swapfile: remove surplus whitespace
[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/mm.h>
9 #include <linux/hugetlb.h>
10 #include <linux/mman.h>
11 #include <linux/slab.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/swap.h>
14 #include <linux/vmalloc.h>
15 #include <linux/pagemap.h>
16 #include <linux/namei.h>
17 #include <linux/shm.h>
18 #include <linux/blkdev.h>
19 #include <linux/writeback.h>
20 #include <linux/proc_fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/init.h>
23 #include <linux/module.h>
24 #include <linux/rmap.h>
25 #include <linux/security.h>
26 #include <linux/backing-dev.h>
27 #include <linux/mutex.h>
28 #include <linux/capability.h>
29 #include <linux/syscalls.h>
30 #include <linux/memcontrol.h>
31
32 #include <asm/pgtable.h>
33 #include <asm/tlbflush.h>
34 #include <linux/swapops.h>
35
36 static DEFINE_SPINLOCK(swap_lock);
37 static unsigned int nr_swapfiles;
38 long nr_swap_pages;
39 long total_swap_pages;
40 static int swap_overflow;
41 static int least_priority;
42
43 static const char Bad_file[] = "Bad swap file entry ";
44 static const char Unused_file[] = "Unused swap file entry ";
45 static const char Bad_offset[] = "Bad swap offset entry ";
46 static const char Unused_offset[] = "Unused swap offset entry ";
47
48 static struct swap_list_t swap_list = {-1, -1};
49
50 static struct swap_info_struct swap_info[MAX_SWAPFILES];
51
52 static DEFINE_MUTEX(swapon_mutex);
53
54 /*
55  * We need this because the bdev->unplug_fn can sleep and we cannot
56  * hold swap_lock while calling the unplug_fn. And swap_lock
57  * cannot be turned into a mutex.
58  */
59 static DECLARE_RWSEM(swap_unplug_sem);
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(page);
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(page) above.
75                  * If 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 #define SWAPFILE_CLUSTER        256
88 #define LATENCY_LIMIT           256
89
90 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
91 {
92         unsigned long offset, last_in_cluster;
93         int latency_ration = LATENCY_LIMIT;
94
95         /*
96          * We try to cluster swap pages by allocating them sequentially
97          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
98          * way, however, we resort to first-free allocation, starting
99          * a new cluster.  This prevents us from scattering swap pages
100          * all over the entire swap partition, so that we reduce
101          * overall disk seek times between swap pages.  -- sct
102          * But we do now try to find an empty cluster.  -Andrea
103          */
104
105         si->flags += SWP_SCANNING;
106         if (unlikely(!si->cluster_nr)) {
107                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
108                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER)
109                         goto lowest;
110                 spin_unlock(&swap_lock);
111
112                 offset = si->lowest_bit;
113                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
114
115                 /* Locate the first empty (unaligned) cluster */
116                 for (; last_in_cluster <= si->highest_bit; offset++) {
117                         if (si->swap_map[offset])
118                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
119                         else if (offset == last_in_cluster) {
120                                 spin_lock(&swap_lock);
121                                 si->cluster_next = offset-SWAPFILE_CLUSTER+1;
122                                 goto cluster;
123                         }
124                         if (unlikely(--latency_ration < 0)) {
125                                 cond_resched();
126                                 latency_ration = LATENCY_LIMIT;
127                         }
128                 }
129                 spin_lock(&swap_lock);
130                 goto lowest;
131         }
132
133         si->cluster_nr--;
134 cluster:
135         offset = si->cluster_next;
136         if (offset > si->highest_bit)
137 lowest:         offset = si->lowest_bit;
138 checks: if (!(si->flags & SWP_WRITEOK))
139                 goto no_page;
140         if (!si->highest_bit)
141                 goto no_page;
142         if (!si->swap_map[offset]) {
143                 if (offset == si->lowest_bit)
144                         si->lowest_bit++;
145                 if (offset == si->highest_bit)
146                         si->highest_bit--;
147                 si->inuse_pages++;
148                 if (si->inuse_pages == si->pages) {
149                         si->lowest_bit = si->max;
150                         si->highest_bit = 0;
151                 }
152                 si->swap_map[offset] = 1;
153                 si->cluster_next = offset + 1;
154                 si->flags -= SWP_SCANNING;
155                 return offset;
156         }
157
158         spin_unlock(&swap_lock);
159         while (++offset <= si->highest_bit) {
160                 if (!si->swap_map[offset]) {
161                         spin_lock(&swap_lock);
162                         goto checks;
163                 }
164                 if (unlikely(--latency_ration < 0)) {
165                         cond_resched();
166                         latency_ration = LATENCY_LIMIT;
167                 }
168         }
169         spin_lock(&swap_lock);
170         goto lowest;
171
172 no_page:
173         si->flags -= SWP_SCANNING;
174         return 0;
175 }
176
177 swp_entry_t get_swap_page(void)
178 {
179         struct swap_info_struct *si;
180         pgoff_t offset;
181         int type, next;
182         int wrapped = 0;
183
184         spin_lock(&swap_lock);
185         if (nr_swap_pages <= 0)
186                 goto noswap;
187         nr_swap_pages--;
188
189         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
190                 si = swap_info + type;
191                 next = si->next;
192                 if (next < 0 ||
193                     (!wrapped && si->prio != swap_info[next].prio)) {
194                         next = swap_list.head;
195                         wrapped++;
196                 }
197
198                 if (!si->highest_bit)
199                         continue;
200                 if (!(si->flags & SWP_WRITEOK))
201                         continue;
202
203                 swap_list.next = next;
204                 offset = scan_swap_map(si);
205                 if (offset) {
206                         spin_unlock(&swap_lock);
207                         return swp_entry(type, offset);
208                 }
209                 next = swap_list.next;
210         }
211
212         nr_swap_pages++;
213 noswap:
214         spin_unlock(&swap_lock);
215         return (swp_entry_t) {0};
216 }
217
218 swp_entry_t get_swap_page_of_type(int type)
219 {
220         struct swap_info_struct *si;
221         pgoff_t offset;
222
223         spin_lock(&swap_lock);
224         si = swap_info + type;
225         if (si->flags & SWP_WRITEOK) {
226                 nr_swap_pages--;
227                 offset = scan_swap_map(si);
228                 if (offset) {
229                         spin_unlock(&swap_lock);
230                         return swp_entry(type, offset);
231                 }
232                 nr_swap_pages++;
233         }
234         spin_unlock(&swap_lock);
235         return (swp_entry_t) {0};
236 }
237
238 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
239 {
240         struct swap_info_struct * p;
241         unsigned long offset, type;
242
243         if (!entry.val)
244                 goto out;
245         type = swp_type(entry);
246         if (type >= nr_swapfiles)
247                 goto bad_nofile;
248         p = & swap_info[type];
249         if (!(p->flags & SWP_USED))
250                 goto bad_device;
251         offset = swp_offset(entry);
252         if (offset >= p->max)
253                 goto bad_offset;
254         if (!p->swap_map[offset])
255                 goto bad_free;
256         spin_lock(&swap_lock);
257         return p;
258
259 bad_free:
260         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
261         goto out;
262 bad_offset:
263         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
264         goto out;
265 bad_device:
266         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
267         goto out;
268 bad_nofile:
269         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
270 out:
271         return NULL;
272 }
273
274 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
275 {
276         int count = p->swap_map[offset];
277
278         if (count < SWAP_MAP_MAX) {
279                 count--;
280                 p->swap_map[offset] = count;
281                 if (!count) {
282                         if (offset < p->lowest_bit)
283                                 p->lowest_bit = offset;
284                         if (offset > p->highest_bit)
285                                 p->highest_bit = offset;
286                         if (p->prio > swap_info[swap_list.next].prio)
287                                 swap_list.next = p - swap_info;
288                         nr_swap_pages++;
289                         p->inuse_pages--;
290                 }
291         }
292         return count;
293 }
294
295 /*
296  * Caller has made sure that the swapdevice corresponding to entry
297  * is still around or has not been recycled.
