i386: Allow KVM on i386 nonpae
[sfrench/cifs-2.6.git] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/capability.h>
28 #include <linux/blkdev.h>
29 #include <linux/file.h>
30 #include <linux/quotaops.h>
31 #include <linux/highmem.h>
32 #include <linux/module.h>
33 #include <linux/writeback.h>
34 #include <linux/hash.h>
35 #include <linux/suspend.h>
36 #include <linux/buffer_head.h>
37 #include <linux/task_io_accounting_ops.h>
38 #include <linux/bio.h>
39 #include <linux/notifier.h>
40 #include <linux/cpu.h>
41 #include <linux/bitops.h>
42 #include <linux/mpage.h>
43 #include <linux/bit_spinlock.h>
44
45 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
46
47 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
48
49 inline void
50 init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
51 {
52         bh->b_end_io = handler;
53         bh->b_private = private;
54 }
55
56 static int sync_buffer(void *word)
57 {
58         struct block_device *bd;
59         struct buffer_head *bh
60                 = container_of(word, struct buffer_head, b_state);
61
62         smp_mb();
63         bd = bh->b_bdev;
64         if (bd)
65                 blk_run_address_space(bd->bd_inode->i_mapping);
66         io_schedule();
67         return 0;
68 }
69
70 void fastcall __lock_buffer(struct buffer_head *bh)
71 {
72         wait_on_bit_lock(&bh->b_state, BH_Lock, sync_buffer,
73                                                         TASK_UNINTERRUPTIBLE);
74 }
75 EXPORT_SYMBOL(__lock_buffer);
76
77 void fastcall unlock_buffer(struct buffer_head *bh)
78 {
79         smp_mb__before_clear_bit();
80         clear_buffer_locked(bh);
81         smp_mb__after_clear_bit();
82         wake_up_bit(&bh->b_state, BH_Lock);
83 }
84
85 /*
86  * Block until a buffer comes unlocked.  This doesn't stop it
87  * from becoming locked again - you have to lock it yourself
88  * if you want to preserve its state.
89  */
90 void __wait_on_buffer(struct buffer_head * bh)
91 {
92         wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
93 }
94
95 static void
96 __clear_page_buffers(struct page *page)
97 {
98         ClearPagePrivate(page);
99         set_page_private(page, 0);
100         page_cache_release(page);
101 }
102
103 static void buffer_io_error(struct buffer_head *bh)
104 {
105         char b[BDEVNAME_SIZE];
106
107         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
108                         bdevname(bh->b_bdev, b),
109                         (unsigned long long)bh->b_blocknr);
110 }
111
112 /*
113  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
114  * unlock the buffer. This is what ll_rw_block uses too.
115  */
116 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
117 {
118         if (uptodate) {
119                 set_buffer_uptodate(bh);
120         } else {
121                 /* This happens, due to failed READA attempts. */
122                 clear_buffer_uptodate(bh);
123         }
124         unlock_buffer(bh);
125         put_bh(bh);
126 }
127
128 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
129 {
130         char b[BDEVNAME_SIZE];
131
132         if (uptodate) {
133                 set_buffer_uptodate(bh);
134         } else {
135                 if (!buffer_eopnotsupp(bh) && printk_ratelimit()) {
136                         buffer_io_error(bh);
137                         printk(KERN_WARNING "lost page write due to "
138                                         "I/O error on %s\n",
139                                        bdevname(bh->b_bdev, b));
140                 }
141                 set_buffer_write_io_error(bh);
142                 clear_buffer_uptodate(bh);
143         }
144         unlock_buffer(bh);
145         put_bh(bh);
146 }
147
148 /*
149  * Write out and wait upon all the dirty data associated with a block
150  * device via its mapping.  Does not take the superblock lock.
151  */
152 int sync_blockdev(struct block_device *bdev)
153 {
154         int ret = 0;
155
156         if (bdev)
157                 ret = filemap_write_and_wait(bdev->bd_inode->i_mapping);
158         return ret;
159 }
160 EXPORT_SYMBOL(sync_blockdev);
161
162 /*
163  * Write out and wait upon all dirty data associated with this
164  * device.   Filesystem data as well as the underlying block
165  * device.  Takes the superblock lock.
166  */
167 int fsync_bdev(struct block_device *bdev)
168 {
169         struct super_block *sb = get_super(bdev);
170         if (sb) {
171                 int res = fsync_super(sb);
172                 drop_super(sb);
173                 return res;
174         }
175         return sync_blockdev(bdev);
176 }
177
178 /**
179  * freeze_bdev  --  lock a filesystem and force it into a consistent state
180  * @bdev:       blockdevice to lock
181  *
182  * This takes the block device bd_mount_sem to make sure no new mounts
183  * happen on bdev until thaw_bdev() is called.
184  * If a superblock is found on this device, we take the s_umount semaphore
185  * on it to make sure nobody unmounts until the snapshot creation is done.
186  */
187 struct super_block *freeze_bdev(struct block_device *bdev)
188 {
189         struct super_block *sb;
190
191         down(&bdev->bd_mount_sem);
192         sb = get_super(bdev);
193         if (sb && !(sb->s_flags & MS_RDONLY)) {
194                 sb->s_frozen = SB_FREEZE_WRITE;
195                 smp_wmb();
196
197                 __fsync_super(sb);
198
199                 sb->s_frozen = SB_FREEZE_TRANS;
200                 smp_wmb();
201
202                 sync_blockdev(sb->s_bdev);
203
204                 if (sb->s_op->write_super_lockfs)
205                         sb->s_op->write_super_lockfs(sb);
206         }
207
208         sync_blockdev(bdev);
209         return sb;      /* thaw_bdev releases s->s_umount and bd_mount_sem */
210 }
211 EXPORT_SYMBOL(freeze_bdev);
212
213 /**
214  * thaw_bdev  -- unlock filesystem
215  * @bdev:       blockdevice to unlock
216  * @sb:         associated superblock
217  *
218  * Unlocks the filesystem and marks it writeable again after freeze_bdev().
219  */
220 void thaw_bdev(struct block_device *bdev, struct super_block *sb)
221 {
222         if (sb) {
223                 BUG_ON(sb->s_bdev != bdev);
224
225                 if (sb->s_op->unlockfs)
226                         sb->s_op->unlockfs(sb);
227                 sb->s_frozen = SB_UNFROZEN;
228                 smp_wmb();
229                 wake_up(&sb->s_wait_unfrozen);
230                 drop_super(sb);
231         }
232
233         up(&bdev->bd_mount_sem);
234 }
235 EXPORT_SYMBOL(thaw_bdev);
236
237 /*
238  * Various filesystems appear to want __find_get_block to be non-blocking.
239  * But it's the page lock which protects the buffers.  To get around this,
240  * we get exclusion from try_to_free_buffers with the blockdev mapping's
241  * private_lock.
242  *
243  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
244  * may be quite high.  This code could TryLock the page, and if that
245  * succeeds, there is no need to take private_lock. (But if
246  * private_lock is contended then so is mapping->tree_lock).
247  */
248 static struct buffer_head *
249 __find_get_block_slow(struct block_device *bdev, sector_t block)
250 {
251         struct inode *bd_inode = bdev->bd_inode;
252         struct address_space *bd_mapping = bd_inode->i_mapping;
253         struct buffer_head *ret = NULL;
254         pgoff_t index;
255         struct buffer_head *bh;
256         struct buffer_head *head;
257         struct page *page;
258         int all_mapped = 1;
259
260         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
261         page = find_get_page(bd_mapping, index);
262         if (!page)
263                 goto out;
264
265         spin_lock(&bd_mapping->private_lock);
266         if (!page_has_buffers(page))
267                 goto out_unlock;
268         head = page_buffers(page);
269         bh = head;
270         do {
271                 if (bh->b_blocknr == block) {
272                         ret = bh;
273                         get_bh(bh);
274                         goto out_unlock;
275                 }
276                 if (!buffer_mapped(bh))
277                         all_mapped = 0;
278                 bh = bh->b_this_page;
279         } while (bh != head);
280
281         /* we might be here because some of the buffers on this page are
282          * not mapped.  This is due to various races between
283          * file io on the block device and getblk.  It gets dealt with
284          * elsewhere, don't buffer_error if we had some unmapped buffers
285          */
286         if (all_mapped) {
287                 printk("__find_get_block_slow() failed. "
288                         "block=%llu, b_blocknr=%llu\n",
289                         (unsigned long long)block,
290                         (unsigned long long)bh->b_blocknr);
291                 printk("b_state=0x%08lx, b_size=%zu\n",
292                         bh->b_state, bh->b_size);
293                 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
294         }
295 out_unlock:
296         spin_unlock(&bd_mapping->private_lock);
297         page_cache_release(page);
298 out:
299         return ret;
300 }
301
302 /* If invalidate_buffers() will trash dirty buffers, it means some kind
303    of fs corruption is going on. Trashing dirty data always imply losing
304    information that was supposed to be just stored on the physical layer
305    by the user.
