new extent, preserving ordering.
921          */
922         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
923         if (new_se == NULL)
924                 return -ENOMEM;
925         new_se->start_page = start_page;
926         new_se->nr_pages = nr_pages;
927         new_se->start_block = start_block;
928
929         list_add_tail(&new_se->list, &sis->extent_list);
930         return 1;
931 }
932
933 /*
934  * A `swap extent' is a simple thing which maps a contiguous range of pages
935  * onto a contiguous range of disk blocks.  An ordered list of swap extents
936  * is built at swapon time and is then used at swap_writepage/swap_readpage
937  * time for locating where on disk a page belongs.
938  *
939  * If the swapfile is an S_ISBLK block device, a single extent is installed.
940  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
941  * swap files identically.
942  *
943  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
944  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
945  * swapfiles are handled *identically* after swapon time.
946  *
947  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
948  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
949  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
950  * requirements, they are simply tossed out - we will never use those blocks
951  * for swapping.
952  *
953  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
954  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
955  * which will scribble on the fs.
956  *
957  * The amount of disk space which a single swap extent represents varies.
958  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
959  * extents in the list.  To avoid much list walking, we cache the previous
960  * search location in `curr_swap_extent', and start new searches from there.
961  * This is extremely effective.  The average number of iterations in
962  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
963  */
964 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
965 {
966         struct inode *inode;
967         unsigned blocks_per_page;
968         unsigned long page_no;
969         unsigned blkbits;
970         sector_t probe_block;
971         sector_t last_block;
972         sector_t lowest_block = -1;
973         sector_t highest_block = 0;
974         int nr_extents = 0;
975         int ret;
976
977         inode = sis->swap_file->f_mapping->host;
978         if (S_ISBLK(inode->i_mode)) {
979                 ret = add_swap_extent(sis, 0, sis->max, 0);
980                 *span = sis->pages;
981                 goto done;
982         }
983
984         blkbits = inode->i_blkbits;
985         blocks_per_page = PAGE_SIZE >> blkbits;
986
987         /*
988          * Map all the blocks into the extent list.  This code doesn't try
989          * to be very smart.
990          */
991         probe_block = 0;
992         page_no = 0;
993         last_block = i_size_read(inode) >> blkbits;
994         while ((probe_block + blocks_per_page) <= last_block &&
995                         page_no < sis->max) {
996                 unsigned block_in_page;
997                 sector_t first_block;
998
999                 first_block = bmap(inode, probe_block);
1000                 if (first_block == 0)
1001                         goto bad_bmap;
1002
1003                 /*
1004                  * It must be PAGE_SIZE aligned on-disk
1005                  */
1006                 if (first_block & (blocks_per_page - 1)) {
1007                         probe_block++;
1008                         goto reprobe;
1009                 }
1010
1011                 for (block_in_page = 1; block_in_page < blocks_per_page;
1012                                         block_in_page++) {
1013                         sector_t block;
1014
1015                         block = bmap(inode, probe_block + block_in_page);
1016                         if (block == 0)
1017                                 goto bad_bmap;
1018                         if (block != first_block + block_in_page) {
1019                                 /* Discontiguity */
1020                                 probe_block++;
1021                                 goto reprobe;
1022                         }
1023                 }
1024
1025                 first_block >>= (PAGE_SHIFT - blkbits);
1026                 if (page_no) {  /* exclude the header page */
1027                         if (first_block < lowest_block)
1028                                 lowest_block = first_block;
1029                         if (first_block > highest_block)
1030                                 highest_block = first_block;
1031                 }
1032
1033                 /*
1034                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1035                  */
1036                 ret = add_swap_extent(sis, page_no, 1, first_block);
1037                 if (ret < 0)
1038                         goto out;
1039                 nr_extents += ret;
1040                 page_no++;
1041                 probe_block += blocks_per_page;
1042 reprobe:
1043                 continue;
1044         }
1045         ret = nr_extents;
1046         *span = 1 + highest_block - lowest_block;
1047         if (page_no == 0)
1048                 page_no = 1;    /* force Empty message */
1049         sis->max = page_no;
1050         sis->pages = page_no - 1;
