[PATCH] Swapless page migration: rip out swap based logic
[linux-2.6.git] / mm / swapfile.c
1 /*
2  *  linux/mm/swapfile.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *  Swap reorganised 29.12.95, Stephen Tweedie
6  */
7
8 #include <linux/config.h>
9 #include <linux/mm.h>
10 #include <linux/hugetlb.h>
11 #include <linux/mman.h>
12 #include <linux/slab.h>
13 #include <linux/kernel_stat.h>
14 #include <linux/swap.h>
15 #include <linux/vmalloc.h>
16 #include <linux/pagemap.h>
17 #include <linux/namei.h>
18 #include <linux/shm.h>
19 #include <linux/blkdev.h>
20 #include <linux/writeback.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/init.h>
24 #include <linux/module.h>
25 #include <linux/rmap.h>
26 #include <linux/security.h>
27 #include <linux/backing-dev.h>
28 #include <linux/mutex.h>
29 #include <linux/capability.h>
30 #include <linux/syscalls.h>
31
32 #include <asm/pgtable.h>
33 #include <asm/tlbflush.h>
34 #include <linux/swapops.h>
35
36 DEFINE_SPINLOCK(swap_lock);
37 unsigned int nr_swapfiles;
38 long total_swap_pages;
39 static int swap_overflow;
40
41 static const char Bad_file[] = "Bad swap file entry ";
42 static const char Unused_file[] = "Unused swap file entry ";
43 static const char Bad_offset[] = "Bad swap offset entry ";
44 static const char Unused_offset[] = "Unused swap offset entry ";
45
46 struct swap_list_t swap_list = {-1, -1};
47
48 static struct swap_info_struct swap_info[MAX_SWAPFILES];
49
50 static DEFINE_MUTEX(swapon_mutex);
51
52 /*
53  * We need this because the bdev->unplug_fn can sleep and we cannot
54  * hold swap_lock while calling the unplug_fn. And swap_lock
55  * cannot be turned into a mutex.
56  */
57 static DECLARE_RWSEM(swap_unplug_sem);
58
59 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
60 {
61         swp_entry_t entry;
62
63         down_read(&swap_unplug_sem);
64         entry.val = page_private(page);
65         if (PageSwapCache(page)) {
66                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
67                 struct backing_dev_info *bdi;
68
69                 /*
70                  * If the page is removed from swapcache from under us (with a
71                  * racy try_to_unuse/swapoff) we need an additional reference
72                  * count to avoid reading garbage from page_private(page) above.
73                  * If the WARN_ON triggers during a swapoff it maybe the race
74                  * condition and it's harmless. However if it triggers without
75                  * swapoff it signals a problem.
76                  */
77                 WARN_ON(page_count(page) <= 1);
78
79                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
80                 blk_run_backing_dev(bdi, page);
81         }
82         up_read(&swap_unplug_sem);
83 }
84
85 #define SWAPFILE_CLUSTER        256
86 #define LATENCY_LIMIT           256
87
88 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
89 {
90         unsigned long offset, last_in_cluster;
91         int latency_ration = LATENCY_LIMIT;
92
93         /* 
94          * We try to cluster swap pages by allocating them sequentially
95          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
96          * way, however, we resort to first-free allocation, starting
97          * a new cluster.  This prevents us from scattering swap pages
98          * all over the entire swap partition, so that we reduce
99          * overall disk seek times between swap pages.  -- sct
100          * But we do now try to find an empty cluster.  -Andrea
101          */
102
103         si->flags += SWP_SCANNING;
104         if (unlikely(!si->cluster_nr)) {
105                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
106                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER)
107                         goto lowest;
108                 spin_unlock(&swap_lock);
109
110                 offset = si->lowest_bit;
111                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
112
113                 /* Locate the first empty (unaligned) cluster */
114                 for (; last_in_cluster <= si->highest_bit; offset++) {
115                         if (si->swap_map[offset])
116                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
117                         else if (offset == last_in_cluster) {
118                                 spin_lock(&swap_lock);
119                                 si->cluster_next = offset-SWAPFILE_CLUSTER+1;
120                                 goto cluster;
121                         }
122                         if (unlikely(--latency_ration < 0)) {
123                                 cond_resched();
124                                 latency_ration = LATENCY_LIMIT;
125                         }
126                 }
127                 spin_lock(&swap_lock);
128                 goto lowest;
129         }
130
131         si->cluster_nr--;
132 cluster:
133         offset = si->cluster_next;
134         if (offset > si->highest_bit)
135 lowest:         offset = si->lowest_bit;
136 checks: if (!(si->flags & SWP_WRITEOK))
137                 goto no_page;
138         if (!si->highest_bit)
139                 goto no_page;
140         if (!si->swap_map[offset]) {
141                 if (offset == si->lowest_bit)
142                         si->lowest_bit++;
143                 if (offset == si->highest_bit)
144                         si->highest_bit--;
145                 si->inuse_pages++;
146                 if (si->inuse_pages == si->pages) {
147                         si->lowest_bit = si->max;
148                         si->highest_bit = 0;
149                 }
150                 si->swap_map[offset] = 1;
151                 si->cluster_next = offset + 1;
152                 si->flags -= SWP_SCANNING;
153                 return offset;
154         }
155
156         spin_unlock(&swap_lock);
157         while (++offset <= si->highest_bit) {
158                 if (!si->swap_map[offset]) {
159                         spin_lock(&swap_lock);
160                         goto checks;
161                 }
162                 if (unlikely(--latency_ration < 0)) {
163                         cond_resched();
164                         latency_ration = LATENCY_LIMIT;
165                 }
166         }
167         spin_lock(&swap_lock);
168         goto lowest;
169
170 no_page:
171         si->flags -= SWP_SCANNING;
172         return 0;
173 }
174
175 swp_entry_t get_swap_page(void)
176 {
177         struct swap_info_struct *si;
178         pgoff_t offset;
179         int type, next;
180         int wrapped = 0;
181
182         spin_lock(&swap_lock);
183         if (nr_swap_pages <= 0)
184                 goto noswap;
185         nr_swap_pages--;
186
187         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
188                 si = swap_info + type;
189                 next = si->next;
190                 if (next < 0 ||
191                     (!wrapped && si->prio != swap_info[next].prio)) {
192                         next = swap_list.head;
193                         wrapped++;
194                 }
195
196                 if (!si->highest_bit)
197                         continue;
198                 if (!(si->flags & SWP_WRITEOK))
199                         continue;
200
201                 swap_list.next = next;
202                 offset = scan_swap_map(si);
203                 if (offset) {
204                         spin_unlock(&swap_lock);
205                         return swp_entry(type, offset);
206                 }
207                 next = swap_list.next;
208         }
209
210         nr_swap_pages++;
211 noswap:
212         spin_unlock(&swap_lock);
213         return (swp_entry_t) {0};
214 }
215
216 swp_entry_t get_swap_page_of_type(int type)
217 {
218         struct swap_info_struct *si;
219         pgoff_t offset;
220
221         spin_lock(&swap_lock);
222         si = swap_info + type;
223         if (si->flags & SWP_WRITEOK) {
224                 nr_swap_pages--;
225                 offset = scan_swap_map(si);
226                 if (offset) {
227                         spin_unlock(&swap_lock);
228                         return swp_entry(type, offset);
229                 }
230                 nr_swap_pages++;
231         }
232         spin_unlock(&swap_lock);
233         return (swp_entry_t) {0};
234 }
235
236 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
237 {
238         struct swap_info_struct * p;
239         unsigned long offset, type;
240
241         if (!entry.val)
242                 goto out;
243         type = swp_type(entry);
244         if (type >= nr_swapfiles)
245                 goto bad_nofile;
246         p = & swap_info[type];
247         if (!(p->flags & SWP_USED))
248                 goto bad_device;
249         offset = swp_offset(entry);
250         if (offset >= p->max)
251                 goto bad_offset;
252         if (!p->swap_map[offset])
253                 goto bad_free;
254         spin_lock(&swap_lock);
255         return p;
256
257 bad_free:
258         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
259         goto out;
260 bad_offset:
261         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
262         goto out;
263 bad_device:
264         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
265         goto out;
266 bad_nofile:
267         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
268 out:
269         return NULL;
270 }       
271
272 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
273 {
274         int count = p->swap_map[offset];
275
276         if (count < SWAP_MAP_MAX) {
277                 count--;
278                 p->swap_map[offset] = count;
279                 if (!count) {
280                         if (offset < p->lowest_bit)
281                                 p->lowest_bit = offset;
282                         if (offset > p->highest_bit)
283                                 p->highest_bit = offset;
284                         if (p->prio > swap_info[swap_list.next].prio)
285                                 swap_list.next = p - swap_info;
286                         nr_swap_pages++;
287                         p->inuse_pages--;
288                 }
289         }
290         return count;
291 }
292
293 /*
294  * Caller has made sure that the swapdevice corresponding to entry
295  * is still around or has not been recycled.
