[PATCH] swap: scan_swap_map restyled
[linux-2.6.git] / mm / swapfile.c
1 /*
2  *  linux/mm/swapfile.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *  Swap reorganised 29.12.95, Stephen Tweedie
6  */
7
8 #include <linux/config.h>
9 #include <linux/mm.h>
10 #include <linux/hugetlb.h>
11 #include <linux/mman.h>
12 #include <linux/slab.h>
13 #include <linux/kernel_stat.h>
14 #include <linux/swap.h>
15 #include <linux/vmalloc.h>
16 #include <linux/pagemap.h>
17 #include <linux/namei.h>
18 #include <linux/shm.h>
19 #include <linux/blkdev.h>
20 #include <linux/writeback.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/init.h>
24 #include <linux/module.h>
25 #include <linux/rmap.h>
26 #include <linux/security.h>
27 #include <linux/backing-dev.h>
28 #include <linux/syscalls.h>
29
30 #include <asm/pgtable.h>
31 #include <asm/tlbflush.h>
32 #include <linux/swapops.h>
33
34 DEFINE_SPINLOCK(swaplock);
35 unsigned int nr_swapfiles;
36 long total_swap_pages;
37 static int swap_overflow;
38
39 EXPORT_SYMBOL(total_swap_pages);
40
41 static const char Bad_file[] = "Bad swap file entry ";
42 static const char Unused_file[] = "Unused swap file entry ";
43 static const char Bad_offset[] = "Bad swap offset entry ";
44 static const char Unused_offset[] = "Unused swap offset entry ";
45
46 struct swap_list_t swap_list = {-1, -1};
47
48 struct swap_info_struct swap_info[MAX_SWAPFILES];
49
50 static DECLARE_MUTEX(swapon_sem);
51
52 /*
53  * We need this because the bdev->unplug_fn can sleep and we cannot
54  * hold swap_list_lock while calling the unplug_fn. And swap_list_lock
55  * cannot be turned into a semaphore.
56  */
57 static DECLARE_RWSEM(swap_unplug_sem);
58
59 #define SWAPFILE_CLUSTER 256
60
61 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
62 {
63         swp_entry_t entry;
64
65         down_read(&swap_unplug_sem);
66         entry.val = page->private;
67         if (PageSwapCache(page)) {
68                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
69                 struct backing_dev_info *bdi;
70
71                 /*
72                  * If the page is removed from swapcache from under us (with a
73                  * racy try_to_unuse/swapoff) we need an additional reference
74                  * count to avoid reading garbage from page->private above. If
75                  * the WARN_ON triggers during a swapoff it maybe the race
76                  * condition and it's harmless. However if it triggers without
77                  * swapoff it signals a problem.
78                  */
79                 WARN_ON(page_count(page) <= 1);
80
81                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
82                 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, last_in_cluster;
90
91         /* 
92          * We try to cluster swap pages by allocating them sequentially
93          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
94          * way, however, we resort to first-free allocation, starting
95          * a new cluster.  This prevents us from scattering swap pages
96          * all over the entire swap partition, so that we reduce
97          * overall disk seek times between swap pages.  -- sct
98          * But we do now try to find an empty cluster.  -Andrea
99          */
100
101         if (unlikely(!si->cluster_nr)) {
102                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
103                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER)
104                         goto lowest;
105
106                 offset = si->lowest_bit;
107                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
108
109                 /* Locate the first empty (unaligned) cluster */
110                 for (; last_in_cluster <= si->highest_bit; offset++) {
111                         if (si->swap_map[offset])
112                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
113                         else if (offset == last_in_cluster) {
114                                 si->cluster_next = offset-SWAPFILE_CLUSTER-1;
115                                 goto cluster;
116                         }
117                 }
118                 goto lowest;
119         }
120
121         si->cluster_nr--;
122 cluster:
123         offset = si->cluster_next;
124         if (offset > si->highest_bit)
125 lowest:         offset = si->lowest_bit;
126         if (!si->highest_bit)
127                 goto no_page;
128         if (!si->swap_map[offset]) {
129 got_page:       if (offset == si->lowest_bit)
130                         si->lowest_bit++;
131                 if (offset == si->highest_bit)
132                         si->highest_bit--;
133                 si->inuse_pages++;
134                 if (si->inuse_pages == si->pages) {
135                         si->lowest_bit = si->max;
136                         si->highest_bit = 0;
137                 }
138                 si->swap_map[offset] = 1;
139                 si->cluster_next = offset + 1;
140                 return offset;
141         }
142
143         while (++offset <= si->highest_bit) {