1051         sis->highest_bit = page_no - 1;
1052 done:
1053         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1054                                         struct swap_extent, list);
1055         goto out;
1056 bad_bmap:
1057         printk(KERN_ERR "swapon: swapfile has holes\n");
1058         ret = -EINVAL;
1059 out:
1060         return ret;
1061 }
1062
1063 #if 0   /* We don't need this yet */
1064 #include <linux/backing-dev.h>
1065 int page_queue_congested(struct page *page)
1066 {
1067         struct backing_dev_info *bdi;
1068
1069         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1070
1071         if (PageSwapCache(page)) {
1072                 swp_entry_t entry = { .val = page_private(page) };
1073                 struct swap_info_struct *sis;
1074
1075                 sis = get_swap_info_struct(swp_type(entry));
1076                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1077         } else
1078                 bdi = page->mapping->backing_dev_info;
1079         return bdi_write_congested(bdi);
1080 }
1081 #endif
1082
1083 asmlinkage long sys_swapoff(const char __user * specialfile)
1084 {
1085         struct swap_info_struct * p = NULL;
1086         unsigned short *swap_map;
1087         struct file *swap_file, *victim;
1088         struct address_space *mapping;
1089         struct inode *inode;
1090         char * pathname;
1091         int i, type, prev;
1092         int err;
1093         
1094         if (!capable(CAP_SYS_ADMIN))
1095                 return -EPERM;
1096
1097         pathname = getname(specialfile);
1098         err = PTR_ERR(pathname);
1099         if (IS_ERR(pathname))
1100                 goto out;
1101
1102         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1103         putname(pathname);
1104         err = PTR_ERR(victim);
1105         if (IS_ERR(victim))
1106                 goto out;
1107
1108         mapping = victim->f_mapping;
1109         prev = -1;
1110         spin_lock(&swap_lock);
1111         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1112                 p = swap_info + type;
1113                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1114                         if (p->swap_file->f_mapping == mapping)
1115                                 break;
1116                 }
1117                 prev = type;
1118         }
1119         if (type < 0) {
1120                 err = -EINVAL;
1121                 spin_unlock(&swap_lock);
1122                 goto out_dput;
1123         }
1124         if (!security_vm_enough_memory(p->pages))
1125                 vm_unacct_memory(p->pages);
1126         else {
1127                 err = -ENOMEM;
1128                 spin_unlock(&swap_lock);
1129                 goto out_dput;
1130         }
1131         if (prev < 0) {
1132                 swap_list.head = p->next;
1133         } else {
1134                 swap_info[prev].next = p->next;
1135         }
1136         if (type == swap_list.next) {
1137                 /* just pick something that's safe... */
1138                 swap_list.next = swap_list.head;
1139         }
1140         nr_swap_pages -= p->pages;
1141         total_swap_pages -= p->pages;
1142         p->flags &= ~SWP_WRITEOK;
1143         spin_unlock(&swap_lock);
1144
1145         current->flags |= PF_SWAPOFF;
1146         err = try_to_unuse(type);
1147         current->flags &= ~PF_SWAPOFF;
1148
1149         if (err) {
1150                 /* re-insert swap space back into swap_list */
1151                 spin_lock(&swap_lock);
1152                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1153                         if (p->prio >= swap_info[i].prio)
1154                                 break;
1155                 p->next = i;
1156                 if (prev < 0)
1157                         swap_list.head = swap_list.next = p - swap_info;
1158                 else
1159                         swap_info[prev].next = p - swap_info;
1160                 nr_swap_pages += p->pages;
1161                 total_swap_pages += p->pages;
1162                 p->flags |= SWP_WRITEOK;
1163                 spin_unlock(&swap_lock);
1164                 goto out_dput;
1165         }
1166
1167         /* wait for any unplug function to finish */
1168         down_write(&swap_unplug_sem);
1169         up_write(&swap_unplug_sem);
1170
1171         destroy_swap_extents(p);
1172         mutex_lock(&swapon_mutex);
1173         spin_lock(&swap_lock);
1174         drain_mmlist();
1175
1176         /* wait for anyone still in scan_swap_map */
1177         p->highest_bit = 0;             /* cuts scans short */
1178         while (p->flags >= SWP_SCANNING) {
1179                 spin_unlock(&swap_lock);
1180                 schedule_timeout_uninterruptible(1);
1181                 spin_lock(&swap_lock);
1182         }
1183
1184         swap_file = p->swap_file;
1185         p->swap_file = NULL;
1186         p->max = 0;
1187         swap_map = p->swap_map;
1188         p->swap_map = NULL;
1189         p->flags = 0;
1190         spin_unlock(&swap_lock);
1191         mutex_unlock(&swapon_mutex);
1192         vfree(swap_map);
1193         inode = mapping->host;
1194         if (S_ISBLK(inode->i_mode)) {
1195                 struct block_device *bdev = I_BDEV(inode);
1196                 set_blocksize(bdev, p->old_block_size);