296  */
297 void swap_free(swp_entry_t entry)
298 {
299         struct swap_info_struct * p;
300
301         p = swap_info_get(entry);
302         if (p) {
303                 swap_entry_free(p, swp_offset(entry));
304                 spin_unlock(&swap_lock);
305         }
306 }
307
308 /*
309  * How many references to page are currently swapped out?
310  */
311 static inline int page_swapcount(struct page *page)
312 {
313         int count = 0;
314         struct swap_info_struct *p;
315         swp_entry_t entry;
316
317         entry.val = page_private(page);
318         p = swap_info_get(entry);
319         if (p) {
320                 /* Subtract the 1 for the swap cache itself */
321                 count = p->swap_map[swp_offset(entry)] - 1;
322                 spin_unlock(&swap_lock);
323         }
324         return count;
325 }
326
327 /*
328  * We can use this swap cache entry directly
329  * if there are no other references to it.
330  */
331 int can_share_swap_page(struct page *page)
332 {
333         int count;
334
335         BUG_ON(!PageLocked(page));
336         count = page_mapcount(page);
337         if (count <= 1 && PageSwapCache(page))
338                 count += page_swapcount(page);
339         return count == 1;
340 }
341
342 /*
343  * Work out if there are any other processes sharing this
344  * swap cache page. Free it if you can. Return success.
345  */
346 int remove_exclusive_swap_page(struct page *page)
347 {
348         int retval;
349         struct swap_info_struct * p;
350         swp_entry_t entry;
351
352         BUG_ON(PagePrivate(page));
353         BUG_ON(!PageLocked(page));
354
355         if (!PageSwapCache(page))
356                 return 0;
357         if (PageWriteback(page))
358                 return 0;
359         if (page_count(page) != 2) /* 2: us + cache */
360                 return 0;
361
362         entry.val = page_private(page);
363         p = swap_info_get(entry);
364         if (!p)
365                 return 0;
366
367         /* Is the only swap cache user the cache itself? */
368         retval = 0;
369         if (p->swap_map[swp_offset(entry)] == 1) {
370                 /* Recheck the page count with the swapcache lock held.. */
371                 write_lock_irq(&swapper_space.tree_lock);
372                 if ((page_count(page) == 2) && !PageWriteback(page)) {
373                         __delete_from_swap_cache(page);
374                         SetPageDirty(page);
375                         retval = 1;
376                 }
377                 write_unlock_irq(&swapper_space.tree_lock);
378         }
379         spin_unlock(&swap_lock);
380
381         if (retval) {
382                 swap_free(entry);
383                 page_cache_release(page);
384         }
385
386         return retval;
387 }
388
389 /*
390  * Free the swap entry like above, but also try to
391  * free the page cache entry if it is the last user.
392  */
393 void free_swap_and_cache(swp_entry_t entry)
394 {
395         struct swap_info_struct * p;
396         struct page *page = NULL;
397
398         if (is_migration_entry(entry))
399                 return;
400
401         p = swap_info_get(entry);
402         if (p) {
403                 if (swap_entry_free(p, swp_offset(entry)) == 1) {
404                         page = find_get_page(&swapper_space, entry.val);
405                         if (page && unlikely(TestSetPageLocked(page))) {
406                                 page_cache_release(page);
407                                 page = NULL;
408                         }
409                 }
410                 spin_unlock(&swap_lock);
411         }
412         if (page) {
413                 int one_user;
414
415                 BUG_ON(PagePrivate(page));
416                 one_user = (page_count(page) == 2);
417                 /* Only cache user (+us), or swap space full? Free it! */
418                 /* Also recheck PageSwapCache after page is locked (above) */
419                 if (PageSwapCache(page) && !PageWriteback(page) &&
420                                         (one_user || vm_swap_full())) {
421                         delete_from_swap_cache(page);
422                         SetPageDirty(page);
423                 }
424                 unlock_page(page);
425                 page_cache_release(page);
426         }
427 }
428
429 #ifdef CONFIG_SOFTWARE_SUSPEND
430 /*
431  * Find the swap type that corresponds to given device (if any)
432  *
433  * This is needed for software suspend and is done in such a way that inode
434  * aliasing is allowed.
435  */
436 int swap_type_of(dev_t device)
437 {
438         int i;
439
440         spin_lock(&swap_lock);
441         for (i = 0; i < nr_swapfiles; i++) {
442                 struct inode *inode;
443
444                 if (!(swap_info[i].flags & SWP_WRITEOK))
445                         continue;
446                 if (!device) {
447                         spin_unlock(&swap_lock);
448                         return i;
449                 }
450                 inode = swap_info->swap_file->f_dentry->d_inode;
451                 if (S_ISBLK(inode->i_mode) &&
452                     device == MKDEV(imajor(inode), iminor(inode))) {
453                         spin_unlock(&swap_lock);
454                         return i;
455                 }
456         }
457         spin_unlock(&swap_lock);
458         return -ENODEV;
459 }
460
461 /*
462  * Return either the total number of swap pages of given type, or the number
463  * of free pages of that type (depending on @free)
464  *
465  * This is needed for software suspend
466  */
467 unsigned int count_swap_pages(int type, int free)
468 {
469         unsigned int n = 0;
470
471         if (type < nr_swapfiles) {
472                 spin_lock(&swap_lock);
473                 if (swap_info[type].flags & SWP_WRITEOK) {
474                         n = swap_info[type].pages;
475                         if (free)
476                                 n -= swap_info[type].inuse_pages;
477                 }
478                 spin_unlock(&swap_lock);
479         }
480         return n;
481 }
482 #endif
483
484 /*
485  * No need to decide whether this PTE shares the swap entry with others,
486  * just let do_wp_page work it out if a write is requested later - to
487  * force COW, vm_page_prot omits write permission from any private vma.