144                 if (!si->swap_map[offset])
145                         goto got_page;
146         }
147         goto lowest;
148
149 no_page:
150         return 0;
151 }
152
153 swp_entry_t get_swap_page(void)
154 {
155         struct swap_info_struct *si;
156         pgoff_t offset;
157         int type, next;
158         int wrapped = 0;
159
160         swap_list_lock();
161         if (nr_swap_pages <= 0)
162                 goto noswap;
163         nr_swap_pages--;
164
165         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
166                 si = swap_info + type;
167                 next = si->next;
168                 if (next < 0 ||
169                     (!wrapped && si->prio != swap_info[next].prio)) {
170                         next = swap_list.head;
171                         wrapped++;
172                 }
173
174                 if (!si->highest_bit)
175                         continue;
176                 if (!(si->flags & SWP_WRITEOK))
177                         continue;
178
179                 swap_list.next = next;
180                 swap_device_lock(si);
181                 swap_list_unlock();
182                 offset = scan_swap_map(si);
183                 swap_device_unlock(si);
184                 if (offset)
185                         return swp_entry(type, offset);
186                 swap_list_lock();
187                 next = swap_list.next;
188         }
189
190         nr_swap_pages++;
191 noswap:
192         swap_list_unlock();
193         return (swp_entry_t) {0};
194 }
195
196 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
197 {
198         struct swap_info_struct * p;
199         unsigned long offset, type;
200
201         if (!entry.val)
202                 goto out;
203         type = swp_type(entry);
204         if (type >= nr_swapfiles)
205                 goto bad_nofile;
206         p = & swap_info[type];
207         if (!(p->flags & SWP_USED))
208                 goto bad_device;
209         offset = swp_offset(entry);
210         if (offset >= p->max)
211                 goto bad_offset;
212         if (!p->swap_map[offset])
213                 goto bad_free;
214         swap_list_lock();
215         swap_device_lock(p);
216         return p;
217
218 bad_free:
219         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
220         goto out;
221 bad_offset:
222         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
223         goto out;
224 bad_device:
225         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
226         goto out;
227 bad_nofile:
228         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
229 out:
230         return NULL;
231 }       
232
233 static void swap_info_put(struct swap_info_struct * p)
234 {
235         swap_device_unlock(p);
236         swap_list_unlock();
237 }
238
239 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
240 {
241         int count = p->swap_map[offset];
242
243         if (count < SWAP_MAP_MAX) {
244                 count--;
245                 p->swap_map[offset] = count;
246                 if (!count) {
247                         if (offset < p->lowest_bit)
248                                 p->lowest_bit = offset;
249                         if (offset > p->highest_bit)
250                                 p->highest_bit = offset;
251                         if (p->prio > swap_info[swap_list.next].prio)
252                                 swap_list.next = p - swap_info;
253                         nr_swap_pages++;
254                         p->inuse_pages--;
255                 }
256         }
257         return count;
258 }
259
260 /*
261  * Caller has made sure that the swapdevice corresponding to entry
262  * is still around or has not been recycled.
263  */
264 void swap_free(swp_entry_t entry)
265 {
266         struct swap_info_struct * p;
267
268         p = swap_info_get(entry);
269         if (p) {
270                 swap_entry_free(p, swp_offset(entry));
271                 swap_info_put(p);
272         }
273 }
274
275 /*
276  * How many references to page are currently swapped out?
277  */
278 static inline int page_swapcount(struct page *page)
279 {
280         int count = 0;
281         struct swap_info_struct *p;
282         swp_entry_t entry;
283
284         entry.val = page->private;
285         p = swap_info_get(entry);
286         if (p) {
287                 /* Subtract the 1 for the swap cache itself */
288                 count = p->swap_map[swp_offset(entry)] - 1;
289                 swap_info_put(p);
290         }
291         return count;
292 }
293
294 /*
295  * We can use this swap cache entry directly
296  * if there are no other references to it.
297  */
298 int can_share_swap_page(struct page *page)
299 {
300         int count;
301
302         BUG_ON(!PageLocked(page));
303         count = page_mapcount(page);
304         if (count <= 1 && PageSwapCache(page))
305                 count += page_swapcount(page);
306         return count == 1;
307 }
308
309 /*
310  * Work out if there are any other processes sharing this
311  * swap cache page. Free it if you can. Return success.