1197                 bd_release(bdev);
1198         } else {
1199                 mutex_lock(&inode->i_mutex);
1200                 inode->i_flags &= ~S_SWAPFILE;
1201                 mutex_unlock(&inode->i_mutex);
1202         }
1203         filp_close(swap_file, NULL);
1204         err = 0;
1205
1206 out_dput:
1207         filp_close(victim, NULL);
1208 out:
1209         return err;
1210 }
1211
1212 #ifdef CONFIG_PROC_FS
1213 /* iterator */
1214 static void *swap_start(struct seq_file *swap, loff_t *pos)
1215 {
1216         struct swap_info_struct *ptr = swap_info;
1217         int i;
1218         loff_t l = *pos;
1219
1220         mutex_lock(&swapon_mutex);
1221
1222         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1223                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1224                         continue;
1225                 if (!l--)
1226                         return ptr;
1227         }
1228
1229         return NULL;
1230 }
1231
1232 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1233 {
1234         struct swap_info_struct *ptr = v;
1235         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1236
1237         for (++ptr; ptr < endptr; ptr++) {
1238                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1239                         continue;
1240                 ++*pos;
1241                 return ptr;
1242         }
1243
1244         return NULL;
1245 }
1246
1247 static void swap_stop(struct seq_file *swap, void *v)
1248 {
1249         mutex_unlock(&swapon_mutex);
1250 }
1251
1252 static int swap_show(struct seq_file *swap, void *v)
1253 {
1254         struct swap_info_struct *ptr = v;
1255         struct file *file;
1256         int len;
1257
1258         if (v == swap_info)
1259                 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1260
1261         file = ptr->swap_file;
1262         len = seq_path(swap, file->f_vfsmnt, file->f_dentry, " \t\n\\");
1263         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1264                        len < 40 ? 40 - len : 1, " ",
1265                        S_ISBLK(file->f_dentry->d_inode->i_mode) ?
1266                                 "partition" : "file\t",
1267                        ptr->pages << (PAGE_SHIFT - 10),
1268                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1269                        ptr->prio);
1270         return 0;
1271 }
1272
1273 static struct seq_operations swaps_op = {
1274         .start =        swap_start,
1275         .next =         swap_next,
1276         .stop =         swap_stop,
1277         .show =         swap_show
1278 };
1279
1280 static int swaps_open(struct inode *inode, struct file *file)
1281 {
1282         return seq_open(file, &swaps_op);
1283 }
1284
1285 static struct file_operations proc_swaps_operations = {
1286         .open           = swaps_open,
1287         .read           = seq_read,
1288         .llseek         = seq_lseek,
1289         .release        = seq_release,
1290 };
1291
1292 static int __init procswaps_init(void)
1293 {
1294         struct proc_dir_entry *entry;
1295
1296         entry = create_proc_entry("swaps", 0, NULL);
1297         if (entry)
1298                 entry->proc_fops = &proc_swaps_operations;
1299         return 0;
1300 }
1301 __initcall(procswaps_init);
1302 #endif /* CONFIG_PROC_FS */
1303
1304 /*
1305  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1306  *
1307  * The swapon system call
1308  */
1309 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1310 {
1311         struct swap_info_struct * p;
1312         char *name = NULL;
1313         struct block_device *bdev = NULL;
1314         struct file *swap_file = NULL;
1315         struct address_space *mapping;
1316         unsigned int type;
1317         int i, prev;
1318         int error;
1319         static int least_priority;
1320         union swap_header *swap_header = NULL;
1321         int swap_header_version;
1322         unsigned int nr_good_pages = 0;
1323         int nr_extents = 0;
1324         sector_t span;
1325         unsigned long maxpages = 1;
1326         int swapfilesize;
1327         unsigned short *swap_map;
1328         struct page *page = NULL;
1329         struct inode *inode = NULL;
1330         int did_down = 0;
1331
1332         if (!capable(CAP_SYS_ADMIN))
1333                 return -EPERM;
1334         spin_lock(&swap_lock);
1335         p = swap_info;
1336         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1337                 if (!(p->flags & SWP_USED))
1338                         break;
1339         error = -EPERM;
1340         /*
1341          * Test if adding another swap device is possible. There are
1342          * two limiting factors: 1) the number of bits for the swap
1343          * type swp_entry_t definition and 2) the number of bits for
1344          * the swap type in the swap ptes as defined by the different
1345          * architectures. To honor both limitations a swap entry
1346          * with swap offset 0 and swap type ~0UL is created, encoded
1347          * to a swap pte, decoded to a swp_entry_t again and finally
1348          * the swap type part is extracted. This will mask all bits
1349          * from the initial ~0UL that can't be encoded in either the
1350          * swp_entry_t or the architecture definition of a swap pte.