306
307    Thus invalidate_buffers in general usage is not allwowed to trash
308    dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
309    be preserved.  These buffers are simply skipped.
310   
311    We also skip buffers which are still in use.  For example this can
312    happen if a userspace program is reading the block device.
313
314    NOTE: In the case where the user removed a removable-media-disk even if
315    there's still dirty data not synced on disk (due a bug in the device driver
316    or due an error of the user), by not destroying the dirty buffers we could
317    generate corruption also on the next media inserted, thus a parameter is
318    necessary to handle this case in the most safe way possible (trying
319    to not corrupt also the new disk inserted with the data belonging to
320    the old now corrupted disk). Also for the ramdisk the natural thing
321    to do in order to release the ramdisk memory is to destroy dirty buffers.
322
323    These are two special cases. Normal usage imply the device driver
324    to issue a sync on the device (without waiting I/O completion) and
325    then an invalidate_buffers call that doesn't trash dirty buffers.
326
327    For handling cache coherency with the blkdev pagecache the 'update' case
328    is been introduced. It is needed to re-read from disk any pinned
329    buffer. NOTE: re-reading from disk is destructive so we can do it only
330    when we assume nobody is changing the buffercache under our I/O and when
331    we think the disk contains more recent information than the buffercache.
332    The update == 1 pass marks the buffers we need to update, the update == 2
333    pass does the actual I/O. */
334 void invalidate_bdev(struct block_device *bdev)
335 {
336         struct address_space *mapping = bdev->bd_inode->i_mapping;
337
338         if (mapping->nrpages == 0)
339                 return;
340
341         invalidate_bh_lrus();
342         invalidate_mapping_pages(mapping, 0, -1);
343 }
344
345 /*
346  * Kick pdflush then try to free up some ZONE_NORMAL memory.
347  */
348 static void free_more_memory(void)
349 {
350         struct zone **zones;
351         pg_data_t *pgdat;
352
353         wakeup_pdflush(1024);
354         yield();
355
356         for_each_online_pgdat(pgdat) {
357                 zones = pgdat->node_zonelists[gfp_zone(GFP_NOFS)].zones;
358                 if (*zones)
359                         try_to_free_pages(zones, 0, GFP_NOFS);
360         }
361 }
362
363 /*
364  * I/O completion handler for block_read_full_page() - pages
365  * which come unlocked at the end of I/O.
366  */
367 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
368 {
369         unsigned long flags;
370         struct buffer_head *first;
371         struct buffer_head *tmp;
372         struct page *page;
373         int page_uptodate = 1;
374
375         BUG_ON(!buffer_async_read(bh));
376
377         page = bh->b_page;
378         if (uptodate) {
379                 set_buffer_uptodate(bh);
380         } else {
381                 clear_buffer_uptodate(bh);
382                 if (printk_ratelimit())
383                         buffer_io_error(bh);
384                 SetPageError(page);
385         }
386
387         /*
388          * Be _very_ careful from here on. Bad things can happen if
389          * two buffer heads end IO at almost the same time and both
390          * decide that the page is now completely done.
391          */
392         first = page_buffers(page);
393         local_irq_save(flags);
394         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
395         clear_buffer_async_read(bh);
396         unlock_buffer(bh);
397         tmp = bh;
398         do {
399                 if (!buffer_uptodate(tmp))
400                         page_uptodate = 0;
401                 if (buffer_async_read(tmp)) {
402                         BUG_ON(!buffer_locked(tmp));
403                         goto still_busy;
404                 }
405                 tmp = tmp->b_this_page;
406         } while (tmp != bh);
407         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
408         local_irq_restore(flags);
409
410         /*
411          * If none of the buffers had errors and they are all
412          * uptodate then we can set the page uptodate.
413          */
414         if (page_uptodate && !PageError(page))
415                 SetPageUptodate(page);
416         unlock_page(page);
417         return;
418
419 still_busy:
420         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
421         local_irq_restore(flags);
422         return;
423 }
424
425 /*
426  * Completion handler for block_write_full_page() - pages which are unlocked
427  * during I/O, and which have PageWriteback cleared upon I/O completion.
428  */
429 static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
430 {
431         char b[BDEVNAME_SIZE];
432         unsigned long flags;
433         struct buffer_head *first;
434         struct buffer_head *tmp;
435         struct page *page;
436
437         BUG_ON(!buffer_async_write(bh));
438
439         page = bh->b_page;
440         if (uptodate) {
441                 set_buffer_uptodate(bh);
442         } else {
443                 if (printk_ratelimit()) {
444                         buffer_io_error(bh);
445                         printk(KERN_WARNING "lost page write due to "
446                                         "I/O error on %s\n",
447                                bdevname(bh->b_bdev, b));
448                 }
449                 set_bit(AS_EIO, &page->mapping->flags);
450                 set_buffer_write_io_error(bh);
451                 clear_buffer_uptodate(bh);
452                 SetPageError(page);
453         }
454
455         first = page_buffers(page);
456         local_irq_save(flags);
457         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
458
459         clear_buffer_async_write(bh);
460         unlock_buffer(bh);
461         tmp = bh->b_this_page;
462         while (tmp != bh) {
463                 if (buffer_async_write(tmp)) {
464                         BUG_ON(!buffer_locked(tmp));
465                         goto still_busy;
466                 }
467                 tmp = tmp->b_this_page;
468         }
469         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
470         local_irq_restore(flags);
471         end_page_writeback(page);
472         return;
473
474 still_busy:
475         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
476         local_irq_restore(flags);
477         return;
478 }
479
480 /*
481  * If a page's buffers are under async readin (end_buffer_async_read
482  * completion) then there is a possibility that another thread of
483  * control could lock one of the buffers after it has completed
484  * but while some of the other buffers have not completed.  This
485  * locked buffer would confuse end_buffer_async_read() into not unlocking
486  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
487  * that this buffer is not under async I/O.
488  *
489  * The page comes unlocked when it has no locked buffer_async buffers
490  * left.
491  *
492  * PageLocked prevents anyone starting new async I/O reads any of
493  * the buffers.
494  *
495  * PageWriteback is used to prevent simultaneous writeout of the same
496  * page.
497  *
498  * PageLocked prevents anyone from starting writeback of a page which is
499  * under read I/O (PageWriteback is only ever set against a locked page).
500  */
501 static void mark_buffer_async_read(struct buffer_head *bh)
502 {
503         bh->b_end_io = end_buffer_async_read;
504         set_buffer_async_read(bh);
505 }
506
507 void mark_buffer_async_write(struct buffer_head *bh)
508 {
509         bh->b_end_io = end_buffer_async_write;
510         set_buffer_async_write(bh);
511 }
512 EXPORT_SYMBOL(mark_buffer_async_write);
513
514
515 /*
516  * fs/buffer.c contains helper functions for buffer-backed address space's
517  * fsync functions.  A common requirement for buffer-based filesystems is
518  * that certain data from the backing blockdev needs to be written out for
519  * a successful fsync().  For example, ext2 indirect blocks need to be
520  * written back and waited upon before fsync() returns.