312  */
313 int remove_exclusive_swap_page(struct page *page)
314 {
315         int retval;
316         struct swap_info_struct * p;
317         swp_entry_t entry;
318
319         BUG_ON(PagePrivate(page));
320         BUG_ON(!PageLocked(page));
321
322         if (!PageSwapCache(page))
323                 return 0;
324         if (PageWriteback(page))
325                 return 0;
326         if (page_count(page) != 2) /* 2: us + cache */
327                 return 0;
328
329         entry.val = page->private;
330         p = swap_info_get(entry);
331         if (!p)
332                 return 0;
333
334         /* Is the only swap cache user the cache itself? */
335         retval = 0;
336         if (p->swap_map[swp_offset(entry)] == 1) {
337                 /* Recheck the page count with the swapcache lock held.. */
338                 write_lock_irq(&swapper_space.tree_lock);
339                 if ((page_count(page) == 2) && !PageWriteback(page)) {
340                         __delete_from_swap_cache(page);
341                         SetPageDirty(page);
342                         retval = 1;
343                 }
344                 write_unlock_irq(&swapper_space.tree_lock);
345         }
346         swap_info_put(p);
347
348         if (retval) {
349                 swap_free(entry);
350                 page_cache_release(page);
351         }
352
353         return retval;
354 }
355
356 /*
357  * Free the swap entry like above, but also try to
358  * free the page cache entry if it is the last user.
359  */
360 void free_swap_and_cache(swp_entry_t entry)
361 {
362         struct swap_info_struct * p;
363         struct page *page = NULL;
364
365         p = swap_info_get(entry);
366         if (p) {
367                 if (swap_entry_free(p, swp_offset(entry)) == 1)
368                         page = find_trylock_page(&swapper_space, entry.val);
369                 swap_info_put(p);
370         }
371         if (page) {
372                 int one_user;
373
374                 BUG_ON(PagePrivate(page));
375                 page_cache_get(page);
376                 one_user = (page_count(page) == 2);
377                 /* Only cache user (+us), or swap space full? Free it! */
378                 if (!PageWriteback(page) && (one_user || vm_swap_full())) {
379                         delete_from_swap_cache(page);
380                         SetPageDirty(page);
381                 }
382                 unlock_page(page);
383                 page_cache_release(page);
384         }
385 }
386
387 /*
388  * Always set the resulting pte to be nowrite (the same as COW pages
389  * after one process has exited).  We don't know just how many PTEs will
390  * share this swap entry, so be cautious and let do_wp_page work out
391  * what to do if a write is requested later.
392  *
393  * vma->vm_mm->page_table_lock is held.
394  */
395 static void unuse_pte(struct vm_area_struct *vma, pte_t *pte,
396                 unsigned long addr, swp_entry_t entry, struct page *page)
397 {
398         inc_mm_counter(vma->vm_mm, rss);
399         get_page(page);
400         set_pte_at(vma->vm_mm, addr, pte,
401                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
402         page_add_anon_rmap(page, vma, addr);
403         swap_free(entry);
404         /*
405          * Move the page to the active list so it is not
406          * immediately swapped out again after swapon.
407          */
408         activate_page(page);
409 }
410
411 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
412                                 unsigned long addr, unsigned long end,
413                                 swp_entry_t entry, struct page *page)
414 {
415         pte_t *pte;
416         pte_t swp_pte = swp_entry_to_pte(entry);
417
418         pte = pte_offset_map(pmd, addr);
419         do {
420                 /*
421                  * swapoff spends a _lot_ of time in this loop!
422                  * Test inline before going to call unuse_pte.