298  */
299 void swap_free(swp_entry_t entry)
300 {
301         struct swap_info_struct * p;
302
303         p = swap_info_get(entry);
304         if (p) {
305                 swap_entry_free(p, swp_offset(entry));
306                 spin_unlock(&swap_lock);
307         }
308 }
309
310 /*
311  * How many references to page are currently swapped out?
312  */
313 static inline int page_swapcount(struct page *page)
314 {
315         int count = 0;
316         struct swap_info_struct *p;
317         swp_entry_t entry;
318
319         entry.val = page_private(page);
320         p = swap_info_get(entry);
321         if (p) {
322                 /* Subtract the 1 for the swap cache itself */
323                 count = p->swap_map[swp_offset(entry)] - 1;
324                 spin_unlock(&swap_lock);
325         }
326         return count;
327 }
328
329 /*
330  * We can write to an anon page without COW if there are no other references
331  * to it.  And as a side-effect, free up its swap: because the old content
332  * on disk will never be read, and seeking back there to write new content
333  * later would only waste time away from clustering.
334  */
335 int reuse_swap_page(struct page *page)
336 {
337         int count;
338
339         VM_BUG_ON(!PageLocked(page));
340         count = page_mapcount(page);
341         if (count <= 1 && PageSwapCache(page)) {
342                 count += page_swapcount(page);
343                 if (count == 1 && !PageWriteback(page)) {
344                         delete_from_swap_cache(page);
345                         SetPageDirty(page);
346                 }
347         }
348         return count == 1;
349 }
350
351 /*
352  * If swap is getting full, or if there are no more mappings of this page,
353  * then try_to_free_swap is called to free its swap space.
354  */
355 int try_to_free_swap(struct page *page)
356 {
357         VM_BUG_ON(!PageLocked(page));
358
359         if (!PageSwapCache(page))
360                 return 0;
361         if (PageWriteback(page))
362                 return 0;
363         if (page_swapcount(page))
364                 return 0;
365
366         delete_from_swap_cache(page);
367         SetPageDirty(page);
368         return 1;
369 }
370
371 /*
372  * Free the swap entry like above, but also try to
373  * free the page cache entry if it is the last user.
374  */
375 void free_swap_and_cache(swp_entry_t entry)
376 {
377         struct swap_info_struct * p;
378         struct page *page = NULL;
379
380         if (is_migration_entry(entry))
381                 return;
382
383         p = swap_info_get(entry);
384         if (p) {
385                 if (swap_entry_free(p, swp_offset(entry)) == 1) {
386                         page = find_get_page(&swapper_space, entry.val);
387                         if (page && !trylock_page(page)) {
388                                 page_cache_release(page);
389                                 page = NULL;
390                         }
391                 }
392                 spin_unlock(&swap_lock);
393         }
394         if (page) {
395                 /*
396                  * Not mapped elsewhere, or swap space full? Free it!
397                  * Also recheck PageSwapCache now page is locked (above).
398                  */
399                 if (PageSwapCache(page) && !PageWriteback(page) &&
400                                 (!page_mapped(page) || vm_swap_full())) {
401                         delete_from_swap_cache(page);
402                         SetPageDirty(page);
403                 }
404                 unlock_page(page);
405                 page_cache_release(page);
406         }
407 }
408
409 #ifdef CONFIG_HIBERNATION
410 /*
411  * Find the swap type that corresponds to given device (if any).
412  *
413  * @offset - number of the PAGE_SIZE-sized block of the device, starting
414  * from 0, in which the swap header is expected to be located.
415  *
416  * This is needed for the suspend to disk (aka swsusp).
417  */
418 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
419 {
420         struct block_device *bdev = NULL;
421         int i;
422
423         if (device)
424                 bdev = bdget(device);
425
426         spin_lock(&swap_lock);
427         for (i = 0; i < nr_swapfiles; i++) {
428                 struct swap_info_struct *sis = swap_info + i;
429
430                 if (!(sis->flags & SWP_WRITEOK))
431                         continue;
432
433                 if (!bdev) {
434                         if (bdev_p)
435                                 *bdev_p = sis->bdev;
436
437                         spin_unlock(&swap_lock);
438                         return i;
439                 }
440                 if (bdev == sis->bdev) {
441                         struct swap_extent *se;
442
443                         se = list_entry(sis->extent_list.next,
444                                         struct swap_extent, list);
445                         if (se->start_block == offset) {
446                                 if (bdev_p)
447                                         *bdev_p = sis->bdev;
448
449                                 spin_unlock(&swap_lock);
450                                 bdput(bdev);
451                                 return i;
452                         }
453                 }
454         }
455         spin_unlock(&swap_lock);
456         if (bdev)
457                 bdput(bdev);
458
459         return -ENODEV;
460 }
461
462 /*
463  * Return either the total number of swap pages of given type, or the number
464  * of free pages of that type (depending on @free)
465  *
466  * This is needed for software suspend
467  */
468 unsigned int count_swap_pages(int type, int free)
469 {
470         unsigned int n = 0;
471
472         if (type < nr_swapfiles) {
473                 spin_lock(&swap_lock);
474                 if (swap_info[type].flags & SWP_WRITEOK) {
475                         n = swap_info[type].pages;
476                         if (free)
477                                 n -= swap_info[type].inuse_pages;
478                 }
479                 spin_unlock(&swap_lock);
480         }
481         return n;
482 }
483 #endif
484
485 /*
486  * No need to decide whether this PTE shares the swap entry with others,
487  * just let do_wp_page work it out if a write is requested later - to
488  * force COW, vm_page_prot omits write permission from any private vma.
489  */
490 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
491                 unsigned long addr, swp_entry_t entry, struct page *page)
492 {
493         spinlock_t *ptl;
494         pte_t *pte;
495         int ret = 1;
496
497         if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
498                 ret = -ENOMEM;
499
500         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
501         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
502                 if (ret > 0)
503                         mem_cgroup_uncharge_page(page);
504                 ret = 0;
505                 goto out;
506         }
507
508         inc_mm_counter(vma->vm_mm, anon_rss);
509         get_page(page);
510         set_pte_at(vma->vm_mm, addr, pte,
511                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
512         page_add_anon_rmap(page, vma, addr);
513         swap_free(entry);
514         /*
515          * Move the page to the active list so it is not
516          * immediately swapped out again after swapon.