521  *
522  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
523  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
524  * management of a list of dependent buffers at ->i_mapping->private_list.
525  *
526  * Locking is a little subtle: try_to_free_buffers() will remove buffers
527  * from their controlling inode's queue when they are being freed.  But
528  * try_to_free_buffers() will be operating against the *blockdev* mapping
529  * at the time, not against the S_ISREG file which depends on those buffers.
530  * So the locking for private_list is via the private_lock in the address_space
531  * which backs the buffers.  Which is different from the address_space 
532  * against which the buffers are listed.  So for a particular address_space,
533  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
534  * mapping->private_list will always be protected by the backing blockdev's
535  * ->private_lock.
536  *
537  * Which introduces a requirement: all buffers on an address_space's
538  * ->private_list must be from the same address_space: the blockdev's.
539  *
540  * address_spaces which do not place buffers at ->private_list via these
541  * utility functions are free to use private_lock and private_list for
542  * whatever they want.  The only requirement is that list_empty(private_list)
543  * be true at clear_inode() time.
544  *
545  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
546  * filesystems should do that.  invalidate_inode_buffers() should just go
547  * BUG_ON(!list_empty).
548  *
549  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
550  * take an address_space, not an inode.  And it should be called
551  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
552  * queued up.
553  *
554  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
555  * list if it is already on a list.  Because if the buffer is on a list,
556  * it *must* already be on the right one.  If not, the filesystem is being
557  * silly.  This will save a ton of locking.  But first we have to ensure
558  * that buffers are taken *off* the old inode's list when they are freed
559  * (presumably in truncate).  That requires careful auditing of all
560  * filesystems (do it inside bforget()).  It could also be done by bringing
561  * b_inode back.
562  */
563
564 /*
565  * The buffer's backing address_space's private_lock must be held
566  */
567 static inline void __remove_assoc_queue(struct buffer_head *bh)
568 {
569         list_del_init(&bh->b_assoc_buffers);
570         WARN_ON(!bh->b_assoc_map);
571         if (buffer_write_io_error(bh))
572                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
573         bh->b_assoc_map = NULL;
574 }
575
576 int inode_has_buffers(struct inode *inode)
577 {
578         return !list_empty(&inode->i_data.private_list);
579 }
580
581 /*
582  * osync is designed to support O_SYNC io.  It waits synchronously for
583  * all already-submitted IO to complete, but does not queue any new
584  * writes to the disk.
585  *
586  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
587  * you dirty the buffers, and then use osync_inode_buffers to wait for
588  * completion.  Any other dirty buffers which are not yet queued for
589  * write will not be flushed to disk by the osync.
590  */
591 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
592 {
593         struct buffer_head *bh;
594         struct list_head *p;
595         int err = 0;
596
597         spin_lock(lock);
598 repeat:
599         list_for_each_prev(p, list) {
600                 bh = BH_ENTRY(p);
601                 if (buffer_locked(bh)) {
602                         get_bh(bh);
603                         spin_unlock(lock);
604                         wait_on_buffer(bh);
605                         if (!buffer_uptodate(bh))
606                                 err = -EIO;
607                         brelse(bh);
608                         spin_lock(lock);
609                         goto repeat;
610                 }
611         }
612         spin_unlock(lock);
613         return err;
614 }
615
616 /**
617  * sync_mapping_buffers - write out and wait upon a mapping's "associated"
618  *                        buffers
619  * @mapping: the mapping which wants those buffers written
620  *
621  * Starts I/O against the buffers at mapping->private_list, and waits upon
622  * that I/O.
623  *
624  * Basically, this is a convenience function for fsync().
625  * @mapping is a file or directory which needs those buffers to be written for
626  * a successful fsync().
627  */
628 int sync_mapping_buffers(struct address_space *mapping)
629 {
630         struct address_space *buffer_mapping = mapping->assoc_mapping;
631
632         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
633                 return 0;
634
635         return fsync_buffers_list(&buffer_mapping->private_lock,
636                                         &mapping->private_list);
637 }
638 EXPORT_SYMBOL(sync_mapping_buffers);
639
640 /*
641  * Called when we've recently written block `bblock', and it is known that
642  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
643  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
644  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
645  */
646 void write_boundary_block(struct block_device *bdev,
647                         sector_t bblock, unsigned blocksize)
648 {
649         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
650         if (bh) {
651                 if (buffer_dirty(bh))
652                         ll_rw_block(WRITE, 1, &bh);
653                 put_bh(bh);
654         }
655 }
656
657 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
658 {
659         struct address_space *mapping = inode->i_mapping;
660         struct address_space *buffer_mapping = bh->b_page->mapping;
661
662         mark_buffer_dirty(bh);
663         if (!mapping->assoc_mapping) {
664                 mapping->assoc_mapping = buffer_mapping;
665         } else {
666                 BUG_ON(mapping->assoc_mapping != buffer_mapping);
667         }
668         if (list_empty(&bh->b_assoc_buffers)) {
669                 spin_lock(&buffer_mapping->private_lock);
670                 list_move_tail(&bh->b_assoc_buffers,
671                                 &mapping->private_list);
672                 bh->b_assoc_map = mapping;
673                 spin_unlock(&buffer_mapping->private_lock);
674         }
675 }
676 EXPORT_SYMBOL(mark_buffer_dirty_inode);
677
678 /*
679  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
680  * dirty.
681  *
682  * If warn is true, then emit a warning if the page is not uptodate and has
683  * not been truncated.
684  */
685 static int __set_page_dirty(struct page *page,
686                 struct address_space *mapping, int warn)
687 {
688         if (unlikely(!mapping))
689                 return !TestSetPageDirty(page);
690
691         if (TestSetPageDirty(page))
692                 return 0;
693
694         write_lock_irq(&mapping->tree_lock);
695         if (page->mapping) {    /* Race with truncate? */
696                 WARN_ON_ONCE(warn && !PageUptodate(page));
697
698                 if (mapping_cap_account_dirty(mapping)) {
699                         __inc_zone_page_state(page, NR_FILE_DIRTY);
700                         task_io_account_write(PAGE_CACHE_SIZE);
701                 }
702                 radix_tree_tag_set(&mapping->page_tree,
703                                 page_index(page), PAGECACHE_TAG_DIRTY);
704         }
705         write_unlock_irq(&mapping->tree_lock);
706         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
707
708         return 1;
709 }
710
711 /*
712  * Add a page to the dirty page list.
713  *
714  * It is a sad fact of life that this function is called from several places
715  * deeply under spinlocking.  It may not sleep.
716  *
717  * If the page has buffers, the uptodate buffers are set dirty, to preserve
718  * dirty-state coherency between the page and the buffers.  It the page does
719  * not have buffers then when they are later attached they will all be set
720  * dirty.
721  *
722  * The buffers are dirtied before the page is dirtied.  There's a small race
723  * window in which a writepage caller may see the page cleanness but not the
724  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
725  * before the buffers, a concurrent writepage caller could clear the page dirty
726  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
727  * page on the dirty page list.
728  *
729  * We use private_lock to lock against try_to_free_buffers while using the
730  * page's buffer list.  Also use this to protect against clean buffers being
731  * added to the page after it was set dirty.
732  *
733  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
734  * address_space though.
735  */
736 int __set_page_dirty_buffers(struct page *page)
737 {
738         struct address_space *mapping = page_mapping(page);
739
740         if (unlikely(!mapping))
741                 return !TestSetPageDirty(page);
742
743         spin_lock(&mapping->private_lock);
744         if (page_has_buffers(page)) {
745                 struct buffer_head *head = page_buffers(page);
746                 struct buffer_head *bh = head;
747
748                 do {
749                         set_buffer_dirty(bh);
750                         bh = bh->b_this_page;
751                 } while (bh != head);
752         }
753         spin_unlock(&mapping->private_lock);
754
755         return __set_page_dirty(page, mapping, 1);
756 }
757 EXPORT_SYMBOL(__set_page_dirty_buffers);
758
759 /*
760  * Write out and wait upon a list of buffers.