1351          */
1352         if (type > swp_type(pte_to_swp_entry(swp_entry_to_pte(swp_entry(~0UL,0))))) {
1353                 spin_unlock(&swap_lock);
1354                 goto out;
1355         }
1356         if (type >= nr_swapfiles)
1357                 nr_swapfiles = type+1;
1358         INIT_LIST_HEAD(&p->extent_list);
1359         p->flags = SWP_USED;
1360         p->swap_file = NULL;
1361         p->old_block_size = 0;
1362         p->swap_map = NULL;
1363         p->lowest_bit = 0;
1364         p->highest_bit = 0;
1365         p->cluster_nr = 0;
1366         p->inuse_pages = 0;
1367         p->next = -1;
1368         if (swap_flags & SWAP_FLAG_PREFER) {
1369                 p->prio =
1370                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1371         } else {
1372                 p->prio = --least_priority;
1373         }
1374         spin_unlock(&swap_lock);
1375         name = getname(specialfile);
1376         error = PTR_ERR(name);
1377         if (IS_ERR(name)) {
1378                 name = NULL;
1379                 goto bad_swap_2;
1380         }
1381         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1382         error = PTR_ERR(swap_file);
1383         if (IS_ERR(swap_file)) {
1384                 swap_file = NULL;
1385                 goto bad_swap_2;
1386         }
1387
1388         p->swap_file = swap_file;
1389         mapping = swap_file->f_mapping;
1390         inode = mapping->host;
1391
1392         error = -EBUSY;
1393         for (i = 0; i < nr_swapfiles; i++) {
1394                 struct swap_info_struct *q = &swap_info[i];
1395
1396                 if (i == type || !q->swap_file)
1397                         continue;
1398                 if (mapping == q->swap_file->f_mapping)
1399                         goto bad_swap;
1400         }
1401
1402         error = -EINVAL;
1403         if (S_ISBLK(inode->i_mode)) {
1404                 bdev = I_BDEV(inode);
1405                 error = bd_claim(bdev, sys_swapon);
1406                 if (error < 0) {
1407                         bdev = NULL;
1408                         error = -EINVAL;
1409                         goto bad_swap;
1410                 }
1411                 p->old_block_size = block_size(bdev);
1412                 error = set_blocksize(bdev, PAGE_SIZE);
1413                 if (error < 0)
1414                         goto bad_swap;
1415                 p->bdev = bdev;
1416         } else if (S_ISREG(inode->i_mode)) {
1417                 p->bdev = inode->i_sb->s_bdev;
1418                 mutex_lock(&inode->i_mutex);
1419                 did_down = 1;
1420                 if (IS_SWAPFILE(inode)) {
1421                         error = -EBUSY;
1422                         goto bad_swap;
1423                 }
1424         } else {
1425                 goto bad_swap;
1426         }
1427
1428         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1429
1430         /*
1431          * Read the swap header.