517          */
518         activate_page(page);
519 out:
520         pte_unmap_unlock(pte, ptl);
521         return ret;
522 }
523
524 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
525                                 unsigned long addr, unsigned long end,
526                                 swp_entry_t entry, struct page *page)
527 {
528         pte_t swp_pte = swp_entry_to_pte(entry);
529         pte_t *pte;
530         int ret = 0;
531
532         /*
533          * We don't actually need pte lock while scanning for swp_pte: since
534          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
535          * page table while we're scanning; though it could get zapped, and on
536          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
537          * of unmatched parts which look like swp_pte, so unuse_pte must
538          * recheck under pte lock.  Scanning without pte lock lets it be
539          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
540          */
541         pte = pte_offset_map(pmd, addr);
542         do {
543                 /*
544                  * swapoff spends a _lot_ of time in this loop!
545                  * Test inline before going to call unuse_pte.
546                  */
547                 if (unlikely(pte_same(*pte, swp_pte))) {
548                         pte_unmap(pte);
549                         ret = unuse_pte(vma, pmd, addr, entry, page);
550                         if (ret)
551                                 goto out;
552                         pte = pte_offset_map(pmd, addr);
553                 }
554         } while (pte++, addr += PAGE_SIZE, addr != end);
555         pte_unmap(pte - 1);
556 out:
557         return ret;
558 }
559
560 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
561                                 unsigned long addr, unsigned long end,
562                                 swp_entry_t entry, struct page *page)
563 {
564         pmd_t *pmd;
565         unsigned long next;
566         int ret;
567
568         pmd = pmd_offset(pud, addr);
569         do {
570                 next = pmd_addr_end(addr, end);
571                 if (pmd_none_or_clear_bad(pmd))
572                         continue;
573                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
574                 if (ret)
575                         return ret;
576         } while (pmd++, addr = next, addr != end);
577         return 0;
578 }
579
580 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
581                                 unsigned long addr, unsigned long end,
582                                 swp_entry_t entry, struct page *page)
583 {
584         pud_t *pud;
585         unsigned long next;
586         int ret;
587
588         pud = pud_offset(pgd, addr);
589         do {
590                 next = pud_addr_end(addr, end);
591                 if (pud_none_or_clear_bad(pud))
592                         continue;
593                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
594                 if (ret)
595                         return ret;
596         } while (pud++, addr = next, addr != end);
597         return 0;
598 }
599
600 static int unuse_vma(struct vm_area_struct *vma,
601                                 swp_entry_t entry, struct page *page)
602 {
603         pgd_t *pgd;
604         unsigned long addr, end, next;
605         int ret;
606
607         if (page->mapping) {
608                 addr = page_address_in_vma(page, vma);
609                 if (addr == -EFAULT)
610                         return 0;
611                 else
612                         end = addr + PAGE_SIZE;
613         } else {
614                 addr = vma->vm_start;
615                 end = vma->vm_end;
616         }
617
618         pgd = pgd_offset(vma->vm_mm, addr);
619         do {
620                 next = pgd_addr_end(addr, end);
621                 if (pgd_none_or_clear_bad(pgd))
622                         continue;
623                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
624                 if (ret)
625                         return ret;
626         } while (pgd++, addr = next, addr != end);
627         return 0;
628 }
629
630 static int unuse_mm(struct mm_struct *mm,
631                                 swp_entry_t entry, struct page *page)
632 {
633         struct vm_area_struct *vma;
634         int ret = 0;
635
636         if (!down_read_trylock(&mm->mmap_sem)) {
637                 /*
638                  * Activate page so shrink_inactive_list is unlikely to unmap
639                  * its ptes while lock is dropped, so swapoff can make progress.
640                  */
641                 activate_page(page);
642                 unlock_page(page);
643                 down_read(&mm->mmap_sem);
644                 lock_page(page);
645         }
646         for (vma = mm->mmap; vma; vma = vma->vm_next) {
647                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
648                         break;
649         }
650         up_read(&mm->mmap_sem);
651         return (ret < 0)? ret: 0;
652 }
653
654 /*
655  * Scan swap_map from current position to next entry still in use.
656  * Recycle to start on reaching the end, returning 0 when empty.
657  */
658 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
659                                         unsigned int prev)
660 {
661         unsigned int max = si->max;
662         unsigned int i = prev;
663         int count;
664
665         /*
666          * No need for swap_lock here: we're just looking
667          * for whether an entry is in use, not modifying it; false
668          * hits are okay, and sys_swapoff() has already prevented new
669          * allocations from this area (while holding swap_lock).
670          */
671         for (;;) {
672                 if (++i >= max) {
673                         if (!prev) {
674                                 i = 0;
675                                 break;
676                         }
677                         /*
678                          * No entries in use at top of swap_map,
679                          * loop back to start and recheck there.
680                          */
681                         max = prev + 1;
682                         prev = 0;
683                         i = 1;
684                 }
685                 count = si->swap_map[i];
686                 if (count && count != SWAP_MAP_BAD)
687                         break;
688         }
689         return i;
690 }
691
692 /*
693  * We completely avoid races by reading each swap page in advance,
694  * and then search for the process using it.  All the necessary
695  * page table adjustments can then be made atomically.
696  */
697 static int try_to_unuse(unsigned int type)
698 {
699         struct swap_info_struct * si = &swap_info[type];
700         struct mm_struct *start_mm;
701         unsigned short *swap_map;
702         unsigned short swcount;
703         struct page *page;
704         swp_entry_t entry;
705         unsigned int i = 0;
706         int retval = 0;
707         int reset_overflow = 0;
708         int shmem;
709
710         /*
711          * When searching mms for an entry, a good strategy is to
712          * start at the first mm we freed the previous entry from
713          * (though actually we don't notice whether we or coincidence
714          * freed the entry).  Initialize this start_mm with a hold.
715          *
716          * A simpler strategy would be to start at the last mm we
717          * freed the previous entry from; but that would take less
718          * advantage of mmlist ordering, which clusters forked mms
719          * together, child after parent.  If we race with dup_mmap(), we
720          * prefer to resolve parent before child, lest we miss entries
721          * duplicated after we scanned child: using last mm would invert
722          * that.  Though it's only a serious concern when an overflowed
723          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
724          */
725         start_mm = &init_mm;
726         atomic_inc(&init_mm.mm_users);
727
728         /*
729          * Keep on scanning until all entries have gone.  Usually,
730          * one pass through swap_map is enough, but not necessarily:
731          * there are races when an instance of an entry might be missed.
732          */
733         while ((i = find_next_to_unuse(si, i)) != 0) {
734                 if (signal_pending(current)) {
735                         retval = -EINTR;
736                         break;
737                 }
738
739                 /*
740                  * Get a page for the entry, using the existing swap
741                  * cache page if there is one.  Otherwise, get a clean
742                  * page and read the swap into it.
743                  */
744                 swap_map = &si->swap_map[i];
745                 entry = swp_entry(type, i);
746                 page = read_swap_cache_async(entry,
747                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
748                 if (!page) {
749                         /*
750                          * Either swap_duplicate() failed because entry
751                          * has been freed independently, and will not be
752                          * reused since sys_swapoff() already disabled
753                          * allocation from here, or alloc_page() failed.
754                          */
755                         if (!*swap_map)
756                                 continue;
757                         retval = -ENOMEM;
758                         break;
759                 }
760
761                 /*
762                  * Don't hold on to start_mm if it looks like exiting.