761  *
762  * We have conflicting pressures: we want to make sure that all
763  * initially dirty buffers get waited on, but that any subsequently
764  * dirtied buffers don't.  After all, we don't want fsync to last
765  * forever if somebody is actively writing to the file.
766  *
767  * Do this in two main stages: first we copy dirty buffers to a
768  * temporary inode list, queueing the writes as we go.  Then we clean
769  * up, waiting for those writes to complete.
770  * 
771  * During this second stage, any subsequent updates to the file may end
772  * up refiling the buffer on the original inode's dirty list again, so
773  * there is a chance we will end up with a buffer queued for write but
774  * not yet completed on that list.  So, as a final cleanup we go through
775  * the osync code to catch these locked, dirty buffers without requeuing
776  * any newly dirty buffers for write.
777  */
778 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
779 {
780         struct buffer_head *bh;
781         struct list_head tmp;
782         int err = 0, err2;
783
784         INIT_LIST_HEAD(&tmp);
785
786         spin_lock(lock);
787         while (!list_empty(list)) {
788                 bh = BH_ENTRY(list->next);
789                 __remove_assoc_queue(bh);
790                 if (buffer_dirty(bh) || buffer_locked(bh)) {
791                         list_add(&bh->b_assoc_buffers, &tmp);
792                         if (buffer_dirty(bh)) {
793                                 get_bh(bh);
794                                 spin_unlock(lock);
795                                 /*
796                                  * Ensure any pending I/O completes so that
797                                  * ll_rw_block() actually writes the current
798                                  * contents - it is a noop if I/O is still in
799                                  * flight on potentially older contents.
800                                  */
801                                 ll_rw_block(SWRITE, 1, &bh);
802                                 brelse(bh);
803                                 spin_lock(lock);
804                         }
805                 }
806         }
807
808         while (!list_empty(&tmp)) {
809                 bh = BH_ENTRY(tmp.prev);
810                 list_del_init(&bh->b_assoc_buffers);
811                 get_bh(bh);
812                 spin_unlock(lock);
813                 wait_on_buffer(bh);
814                 if (!buffer_uptodate(bh))
815                         err = -EIO;
816                 brelse(bh);
817                 spin_lock(lock);
818         }
819         
820         spin_unlock(lock);
821         err2 = osync_buffers_list(lock, list);
822         if (err)
823                 return err;
824         else
825                 return err2;
826 }
827
828 /*
829  * Invalidate any and all dirty buffers on a given inode.  We are
830  * probably unmounting the fs, but that doesn't mean we have already
831  * done a sync().  Just drop the buffers from the inode list.
832  *
833  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
834  * assumes that all the buffers are against the blockdev.  Not true
835  * for reiserfs.
836  */
837 void invalidate_inode_buffers(struct inode *inode)
838 {
839         if (inode_has_buffers(inode)) {
840                 struct address_space *mapping = &inode->i_data;
841                 struct list_head *list = &mapping->private_list;
842                 struct address_space *buffer_mapping = mapping->assoc_mapping;
843
844                 spin_lock(&buffer_mapping->private_lock);
845                 while (!list_empty(list))
846                         __remove_assoc_queue(BH_ENTRY(list->next));
847                 spin_unlock(&buffer_mapping->private_lock);
848         }
849 }
850
851 /*
852  * Remove any clean buffers from the inode's buffer list.  This is called
853  * when we're trying to free the inode itself.  Those buffers can pin it.
854  *
855  * Returns true if all buffers were removed.
856  */
857 int remove_inode_buffers(struct inode *inode)
858 {
859         int ret = 1;
860
861         if (inode_has_buffers(inode)) {
862                 struct address_space *mapping = &inode->i_data;
863                 struct list_head *list = &mapping->private_list;
864                 struct address_space *buffer_mapping = mapping->assoc_mapping;
865
866                 spin_lock(&buffer_mapping->private_lock);
867                 while (!list_empty(list)) {
868                         struct buffer_head *bh = BH_ENTRY(list->next);
869                         if (buffer_dirty(bh)) {
870                                 ret = 0;
871                                 break;
872                         }
873                         __remove_assoc_queue(bh);
874                 }
875                 spin_unlock(&buffer_mapping->private_lock);
876         }
877         return ret;
878 }
879
880 /*
881  * Create the appropriate buffers when given a page for data area and
882  * the size of each buffer.. Use the bh->b_this_page linked list to
883  * follow the buffers created.  Return NULL if unable to create more
884  * buffers.
885  *
886  * The retry flag is used to differentiate async IO (paging, swapping)
887  * which may not fail from ordinary buffer allocations.
888  */
889 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
890                 int retry)
891 {
892         struct buffer_head *bh, *head;
893         long offset;
894
895 try_again:
896         head = NULL;
897         offset = PAGE_SIZE;
898         while ((offset -= size) >= 0) {
899                 bh = alloc_buffer_head(GFP_NOFS);
900                 if (!bh)
901                         goto no_grow;
902
903                 bh->b_bdev = NULL;
904                 bh->b_this_page = head;
905                 bh->b_blocknr = -1;
906                 head = bh;
907
908                 bh->b_state = 0;
909                 atomic_set(&bh->b_count, 0);
910                 bh->b_private = NULL;
911                 bh->b_size = size;
912
913                 /* Link the buffer to its page */
914                 set_bh_page(bh, page, offset);
915
916                 init_buffer(bh, NULL, NULL);
917         }
918         return head;
919 /*
920  * In case anything failed, we just free everything we got.
921  */
922 no_grow:
923         if (head) {
924                 do {
925                         bh = head;
926                         head = head->b_this_page;
927                         free_buffer_head(bh);
928                 } while (head);
929         }
930
931         /*
932          * Return failure for non-async IO requests.  Async IO requests
933          * are not allowed to fail, so we have to wait until buffer heads
934          * become available.  But we don't want tasks sleeping with 
935          * partially complete buffers, so all were released above.
936          */
937         if (!retry)
938                 return NULL;
939
940         /* We're _really_ low on memory. Now we just
941          * wait for old buffer heads to become free due to
942          * finishing IO.  Since this is an async request and
943          * the reserve list is empty, we're sure there are 
944          * async buffer heads in use.
945          */
946         free_more_memory();
947         goto try_again;
948 }
949 EXPORT_SYMBOL_GPL(alloc_page_buffers);
950
951 static inline void
952 link_dev_buffers(struct page *page, struct buffer_head *head)
953 {
954         struct buffer_head *bh, *tail;
955
956         bh = head;
957         do {
958                 tail = bh;
959                 bh = bh->b_this_page;
960         } while (bh);
961         tail->b_this_page = head;
962         attach_page_buffers(page, head);
963 }
964
965 /*
966  * Initialise the state of a blockdev page's buffers.
967  */ 
968 static void
969 init_page_buffers(struct page *page, struct block_device *bdev,
970                         sector_t block, int size)
971 {
972         struct buffer_head *head = page_buffers(page);
973         struct buffer_head *bh = head;
974         int uptodate = PageUptodate(page);
975
976         do {
977                 if (!buffer_mapped(bh)) {
978                         init_buffer(bh, NULL, NULL);
979                         bh->b_bdev = bdev;
980                         bh->b_blocknr = block;
981                         if (uptodate)
982                                 set_buffer_uptodate(bh);
983                         set_buffer_mapped(bh);
984                 }
985                 block++;
986                 bh = bh->b_this_page;
987         } while (bh != head);
988 }
989
990 /*
991  * Create the page-cache page that contains the requested block.
992  *
993  * This is user purely for blockdev mappings.
