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