763                  */
764                 if (atomic_read(&start_mm->mm_users) == 1) {
765                         mmput(start_mm);
766                         start_mm = &init_mm;
767                         atomic_inc(&init_mm.mm_users);
768                 }
769
770                 /*
771                  * Wait for and lock page.  When do_swap_page races with
772                  * try_to_unuse, do_swap_page can handle the fault much
773                  * faster than try_to_unuse can locate the entry.  This
774                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
775                  * defer to do_swap_page in such a case - in some tests,
776                  * do_swap_page and try_to_unuse repeatedly compete.
777                  */
778                 wait_on_page_locked(page);
779                 wait_on_page_writeback(page);
780                 lock_page(page);
781                 wait_on_page_writeback(page);
782
783                 /*
784                  * Remove all references to entry.
785                  * Whenever we reach init_mm, there's no address space
786                  * to search, but use it as a reminder to search shmem.
787                  */
788                 shmem = 0;
789                 swcount = *swap_map;
790                 if (swcount > 1) {
791                         if (start_mm == &init_mm)
792                                 shmem = shmem_unuse(entry, page);
793                         else
794                                 retval = unuse_mm(start_mm, entry, page);
795                 }
796                 if (*swap_map > 1) {
797                         int set_start_mm = (*swap_map >= swcount);
798                         struct list_head *p = &start_mm->mmlist;
799                         struct mm_struct *new_start_mm = start_mm;
800                         struct mm_struct *prev_mm = start_mm;
801                         struct mm_struct *mm;
802
803                         atomic_inc(&new_start_mm->mm_users);
804                         atomic_inc(&prev_mm->mm_users);
805                         spin_lock(&mmlist_lock);
806                         while (*swap_map > 1 && !retval && !shmem &&
807                                         (p = p->next) != &start_mm->mmlist) {
808                                 mm = list_entry(p, struct mm_struct, mmlist);
809                                 if (!atomic_inc_not_zero(&mm->mm_users))
810                                         continue;
811                                 spin_unlock(&mmlist_lock);
812                                 mmput(prev_mm);
813                                 prev_mm = mm;
814
815                                 cond_resched();
816
817                                 swcount = *swap_map;
818                                 if (swcount <= 1)
819                                         ;
820                                 else if (mm == &init_mm) {
821                                         set_start_mm = 1;
822                                         shmem = shmem_unuse(entry, page);
823                                 } else
824                                         retval = unuse_mm(mm, entry, page);
825                                 if (set_start_mm && *swap_map < swcount) {
826                                         mmput(new_start_mm);
827                                         atomic_inc(&mm->mm_users);
828                                         new_start_mm = mm;
829                                         set_start_mm = 0;
830                                 }
831                                 spin_lock(&mmlist_lock);
832                         }
833                         spin_unlock(&mmlist_lock);
834                         mmput(prev_mm);
835                         mmput(start_mm);
836                         start_mm = new_start_mm;
837                 }
838                 if (shmem) {
839                         /* page has already been unlocked and released */
840                         if (shmem > 0)
841                                 continue;
842                         retval = shmem;
843                         break;
844                 }
845                 if (retval) {
846                         unlock_page(page);
847                         page_cache_release(page);
848                         break;
849                 }
850
851                 /*
852                  * How could swap count reach 0x7fff when the maximum
853                  * pid is 0x7fff, and there's no way to repeat a swap
854                  * page within an mm (except in shmem, where it's the
855                  * shared object which takes the reference count)?
856                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
857                  *
858                  * If that's wrong, then we should worry more about
859                  * exit_mmap() and do_munmap() cases described above:
860                  * we might be resetting SWAP_MAP_MAX too early here.
861                  * We know "Undead"s can happen, they're okay, so don't
862                  * report them; but do report if we reset SWAP_MAP_MAX.
863                  */
864                 if (*swap_map == SWAP_MAP_MAX) {
865                         spin_lock(&swap_lock);
866                         *swap_map = 1;
867                         spin_unlock(&swap_lock);
868                         reset_overflow = 1;
869                 }
870
871                 /*
872                  * If a reference remains (rare), we would like to leave
873                  * the page in the swap cache; but try_to_unmap could
874                  * then re-duplicate the entry once we drop page lock,
875                  * so we might loop indefinitely; also, that page could
876                  * not be swapped out to other storage meanwhile.  So:
877                  * delete from cache even if there's another reference,
878                  * after ensuring that the data has been saved to disk -
879                  * since if the reference remains (rarer), it will be
880                  * read from disk into another page.  Splitting into two
881                  * pages would be incorrect if swap supported "shared
882                  * private" pages, but they are handled by tmpfs files.
883                  */
884                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
885                         struct writeback_control wbc = {
886                                 .sync_mode = WB_SYNC_NONE,
887                         };
888
889                         swap_writepage(page, &wbc);
890                         lock_page(page);
891                         wait_on_page_writeback(page);
892                 }
893
894                 /*
895                  * It is conceivable that a racing task removed this page from
896                  * swap cache just before we acquired the page lock at the top,
897                  * or while we dropped it in unuse_mm().  The page might even
898                  * be back in swap cache on another swap area: that we must not
899                  * delete, since it may not have been written out to swap yet.
900                  */
901                 if (PageSwapCache(page) &&
902                     likely(page_private(page) == entry.val))
903                         delete_from_swap_cache(page);
904
905                 /*
906                  * So we could skip searching mms once swap count went
907                  * to 1, we did not mark any present ptes as dirty: must
908                  * mark page dirty so shrink_page_list will preserve it.
909                  */
910                 SetPageDirty(page);
911                 unlock_page(page);
912                 page_cache_release(page);
913
914                 /*
915                  * Make sure that we aren't completely killing
916                  * interactive performance.
917                  */
918                 cond_resched();
919         }
920
921         mmput(start_mm);
922         if (reset_overflow) {
923                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
924                 swap_overflow = 0;
925         }
926         return retval;
927 }
928
929 /*
930  * After a successful try_to_unuse, if no swap is now in use, we know
931  * we can empty the mmlist.  swap_lock must be held on entry and exit.
932  * Note that mmlist_lock nests inside swap_lock, and an mm must be
933  * added to the mmlist just after page_duplicate - before would be racy.
934  */
935 static void drain_mmlist(void)
936 {
937         struct list_head *p, *next;
938         unsigned int i;
939
940         for (i = 0; i < nr_swapfiles; i++)
941                 if (swap_info[i].inuse_pages)
942                         return;
943         spin_lock(&mmlist_lock);
944         list_for_each_safe(p, next, &init_mm.mmlist)
945                 list_del_init(p);
946         spin_unlock(&mmlist_lock);
947 }
948
949 /*
950  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
951  * corresponds to page offset `offset'.
952  */
953 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
954 {
955         struct swap_extent *se = sis->curr_swap_extent;
956         struct swap_extent *start_se = se;
957
958         for ( ; ; ) {
959                 struct list_head *lh;
960
961                 if (se->start_page <= offset &&
962                                 offset < (se->start_page + se->nr_pages)) {
963                         return se->start_block + (offset - se->start_page);
964                 }
965                 lh = se->list.next;
966                 if (lh == &sis->extent_list)
967                         lh = lh->next;
968                 se = list_entry(lh, struct swap_extent, list);
969                 sis->curr_swap_extent = se;
970                 BUG_ON(se == start_se);         /* It *must* be present */
971         }
972 }
973
974 #ifdef CONFIG_HIBERNATION
975 /*
976  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
977  * corresponding to given index in swap_info (swap type).