994  */
995 static struct page *
996 grow_dev_page(struct block_device *bdev, sector_t block,
997                 pgoff_t index, int size)
998 {
999         struct inode *inode = bdev->bd_inode;
1000         struct page *page;
1001         struct buffer_head *bh;
1002
1003         page = find_or_create_page(inode->i_mapping, index,
1004                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
1005         if (!page)
1006                 return NULL;
1007
1008         BUG_ON(!PageLocked(page));
1009
1010         if (page_has_buffers(page)) {
1011                 bh = page_buffers(page);
1012                 if (bh->b_size == size) {
1013                         init_page_buffers(page, bdev, block, size);
1014                         return page;
1015                 }
1016                 if (!try_to_free_buffers(page))
1017                         goto failed;
1018         }
1019
1020         /*
1021          * Allocate some buffers for this page
1022          */
1023         bh = alloc_page_buffers(page, size, 0);
1024         if (!bh)
1025                 goto failed;
1026
1027         /*
1028          * Link the page to the buffers and initialise them.  Take the
1029          * lock to be atomic wrt __find_get_block(), which does not
1030          * run under the page lock.
1031          */
1032         spin_lock(&inode->i_mapping->private_lock);
1033         link_dev_buffers(page, bh);
1034         init_page_buffers(page, bdev, block, size);
1035         spin_unlock(&inode->i_mapping->private_lock);
1036         return page;
1037
1038 failed:
1039         BUG();
1040         unlock_page(page);
1041         page_cache_release(page);
1042         return NULL;
1043 }
1044
1045 /*
1046  * Create buffers for the specified block device block's page.  If
1047  * that page was dirty, the buffers are set dirty also.
1048  */
1049 static int
1050 grow_buffers(struct block_device *bdev, sector_t block, int size)
1051 {
1052         struct page *page;
1053         pgoff_t index;
1054         int sizebits;
1055
1056         sizebits = -1;
1057         do {
1058                 sizebits++;
1059         } while ((size << sizebits) < PAGE_SIZE);
1060
1061         index = block >> sizebits;
1062
1063         /*
1064          * Check for a block which wants to lie outside our maximum possible
1065          * pagecache index.  (this comparison is done using sector_t types).
1066          */
1067         if (unlikely(index != block >> sizebits)) {
1068                 char b[BDEVNAME_SIZE];
1069
1070                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1071                         "device %s\n",
1072                         __FUNCTION__, (unsigned long long)block,
1073                         bdevname(bdev, b));
1074                 return -EIO;
1075         }
1076         block = index << sizebits;
1077         /* Create a page with the proper size buffers.. */
1078         page = grow_dev_page(bdev, block, index, size);
1079         if (!page)
1080                 return 0;
1081         unlock_page(page);
1082         page_cache_release(page);
1083         return 1;
1084 }
1085
1086 static struct buffer_head *
1087 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1088 {
1089         /* Size must be multiple of hard sectorsize */
1090         if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1091                         (size < 512 || size > PAGE_SIZE))) {
1092                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1093                                         size);
1094                 printk(KERN_ERR "hardsect size: %d\n",
1095                                         bdev_hardsect_size(bdev));
1096
1097                 dump_stack();
1098                 return NULL;
1099         }
1100
1101         for (;;) {
1102                 struct buffer_head * bh;
1103                 int ret;
1104
1105                 bh = __find_get_block(bdev, block, size);
1106                 if (bh)
1107                         return bh;
1108
1109                 ret = grow_buffers(bdev, block, size);
1110                 if (ret < 0)
1111                         return NULL;
1112                 if (ret == 0)
1113                         free_more_memory();
1114         }
1115 }
1116
1117 /*
1118  * The relationship between dirty buffers and dirty pages:
1119  *
1120  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1121  * the page is tagged dirty in its radix tree.
1122  *
1123  * At all times, the dirtiness of the buffers represents the dirtiness of
1124  * subsections of the page.  If the page has buffers, the page dirty bit is
1125  * merely a hint about the true dirty state.
1126  *
1127  * When a page is set dirty in its entirety, all its buffers are marked dirty
1128  * (if the page has buffers).
1129  *
1130  * When a buffer is marked dirty, its page is dirtied, but the page's other
1131  * buffers are not.
1132  *
1133  * Also.  When blockdev buffers are explicitly read with bread(), they
1134  * individually become uptodate.  But their backing page remains not
1135  * uptodate - even if all of its buffers are uptodate.  A subsequent
1136  * block_read_full_page() against that page will discover all the uptodate
1137  * buffers, will set the page uptodate and will perform no I/O.
1138  */
1139
1140 /**
1141  * mark_buffer_dirty - mark a buffer_head as needing writeout
1142  * @bh: the buffer_head to mark dirty
1143  *
1144  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1145  * backing page dirty, then tag the page as dirty in its address_space's radix
1146  * tree and then attach the address_space's inode to its superblock's dirty
1147  * inode list.
1148  *
1149  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1150  * mapping->tree_lock and the global inode_lock.
1151  */
1152 void fastcall mark_buffer_dirty(struct buffer_head *bh)
1153 {
1154         WARN_ON_ONCE(!buffer_uptodate(bh));
1155         if (!buffer_dirty(bh) && !test_set_buffer_dirty(bh))
1156                 __set_page_dirty(bh->b_page, page_mapping(bh->b_page), 0);
1157 }
1158
1159 /*
1160  * Decrement a buffer_head's reference count.  If all buffers against a page
1161  * have zero reference count, are clean and unlocked, and if the page is clean
1162  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1163  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1164  * a page but it ends up not being freed, and buffers may later be reattached).
1165  */
1166 void __brelse(struct buffer_head * buf)
1167 {
1168         if (atomic_read(&buf->b_count)) {
1169                 put_bh(buf);
1170                 return;
1171         }
1172         printk(KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1173         WARN_ON(1);
1174 }
1175
1176 /*
1177  * bforget() is like brelse(), except it discards any
1178  * potentially dirty data.
1179  */
1180 void __bforget(struct buffer_head *bh)
1181 {
1182         clear_buffer_dirty(bh);
1183         if (!list_empty(&bh->b_assoc_buffers)) {
1184                 struct address_space *buffer_mapping = bh->b_page->mapping;
1185
1186                 spin_lock(&buffer_mapping->private_lock);
1187                 list_del_init(&bh->b_assoc_buffers);
1188                 bh->b_assoc_map = NULL;
1189                 spin_unlock(&buffer_mapping->private_lock);
1190         }
1191         __brelse(bh);
1192 }
1193
1194 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1195 {
1196         lock_buffer(bh);
1197         if (buffer_uptodate(bh)) {
1198                 unlock_buffer(bh);
1199                 return bh;
1200         } else {
1201                 get_bh(bh);
1202                 bh->b_end_io = end_buffer_read_sync;
1203                 submit_bh(READ, bh);
1204                 wait_on_buffer(bh);
1205                 if (buffer_uptodate(bh))
1206                         return bh;
1207         }
1208         brelse(bh);
1209         return NULL;
1210 }
1211
1212 /*
1213  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1214  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1215  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1216  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1217  * CPU's LRUs at the same time.
1218  *
1219  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1220  * sb_find_get_block().
1221  *
1222  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1223  * a local interrupt disable for that.
1224  */
1225
1226 #define BH_LRU_SIZE     8
1227
1228 struct bh_lru {
1229         struct buffer_head *bhs[BH_LRU_SIZE];
1230 };
1231
1232 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1233
1234 #ifdef CONFIG_SMP
1235 #define bh_lru_lock()   local_irq_disable()
1236 #define bh_lru_unlock() local_irq_enable()
1237 #else
1238 #define bh_lru_lock()   preempt_disable()
1239 #define bh_lru_unlock() preempt_enable()
1240 #endif
1241
1242 static inline void check_irqs_on(void)
1243 {
1244 #ifdef irqs_disabled
1245         BUG_ON(irqs_disabled());
1246 #endif
1247 }
1248
1249 /*
1250  * The LRU management algorithm is dopey-but-simple.  Sorry.