423                  */
424                 if (unlikely(pte_same(*pte, swp_pte))) {
425                         unuse_pte(vma, pte, addr, entry, page);
426                         pte_unmap(pte);
427                         return 1;
428                 }
429         } while (pte++, addr += PAGE_SIZE, addr != end);
430         pte_unmap(pte - 1);
431         return 0;
432 }
433
434 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
435                                 unsigned long addr, unsigned long end,
436                                 swp_entry_t entry, struct page *page)
437 {
438         pmd_t *pmd;
439         unsigned long next;
440
441         pmd = pmd_offset(pud, addr);
442         do {
443                 next = pmd_addr_end(addr, end);
444                 if (pmd_none_or_clear_bad(pmd))
445                         continue;
446                 if (unuse_pte_range(vma, pmd, addr, next, entry, page))
447                         return 1;
448         } while (pmd++, addr = next, addr != end);
449         return 0;
450 }
451
452 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
453                                 unsigned long addr, unsigned long end,
454                                 swp_entry_t entry, struct page *page)
455 {
456         pud_t *pud;
457         unsigned long next;
458
459         pud = pud_offset(pgd, addr);
460         do {
461                 next = pud_addr_end(addr, end);
462                 if (pud_none_or_clear_bad(pud))
463                         continue;
464                 if (unuse_pmd_range(vma, pud, addr, next, entry, page))
465                         return 1;
466         } while (pud++, addr = next, addr != end);
467         return 0;
468 }
469
470 static int unuse_vma(struct vm_area_struct *vma,
471                                 swp_entry_t entry, struct page *page)
472 {
473         pgd_t *pgd;
474         unsigned long addr, end, next;
475
476         if (page->mapping) {
477                 addr = page_address_in_vma(page, vma);
478                 if (addr == -EFAULT)
479                         return 0;
480                 else
481                         end = addr + PAGE_SIZE;
482         } else {
483                 addr = vma->vm_start;
484                 end = vma->vm_end;
485         }
486
487         pgd = pgd_offset(vma->vm_mm, addr);
488         do {
489                 next = pgd_addr_end(addr, end);
490                 if (pgd_none_or_clear_bad(pgd))
491                         continue;
492                 if (unuse_pud_range(vma, pgd, addr, next, entry, page))
493                         return 1;
494         } while (pgd++, addr = next, addr != end);
495         return 0;
496 }
497
498 static int unuse_mm(struct mm_struct *mm,
499                                 swp_entry_t entry, struct page *page)
500 {
501         struct vm_area_struct *vma;
502
503         if (!down_read_trylock(&mm->mmap_sem)) {
504                 /*
505                  * Activate page so shrink_cache is unlikely to unmap its
506                  * ptes while lock is dropped, so swapoff can make progress.
507                  */
508                 activate_page(page);
509                 unlock_page(page);
510                 down_read(&mm->mmap_sem);
511                 lock_page(page);
512         }
513         spin_lock(&mm->page_table_lock);
514         for (vma = mm->mmap; vma; vma = vma->vm_next) {
515                 if (vma->anon_vma && unuse_vma(vma, entry, page))
516                         break;
517         }
518         spin_unlock(&mm->page_table_lock);
519         up_read(&mm->mmap_sem);
520         /*
521          * Currently unuse_mm cannot fail, but leave error handling
522          * at call sites for now, since we change it from time to time.
523          */
524         return 0;
525 }
526
527 /*
528  * Scan swap_map from current position to next entry still in use.
529  * Recycle to start on reaching the end, returning 0 when empty.
530  */
531 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
532                                         unsigned int prev)
533 {
534         unsigned int max = si->max;
535         unsigned int i = prev;
536         int count;
537
538         /*
539          * No need for swap_device_lock(si) here: we're just looking
540          * for whether an entry is in use, not modifying it; false
541          * hits are okay, and sys_swapoff() has already prevented new
542          * allocations from this area (while holding swap_list_lock()).
543          */
544         for (;;) {
545                 if (++i >= max) {
546                         if (!prev) {
547                                 i = 0;
548                                 break;
549                         }
550                         /*
551                          * No entries in use at top of swap_map,
552                          * loop back to start and recheck there.
553                          */
554                         max = prev + 1;
555                         prev = 0;
556                         i = 1;
557                 }
558                 count = si->swap_map[i];
559                 if (count && count != SWAP_MAP_BAD)
560                         break;
561         }
562         return i;
563 }
564
565 /*
566  * We completely avoid races by reading each swap page in advance,
567  * and then search for the process using it.  All the necessary
568  * page table adjustments can then be made atomically.
569  */
570 static int try_to_unuse(unsigned int type)
571 {
572         struct swap_info_struct * si = &swap_info[type];
573         struct mm_struct *start_mm;
574         unsigned short *swap_map;
575         unsigned short swcount;
576         struct page *page;
577         swp_entry_t entry;
578         unsigned int i = 0;
579         int retval = 0;
580         int reset_overflow = 0;
581         int shmem;
582
583         /*
584          * When searching mms for an entry, a good strategy is to
585          * start at the first mm we freed the previous entry from
586          * (though actually we don't notice whether we or coincidence
587          * freed the entry).  Initialize this start_mm with a hold.