488  */
489 static void unuse_pte(struct vm_area_struct *vma, pte_t *pte,
490                 unsigned long addr, swp_entry_t entry, struct page *page)
491 {
492         inc_mm_counter(vma->vm_mm, anon_rss);
493         get_page(page);
494         set_pte_at(vma->vm_mm, addr, pte,
495                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
496         page_add_anon_rmap(page, vma, addr);
497         swap_free(entry);
498         /*
499          * Move the page to the active list so it is not
500          * immediately swapped out again after swapon.
501          */
502         activate_page(page);
503 }
504
505 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
506                                 unsigned long addr, unsigned long end,
507                                 swp_entry_t entry, struct page *page)
508 {
509         pte_t swp_pte = swp_entry_to_pte(entry);
510         pte_t *pte;
511         spinlock_t *ptl;
512         int found = 0;
513
514         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
515         do {
516                 /*
517                  * swapoff spends a _lot_ of time in this loop!
518                  * Test inline before going to call unuse_pte.
519                  */
520                 if (unlikely(pte_same(*pte, swp_pte))) {
521                         unuse_pte(vma, pte++, addr, entry, page);
522                         found = 1;
523                         break;
524                 }
525         } while (pte++, addr += PAGE_SIZE, addr != end);
526         pte_unmap_unlock(pte - 1, ptl);
527         return found;
528 }
529
530 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
531                                 unsigned long addr, unsigned long end,
532                                 swp_entry_t entry, struct page *page)
533 {
534         pmd_t *pmd;
535         unsigned long next;
536
537         pmd = pmd_offset(pud, addr);
538         do {
539                 next = pmd_addr_end(addr, end);
540                 if (pmd_none_or_clear_bad(pmd))
541                         continue;
542                 if (unuse_pte_range(vma, pmd, addr, next, entry, page))
543                         return 1;
544         } while (pmd++, addr = next, addr != end);
545         return 0;
546 }
547
548 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
549                                 unsigned long addr, unsigned long end,
550                                 swp_entry_t entry, struct page *page)
551 {
552         pud_t *pud;
553         unsigned long next;
554
555         pud = pud_offset(pgd, addr);
556         do {
557                 next = pud_addr_end(addr, end);
558                 if (pud_none_or_clear_bad(pud))
559                         continue;
560                 if (unuse_pmd_range(vma, pud, addr, next, entry, page))
561                         return 1;
562         } while (pud++, addr = next, addr != end);
563         return 0;
564 }
565
566 static int unuse_vma(struct vm_area_struct *vma,
567                                 swp_entry_t entry, struct page *page)
568 {
569         pgd_t *pgd;
570         unsigned long addr, end, next;
571
572         if (page->mapping) {
573                 addr = page_address_in_vma(page, vma);
574                 if (addr == -EFAULT)
575                         return 0;
576                 else
577                         end = addr + PAGE_SIZE;
578         } else {
579                 addr = vma->vm_start;
580                 end = vma->vm_end;
581         }
582
583         pgd = pgd_offset(vma->vm_mm, addr);
584         do {
585                 next = pgd_addr_end(addr, end);
586                 if (pgd_none_or_clear_bad(pgd))
587                         continue;
588                 if (unuse_pud_range(vma, pgd, addr, next, entry, page))
589                         return 1;
590         } while (pgd++, addr = next, addr != end);
591         return 0;
592 }
593
594 static int unuse_mm(struct mm_struct *mm,
595                                 swp_entry_t entry, struct page *page)
596 {
597         struct vm_area_struct *vma;
598
599         if (!down_read_trylock(&mm->mmap_sem)) {
600                 /*
601                  * Activate page so shrink_cache is unlikely to unmap its
602                  * ptes while lock is dropped, so swapoff can make progress.
603                  */
604                 activate_page(page);
605                 unlock_page(page);
606                 down_read(&mm->mmap_sem);
607                 lock_page(page);
608         }
609         for (vma = mm->mmap; vma; vma = vma->vm_next) {
610                 if (vma->anon_vma && unuse_vma(vma, entry, page))
611                         break;
612         }
613         up_read(&mm->mmap_sem);
614         /*
615          * Currently unuse_mm cannot fail, but leave error handling
616          * at call sites for now, since we change it from time to time.
617          */
618         return 0;
619 }
620
621 /*
622  * Scan swap_map from current position to next entry still in use.
623  * Recycle to start on reaching the end, returning 0 when empty.
624  */
625 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
626                                         unsigned int prev)
627 {
628         unsigned int max = si->max;
629         unsigned int i = prev;
630         int count;
631
632         /*
633          * No need for swap_lock here: we're just looking
634          * for whether an entry is in use, not modifying it; false
635          * hits are okay, and sys_swapoff() has already prevented new
636          * allocations from this area (while holding swap_lock).
637          */
638         for (;;) {
639                 if (++i >= max) {
640                         if (!prev) {
641                                 i = 0;
642                                 break;
643                         }
644                         /*
645                          * No entries in use at top of swap_map,
646                          * loop back to start and recheck there.
647                          */
648                         max = prev + 1;
649                         prev = 0;
650                         i = 1;
651                 }
652                 count = si->swap_map[i];
653                 if (count && count != SWAP_MAP_BAD)
654                         break;
655         }
656         return i;
657 }
658
659 /*
660  * We completely avoid races by reading each swap page in advance,
661  * and then search for the process using it.  All the necessary
662  * page table adjustments can then be made atomically.
663  */
664 static int try_to_unuse(unsigned int type)
665 {
666         struct swap_info_struct * si = &swap_info[type];
667         struct mm_struct *start_mm;
668         unsigned short *swap_map;
669         unsigned short swcount;
670         struct page *page;
671         swp_entry_t entry;
672         unsigned int i = 0;
673         int retval = 0;
674         int reset_overflow = 0;
675         int shmem;
676
677         /*
678          * When searching mms for an entry, a good strategy is to
679          * start at the first mm we freed the previous entry from
680          * (though actually we don't notice whether we or coincidence
681          * freed the entry).  Initialize this start_mm with a hold.
682          *
683          * A simpler strategy would be to start at the last mm we
684          * freed the previous entry from; but that would take less
685          * advantage of mmlist ordering, which clusters forked mms
686          * together, child after parent.  If we race with dup_mmap(), we
687          * prefer to resolve parent before child, lest we miss entries
688          * duplicated after we scanned child: using last mm would invert
689          * that.  Though it's only a serious concern when an overflowed
690          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
691          */
692         start_mm = &init_mm;
693         atomic_inc(&init_mm.mm_users);
694
695         /*
696          * Keep on scanning until all entries have gone.  Usually,
697          * one pass through swap_map is enough, but not necessarily:
698          * there are races when an instance of an entry might be missed.