1432          */
1433         if (!mapping->a_ops->readpage) {
1434                 error = -EINVAL;
1435                 goto bad_swap;
1436         }
1437         page = read_cache_page(mapping, 0,
1438                         (filler_t *)mapping->a_ops->readpage, swap_file);
1439         if (IS_ERR(page)) {
1440                 error = PTR_ERR(page);
1441                 goto bad_swap;
1442         }
1443         wait_on_page_locked(page);
1444         if (!PageUptodate(page))
1445                 goto bad_swap;
1446         kmap(page);
1447         swap_header = page_address(page);
1448
1449         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1450                 swap_header_version = 1;
1451         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1452                 swap_header_version = 2;
1453         else {
1454                 printk(KERN_ERR "Unable to find swap-space signature\n");
1455                 error = -EINVAL;
1456                 goto bad_swap;
1457         }
1458         
1459         switch (swap_header_version) {
1460         case 1:
1461                 printk(KERN_ERR "version 0 swap is no longer supported. "
1462                         "Use mkswap -v1 %s\n", name);
1463                 error = -EINVAL;
1464                 goto bad_swap;
1465         case 2:
1466                 /* Check the swap header's sub-version and the size of
1467                    the swap file and bad block lists */
1468                 if (swap_header->info.version != 1) {
1469                         printk(KERN_WARNING
1470                                "Unable to handle swap header version %d\n",
1471                                swap_header->info.version);
1472                         error = -EINVAL;
1473                         goto bad_swap;
1474                 }
1475
1476                 p->lowest_bit  = 1;
1477                 p->cluster_next = 1;
1478
1479                 /*
1480                  * Find out how many pages are allowed for a single swap
1481                  * device. There are two limiting factors: 1) the number of
1482                  * bits for the swap offset in the swp_entry_t type and
1483                  * 2) the number of bits in the a swap pte as defined by
1484                  * the different architectures. In order to find the
1485                  * largest possible bit mask a swap entry with swap type 0
1486                  * and swap offset ~0UL is created, encoded to a swap pte,
1487                  * decoded to a swp_entry_t again and finally the swap
1488                  * offset is extracted. This will mask all the bits from
1489                  * the initial ~0UL mask that can't be encoded in either
1490                  * the swp_entry_t or the architecture definition of a
1491                  * swap pte.
1492                  */
1493                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1494                 if (maxpages > swap_header->info.last_page)
1495                         maxpages = swap_header->info.last_page;
1496                 p->highest_bit = maxpages - 1;
1497
1498                 error = -EINVAL;
1499                 if (!maxpages)
1500                         goto bad_swap;
1501                 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1502                         goto bad_swap;
1503                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1504                         goto bad_swap;
1505
1506                 /* OK, set up the swap map and apply the bad block list */
1507                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1508                         error = -ENOMEM;
1509                         goto bad_swap;
1510                 }
1511
1512                 error = 0;
1513                 memset(p->swap_map, 0, maxpages * sizeof(short));
1514                 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1515                         int page_nr = swap_header->info.badpages[i];
1516                         if (page_nr <= 0 || page_nr >= swap_header->info.last_page)
1517                                 error = -EINVAL;
1518                         else
1519                                 p->swap_map[page_nr] = SWAP_MAP_BAD;
1520                 }
1521                 nr_good_pages = swap_header->info.last_page -
1522                                 swap_header->info.nr_badpages -
1523                                 1 /* header page */;
1524                 if (error)
1525                         goto bad_swap;
1526         }
1527
1528         if (swapfilesize && maxpages > swapfilesize) {
1529                 printk(KERN_WARNING
1530                        "Swap area shorter than signature indicates\n");
1531                 error = -EINVAL;
1532                 goto bad_swap;
1533         }
1534         if (nr_good_pages) {
1535                 p->swap_map[0] = SWAP_MAP_BAD;
1536                 p->max = maxpages;
1537                 p->pages = nr_good_pages;
1538                 nr_extents = setup_swap_extents(p, &span);
1539                 if (nr_extents < 0) {
1540                         error = nr_extents;
1541                         goto bad_swap;
1542                 }
1543                 nr_good_pages = p->pages;
1544         }
1545         if (!nr_good_pages) {
1546                 printk(KERN_WARNING "Empty swap-file\n");
1547                 error = -EINVAL;
1548                 goto bad_swap;
1549         }
1550
1551         mutex_lock(&swapon_mutex);
1552         spin_lock(&swap_lock);