588          *
589          * A simpler strategy would be to start at the last mm we
590          * freed the previous entry from; but that would take less
591          * advantage of mmlist ordering, which clusters forked mms
592          * together, child after parent.  If we race with dup_mmap(), we
593          * prefer to resolve parent before child, lest we miss entries
594          * duplicated after we scanned child: using last mm would invert
595          * that.  Though it's only a serious concern when an overflowed
596          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
597          */
598         start_mm = &init_mm;
599         atomic_inc(&init_mm.mm_users);
600
601         /*
602          * Keep on scanning until all entries have gone.  Usually,
603          * one pass through swap_map is enough, but not necessarily:
604          * there are races when an instance of an entry might be missed.
605          */
606         while ((i = find_next_to_unuse(si, i)) != 0) {
607                 if (signal_pending(current)) {
608                         retval = -EINTR;
609                         break;
610                 }
611
612                 /* 
613                  * Get a page for the entry, using the existing swap
614                  * cache page if there is one.  Otherwise, get a clean
615                  * page and read the swap into it. 
616                  */
617                 swap_map = &si->swap_map[i];
618                 entry = swp_entry(type, i);
619                 page = read_swap_cache_async(entry, NULL, 0);
620                 if (!page) {
621                         /*
622                          * Either swap_duplicate() failed because entry
623                          * has been freed independently, and will not be
624                          * reused since sys_swapoff() already disabled
625                          * allocation from here, or alloc_page() failed.
626                          */
627                         if (!*swap_map)
628                                 continue;
629                         retval = -ENOMEM;
630                         break;
631                 }
632
633                 /*
634                  * Don't hold on to start_mm if it looks like exiting.
635                  */
636                 if (atomic_read(&start_mm->mm_users) == 1) {
637                         mmput(start_mm);
638                         start_mm = &init_mm;
639                         atomic_inc(&init_mm.mm_users);
640                 }
641
642                 /*
643                  * Wait for and lock page.  When do_swap_page races with
644                  * try_to_unuse, do_swap_page can handle the fault much
645                  * faster than try_to_unuse can locate the entry.  This
646                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
647                  * defer to do_swap_page in such a case - in some tests,
648                  * do_swap_page and try_to_unuse repeatedly compete.
649                  */
650                 wait_on_page_locked(page);
651                 wait_on_page_writeback(page);
652                 lock_page(page);
653                 wait_on_page_writeback(page);
654
655                 /*
656                  * Remove all references to entry.
657                  * Whenever we reach init_mm, there's no address space
658                  * to search, but use it as a reminder to search shmem.
659                  */
660                 shmem = 0;
661                 swcount = *swap_map;
662                 if (swcount > 1) {
663                         if (start_mm == &init_mm)
664                                 shmem = shmem_unuse(entry, page);
665                         else
666                                 retval = unuse_mm(start_mm, entry, page);
667                 }
668                 if (*swap_map > 1) {
669                         int set_start_mm = (*swap_map >= swcount);
670                         struct list_head *p = &start_mm->mmlist;
671                         struct mm_struct *new_start_mm = start_mm;
672                         struct mm_struct *prev_mm = start_mm;
673                         struct mm_struct *mm;
674
675                         atomic_inc(&new_start_mm->mm_users);
676                         atomic_inc(&prev_mm->mm_users);
677                         spin_lock(&mmlist_lock);
678                         while (*swap_map > 1 && !retval &&
679                                         (p = p->next) != &start_mm->mmlist) {
680                                 mm = list_entry(p, struct mm_struct, mmlist);
681                                 if (atomic_inc_return(&mm->mm_users) == 1) {
682                                         atomic_dec(&mm->mm_users);
683                                         continue;
684                                 }
685                                 spin_unlock(&mmlist_lock);
686                                 mmput(prev_mm);
687                                 prev_mm = mm;
688
689                                 cond_resched();
690
691                                 swcount = *swap_map;
692                                 if (swcount <= 1)
693                                         ;
694                                 else if (mm == &init_mm) {
695                                         set_start_mm = 1;
696                                         shmem = shmem_unuse(entry, page);
697                                 } else
698                                         retval = unuse_mm(mm, entry, page);
699                                 if (set_start_mm && *swap_map < swcount) {
700                                         mmput(new_start_mm);
701                                         atomic_inc(&mm->mm_users);
702                                         new_start_mm = mm;
703                                         set_start_mm = 0;
704                                 }
705                                 spin_lock(&mmlist_lock);
706                         }
707                         spin_unlock(&mmlist_lock);
708                         mmput(prev_mm);
709                         mmput(start_mm);
710                         start_mm = new_start_mm;
711                 }
712                 if (retval) {
713                         unlock_page(page);
714                         page_cache_release(page);
715                         break;
716                 }
717
718                 /*
719                  * How could swap count reach 0x7fff when the maximum
720                  * pid is 0x7fff, and there's no way to repeat a swap
721                  * page within an mm (except in shmem, where it's the
722                  * shared object which takes the reference count)?