1251  */
1252 static void bh_lru_install(struct buffer_head *bh)
1253 {
1254         struct buffer_head *evictee = NULL;
1255         struct bh_lru *lru;
1256
1257         check_irqs_on();
1258         bh_lru_lock();
1259         lru = &__get_cpu_var(bh_lrus);
1260         if (lru->bhs[0] != bh) {
1261                 struct buffer_head *bhs[BH_LRU_SIZE];
1262                 int in;
1263                 int out = 0;
1264
1265                 get_bh(bh);
1266                 bhs[out++] = bh;
1267                 for (in = 0; in < BH_LRU_SIZE; in++) {
1268                         struct buffer_head *bh2 = lru->bhs[in];
1269
1270                         if (bh2 == bh) {
1271                                 __brelse(bh2);
1272                         } else {
1273                                 if (out >= BH_LRU_SIZE) {
1274                                         BUG_ON(evictee != NULL);
1275                                         evictee = bh2;
1276                                 } else {
1277                                         bhs[out++] = bh2;
1278                                 }
1279                         }
1280                 }
1281                 while (out < BH_LRU_SIZE)
1282                         bhs[out++] = NULL;
1283                 memcpy(lru->bhs, bhs, sizeof(bhs));
1284         }
1285         bh_lru_unlock();
1286
1287         if (evictee)
1288                 __brelse(evictee);
1289 }
1290
1291 /*
1292  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1293  */
1294 static struct buffer_head *
1295 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1296 {
1297         struct buffer_head *ret = NULL;
1298         struct bh_lru *lru;
1299         unsigned int i;
1300
1301         check_irqs_on();
1302         bh_lru_lock();
1303         lru = &__get_cpu_var(bh_lrus);
1304         for (i = 0; i < BH_LRU_SIZE; i++) {
1305                 struct buffer_head *bh = lru->bhs[i];
1306
1307                 if (bh && bh->b_bdev == bdev &&
1308                                 bh->b_blocknr == block && bh->b_size == size) {
1309                         if (i) {
1310                                 while (i) {
1311                                         lru->bhs[i] = lru->bhs[i - 1];
1312                                         i--;
1313                                 }
1314                                 lru->bhs[0] = bh;
1315                         }
1316                         get_bh(bh);
1317                         ret = bh;
1318                         break;
1319                 }
1320         }
1321         bh_lru_unlock();
1322         return ret;
1323 }
1324
1325 /*
1326  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1327  * it in the LRU and mark it as accessed.  If it is not present then return
1328  * NULL
1329  */
1330 struct buffer_head *
1331 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1332 {
1333         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1334
1335         if (bh == NULL) {
1336                 bh = __find_get_block_slow(bdev, block);
1337                 if (bh)
1338                         bh_lru_install(bh);
1339         }
1340         if (bh)
1341                 touch_buffer(bh);
1342         return bh;
1343 }
1344 EXPORT_SYMBOL(__find_get_block);
1345
1346 /*
1347  * __getblk will locate (and, if necessary, create) the buffer_head
1348  * which corresponds to the passed block_device, block and size. The
1349  * returned buffer has its reference count incremented.
1350  *
1351  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1352  * illegal block number, __getblk() will happily return a buffer_head
1353  * which represents the non-existent block.  Very weird.
1354  *
1355  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1356  * attempt is failing.  FIXME, perhaps?
1357  */
1358 struct buffer_head *
1359 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1360 {
1361         struct buffer_head *bh = __find_get_block(bdev, block, size);
1362
1363         might_sleep();
1364         if (bh == NULL)
1365                 bh = __getblk_slow(bdev, block, size);
1366         return bh;
1367 }
1368 EXPORT_SYMBOL(__getblk);
1369
1370 /*
1371  * Do async read-ahead on a buffer..
1372  */
1373 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1374 {
1375         struct buffer_head *bh = __getblk(bdev, block, size);
1376         if (likely(bh)) {
1377                 ll_rw_block(READA, 1, &bh);
1378                 brelse(bh);
1379         }
1380 }
1381 EXPORT_SYMBOL(__breadahead);
1382
1383 /**
1384  *  __bread() - reads a specified block and returns the bh
1385  *  @bdev: the block_device to read from
1386  *  @block: number of block
1387  *  @size: size (in bytes) to read
1388  * 
1389  *  Reads a specified block, and returns buffer head that contains it.
1390  *  It returns NULL if the block was unreadable.
1391  */
1392 struct buffer_head *
1393 __bread(struct block_device *bdev, sector_t block, unsigned size)
1394 {
1395         struct buffer_head *bh = __getblk(bdev, block, size);
1396
1397         if (likely(bh) && !buffer_uptodate(bh))
1398                 bh = __bread_slow(bh);
1399         return bh;
1400 }
1401 EXPORT_SYMBOL(__bread);
1402
1403 /*
1404  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1405  * This doesn't race because it runs in each cpu either in irq
1406  * or with preempt disabled.
1407  */
1408 static void invalidate_bh_lru(void *arg)
1409 {
1410         struct bh_lru *b = &get_cpu_var(bh_lrus);
1411         int i;
1412
1413         for (i = 0; i < BH_LRU_SIZE; i++) {
1414                 brelse(b->bhs[i]);
1415                 b->bhs[i] = NULL;
1416         }
1417         put_cpu_var(bh_lrus);
1418 }
1419         
1420 void invalidate_bh_lrus(void)
1421 {
1422         on_each_cpu(invalidate_bh_lru, NULL, 1, 1);
1423 }
1424
1425 void set_bh_page(struct buffer_head *bh,
1426                 struct page *page, unsigned long offset)
1427 {
1428         bh->b_page = page;
1429         BUG_ON(offset >= PAGE_SIZE);
1430         if (PageHighMem(page))
1431                 /*
1432                  * This catches illegal uses and preserves the offset:
1433                  */
1434                 bh->b_data = (char *)(0 + offset);
1435         else
1436                 bh->b_data = page_address(page) + offset;
1437 }
1438 EXPORT_SYMBOL(set_bh_page);
1439
1440 /*
1441  * Called when truncating a buffer on a page completely.
1442  */
1443 static void discard_buffer(struct buffer_head * bh)
1444 {
1445         lock_buffer(bh);
1446         clear_buffer_dirty(bh);
1447         bh->b_bdev = NULL;
1448         clear_buffer_mapped(bh);
1449         clear_buffer_req(bh);
1450         clear_buffer_new(bh);
1451         clear_buffer_delay(bh);
1452         clear_buffer_unwritten(bh);
1453         unlock_buffer(bh);
1454 }
1455
1456 /**
1457  * block_invalidatepage - invalidate part of all of a buffer-backed page
1458  *
1459  * @page: the page which is affected
1460  * @offset: the index of the truncation point
1461  *
1462  * block_invalidatepage() is called when all or part of the page has become
1463  * invalidatedby a truncate operation.
1464  *
1465  * block_invalidatepage() does not have to release all buffers, but it must
1466  * ensure that no dirty buffer is left outside @offset and that no I/O
1467  * is underway against any of the blocks which are outside the truncation
1468  * point.  Because the caller is about to free (and possibly reuse) those
1469  * blocks on-disk.
1470  */
1471 void block_invalidatepage(struct page *page, unsigned long offset)
1472 {
1473         struct buffer_head *head, *bh, *next;
1474         unsigned int curr_off = 0;
1475
1476         BUG_ON(!PageLocked(page));
1477         if (!page_has_buffers(page))
1478                 goto out;
1479
1480         head = page_buffers(page);
1481         bh = head;
1482         do {
1483                 unsigned int next_off = curr_off + bh->b_size;
1484                 next = bh->b_this_page;
1485
1486                 /*
1487                  * is this block fully invalidated?
1488                  */
1489                 if (offset <= curr_off)
1490                         discard_buffer(bh);
1491                 curr_off = next_off;
1492                 bh = next;
1493         } while (bh != head);
1494
1495         /*
1496          * We release buffers only if the entire page is being invalidated.
1497          * The get_block cached value has been unconditionally invalidated,
1498          * so real IO is not possible anymore.
1499          */
1500         if (offset == 0)
1501                 try_to_release_page(page, 0);
1502 out:
1503         return;
1504 }
1505 EXPORT_SYMBOL(block_invalidatepage);
1506
1507 /*
1508  * We attach and possibly dirty the buffers atomically wrt
1509  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1510  * is already excluded via the page lock.