978  */
979 sector_t swapdev_block(int swap_type, pgoff_t offset)
980 {
981         struct swap_info_struct *sis;
982
983         if (swap_type >= nr_swapfiles)
984                 return 0;
985
986         sis = swap_info + swap_type;
987         return (sis->flags & SWP_WRITEOK) ? map_swap_page(sis, offset) : 0;
988 }
989 #endif /* CONFIG_HIBERNATION */
990
991 /*
992  * Free all of a swapdev's extent information
993  */
994 static void destroy_swap_extents(struct swap_info_struct *sis)
995 {
996         while (!list_empty(&sis->extent_list)) {
997                 struct swap_extent *se;
998
999                 se = list_entry(sis->extent_list.next,
1000                                 struct swap_extent, list);
1001                 list_del(&se->list);
1002                 kfree(se);
1003         }
1004 }
1005
1006 /*
1007  * Add a block range (and the corresponding page range) into this swapdev's
1008  * extent list.  The extent list is kept sorted in page order.
1009  *
1010  * This function rather assumes that it is called in ascending page order.
1011  */
1012 static int
1013 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1014                 unsigned long nr_pages, sector_t start_block)
1015 {
1016         struct swap_extent *se;
1017         struct swap_extent *new_se;
1018         struct list_head *lh;
1019
1020         lh = sis->extent_list.prev;     /* The highest page extent */
1021         if (lh != &sis->extent_list) {
1022                 se = list_entry(lh, struct swap_extent, list);
1023                 BUG_ON(se->start_page + se->nr_pages != start_page);
1024                 if (se->start_block + se->nr_pages == start_block) {
1025                         /* Merge it */
1026                         se->nr_pages += nr_pages;
1027                         return 0;
1028                 }
1029         }
1030
1031         /*
1032          * No merge.  Insert a new extent, preserving ordering.
1033          */
1034         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1035         if (new_se == NULL)
1036                 return -ENOMEM;
1037         new_se->start_page = start_page;
1038         new_se->nr_pages = nr_pages;
1039         new_se->start_block = start_block;
1040
1041         list_add_tail(&new_se->list, &sis->extent_list);
1042         return 1;
1043 }
1044
1045 /*
1046  * A `swap extent' is a simple thing which maps a contiguous range of pages
1047  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1048  * is built at swapon time and is then used at swap_writepage/swap_readpage
1049  * time for locating where on disk a page belongs.
1050  *
1051  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1052  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1053  * swap files identically.
1054  *
1055  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1056  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1057  * swapfiles are handled *identically* after swapon time.
1058  *
1059  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1060  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1061  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1062  * requirements, they are simply tossed out - we will never use those blocks
1063  * for swapping.
1064  *
1065  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1066  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1067  * which will scribble on the fs.
1068  *
1069  * The amount of disk space which a single swap extent represents varies.
1070  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1071  * extents in the list.  To avoid much list walking, we cache the previous
1072  * search location in `curr_swap_extent', and start new searches from there.
1073  * This is extremely effective.  The average number of iterations in
1074  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1075  */
1076 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1077 {
1078         struct inode *inode;
1079         unsigned blocks_per_page;
1080         unsigned long page_no;
1081         unsigned blkbits;
1082         sector_t probe_block;
1083         sector_t last_block;
1084         sector_t lowest_block = -1;
1085         sector_t highest_block = 0;
1086         int nr_extents = 0;
1087         int ret;
1088
1089         inode = sis->swap_file->f_mapping->host;
1090         if (S_ISBLK(inode->i_mode)) {
1091                 ret = add_swap_extent(sis, 0, sis->max, 0);
1092                 *span = sis->pages;
1093                 goto done;
1094         }
1095
1096         blkbits = inode->i_blkbits;
1097         blocks_per_page = PAGE_SIZE >> blkbits;
1098
1099         /*
1100          * Map all the blocks into the extent list.  This code doesn't try
1101          * to be very smart.
1102          */
1103         probe_block = 0;
1104         page_no = 0;
1105         last_block = i_size_read(inode) >> blkbits;
1106         while ((probe_block + blocks_per_page) <= last_block &&
1107                         page_no < sis->max) {
1108                 unsigned block_in_page;
1109                 sector_t first_block;
1110
1111                 first_block = bmap(inode, probe_block);
1112                 if (first_block == 0)
1113                         goto bad_bmap;
1114
1115                 /*
1116                  * It must be PAGE_SIZE aligned on-disk
1117                  */
1118                 if (first_block & (blocks_per_page - 1)) {
1119                         probe_block++;
1120                         goto reprobe;
1121                 }
1122
1123                 for (block_in_page = 1; block_in_page < blocks_per_page;
1124                                         block_in_page++) {
1125                         sector_t block;
1126
1127                         block = bmap(inode, probe_block + block_in_page);
1128                         if (block == 0)
1129                                 goto bad_bmap;
1130                         if (block != first_block + block_in_page) {
1131                                 /* Discontiguity */
1132                                 probe_block++;
1133                                 goto reprobe;
1134                         }
1135                 }
1136
1137                 first_block >>= (PAGE_SHIFT - blkbits);
1138                 if (page_no) {  /* exclude the header page */
1139                         if (first_block < lowest_block)
1140                                 lowest_block = first_block;
1141                         if (first_block > highest_block)
1142                                 highest_block = first_block;
1143                 }
1144
1145                 /*
1146                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1147                  */
1148                 ret = add_swap_extent(sis, page_no, 1, first_block);
1149                 if (ret < 0)
1150                         goto out;
1151                 nr_extents += ret;
1152                 page_no++;
1153                 probe_block += blocks_per_page;
1154 reprobe:
1155                 continue;
1156         }
1157         ret = nr_extents;
1158         *span = 1 + highest_block - lowest_block;
1159         if (page_no == 0)
1160                 page_no = 1;    /* force Empty message */
1161         sis->max = page_no;
1162         sis->pages = page_no - 1;
1163         sis->highest_bit = page_no - 1;
1164 done:
1165         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1166                                         struct swap_extent, list);
1167         goto out;
1168 bad_bmap:
1169         printk(KERN_ERR "swapon: swapfile has holes\n");
1170         ret = -EINVAL;
1171 out:
1172         return ret;
1173 }
1174
1175 #if 0   /* We don't need this yet */
1176 #include <linux/backing-dev.h>
1177 int page_queue_congested(struct page *page)
1178 {
1179         struct backing_dev_info *bdi;
1180