699          */
700         while ((i = find_next_to_unuse(si, i)) != 0) {
701                 if (signal_pending(current)) {
702                         retval = -EINTR;
703                         break;
704                 }
705
706                 /* 
707                  * Get a page for the entry, using the existing swap
708                  * cache page if there is one.  Otherwise, get a clean
709                  * page and read the swap into it. 
710                  */
711                 swap_map = &si->swap_map[i];
712                 entry = swp_entry(type, i);
713                 page = read_swap_cache_async(entry, NULL, 0);
714                 if (!page) {
715                         /*
716                          * Either swap_duplicate() failed because entry
717                          * has been freed independently, and will not be
718                          * reused since sys_swapoff() already disabled
719                          * allocation from here, or alloc_page() failed.
720                          */
721                         if (!*swap_map)
722                                 continue;
723                         retval = -ENOMEM;
724                         break;
725                 }
726
727                 /*
728                  * Don't hold on to start_mm if it looks like exiting.
729                  */
730                 if (atomic_read(&start_mm->mm_users) == 1) {
731                         mmput(start_mm);
732                         start_mm = &init_mm;
733                         atomic_inc(&init_mm.mm_users);
734                 }
735
736                 /*
737                  * Wait for and lock page.  When do_swap_page races with
738                  * try_to_unuse, do_swap_page can handle the fault much
739                  * faster than try_to_unuse can locate the entry.  This
740                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
741                  * defer to do_swap_page in such a case - in some tests,
742                  * do_swap_page and try_to_unuse repeatedly compete.
743                  */
744                 wait_on_page_locked(page);
745                 wait_on_page_writeback(page);
746                 lock_page(page);
747                 wait_on_page_writeback(page);
748
749                 /*
750                  * Remove all references to entry.
751                  * Whenever we reach init_mm, there's no address space
752                  * to search, but use it as a reminder to search shmem.
753                  */
754                 shmem = 0;
755                 swcount = *swap_map;
756                 if (swcount > 1) {
757                         if (start_mm == &init_mm)
758                                 shmem = shmem_unuse(entry, page);
759                         else
760                                 retval = unuse_mm(start_mm, entry, page);
761                 }
762                 if (*swap_map > 1) {
763                         int set_start_mm = (*swap_map >= swcount);
764                         struct list_head *p = &start_mm->mmlist;
765                         struct mm_struct *new_start_mm = start_mm;
766                         struct mm_struct *prev_mm = start_mm;
767                         struct mm_struct *mm;
768
769                         atomic_inc(&new_start_mm->mm_users);
770                         atomic_inc(&prev_mm->mm_users);
771                         spin_lock(&mmlist_lock);
772                         while (*swap_map > 1 && !retval &&
773                                         (p = p->next) != &start_mm->mmlist) {
774                                 mm = list_entry(p, struct mm_struct, mmlist);
775                                 if (atomic_inc_return(&mm->mm_users) == 1) {
776                                         atomic_dec(&mm->mm_users);
777                                         continue;
778                                 }
779                                 spin_unlock(&mmlist_lock);
780                                 mmput(prev_mm);
781                                 prev_mm = mm;
782
783                                 cond_resched();
784
785                                 swcount = *swap_map;
786                                 if (swcount <= 1)
787                                         ;
788                                 else if (mm == &init_mm) {
789                                         set_start_mm = 1;
790                                         shmem = shmem_unuse(entry, page);
791                                 } else
792                                         retval = unuse_mm(mm, entry, page);
793                                 if (set_start_mm && *swap_map < swcount) {
794                                         mmput(new_start_mm);
795                                         atomic_inc(&mm->mm_users);
796                                         new_start_mm = mm;
797                                         set_start_mm = 0;
798                                 }
799                                 spin_lock(&mmlist_lock);
800                         }
801                         spin_unlock(&mmlist_lock);
802                         mmput(prev_mm);
803                         mmput(start_mm);
804                         start_mm = new_start_mm;
805                 }
806                 if (retval) {
807                         unlock_page(page);
808                         page_cache_release(page);
809                         break;
810                 }
811
812                 /*
813                  * How could swap count reach 0x7fff when the maximum
814                  * pid is 0x7fff, and there's no way to repeat a swap
815                  * page within an mm (except in shmem, where it's the
816                  * shared object which takes the reference count)?
817                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
818                  *
819                  * If that's wrong, then we should worry more about
820                  * exit_mmap() and do_munmap() cases described above:
821                  * we might be resetting SWAP_MAP_MAX too early here.
822                  * We know "Undead"s can happen, they're okay, so don't
823                  * report them; but do report if we reset SWAP_MAP_MAX.
824                  */
825                 if (*swap_map == SWAP_MAP_MAX) {
826                         spin_lock(&swap_lock);
827                         *swap_map = 1;
828                         spin_unlock(&swap_lock);
829                         reset_overflow = 1;
830                 }
831
832                 /*
833                  * If a reference remains (rare), we would like to leave
834                  * the page in the swap cache; but try_to_unmap could
835                  * then re-duplicate the entry once we drop page lock,
836                  * so we might loop indefinitely; also, that page could
837                  * not be swapped out to other storage meanwhile.  So:
838                  * delete from cache even if there's another reference,
839                  * after ensuring that the data has been saved to disk -
840                  * since if the reference remains (rarer), it will be
841                  * read from disk into another page.  Splitting into two
842                  * pages would be incorrect if swap supported "shared
843                  * private" pages, but they are handled by tmpfs files.
844                  *
845                  * Note shmem_unuse already deleted a swappage from
846                  * the swap cache, unless the move to filepage failed:
847                  * in which case it left swappage in cache, lowered its
848                  * swap count to pass quickly through the loops above,
849                  * and now we must reincrement count to try again later.
850                  */
851                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
852                         struct writeback_control wbc = {
853                                 .sync_mode = WB_SYNC_NONE,
854                         };
855
856                         swap_writepage(page, &wbc);
857                         lock_page(page);
858                         wait_on_page_writeback(page);
859                 }
860                 if (PageSwapCache(page)) {
861                         if (shmem)
862                                 swap_duplicate(entry);
863                         else
864                                 delete_from_swap_cache(page);
865                 }
866
867                 /*
868                  * So we could skip searching mms once swap count went
869                  * to 1, we did not mark any present ptes as dirty: must
870                  * mark page dirty so shrink_list will preserve it.
871                  */
872                 SetPageDirty(page);
873                 unlock_page(page);
874                 page_cache_release(page);
875
876                 /*
877                  * Make sure that we aren't completely killing
878                  * interactive performance.
879                  */
880                 cond_resched();
881         }
882
883         mmput(start_mm);
884         if (reset_overflow) {
885                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
886                 swap_overflow = 0;
887         }
888         return retval;
889 }
890
891 /*
892  * After a successful try_to_unuse, if no swap is now in use, we know
893  * we can empty the mmlist.  swap_lock must be held on entry and exit.