1511  */
1512 void create_empty_buffers(struct page *page,
1513                         unsigned long blocksize, unsigned long b_state)
1514 {
1515         struct buffer_head *bh, *head, *tail;
1516
1517         head = alloc_page_buffers(page, blocksize, 1);
1518         bh = head;
1519         do {
1520                 bh->b_state |= b_state;
1521                 tail = bh;
1522                 bh = bh->b_this_page;
1523         } while (bh);
1524         tail->b_this_page = head;
1525
1526         spin_lock(&page->mapping->private_lock);
1527         if (PageUptodate(page) || PageDirty(page)) {
1528                 bh = head;
1529                 do {
1530                         if (PageDirty(page))
1531                                 set_buffer_dirty(bh);
1532                         if (PageUptodate(page))
1533                                 set_buffer_uptodate(bh);
1534                         bh = bh->b_this_page;
1535                 } while (bh != head);
1536         }
1537         attach_page_buffers(page, head);
1538         spin_unlock(&page->mapping->private_lock);
1539 }
1540 EXPORT_SYMBOL(create_empty_buffers);
1541
1542 /*
1543  * We are taking a block for data and we don't want any output from any
1544  * buffer-cache aliases starting from return from that function and
1545  * until the moment when something will explicitly mark the buffer
1546  * dirty (hopefully that will not happen until we will free that block ;-)
1547  * We don't even need to mark it not-uptodate - nobody can expect
1548  * anything from a newly allocated buffer anyway. We used to used
1549  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1550  * don't want to mark the alias unmapped, for example - it would confuse
1551  * anyone who might pick it with bread() afterwards...
1552  *
1553  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1554  * be writeout I/O going on against recently-freed buffers.  We don't
1555  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1556  * only if we really need to.  That happens here.
1557  */
1558 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1559 {
1560         struct buffer_head *old_bh;
1561
1562         might_sleep();
1563
1564         old_bh = __find_get_block_slow(bdev, block);
1565         if (old_bh) {
1566                 clear_buffer_dirty(old_bh);
1567                 wait_on_buffer(old_bh);
1568                 clear_buffer_req(old_bh);
1569                 __brelse(old_bh);
1570         }
1571 }
1572 EXPORT_SYMBOL(unmap_underlying_metadata);
1573
1574 /*
1575  * NOTE! All mapped/uptodate combinations are valid:
1576  *
1577  *      Mapped  Uptodate        Meaning
1578  *
1579  *      No      No              "unknown" - must do get_block()
1580  *      No      Yes             "hole" - zero-filled
1581  *      Yes     No              "allocated" - allocated on disk, not read in
1582  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1583  *
1584  * "Dirty" is valid only with the last case (mapped+uptodate).
1585  */
1586
1587 /*
1588  * While block_write_full_page is writing back the dirty buffers under
1589  * the page lock, whoever dirtied the buffers may decide to clean them
1590  * again at any time.  We handle that by only looking at the buffer
1591  * state inside lock_buffer().
1592  *
1593  * If block_write_full_page() is called for regular writeback
1594  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1595  * locked buffer.   This only can happen if someone has written the buffer
1596  * directly, with submit_bh().  At the address_space level PageWriteback
1597  * prevents this contention from occurring.
1598  */
1599 static int __block_write_full_page(struct inode *inode, struct page *page,
1600                         get_block_t *get_block, struct writeback_control *wbc)
1601 {
1602         int err;
1603         sector_t block;
1604         sector_t last_block;
1605         struct buffer_head *bh, *head;
1606         const unsigned blocksize = 1 << inode->i_blkbits;
1607         int nr_underway = 0;
1608
1609         BUG_ON(!PageLocked(page));
1610
1611         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1612
1613         if (!page_has_buffers(page)) {
1614                 create_empty_buffers(page, blocksize,
1615                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1616         }
1617
1618         /*
1619          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1620          * here, and the (potentially unmapped) buffers may become dirty at
1621          * any time.  If a buffer becomes dirty here after we've inspected it
1622          * then we just miss that fact, and the page stays dirty.
1623          *
1624          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1625          * handle that here by just cleaning them.
1626          */
1627
1628         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1629         head = page_buffers(page);
1630         bh = head;
1631
1632         /*
1633          * Get all the dirty buffers mapped to disk addresses and
1634          * handle any aliases from the underlying blockdev's mapping.
1635          */
1636         do {
1637                 if (block > last_block) {
1638                         /*
1639                          * mapped buffers outside i_size will occur, because
1640                          * this page can be outside i_size when there is a
1641                          * truncate in progress.
1642                          */
1643                         /*
1644                          * The buffer was zeroed by block_write_full_page()
1645                          */
1646                         clear_buffer_dirty(bh);
1647                         set_buffer_uptodate(bh);
1648                 } else if (!buffer_mapped(bh) && buffer_dirty(bh)) {
1649                         WARN_ON(bh->b_size != blocksize);
1650                         err = get_block(inode, block, bh, 1);
1651                         if (err)
1652                                 goto recover;
1653                         if (buffer_new(bh)) {
1654                                 /* blockdev mappings never come here */
1655                                 clear_buffer_new(bh);
1656                                 unmap_underlying_metadata(bh->b_bdev,
1657                                                         bh->b_blocknr);
1658                         }
1659                 }
1660                 bh = bh->b_this_page;
1661                 block++;
1662         } while (bh != head);
1663
1664         do {
1665                 if (!buffer_mapped(bh))
1666                         continue;
1667                 /*
1668                  * If it's a fully non-blocking write attempt and we cannot
1669                  * lock the buffer then redirty the page.  Note that this can
1670                  * potentially cause a busy-wait loop from pdflush and kswapd
1671                  * activity, but those code paths have their own higher-level
1672                  * throttling.
1673                  */
1674                 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1675                         lock_buffer(bh);
1676                 } else if (test_set_buffer_locked(bh)) {
1677                         redirty_page_for_writepage(wbc, page);
1678                         continue;
1679                 }
1680                 if (test_clear_buffer_dirty(bh)) {
1681                         mark_buffer_async_write(bh);
1682                 } else {
1683                         unlock_buffer(bh);
1684                 }
1685         } while ((bh = bh->b_this_page) != head);
1686
1687         /*
1688          * The page and its buffers are protected by PageWriteback(), so we can
1689          * drop the bh refcounts early.
1690          */
1691         BUG_ON(PageWriteback(page));
1692         set_page_writeback(page);
1693
1694         do {
1695                 struct buffer_head *next = bh->b_this_page;
1696                 if (buffer_async_write(bh)) {
1697                         submit_bh(WRITE, bh);
1698                         nr_underway++;
1699                 }
1700                 bh = next;
1701         } while (bh != head);
1702         unlock_page(page);
1703
1704         err = 0;
1705 done:
1706         if (nr_underway == 0) {
1707                 /*
1708                  * The page was marked dirty, but the buffers were
1709                  * clean.  Someone wrote them back by hand with
1710                  * ll_rw_block/submit_bh.  A rare case.
1711                  */
1712                 end_page_writeback(page);
1713
1714                 /*
1715                  * The page and buffer_heads can be released at any time from
1716                  * here on.
1717                  */
1718                 wbc->pages_skipped++;   /* We didn't write this page */
1719         }
1720         return err;
1721
1722 recover:
1723         /*
1724          * ENOSPC, or some other error.  We may already have added some
1725          * blocks to the file, so we need to write these out to avoid
1726          * exposing stale data.
1727          * The page is currently locked and not marked for writeback
1728          */
1729         bh = head;
1730         /* Recovery: lock and submit the mapped buffers */
1731         do {
1732                 if (buffer_mapped(bh) && buffer_dirty(bh)) {
1733                         lock_buffer(bh);
1734                         mark_buffer_async_write(bh);
1735                 } else {
1736                         /*
1737                          * The buffer may have been set dirty during
1738                          * attachment to a dirty page.