1181         VM_BUG_ON(!PageLocked(page));   /* It pins the swap_info_struct */
1182
1183         if (PageSwapCache(page)) {
1184                 swp_entry_t entry = { .val = page_private(page) };
1185                 struct swap_info_struct *sis;
1186
1187                 sis = get_swap_info_struct(swp_type(entry));
1188                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1189         } else
1190                 bdi = page->mapping->backing_dev_info;
1191         return bdi_write_congested(bdi);
1192 }
1193 #endif
1194
1195 asmlinkage long sys_swapoff(const char __user * specialfile)
1196 {
1197         struct swap_info_struct * p = NULL;
1198         unsigned short *swap_map;
1199         struct file *swap_file, *victim;
1200         struct address_space *mapping;
1201         struct inode *inode;
1202         char * pathname;
1203         int i, type, prev;
1204         int err;
1205
1206         if (!capable(CAP_SYS_ADMIN))
1207                 return -EPERM;
1208
1209         pathname = getname(specialfile);
1210         err = PTR_ERR(pathname);
1211         if (IS_ERR(pathname))
1212                 goto out;
1213
1214         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1215         putname(pathname);
1216         err = PTR_ERR(victim);
1217         if (IS_ERR(victim))
1218                 goto out;
1219
1220         mapping = victim->f_mapping;
1221         prev = -1;
1222         spin_lock(&swap_lock);
1223         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1224                 p = swap_info + type;
1225                 if (p->flags & SWP_WRITEOK) {
1226                         if (p->swap_file->f_mapping == mapping)
1227                                 break;
1228                 }
1229                 prev = type;
1230         }
1231         if (type < 0) {
1232                 err = -EINVAL;
1233                 spin_unlock(&swap_lock);
1234                 goto out_dput;
1235         }
1236         if (!security_vm_enough_memory(p->pages))
1237                 vm_unacct_memory(p->pages);
1238         else {
1239                 err = -ENOMEM;
1240                 spin_unlock(&swap_lock);
1241                 goto out_dput;
1242         }
1243         if (prev < 0) {
1244                 swap_list.head = p->next;
1245         } else {
1246                 swap_info[prev].next = p->next;
1247         }
1248         if (type == swap_list.next) {
1249                 /* just pick something that's safe... */
1250                 swap_list.next = swap_list.head;
1251         }
1252         if (p->prio < 0) {
1253                 for (i = p->next; i >= 0; i = swap_info[i].next)
1254                         swap_info[i].prio = p->prio--;
1255                 least_priority++;
1256         }
1257         nr_swap_pages -= p->pages;
1258         total_swap_pages -= p->pages;
1259         p->flags &= ~SWP_WRITEOK;
1260         spin_unlock(&swap_lock);
1261
1262         current->flags |= PF_SWAPOFF;
1263         err = try_to_unuse(type);
1264         current->flags &= ~PF_SWAPOFF;
1265
1266         if (err) {
1267                 /* re-insert swap space back into swap_list */
1268                 spin_lock(&swap_lock);
1269                 if (p->prio < 0)
1270                         p->prio = --least_priority;
1271                 prev = -1;
1272                 for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1273                         if (p->prio >= swap_info[i].prio)
1274                                 break;
1275                         prev = i;
1276                 }
1277                 p->next = i;
1278                 if (prev < 0)
1279                         swap_list.head = swap_list.next = p - swap_info;
1280                 else
1281                         swap_info[prev].next = p - swap_info;
1282                 nr_swap_pages += p->pages;
1283                 total_swap_pages += p->pages;
1284                 p->flags |= SWP_WRITEOK;
1285                 spin_unlock(&swap_lock);
1286                 goto out_dput;
1287         }
1288
1289         /* wait for any unplug function to finish */
1290         down_write(&swap_unplug_sem);
1291         up_write(&swap_unplug_sem);
1292
1293         destroy_swap_extents(p);
1294         mutex_lock(&swapon_mutex);
1295         spin_lock(&swap_lock);
1296         drain_mmlist();
1297
1298         /* wait for anyone still in scan_swap_map */
1299         p->highest_bit = 0;             /* cuts scans short */
1300         while (p->flags >= SWP_SCANNING) {
1301                 spin_unlock(&swap_lock);
1302                 schedule_timeout_uninterruptible(1);
1303                 spin_lock(&swap_lock);
1304         }
1305
1306         swap_file = p->swap_file;
1307         p->swap_file = NULL;
1308         p->max = 0;
1309         swap_map = p->swap_map;
1310         p->swap_map = NULL;
1311         p->flags = 0;
1312         spin_unlock(&swap_lock);
1313         mutex_unlock(&swapon_mutex);
1314         vfree(swap_map);
1315         inode = mapping->host;
1316         if (S_ISBLK(inode->i_mode)) {
1317                 struct block_device *bdev = I_BDEV(inode);
1318                 set_blocksize(bdev, p->old_block_size);
1319                 bd_release(bdev);
1320         } else {
1321                 mutex_lock(&inode->i_mutex);
1322                 inode->i_flags &= ~S_SWAPFILE;
1323                 mutex_unlock(&inode->i_mutex);
1324         }
1325         filp_close(swap_file, NULL);
1326         err = 0;
1327
1328 out_dput:
1329         filp_close(victim, NULL);
1330 out:
1331         return err;
1332 }
1333
1334 #ifdef CONFIG_PROC_FS
1335 /* iterator */
1336 static void *swap_start(struct seq_file *swap, loff_t *pos)
1337 {
1338         struct swap_info_struct *ptr = swap_info;
1339         int i;
1340         loff_t l = *pos;
1341
1342         mutex_lock(&swapon_mutex);
1343
1344         if (!l)
1345                 return SEQ_START_TOKEN;
1346
1347         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1348                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1349                         continue;
1350                 if (!--l)
1351                         return ptr;
1352         }
1353
1354         return NULL;
1355 }
1356
1357 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1358 {
1359         struct swap_info_struct *ptr;
1360         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1361
1362         if (v == SEQ_START_TOKEN)
1363                 ptr = swap_info;
1364         else {
1365                 ptr = v;
1366                 ptr++;
1367         }
1368
1369         for (; ptr < endptr; ptr++) {
1370                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1371                         continue;
1372                 ++*pos;
1373                 return ptr;
1374         }
1375
1376         return NULL;
1377 }
1378
1379 static void swap_stop(struct seq_file *swap, void *v)
1380 {
1381         mutex_unlock(&swapon_mutex);
1382 }
1383
1384 static int swap_show(struct seq_file *swap, void *v)
1385 {
1386         struct swap_info_struct *ptr = v;
1387         struct file *file;
1388         int len;
1389
1390         if (ptr == SEQ_START_TOKEN) {
1391                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1392                 return 0;
1393         }
1394
1395         file = ptr->swap_file;
1396         len = seq_path(swap, &file->f_path, " \t\n\\");
1397         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1398                         len < 40 ? 40 - len : 1, " ",
1399                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1400                                 "partition" : "file\t",
1401                         ptr->pages << (PAGE_SHIFT - 10),
1402                         ptr->inuse_pages << (PAGE_SHIFT - 10),
1403                         ptr->prio);
1404         return 0;
1405 }
1406