723                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
724                  *
725                  * If that's wrong, then we should worry more about
726                  * exit_mmap() and do_munmap() cases described above:
727                  * we might be resetting SWAP_MAP_MAX too early here.
728                  * We know "Undead"s can happen, they're okay, so don't
729                  * report them; but do report if we reset SWAP_MAP_MAX.
730                  */
731                 if (*swap_map == SWAP_MAP_MAX) {
732                         swap_device_lock(si);
733                         *swap_map = 1;
734                         swap_device_unlock(si);
735                         reset_overflow = 1;
736                 }
737
738                 /*
739                  * If a reference remains (rare), we would like to leave
740                  * the page in the swap cache; but try_to_unmap could
741                  * then re-duplicate the entry once we drop page lock,
742                  * so we might loop indefinitely; also, that page could
743                  * not be swapped out to other storage meanwhile.  So:
744                  * delete from cache even if there's another reference,
745                  * after ensuring that the data has been saved to disk -
746                  * since if the reference remains (rarer), it will be
747                  * read from disk into another page.  Splitting into two
748                  * pages would be incorrect if swap supported "shared
749                  * private" pages, but they are handled by tmpfs files.
750                  *
751                  * Note shmem_unuse already deleted a swappage from
752                  * the swap cache, unless the move to filepage failed:
753                  * in which case it left swappage in cache, lowered its
754                  * swap count to pass quickly through the loops above,
755                  * and now we must reincrement count to try again later.
756                  */
757                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
758                         struct writeback_control wbc = {
759                                 .sync_mode = WB_SYNC_NONE,
760                         };
761
762                         swap_writepage(page, &wbc);
763                         lock_page(page);
764                         wait_on_page_writeback(page);
765                 }
766                 if (PageSwapCache(page)) {
767                         if (shmem)
768                                 swap_duplicate(entry);
769                         else
770                                 delete_from_swap_cache(page);
771                 }
772
773                 /*
774                  * So we could skip searching mms once swap count went
775                  * to 1, we did not mark any present ptes as dirty: must
776                  * mark page dirty so shrink_list will preserve it.
777                  */
778                 SetPageDirty(page);
779                 unlock_page(page);
780                 page_cache_release(page);
781
782                 /*
783                  * Make sure that we aren't completely killing
784                  * interactive performance.
785                  */
786                 cond_resched();
787         }
788
789         mmput(start_mm);
790         if (reset_overflow) {
791                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
792                 swap_overflow = 0;
793         }
794         return retval;
795 }
796
797 /*
798  * After a successful try_to_unuse, if no swap is now in use, we know we
799  * can empty the mmlist.  swap_list_lock must be held on entry and exit.
800  * Note that mmlist_lock nests inside swap_list_lock, and an mm must be
801  * added to the mmlist just after page_duplicate - before would be racy.
802  */
803 static void drain_mmlist(void)
804 {
805         struct list_head *p, *next;
806         unsigned int i;
807
808         for (i = 0; i < nr_swapfiles; i++)
809                 if (swap_info[i].inuse_pages)
810                         return;
811         spin_lock(&mmlist_lock);
812         list_for_each_safe(p, next, &init_mm.mmlist)
813                 list_del_init(p);
814         spin_unlock(&mmlist_lock);
815 }
816
817 /*
818  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
819  * corresponds to page offset `offset'.