894  * Note that mmlist_lock nests inside swap_lock, and an mm must be
895  * added to the mmlist just after page_duplicate - before would be racy.
896  */
897 static void drain_mmlist(void)
898 {
899         struct list_head *p, *next;
900         unsigned int i;
901
902         for (i = 0; i < nr_swapfiles; i++)
903                 if (swap_info[i].inuse_pages)
904                         return;
905         spin_lock(&mmlist_lock);
906         list_for_each_safe(p, next, &init_mm.mmlist)
907                 list_del_init(p);
908         spin_unlock(&mmlist_lock);
909 }
910
911 /*
912  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
913  * corresponds to page offset `offset'.
914  */
915 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
916 {
917         struct swap_extent *se = sis->curr_swap_extent;
918         struct swap_extent *start_se = se;
919
920         for ( ; ; ) {
921                 struct list_head *lh;
922
923                 if (se->start_page <= offset &&
924                                 offset < (se->start_page + se->nr_pages)) {
925                         return se->start_block + (offset - se->start_page);
926                 }
927                 lh = se->list.next;
928                 if (lh == &sis->extent_list)
929                         lh = lh->next;
930                 se = list_entry(lh, struct swap_extent, list);
931                 sis->curr_swap_extent = se;
932                 BUG_ON(se == start_se);         /* It *must* be present */
933         }
934 }
935
936 /*
937  * Free all of a swapdev's extent information
938  */
939 static void destroy_swap_extents(struct swap_info_struct *sis)
940 {
941         while (!list_empty(&sis->extent_list)) {
942                 struct swap_extent *se;
943
944                 se = list_entry(sis->extent_list.next,
945                                 struct swap_extent, list);
946                 list_del(&se->list);
947                 kfree(se);
948         }
949 }
950
951 /*
952  * Add a block range (and the corresponding page range) into this swapdev's
953  * extent list.  The extent list is kept sorted in page order.
954  *
955  * This function rather assumes that it is called in ascending page order.
956  */
957 static int
958 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
959                 unsigned long nr_pages, sector_t start_block)
960 {
961         struct swap_extent *se;
962         struct swap_extent *new_se;
963         struct list_head *lh;
964
965         lh = sis->extent_list.prev;     /* The highest page extent */
966         if (lh != &sis->extent_list) {
967                 se = list_entry(lh, struct swap_extent, list);
968                 BUG_ON(se->start_page + se->nr_pages != start_page);
969                 if (se->start_block + se->nr_pages == start_block) {
970                         /* Merge it */
971                         se->nr_pages += nr_pages;
972                         return 0;
973                 }
974         }
975
976         /*
977          * No merge.  Insert a new extent, preserving ordering.
978          */
979         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
980         if (new_se == NULL)
981                 return -ENOMEM;
982         new_se->start_page = start_page;
983         new_se->nr_pages = nr_pages;
984         new_se->start_block = start_block;
985
986         list_add_tail(&new_se->list, &sis->extent_list);
987         return 1;
988 }
989
990 /*
991  * A `swap extent' is a simple thing which maps a contiguous range of pages
992  * onto a contiguous range of disk blocks.  An ordered list of swap extents
993  * is built at swapon time and is then used at swap_writepage/swap_readpage
994  * time for locating where on disk a page belongs.
995  *
996  * If the swapfile is an S_ISBLK block device, a single extent is installed.
997  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
998  * swap files identically.
999  *
1000  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1001  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1002  * swapfiles are handled *identically* after swapon time.
1003  *
1004  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1005  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1006  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1007  * requirements, they are simply tossed out - we will never use those blocks
1008  * for swapping.
1009  *
1010  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1011  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1012  * which will scribble on the fs.
1013  *
1014  * The amount of disk space which a single swap extent represents varies.
1015  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1016  * extents in the list.  To avoid much list walking, we cache the previous
1017  * search location in `curr_swap_extent', and start new searches from there.
1018  * This is extremely effective.  The average number of iterations in
1019  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1020  */
1021 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1022 {
1023         struct inode *inode;
1024         unsigned blocks_per_page;
1025         unsigned long page_no;
1026         unsigned blkbits;
1027         sector_t probe_block;
1028         sector_t last_block;
1029         sector_t lowest_block = -1;
1030         sector_t highest_block = 0;
1031         int nr_extents = 0;
1032         int ret;
1033
1034         inode = sis->swap_file->f_mapping->host;
1035         if (S_ISBLK(inode->i_mode)) {
1036                 ret = add_swap_extent(sis, 0, sis->max, 0);
1037                 *span = sis->pages;
1038                 goto done;
1039         }
1040
1041         blkbits = inode->i_blkbits;
1042         blocks_per_page = PAGE_SIZE >> blkbits;
1043
1044         /*
1045          * Map all the blocks into the extent list.  This code doesn't try
1046          * to be very smart.
1047          */
1048         probe_block = 0;
1049         page_no = 0;
1050         last_block = i_size_read(inode) >> blkbits;
1051         while ((probe_block + blocks_per_page) <= last_block &&
1052                         page_no < sis->max) {
1053                 unsigned block_in_page;
1054                 sector_t first_block;
1055
1056                 first_block = bmap(inode, probe_block);
1057                 if (first_block == 0)
1058                         goto bad_bmap;
1059
1060                 /*
1061                  * It must be PAGE_SIZE aligned on-disk
1062                  */
1063                 if (first_block & (blocks_per_page - 1)) {
1064                         probe_block++;
1065                         goto reprobe;
1066                 }
1067
1068                 for (block_in_page = 1; block_in_page < blocks_per_page;
1069                                         block_in_page++) {
1070                         sector_t block;
1071
1072                         block = bmap(inode, probe_block + block_in_page);
1073                         if (block == 0)
1074                                 goto bad_bmap;
1075                         if (block != first_block + block_in_page) {
1076                                 /* Discontiguity */
1077                                 probe_block++;
1078                                 goto reprobe;
1079                         }
1080                 }
1081
1082                 first_block >>= (PAGE_SHIFT - blkbits);
1083                 if (page_no) {  /* exclude the header page */
1084                         if (first_block < lowest_block)
1085                                 lowest_block = first_block;
1086                         if (first_block > highest_block)
1087                                 highest_block = first_block;
1088                 }
1089
1090                 /*
1091                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1092                  */
1093                 ret = add_swap_extent(sis, page_no, 1, first_block);