1407 static const struct seq_operations swaps_op = {
1408         .start =        swap_start,
1409         .next =         swap_next,
1410         .stop =         swap_stop,
1411         .show =         swap_show
1412 };
1413
1414 static int swaps_open(struct inode *inode, struct file *file)
1415 {
1416         return seq_open(file, &swaps_op);
1417 }
1418
1419 static const struct file_operations proc_swaps_operations = {
1420         .open           = swaps_open,
1421         .read           = seq_read,
1422         .llseek         = seq_lseek,
1423         .release        = seq_release,
1424 };
1425
1426 static int __init procswaps_init(void)
1427 {
1428         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1429         return 0;
1430 }
1431 __initcall(procswaps_init);
1432 #endif /* CONFIG_PROC_FS */
1433
1434 #ifdef MAX_SWAPFILES_CHECK
1435 static int __init max_swapfiles_check(void)
1436 {
1437         MAX_SWAPFILES_CHECK();
1438         return 0;
1439 }
1440 late_initcall(max_swapfiles_check);
1441 #endif
1442
1443 /*
1444  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1445  *
1446  * The swapon system call
1447  */
1448 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1449 {
1450         struct swap_info_struct * p;
1451         char *name = NULL;
1452         struct block_device *bdev = NULL;
1453         struct file *swap_file = NULL;
1454         struct address_space *mapping;
1455         unsigned int type;
1456         int i, prev;
1457         int error;
1458         union swap_header *swap_header = NULL;
1459         int swap_header_version;
1460         unsigned int nr_good_pages = 0;
1461         int nr_extents = 0;
1462         sector_t span;
1463         unsigned long maxpages = 1;
1464         unsigned long swapfilepages;
1465         unsigned short *swap_map = NULL;
1466         struct page *page = NULL;
1467         struct inode *inode = NULL;
1468         int did_down = 0;
1469
1470         if (!capable(CAP_SYS_ADMIN))
1471                 return -EPERM;
1472         spin_lock(&swap_lock);
1473         p = swap_info;
1474         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1475                 if (!(p->flags & SWP_USED))
1476                         break;
1477         error = -EPERM;
1478         if (type >= MAX_SWAPFILES) {
1479                 spin_unlock(&swap_lock);
1480                 goto out;
1481         }
1482         if (type >= nr_swapfiles)
1483                 nr_swapfiles = type+1;
1484         memset(p, 0, sizeof(*p));
1485         INIT_LIST_HEAD(&p->extent_list);
1486         p->flags = SWP_USED;
1487         p->next = -1;
1488         spin_unlock(&swap_lock);
1489         name = getname(specialfile);
1490         error = PTR_ERR(name);
1491         if (IS_ERR(name)) {
1492                 name = NULL;
1493                 goto bad_swap_2;
1494         }
1495         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1496         error = PTR_ERR(swap_file);
1497         if (IS_ERR(swap_file)) {
1498                 swap_file = NULL;
1499                 goto bad_swap_2;
1500         }
1501
1502         p->swap_file = swap_file;
1503         mapping = swap_file->f_mapping;
1504         inode = mapping->host;
1505
1506         error = -EBUSY;
1507         for (i = 0; i < nr_swapfiles; i++) {
1508                 struct swap_info_struct *q = &swap_info[i];
1509
1510                 if (i == type || !q->swap_file)
1511                         continue;
1512                 if (mapping == q->swap_file->f_mapping)
1513                         goto bad_swap;
1514         }
1515
1516         error = -EINVAL;
1517         if (S_ISBLK(inode->i_mode)) {
1518                 bdev = I_BDEV(inode);
1519                 error = bd_claim(bdev, sys_swapon);
1520                 if (error < 0) {
1521                         bdev = NULL;
1522                         error = -EINVAL;
1523                         goto bad_swap;
1524                 }
1525                 p->old_block_size = block_size(bdev);
1526                 error = set_blocksize(bdev, PAGE_SIZE);
1527                 if (error < 0)
1528                         goto bad_swap;
1529                 p->bdev = bdev;
1530         } else if (S_ISREG(inode->i_mode)) {
1531                 p->bdev = inode->i_sb->s_bdev;
1532                 mutex_lock(&inode->i_mutex);
1533                 did_down = 1;
1534                 if (IS_SWAPFILE(inode)) {
1535                         error = -EBUSY;
1536                         goto bad_swap;
1537                 }
1538         } else {
1539                 goto bad_swap;
1540         }
1541
1542         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1543
1544         /*
1545          * Read the swap header.
1546          */
1547         if (!mapping->a_ops->readpage) {
1548                 error = -EINVAL;
1549                 goto bad_swap;
1550         }
1551         page = read_mapping_page(mapping, 0, swap_file);
1552         if (IS_ERR(page)) {
1553                 error = PTR_ERR(page);
1554                 goto bad_swap;
1555         }
1556         kmap(page);
1557         swap_header = page_address(page);
1558
1559         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1560                 swap_header_version = 1;
1561         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1562                 swap_header_version = 2;
1563         else {
1564                 printk(KERN_ERR "Unable to find swap-space signature\n");
1565                 error = -EINVAL;
1566                 goto bad_swap;
1567         }
1568
1569         switch (swap_header_version) {
1570         case 1:
1571                 printk(KERN_ERR "version 0 swap is no longer supported. "
1572                         "Use mkswap -v1 %s\n", name);
1573                 error = -EINVAL;
1574                 goto bad_swap;
1575         case 2:
1576                 /* swap partition endianess hack... */
1577                 if (swab32(swap_header->info.version) == 1) {
1578                         swab32s(&swap_header->info.version);
1579                         swab32s(&swap_header->info.last_page);
1580                         swab32s(&swap_header->info.nr_badpages);
1581                         for (i = 0; i < swap_header->info.nr_badpages; i++)
1582                                 swab32s(&swap_header->info.badpages[i]);
1583                 }
1584                 /* Check the swap header's sub-version and the size of
1585                    the swap file and bad block lists */
1586                 if (swap_header->info.version != 1) {
1587                         printk(KERN_WARNING
1588                                "Unable to handle swap header version %d\n",
1589                                swap_header->info.version);
1590                         error = -EINVAL;
1591                         goto bad_swap;
1592                 }
1593
1594                 p->lowest_bit  = 1;
1595                 p->cluster_next = 1;
1596
1597                 /*
1598                  * Find out how many pages are allowed for a single swap
1599                  * device. There are two limiting factors: 1) the number of
1600                  * bits for the swap offset in the swp_entry_t type and
1601                  * 2) the number of bits in the a swap pte as defined by
1602                  * the different architectures. In order to find the
1603                  * largest possible bit mask a swap entry with swap type 0
1604                  * and swap offset ~0UL is created, encoded to a swap pte,
1605                  * decoded to a swp_entry_t again and finally the swap
1606                  * offset is extracted. This will mask all the bits from
1607                  * the initial ~0UL mask that can't be encoded in either
1608                  * the swp_entry_t or the architecture definition of a
1609                  * swap pte.