820  */
821 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
822 {
823         struct swap_extent *se = sis->curr_swap_extent;
824         struct swap_extent *start_se = se;
825
826         for ( ; ; ) {
827                 struct list_head *lh;
828
829                 if (se->start_page <= offset &&
830                                 offset < (se->start_page + se->nr_pages)) {
831                         return se->start_block + (offset - se->start_page);
832                 }
833                 lh = se->list.next;
834                 if (lh == &sis->extent_list)
835                         lh = lh->next;
836                 se = list_entry(lh, struct swap_extent, list);
837                 sis->curr_swap_extent = se;
838                 BUG_ON(se == start_se);         /* It *must* be present */
839         }
840 }
841
842 /*
843  * Free all of a swapdev's extent information
844  */
845 static void destroy_swap_extents(struct swap_info_struct *sis)
846 {
847         while (!list_empty(&sis->extent_list)) {
848                 struct swap_extent *se;
849
850                 se = list_entry(sis->extent_list.next,
851                                 struct swap_extent, list);
852                 list_del(&se->list);
853                 kfree(se);
854         }
855 }
856
857 /*
858  * Add a block range (and the corresponding page range) into this swapdev's
859  * extent list.  The extent list is kept sorted in page order.
860  *
861  * This function rather assumes that it is called in ascending page order.
862  */
863 static int
864 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
865                 unsigned long nr_pages, sector_t start_block)
866 {
867         struct swap_extent *se;
868         struct swap_extent *new_se;
869         struct list_head *lh;
870
871         lh = sis->extent_list.prev;     /* The highest page extent */
872         if (lh != &sis->extent_list) {
873                 se = list_entry(lh, struct swap_extent, list);
874                 BUG_ON(se->start_page + se->nr_pages != start_page);
875                 if (se->start_block + se->nr_pages == start_block) {
876                         /* Merge it */
877                         se->nr_pages += nr_pages;
878                         return 0;
879                 }
880         }
881
882         /*
883          * No merge.  Insert a new extent, preserving ordering.
884          */
885         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
886         if (new_se == NULL)
887                 return -ENOMEM;
888         new_se->start_page = start_page;
889         new_se->nr_pages = nr_pages;
890         new_se->start_block = start_block;
891
892         list_add_tail(&new_se->list, &sis->extent_list);
893         return 1;
894 }
895
896 /*
897  * A `swap extent' is a simple thing which maps a contiguous range of pages
898  * onto a contiguous range of disk blocks.  An ordered list of swap extents
899  * is built at swapon time and is then used at swap_writepage/swap_readpage
900  * time for locating where on disk a page belongs.
901  *
902  * If the swapfile is an S_ISBLK block device, a single extent is installed.
903  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
904  * swap files identically.
905  *
906  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
907  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
908  * swapfiles are handled *identically* after swapon time.
909  *
910  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
911  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
912  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
913  * requirements, they are simply tossed out - we will never use those blocks
914  * for swapping.
915  *
916  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
917  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
918  * which will scribble on the fs.
919  *
920  * The amount of disk space which a single swap extent represents varies.
921  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
922  * extents in the list.  To avoid much list walking, we cache the previous
923  * search location in `curr_swap_extent', and start new searches from there.
924  * This is extremely effective.  The average number of iterations in
925  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
926  */
927 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
928 {
929         struct inode *inode;
930         unsigned blocks_per_page;
931         unsigned long page_no;
932         unsigned blkbits;
933         sector_t probe_block;
934         sector_t last_block;
935         sector_t lowest_block = -1;
936         sector_t highest_block = 0;
937         int nr_extents = 0;
938         int ret;
939
940         inode = sis->swap_file->f_mapping->host;
941         if (S_ISBLK(inode->i_mode)) {
942                 ret = add_swap_extent(sis, 0, sis->max, 0);
943                 *span = sis->pages;
944                 goto done;
945         }
946
947         blkbits = inode->i_blkbits;
948         blocks_per_page = PAGE_SIZE >> blkbits;
949
950         /*
951          * Map all the blocks into the extent list.  This code doesn't try
952          * to be very smart.