1094                 if (ret < 0)
1095                         goto out;
1096                 nr_extents += ret;
1097                 page_no++;
1098                 probe_block += blocks_per_page;
1099 reprobe:
1100                 continue;
1101         }
1102         ret = nr_extents;
1103         *span = 1 + highest_block - lowest_block;
1104         if (page_no == 0)
1105                 page_no = 1;    /* force Empty message */
1106         sis->max = page_no;
1107         sis->pages = page_no - 1;
1108         sis->highest_bit = page_no - 1;
1109 done:
1110         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1111                                         struct swap_extent, list);
1112         goto out;
1113 bad_bmap:
1114         printk(KERN_ERR "swapon: swapfile has holes\n");
1115         ret = -EINVAL;
1116 out:
1117         return ret;
1118 }
1119
1120 #if 0   /* We don't need this yet */
1121 #include <linux/backing-dev.h>
1122 int page_queue_congested(struct page *page)
1123 {
1124         struct backing_dev_info *bdi;
1125
1126         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1127
1128         if (PageSwapCache(page)) {
1129                 swp_entry_t entry = { .val = page_private(page) };
1130                 struct swap_info_struct *sis;
1131
1132                 sis = get_swap_info_struct(swp_type(entry));
1133                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1134         } else
1135                 bdi = page->mapping->backing_dev_info;
1136         return bdi_write_congested(bdi);
1137 }
1138 #endif
1139
1140 asmlinkage long sys_swapoff(const char __user * specialfile)
1141 {
1142         struct swap_info_struct * p = NULL;
1143         unsigned short *swap_map;
1144         struct file *swap_file, *victim;
1145         struct address_space *mapping;
1146         struct inode *inode;
1147         char * pathname;
1148         int i, type, prev;
1149         int err;
1150         
1151         if (!capable(CAP_SYS_ADMIN))
1152                 return -EPERM;
1153
1154         pathname = getname(specialfile);
1155         err = PTR_ERR(pathname);
1156         if (IS_ERR(pathname))
1157                 goto out;
1158
1159         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1160         putname(pathname);
1161         err = PTR_ERR(victim);
1162         if (IS_ERR(victim))
1163                 goto out;
1164
1165         mapping = victim->f_mapping;
1166         prev = -1;
1167         spin_lock(&swap_lock);
1168         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1169                 p = swap_info + type;
1170                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1171                         if (p->swap_file->f_mapping == mapping)
1172                                 break;
1173                 }
1174                 prev = type;
1175         }
1176         if (type < 0) {
1177                 err = -EINVAL;
1178                 spin_unlock(&swap_lock);
1179                 goto out_dput;
1180         }
1181         if (!security_vm_enough_memory(p->pages))
1182                 vm_unacct_memory(p->pages);
1183         else {
1184                 err = -ENOMEM;
1185                 spin_unlock(&swap_lock);
1186                 goto out_dput;
1187         }
1188         if (prev < 0) {
1189                 swap_list.head = p->next;
1190         } else {
1191                 swap_info[prev].next = p->next;
1192         }
1193         if (type == swap_list.next) {
1194                 /* just pick something that's safe... */
1195                 swap_list.next = swap_list.head;
1196         }
1197         nr_swap_pages -= p->pages;
1198         total_swap_pages -= p->pages;
1199         p->flags &= ~SWP_WRITEOK;
1200         spin_unlock(&swap_lock);
1201
1202         current->flags |= PF_SWAPOFF;
1203         err = try_to_unuse(type);
1204         current->flags &= ~PF_SWAPOFF;
1205
1206         if (err) {
1207                 /* re-insert swap space back into swap_list */
1208                 spin_lock(&swap_lock);
1209                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1210                         if (p->prio >= swap_info[i].prio)
1211                                 break;
1212                 p->next = i;
1213                 if (prev < 0)
1214                         swap_list.head = swap_list.next = p - swap_info;
1215                 else
1216                         swap_info[prev].next = p - swap_info;
1217                 nr_swap_pages += p->pages;
1218                 total_swap_pages += p->pages;
1219                 p->flags |= SWP_WRITEOK;
1220                 spin_unlock(&swap_lock);
1221                 goto out_dput;
1222         }
1223
1224         /* wait for any unplug function to finish */
1225         down_write(&swap_unplug_sem);
1226         up_write(&swap_unplug_sem);
1227
1228         destroy_swap_extents(p);
1229         mutex_lock(&swapon_mutex);
1230         spin_lock(&swap_lock);
1231         drain_mmlist();
1232
1233         /* wait for anyone still in scan_swap_map */
1234         p->highest_bit = 0;             /* cuts scans short */
1235         while (p->flags >= SWP_SCANNING) {
1236                 spin_unlock(&swap_lock);
1237                 schedule_timeout_uninterruptible(1);
1238                 spin_lock(&swap_lock);
1239         }
1240
1241         swap_file = p->swap_file;
1242         p->swap_file = NULL;
1243         p->max = 0;
1244         swap_map = p->swap_map;
1245         p->swap_map = NULL;
1246         p->flags = 0;
1247         spin_unlock(&swap_lock);
1248         mutex_unlock(&swapon_mutex);
1249         vfree(swap_map);
1250         inode = mapping->host;
1251         if (S_ISBLK(inode->i_mode)) {
1252                 struct block_device *bdev = I_BDEV(inode);
1253                 set_blocksize(bdev, p->old_block_size);
1254                 bd_release(bdev);
1255         } else {
1256                 mutex_lock(&inode->i_mutex);
1257                 inode->i_flags &= ~S_SWAPFILE;
1258                 mutex_unlock(&inode->i_mutex);
1259         }
1260         filp_close(swap_file, NULL);
1261         err = 0;
1262
1263 out_dput:
1264         filp_close(victim, NULL);
1265 out:
1266         return err;
1267 }
1268
1269 #ifdef CONFIG_PROC_FS
1270 /* iterator */
1271 static void *swap_start(struct seq_file *swap, loff_t *pos)
1272 {
1273         struct swap_info_struct *ptr = swap_info;
1274         int i;
1275         loff_t l = *pos;
1276
1277         mutex_lock(&swapon_mutex);
1278
1279         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1280                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1281                         continue;
1282                 if (!l--)
1283                         return ptr;
1284         }
1285
1286         return NULL;
1287 }
1288
1289 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1290 {
1291         struct swap_info_struct *ptr = v;
1292         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1293
1294         for (++ptr; ptr < endptr; ptr++) {
1295                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1296                         continue;
1297                 ++*pos;
1298                 return ptr;
1299         }
1300
1301         return NULL;
1302 }
1303
1304 static void swap_stop(struct seq_file *swap, void *v)
1305 {
1306         mutex_unlock(&swapon_mutex);
1307 }
1308
1309 static int swap_show(struct seq_file *swap, void *v)
1310 {
1311         struct swap_info_struct *ptr = v;
1312         struct file *file;
1313         int len;
1314
1315         if (v == swap_info)
1316                 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1317
1318         file = ptr->swap_file;
1319         len = seq_path(swap, file->f_vfsmnt, file->f_dentry, " \t\n\\");
1320         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1321                        len < 40 ? 40 - len : 1, " ",
