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