953          */
954         probe_block = 0;
955         page_no = 0;
956         last_block = i_size_read(inode) >> blkbits;
957         while ((probe_block + blocks_per_page) <= last_block &&
958                         page_no < sis->max) {
959                 unsigned block_in_page;
960                 sector_t first_block;
961
962                 first_block = bmap(inode, probe_block);
963                 if (first_block == 0)
964                         goto bad_bmap;
965
966                 /*
967                  * It must be PAGE_SIZE aligned on-disk
968                  */
969                 if (first_block & (blocks_per_page - 1)) {
970                         probe_block++;
971                         goto reprobe;
972                 }
973
974                 for (block_in_page = 1; block_in_page < blocks_per_page;
975                                         block_in_page++) {
976                         sector_t block;
977
978                         block = bmap(inode, probe_block + block_in_page);
979                         if (block == 0)
980                                 goto bad_bmap;
981                         if (block != first_block + block_in_page) {
982                                 /* Discontiguity */
983                                 probe_block++;
984                                 goto reprobe;
985                         }
986                 }
987
988                 first_block >>= (PAGE_SHIFT - blkbits);
989                 if (page_no) {  /* exclude the header page */
990                         if (first_block < lowest_block)
991                                 lowest_block = first_block;
992                         if (first_block > highest_block)
993                                 highest_block = first_block;
994                 }
995
996                 /*
997                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
998                  */
999                 ret = add_swap_extent(sis, page_no, 1, first_block);
1000                 if (ret < 0)
1001                         goto out;
1002                 nr_extents += ret;
1003                 page_no++;
1004                 probe_block += blocks_per_page;
1005 reprobe:
1006                 continue;
1007         }
1008         ret = nr_extents;
1009         *span = 1 + highest_block - lowest_block;
1010         if (page_no == 0)
1011                 page_no = 1;    /* force Empty message */
1012         sis->max = page_no;
1013         sis->pages = page_no - 1;
1014         sis->highest_bit = page_no - 1;
1015 done:
1016         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1017                                         struct swap_extent, list);
1018         goto out;
1019 bad_bmap:
1020         printk(KERN_ERR "swapon: swapfile has holes\n");
1021         ret = -EINVAL;
1022 out:
1023         return ret;
1024 }
1025
1026 #if 0   /* We don't need this yet */
1027 #include <linux/backing-dev.h>
1028 int page_queue_congested(struct page *page)
1029 {
1030         struct backing_dev_info *bdi;
1031
1032         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1033
1034         if (PageSwapCache(page)) {
1035                 swp_entry_t entry = { .val = page->private };
1036                 struct swap_info_struct *sis;
1037
1038                 sis = get_swap_info_struct(swp_type(entry));
1039                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1040         } else
1041                 bdi = page->mapping->backing_dev_info;
1042         return bdi_write_congested(bdi);
1043 }
1044 #endif
1045
1046 asmlinkage long sys_swapoff(const char __user * specialfile)
1047 {
1048         struct swap_info_struct * p = NULL;
1049         unsigned short *swap_map;
1050         struct file *swap_file, *victim;
1051         struct address_space *mapping;
1052         struct inode *inode;
1053         char * pathname;
1054         int i, type, prev;
1055         int err;
1056         
1057         if (!capable(CAP_SYS_ADMIN))
1058                 return -EPERM;
1059
1060         pathname = getname(specialfile);
1061         err = PTR_ERR(pathname);
1062         if (IS_ERR(pathname))
1063                 goto out;
1064
1065         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1066         putname(pathname);
1067         err = PTR_ERR(victim);
1068         if (IS_ERR(victim))
1069                 goto out;
1070
1071         mapping = victim->f_mapping;
1072         prev = -1;
1073         swap_list_lock();
1074         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1075                 p = swap_info + type;
1076                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1077                         if (p->swap_file->f_mapping == mapping)
1078                                 break;
1079                 }
1080                 prev = type;
1081         }
1082         if (type < 0) {
1083                 err = -EINVAL;
1084                 swap_list_unlock();
1085                 goto out_dput;
1086         }
1087         if (!security_vm_enough_memory(p->pages))
1088                 vm_unacct_memory(p->pages);
1089         else {
1090                 err = -ENOMEM;
1091                 swap_list_unlock();
1092                 goto out_dput;
1093         }
1094         if (prev < 0) {
1095                 swap_list.head = p->next;
1096         } else {
1097                 swap_info[prev].next = p->next;
1098         }
1099         if (type == swap_list.next) {
1100                 /* just pick something that's safe... */
1101                 swap_list.next = swap_list.head;
1102         }
1103         nr_swap_pages -= p->pages;
1104         total_swap_pages -= p->pages;
1105         swap_device_lock(p);
1106         p->flags &= ~SWP_WRITEOK;
1107         swap_device_unlock(p);
1108         swap_list_unlock();
1109
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 }