1322                        S_ISBLK(file->f_dentry->d_inode->i_mode) ?
1323                                 "partition" : "file\t",
1324                        ptr->pages << (PAGE_SHIFT - 10),
1325                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1326                        ptr->prio);
1327         return 0;
1328 }
1329
1330 static struct seq_operations swaps_op = {
1331         .start =        swap_start,
1332         .next =         swap_next,
1333         .stop =         swap_stop,
1334         .show =         swap_show
1335 };
1336
1337 static int swaps_open(struct inode *inode, struct file *file)
1338 {
1339         return seq_open(file, &swaps_op);
1340 }
1341
1342 static struct file_operations proc_swaps_operations = {
1343         .open           = swaps_open,
1344         .read           = seq_read,
1345         .llseek         = seq_lseek,
1346         .release        = seq_release,
1347 };
1348
1349 static int __init procswaps_init(void)
1350 {
1351         struct proc_dir_entry *entry;
1352
1353         entry = create_proc_entry("swaps", 0, NULL);
1354         if (entry)
1355                 entry->proc_fops = &proc_swaps_operations;
1356         return 0;
1357 }
1358 __initcall(procswaps_init);
1359 #endif /* CONFIG_PROC_FS */
1360
1361 /*
1362  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1363  *
1364  * The swapon system call
1365  */
1366 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1367 {
1368         struct swap_info_struct * p;
1369         char *name = NULL;
1370         struct block_device *bdev = NULL;
1371         struct file *swap_file = NULL;
1372         struct address_space *mapping;
1373         unsigned int type;
1374         int i, prev;
1375         int error;
1376         static int least_priority;
1377         union swap_header *swap_header = NULL;
1378         int swap_header_version;
1379         unsigned int nr_good_pages = 0;
1380         int nr_extents = 0;
1381         sector_t span;
1382         unsigned long maxpages = 1;
1383         int swapfilesize;
1384         unsigned short *swap_map;
1385         struct page *page = NULL;
1386         struct inode *inode = NULL;
1387         int did_down = 0;
1388
1389         if (!capable(CAP_SYS_ADMIN))
1390                 return -EPERM;
1391         spin_lock(&swap_lock);
1392         p = swap_info;
1393         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1394                 if (!(p->flags & SWP_USED))
1395                         break;
1396         error = -EPERM;
1397         if (type >= MAX_SWAPFILES) {
1398                 spin_unlock(&swap_lock);
1399                 goto out;
1400         }
1401         if (type >= nr_swapfiles)
1402                 nr_swapfiles = type+1;
1403         INIT_LIST_HEAD(&p->extent_list);
1404         p->flags = SWP_USED;
1405         p->swap_file = NULL;
1406         p->old_block_size = 0;
1407         p->swap_map = NULL;
1408         p->lowest_bit = 0;
1409         p->highest_bit = 0;
1410         p->cluster_nr = 0;
1411         p->inuse_pages = 0;
1412         p->next = -1;
1413         if (swap_flags & SWAP_FLAG_PREFER) {
1414                 p->prio =
1415                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1416         } else {
1417                 p->prio = --least_priority;
1418         }
1419         spin_unlock(&swap_lock);
1420         name = getname(specialfile);
1421         error = PTR_ERR(name);
1422         if (IS_ERR(name)) {
1423                 name = NULL;
1424                 goto bad_swap_2;
1425         }
1426         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1427         error = PTR_ERR(swap_file);
1428         if (IS_ERR(swap_file)) {
1429                 swap_file = NULL;
1430                 goto bad_swap_2;
1431         }
1432
1433         p->swap_file = swap_file;
1434         mapping = swap_file->f_mapping;
1435         inode = mapping->host;
1436
1437         error = -EBUSY;
1438         for (i = 0; i < nr_swapfiles; i++) {
1439                 struct swap_info_struct *q = &swap_info[i];
1440
1441                 if (i == type || !q->swap_file)
1442                         continue;
1443                 if (mapping == q->swap_file->f_mapping)
1444                         goto bad_swap;
1445         }
1446
1447         error = -EINVAL;
1448         if (S_ISBLK(inode->i_mode)) {
1449                 bdev = I_BDEV(inode);
1450                 error = bd_claim(bdev, sys_swapon);
1451                 if (error < 0) {
1452                         bdev = NULL;
1453                         error = -EINVAL;
1454                         goto bad_swap;
1455                 }
1456                 p->old_block_size = block_size(bdev);
1457                 error = set_blocksize(bdev, PAGE_SIZE);
1458                 if (error < 0)
1459                         goto bad_swap;
1460                 p->bdev = bdev;
1461         } else if (S_ISREG(inode->i_mode)) {
1462                 p->bdev = inode->i_sb->s_bdev;
1463                 mutex_lock(&inode->i_mutex);
1464                 did_down = 1;
1465                 if (IS_SWAPFILE(inode)) {
1466                         error = -EBUSY;
1467                         goto bad_swap;
1468                 }
1469         } else {
1470                 goto bad_swap;
1471         }
1472
1473         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1474
1475         /*
1476          * Read the swap header.
1477          */
1478         if (!mapping->a_ops->readpage) {
1479                 error = -EINVAL;
1480                 goto bad_swap;
1481         }
1482         page = read_cache_page(mapping, 0,
1483                         (filler_t *)mapping->a_ops->readpage, swap_file);
1484         if (IS_ERR(page)) {
1485                 error = PTR_ERR(page);
1486                 goto bad_swap;
1487         }
1488         wait_on_page_locked(page);
1489         if (!PageUptodate(page))
1490                 goto bad_swap;
1491         kmap(page);
1492         swap_header = page_address(page);
1493
1494         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1495                 swap_header_version = 1;
1496         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1497                 swap_header_version = 2;
1498         else {
1499                 printk(KERN_ERR "Unable to find swap-space signature\n");
1500                 error = -EINVAL;
1501                 goto bad_swap;
1502         }
1503         
1504         switch (swap_header_version) {
1505         case 1:
1506                 printk(KERN_ERR "version 0 swap is no longer supported. "
1507                         "Use mkswap -v1 %s\n", name);
1508                 error = -EINVAL;
1509                 goto bad_swap;
1510         case 2:
1511                 /* Check the swap header's sub-version and the size of
1512                    the swap file and bad block lists */
1513                 if (swap_header->info.version != 1) {
1514                         printk(KERN_WARNING
1515                                "Unable to handle swap header version %d\n",
1516                                swap_header->info.version);
1517                         error = -EINVAL;
1518                         goto bad_swap;
1519                 }
1520
1521                 p->lowest_bit  = 1;
1522                 p->cluster_next = 1;
1523
1524                 /*
1525                  * Find out how many pages are allowed for a single swap
1526                  * device. There are two limiting factors: 1) the number of
1527                  * bits for the swap offset in the swp_entry_t type and
1528                  * 2) the number of bits in the a swap pte as defined by
1529                  * the different architectures. In order to find the
1530                  * largest possible bit mask a swap entry with swap type 0
1531                  * and swap offset ~0UL is created, encoded to a swap pte,
1532                  * decoded to a swp_entry_t again and finally the swap
1533                  * offset is extracted. This will mask all the bits from
1534                  * the initial ~0UL mask that can't be encoded in either
1535                  * the swp_entry_t or the architecture definition of a
1536                  * swap pte.