1553         p->flags = SWP_ACTIVE;
1554         nr_swap_pages += nr_good_pages;
1555         total_swap_pages += nr_good_pages;
1556
1557         printk(KERN_INFO "Adding %uk swap on %s.  "
1558                         "Priority:%d extents:%d across:%lluk\n",
1559                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1560                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1561
1562         /* insert swap space into swap_list: */
1563         prev = -1;
1564         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1565                 if (p->prio >= swap_info[i].prio) {
1566                         break;
1567                 }
1568                 prev = i;
1569         }
1570         p->next = i;
1571         if (prev < 0) {
1572                 swap_list.head = swap_list.next = p - swap_info;
1573         } else {
1574                 swap_info[prev].next = p - swap_info;
1575         }
1576         spin_unlock(&swap_lock);
1577         mutex_unlock(&swapon_mutex);
1578         error = 0;
1579         goto out;
1580 bad_swap:
1581         if (bdev) {
1582                 set_blocksize(bdev, p->old_block_size);
1583                 bd_release(bdev);
1584         }
1585         destroy_swap_extents(p);
1586 bad_swap_2:
1587         spin_lock(&swap_lock);
1588         swap_map = p->swap_map;
1589         p->swap_file = NULL;
1590         p->swap_map = NULL;
1591         p->flags = 0;
1592         if (!(swap_flags & SWAP_FLAG_PREFER))
1593                 ++least_priority;
1594         spin_unlock(&swap_lock);
1595         vfree(swap_map);
1596         if (swap_file)
1597                 filp_close(swap_file, NULL);
1598 out:
1599         if (page && !IS_ERR(page)) {
1600                 kunmap(page);
1601                 page_cache_release(page);
1602         }
1603         if (name)
1604                 putname(name);
1605         if (did_down) {
1606                 if (!error)
1607                         inode->i_flags |= S_SWAPFILE;
1608                 mutex_unlock(&inode->i_mutex);
1609         }
1610         return error;
1611 }
1612
1613 void si_swapinfo(struct sysinfo *val)
1614 {
1615         unsigned int i;
1616         unsigned long nr_to_be_unused = 0;
1617
1618         spin_lock(&swap_lock);
1619         for (i = 0; i < nr_swapfiles; i++) {
1620                 if (!(swap_info[i].flags & SWP_USED) ||
1621                      (swap_info[i].flags & SWP_WRITEOK))
1622                         continue;
1623                 nr_to_be_unused += swap_info[i].inuse_pages;
1624         }
1625         val->freeswap = nr_swap_pages + nr_to_be_unused;
1626         val->totalswap = total_swap_pages + nr_to_be_unused;
1627         spin_unlock(&swap_lock);
1628 }
1629
1630 /*
1631  * Verify that a swap entry is valid and increment its swap map count.
1632  *
1633  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1634  * "permanent", but will be reclaimed by the next swapoff.
1635  */
1636 int swap_duplicate(swp_entry_t entry)
1637 {
1638         struct swap_info_struct * p;
1639         unsigned long offset, type;
1640         int result = 0;
1641
1642         type = swp_type(entry);
1643         if (type >= nr_swapfiles)
1644                 goto bad_file;
1645         p = type + swap_info;
1646         offset = swp_offset(entry);
1647
1648         spin_lock(&swap_lock);
1649         if (offset < p->max && p->swap_map[offset]) {
1650                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1651                         p->swap_map[offset]++;
1652                         result = 1;
1653                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1654                         if (swap_overflow++ < 5)
1655                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1656                         p->swap_map[offset] = SWAP_MAP_MAX;
1657                         result = 1;
1658                 }
1659         }
1660         spin_unlock(&swap_lock);
1661 out:
1662         return result;
1663
1664 bad_file:
1665         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1666         goto out;
1667 }
1668
1669 struct swap_info_struct *
1670 get_swap_info_struct(unsigned type)
1671 {
1672         return &swap_info[type];
1673 }
1674
1675 /*
1676  * swap_lock prevents swap_map being freed. Don't grab an extra
1677  * reference on the swaphandle, it doesn't matter if it becomes unused.
1678  */
1679 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1680 {
1681         int ret = 0, i = 1 << page_cluster;
1682         unsigned long toff;
1683         struct swap_info_struct *swapdev = swp_type(entry) + swap_info;
1684
1685         if (!page_cluster)      /* no readahead */
1686                 return 0;
1687         toff = (swp_offset(entry) >> page_cluster) << page_cluster;
1688         if (!toff)              /* first page is swap header */
1689                 toff++, i--;
1690         *offset = toff;
1691
1692         spin_lock(&swap_lock);
1693         do {
1694                 /* Don't read-ahead past the end of the swap area */
1695                 if (toff >= swapdev->max)
1696                         break;
1697                 /* Don't read in free or bad pages */
1698                 if (!swapdev->swap_map[toff])
1699                         break;
1700                 if (swapdev->swap_map[toff] == SWAP_MAP_BAD)
1701                         break;
1702                 toff++;
1703                 ret++;
1704         } while (--i);
1705         spin_unlock(&swap_lock);
1706         return ret;
1707 }