1610                  */
1611                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1612                 if (maxpages > swap_header->info.last_page)
1613                         maxpages = swap_header->info.last_page;
1614                 p->highest_bit = maxpages - 1;
1615
1616                 error = -EINVAL;
1617                 if (!maxpages)
1618                         goto bad_swap;
1619                 if (swapfilepages && maxpages > swapfilepages) {
1620                         printk(KERN_WARNING
1621                                "Swap area shorter than signature indicates\n");
1622                         goto bad_swap;
1623                 }
1624                 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1625                         goto bad_swap;
1626                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1627                         goto bad_swap;
1628
1629                 /* OK, set up the swap map and apply the bad block list */
1630                 swap_map = vmalloc(maxpages * sizeof(short));
1631                 if (!swap_map) {
1632                         error = -ENOMEM;
1633                         goto bad_swap;
1634                 }
1635
1636                 error = 0;
1637                 memset(swap_map, 0, maxpages * sizeof(short));
1638                 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1639                         int page_nr = swap_header->info.badpages[i];
1640                         if (page_nr <= 0 || page_nr >= swap_header->info.last_page)
1641                                 error = -EINVAL;
1642                         else
1643                                 swap_map[page_nr] = SWAP_MAP_BAD;
1644                 }
1645                 nr_good_pages = swap_header->info.last_page -
1646                                 swap_header->info.nr_badpages -
1647                                 1 /* header page */;
1648                 if (error)
1649                         goto bad_swap;
1650         }
1651
1652         if (nr_good_pages) {
1653                 swap_map[0] = SWAP_MAP_BAD;
1654                 p->max = maxpages;
1655                 p->pages = nr_good_pages;
1656                 nr_extents = setup_swap_extents(p, &span);
1657                 if (nr_extents < 0) {
1658                         error = nr_extents;
1659                         goto bad_swap;
1660                 }
1661                 nr_good_pages = p->pages;
1662         }
1663         if (!nr_good_pages) {
1664                 printk(KERN_WARNING "Empty swap-file\n");
1665                 error = -EINVAL;
1666                 goto bad_swap;
1667         }
1668
1669         mutex_lock(&swapon_mutex);
1670         spin_lock(&swap_lock);
1671         if (swap_flags & SWAP_FLAG_PREFER)
1672                 p->prio =
1673                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
1674         else
1675                 p->prio = --least_priority;
1676         p->swap_map = swap_map;
1677         p->flags |= SWP_WRITEOK;
1678         nr_swap_pages += nr_good_pages;
1679         total_swap_pages += nr_good_pages;
1680
1681         printk(KERN_INFO "Adding %uk swap on %s.  "
1682                         "Priority:%d extents:%d across:%lluk\n",
1683                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1684                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1685
1686         /* insert swap space into swap_list: */
1687         prev = -1;
1688         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1689                 if (p->prio >= swap_info[i].prio) {
1690                         break;
1691                 }
1692                 prev = i;
1693         }
1694         p->next = i;
1695         if (prev < 0) {
1696                 swap_list.head = swap_list.next = p - swap_info;
1697         } else {
1698                 swap_info[prev].next = p - swap_info;
1699         }
1700         spin_unlock(&swap_lock);
1701         mutex_unlock(&swapon_mutex);
1702         error = 0;
1703         goto out;
1704 bad_swap:
1705         if (bdev) {
1706                 set_blocksize(bdev, p->old_block_size);
1707                 bd_release(bdev);
1708         }
1709         destroy_swap_extents(p);
1710 bad_swap_2:
1711         spin_lock(&swap_lock);
1712         p->swap_file = NULL;
1713         p->flags = 0;
1714         spin_unlock(&swap_lock);
1715         vfree(swap_map);
1716         if (swap_file)
1717                 filp_close(swap_file, NULL);
1718 out:
1719         if (page && !IS_ERR(page)) {
1720                 kunmap(page);
1721                 page_cache_release(page);
1722         }
1723         if (name)
1724                 putname(name);
1725         if (did_down) {
1726                 if (!error)
1727                         inode->i_flags |= S_SWAPFILE;
1728                 mutex_unlock(&inode->i_mutex);
1729         }
1730         return error;
1731 }
1732
1733 void si_swapinfo(struct sysinfo *val)
1734 {
1735         unsigned int i;
1736         unsigned long nr_to_be_unused = 0;
1737
1738         spin_lock(&swap_lock);
1739         for (i = 0; i < nr_swapfiles; i++) {
1740                 if (!(swap_info[i].flags & SWP_USED) ||
1741                      (swap_info[i].flags & SWP_WRITEOK))
1742                         continue;
1743                 nr_to_be_unused += swap_info[i].inuse_pages;
1744         }
1745         val->freeswap = nr_swap_pages + nr_to_be_unused;
1746         val->totalswap = total_swap_pages + nr_to_be_unused;
1747         spin_unlock(&swap_lock);
1748 }
1749
1750 /*
1751  * Verify that a swap entry is valid and increment its swap map count.
1752  *
1753  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1754  * "permanent", but will be reclaimed by the next swapoff.
1755  */
1756 int swap_duplicate(swp_entry_t entry)
1757 {
1758         struct swap_info_struct * p;
1759         unsigned long offset, type;
1760         int result = 0;
1761
1762         if (is_migration_entry(entry))
1763                 return 1;
1764
1765         type = swp_type(entry);
1766         if (type >= nr_swapfiles)
1767                 goto bad_file;
1768         p = type + swap_info;
1769         offset = swp_offset(entry);
1770
1771         spin_lock(&swap_lock);
1772         if (offset < p->max && p->swap_map[offset]) {
1773                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1774                         p->swap_map[offset]++;
1775                         result = 1;
1776                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1777                         if (swap_overflow++ < 5)
1778                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1779                         p->swap_map[offset] = SWAP_MAP_MAX;
1780                         result = 1;
1781                 }
1782         }
1783         spin_unlock(&swap_lock);
1784 out:
1785         return result;
1786
1787 bad_file:
1788         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1789         goto out;
1790 }
1791
1792 struct swap_info_struct *
1793 get_swap_info_struct(unsigned type)
1794 {
1795         return &swap_info[type];
1796 }
1797
1798 /*
1799  * swap_lock prevents swap_map being freed. Don't grab an extra
1800  * reference on the swaphandle, it doesn't matter if it becomes unused.
1801  */
1802 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1803 {
1804         struct swap_info_struct *si;
1805         int our_page_cluster = page_cluster;
1806         pgoff_t target, toff;
1807         pgoff_t base, end;
1808         int nr_pages = 0;
1809
1810         if (!our_page_cluster)  /* no readahead */
1811                 return 0;
1812
1813         si = &swap_info[swp_type(entry)];
1814         target = swp_offset(entry);
1815         base = (target >> our_page_cluster) << our_page_cluster;
1816         end = base + (1 << our_page_cluster);
1817         if (!base)              /* first page is swap header */
1818                 base++;
1819
1820         spin_lock(&swap_lock);
1821         if (end > si->max)      /* don't go beyond end of map */
1822                 end = si->max;
1823
1824         /* Count contiguous allocated slots above our target */
1825         for (toff = target; ++toff < end; nr_pages++) {
1826                 /* Don't read in free or bad pages */
1827                 if (!si->swap_map[toff])
1828                         break;
1829                 if (si->swap_map[toff] == SWAP_MAP_BAD)
1830                         break;
1831         }
1832         /* Count contiguous allocated slots below our target */
1833         for (toff = target; --toff >= base; nr_pages++) {
1834                 /* Don't read in free or bad pages */
1835                 if (!si->swap_map[toff])
1836                         break;
1837                 if (si->swap_map[toff] == SWAP_MAP_BAD)
1838                         break;
1839         }
1840         spin_unlock(&swap_lock);
1841
1842         /*
1843          * Indicate starting offset, and return number of pages to get:
1844          * if only 1, say 0, since there's then no readahead to be done.
1845          */
1846         *offset = ++toff;
1847         return nr_pages? ++nr_pages: 0;
1848 }