1537                  */
1538                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1539                 if (maxpages > swap_header->info.last_page)
1540                         maxpages = swap_header->info.last_page;
1541                 p->highest_bit = maxpages - 1;
1542
1543                 error = -EINVAL;
1544                 if (!maxpages)
1545                         goto bad_swap;
1546                 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1547                         goto bad_swap;
1548                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1549                         goto bad_swap;
1550
1551                 /* OK, set up the swap map and apply the bad block list */
1552                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1553                         error = -ENOMEM;
1554                         goto bad_swap;
1555                 }
1556
1557                 error = 0;
1558                 memset(p->swap_map, 0, maxpages * sizeof(short));
1559                 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1560                         int page_nr = swap_header->info.badpages[i];
1561                         if (page_nr <= 0 || page_nr >= swap_header->info.last_page)
1562                                 error = -EINVAL;
1563                         else
1564                                 p->swap_map[page_nr] = SWAP_MAP_BAD;
1565                 }
1566                 nr_good_pages = swap_header->info.last_page -
1567                                 swap_header->info.nr_badpages -
1568                                 1 /* header page */;
1569                 if (error)
1570                         goto bad_swap;
1571         }
1572
1573         if (swapfilesize && maxpages > swapfilesize) {
1574                 printk(KERN_WARNING
1575                        "Swap area shorter than signature indicates\n");
1576                 error = -EINVAL;
1577                 goto bad_swap;
1578         }
1579         if (nr_good_pages) {
1580                 p->swap_map[0] = SWAP_MAP_BAD;
1581                 p->max = maxpages;
1582                 p->pages = nr_good_pages;
1583                 nr_extents = setup_swap_extents(p, &span);
1584                 if (nr_extents < 0) {
1585                         error = nr_extents;
1586                         goto bad_swap;
1587                 }
1588                 nr_good_pages = p->pages;
1589         }
1590         if (!nr_good_pages) {
1591                 printk(KERN_WARNING "Empty swap-file\n");
1592                 error = -EINVAL;
1593                 goto bad_swap;
1594         }
1595
1596         mutex_lock(&swapon_mutex);
1597         spin_lock(&swap_lock);
1598         p->flags = SWP_ACTIVE;
1599         nr_swap_pages += nr_good_pages;
1600         total_swap_pages += nr_good_pages;
1601
1602         printk(KERN_INFO "Adding %uk swap on %s.  "
1603                         "Priority:%d extents:%d across:%lluk\n",
1604                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1605                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1606
1607         /* insert swap space into swap_list: */
1608         prev = -1;
1609         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1610                 if (p->prio >= swap_info[i].prio) {
1611                         break;
1612                 }
1613                 prev = i;
1614         }
1615         p->next = i;
1616         if (prev < 0) {
1617                 swap_list.head = swap_list.next = p - swap_info;
1618         } else {
1619                 swap_info[prev].next = p - swap_info;
1620         }
1621         spin_unlock(&swap_lock);
1622         mutex_unlock(&swapon_mutex);
1623         error = 0;
1624         goto out;
1625 bad_swap:
1626         if (bdev) {
1627                 set_blocksize(bdev, p->old_block_size);
1628                 bd_release(bdev);
1629         }
1630         destroy_swap_extents(p);
1631 bad_swap_2:
1632         spin_lock(&swap_lock);
1633         swap_map = p->swap_map;
1634         p->swap_file = NULL;
1635         p->swap_map = NULL;
1636         p->flags = 0;
1637         if (!(swap_flags & SWAP_FLAG_PREFER))
1638                 ++least_priority;
1639         spin_unlock(&swap_lock);
1640         vfree(swap_map);
1641         if (swap_file)
1642                 filp_close(swap_file, NULL);
1643 out:
1644         if (page && !IS_ERR(page)) {
1645                 kunmap(page);
1646                 page_cache_release(page);
1647         }
1648         if (name)
1649                 putname(name);
1650         if (did_down) {
1651                 if (!error)
1652                         inode->i_flags |= S_SWAPFILE;
1653                 mutex_unlock(&inode->i_mutex);
1654         }
1655         return error;
1656 }
1657
1658 void si_swapinfo(struct sysinfo *val)
1659 {
1660         unsigned int i;
1661         unsigned long nr_to_be_unused = 0;
1662
1663         spin_lock(&swap_lock);
1664         for (i = 0; i < nr_swapfiles; i++) {
1665                 if (!(swap_info[i].flags & SWP_USED) ||
1666                      (swap_info[i].flags & SWP_WRITEOK))
1667                         continue;
1668                 nr_to_be_unused += swap_info[i].inuse_pages;
1669         }
1670         val->freeswap = nr_swap_pages + nr_to_be_unused;
1671         val->totalswap = total_swap_pages + nr_to_be_unused;
1672         spin_unlock(&swap_lock);
1673 }
1674
1675 /*
1676  * Verify that a swap entry is valid and increment its swap map count.
1677  *
1678  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1679  * "permanent", but will be reclaimed by the next swapoff.
1680  */
1681 int swap_duplicate(swp_entry_t entry)
1682 {
1683         struct swap_info_struct * p;
1684         unsigned long offset, type;
1685         int result = 0;
1686
1687         if (is_migration_entry(entry))
1688                 return 1;
1689
1690         type = swp_type(entry);
1691         if (type >= nr_swapfiles)
1692                 goto bad_file;
1693         p = type + swap_info;
1694         offset = swp_offset(entry);
1695
1696         spin_lock(&swap_lock);
1697         if (offset < p->max && p->swap_map[offset]) {
1698                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1699                         p->swap_map[offset]++;
1700                         result = 1;
1701                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1702                         if (swap_overflow++ < 5)
1703                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1704                         p->swap_map[offset] = SWAP_MAP_MAX;
1705                         result = 1;
1706                 }
1707         }
1708         spin_unlock(&swap_lock);
1709 out:
1710         return result;
1711
1712 bad_file:
1713         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1714         goto out;
1715 }
1716
1717 struct swap_info_struct *
1718 get_swap_info_struct(unsigned type)
1719 {
1720         return &swap_info[type];
1721 }
1722
1723 /*
1724  * swap_lock prevents swap_map being freed. Don't grab an extra
1725  * reference on the swaphandle, it doesn't matter if it becomes unused.
1726  */
1727 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1728 {
1729         int ret = 0, i = 1 << page_cluster;
1730         unsigned long toff;
1731         struct swap_info_struct *swapdev = swp_type(entry) + swap_info;
1732
1733         if (!page_cluster)      /* no readahead */
1734                 return 0;
1735         toff = (swp_offset(entry) >> page_cluster) << page_cluster;
1736         if (!toff)              /* first page is swap header */
1737                 toff++, i--;
1738         *offset = toff;
1739
1740         spin_lock(&swap_lock);
1741         do {
1742                 /* Don't read-ahead past the end of the swap area */
1743                 if (toff >= swapdev->max)
1744                         break;
1745                 /* Don't read in free or bad pages */
1746                 if (!swapdev->swap_map[toff])
1747                         break;
1748                 if (swapdev->swap_map[toff] == SWAP_MAP_BAD)
1749                         break;
1750                 toff++;
1751                 ret++;
1752         } while (--i);
1753         spin_unlock(&swap_lock);
1754         return ret;
1755 }