1739                          */
1740                         clear_buffer_dirty(bh);
1741                 }
1742         } while ((bh = bh->b_this_page) != head);
1743         SetPageError(page);
1744         BUG_ON(PageWriteback(page));
1745         mapping_set_error(page->mapping, err);
1746         set_page_writeback(page);
1747         do {
1748                 struct buffer_head *next = bh->b_this_page;
1749                 if (buffer_async_write(bh)) {
1750                         clear_buffer_dirty(bh);
1751                         submit_bh(WRITE, bh);
1752                         nr_underway++;
1753                 }
1754                 bh = next;
1755         } while (bh != head);
1756         unlock_page(page);
1757         goto done;
1758 }
1759
1760 static int __block_prepare_write(struct inode *inode, struct page *page,
1761                 unsigned from, unsigned to, get_block_t *get_block)
1762 {
1763         unsigned block_start, block_end;
1764         sector_t block;
1765         int err = 0;
1766         unsigned blocksize, bbits;
1767         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1768
1769         BUG_ON(!PageLocked(page));
1770         BUG_ON(from > PAGE_CACHE_SIZE);
1771         BUG_ON(to > PAGE_CACHE_SIZE);
1772         BUG_ON(from > to);
1773
1774         blocksize = 1 << inode->i_blkbits;
1775         if (!page_has_buffers(page))
1776                 create_empty_buffers(page, blocksize, 0);
1777         head = page_buffers(page);
1778
1779         bbits = inode->i_blkbits;
1780         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1781
1782         for(bh = head, block_start = 0; bh != head || !block_start;
1783             block++, block_start=block_end, bh = bh->b_this_page) {
1784                 block_end = block_start + blocksize;
1785                 if (block_end <= from || block_start >= to) {
1786                         if (PageUptodate(page)) {
1787                                 if (!buffer_uptodate(bh))
1788                                         set_buffer_uptodate(bh);
1789                         }
1790                         continue;
1791                 }
1792                 if (buffer_new(bh))
1793                         clear_buffer_new(bh);
1794                 if (!buffer_mapped(bh)) {
1795                         WARN_ON(bh->b_size != blocksize);
1796                         err = get_block(inode, block, bh, 1);
1797                         if (err)
1798                                 break;
1799                         if (buffer_new(bh)) {
1800                                 unmap_underlying_metadata(bh->b_bdev,
1801                                                         bh->b_blocknr);
1802                                 if (PageUptodate(page)) {
1803                                         set_buffer_uptodate(bh);
1804                                         continue;
1805                                 }
1806                                 if (block_end > to || block_start < from) {
1807                                         void *kaddr;
1808
1809                                         kaddr = kmap_atomic(page, KM_USER0);
1810                                         if (block_end > to)
1811                                                 memset(kaddr+to, 0,
1812                                                         block_end-to);
1813                                         if (block_start < from)
1814                                                 memset(kaddr+block_start,
1815                                                         0, from-block_start);
1816                                         flush_dcache_page(page);
1817                                         kunmap_atomic(kaddr, KM_USER0);
1818                                 }
1819                                 continue;
1820                         }
1821                 }
1822                 if (PageUptodate(page)) {
1823                         if (!buffer_uptodate(bh))
1824                                 set_buffer_uptodate(bh);
1825                         continue; 
1826                 }
1827                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1828                     !buffer_unwritten(bh) &&
1829                      (block_start < from || block_end > to)) {
1830                         ll_rw_block(READ, 1, &bh);
1831                         *wait_bh++=bh;
1832                 }
1833         }
1834         /*
1835          * If we issued read requests - let them complete.
1836          */
1837         while(wait_bh > wait) {
1838                 wait_on_buffer(*--wait_bh);
1839                 if (!buffer_uptodate(*wait_bh))
1840                         err = -EIO;
1841         }
1842         if (!err) {
1843                 bh = head;
1844                 do {
1845                         if (buffer_new(bh))
1846                                 clear_buffer_new(bh);
1847                 } while ((bh = bh->b_this_page) != head);
1848                 return 0;
1849         }
1850         /* Error case: */
1851         /*
1852          * Zero out any newly allocated blocks to avoid exposing stale
1853          * data.  If BH_New is set, we know that the block was newly
1854          * allocated in the above loop.
1855          */
1856         bh = head;
1857         block_start = 0;
1858         do {
1859                 block_end = block_start+blocksize;
1860                 if (block_end <= from)
1861                         goto next_bh;
1862                 if (block_start >= to)
1863                         break;
1864                 if (buffer_new(bh)) {
1865                         clear_buffer_new(bh);
1866                         zero_user_page(page, block_start, bh->b_size, KM_USER0);
1867                         set_buffer_uptodate(bh);
1868                         mark_buffer_dirty(bh);
1869                 }
1870 next_bh:
1871                 block_start = block_end;
1872                 bh = bh->b_this_page;
1873         } while (bh != head);
1874         return err;
1875 }
1876
1877 static int __block_commit_write(struct inode *inode, struct page *page,
1878                 unsigned from, unsigned to)
1879 {
1880         unsigned block_start, block_end;
1881         int partial = 0;
1882         unsigned blocksize;
1883         struct buffer_head *bh, *head;
1884
1885         blocksize = 1 << inode->i_blkbits;
1886
1887         for(bh = head = page_buffers(page), block_start = 0;
1888             bh != head || !block_start;
1889             block_start=block_end, bh = bh->b_this_page) {
1890                 block_end = block_start + blocksize;
1891                 if (block_end <= from || block_start >= to) {
1892                         if (!buffer_uptodate(bh))
1893                                 partial = 1;
1894                 } else {
1895                         set_buffer_uptodate(bh);
1896                         mark_buffer_dirty(bh);
1897                 }
1898         }
1899
1900         /*
1901          * If this is a partial write which happened to make all buffers
1902          * uptodate then we can optimize away a bogus readpage() for
1903          * the next read(). Here we 'discover' whether the page went
1904          * uptodate as a result of this (potentially partial) write.
1905          */
1906         if (!partial)
1907                 SetPageUptodate(page);
1908         return 0;
1909 }
1910
1911 /*
1912  * Generic "read page" function for block devices that have the normal
1913  * get_block functionality. This is most of the block device filesystems.
1914  * Reads the page asynchronously --- the unlock_buffer() and
1915  * set/clear_buffer_uptodate() functions propagate buffer state into the
1916  * page struct once IO has completed.
1917  */
1918 int block_read_full_page(struct page *page, get_block_t *get_block)
1919 {
1920         struct inode *inode = page->mapping->host;
1921         sector_t iblock, lblock;
1922         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
1923         unsigned int blocksize;
1924         int nr, i;
1925         int fully_mapped = 1;
1926
1927         BUG_ON(!PageLocked(page));
1928         blocksize = 1 << inode->i_blkbits;
1929         if (!page_has_buffers(page))
1930                 create_empty_buffers(page, blocksize, 0);
1931         head = page_buffers(page);
1932
1933         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1934         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
1935         bh = head;
1936         nr = 0;
1937         i = 0;
1938
1939         do {
1940                 if (buffer_uptodate(bh))
1941                         continue;
1942
1943                 if (!buffer_mapped(bh)) {
1944                         int err = 0;
1945
1946                         fully_mapped = 0;
1947                         if (iblock < lblock) {
1948                                 WARN_ON(bh->b_size != blocksize);
1949                                 err = get_block(inode, iblock, bh, 0);
1950                                 if (err)
1951                                         SetPageError(page);
1952                         }
1953                         if (!buffer_mapped(bh)) {
1954                                 zero_user_page(page, i * blocksize, blocksize,
1955                                                 KM_USER0);
1956                                 if (!err)
1957                                         set_buffer_uptodate(bh);
1958                                 continue;
1959                         }
1960                         /*
1961                          * get_block() might have updated the buffer
1962                          * synchronously
1963                          */
1964                         if (buffer_uptodate(bh))
1965                                 continue;
1966                 }
1967                 arr[nr++] = bh;
1968         } while (i++, iblock++, (bh = bh->b_this_page) != head);
1969
1970         if (fully_mapped)
1971                 SetPageMappedToDisk(page);
1972
1973         if (!nr) {
1974                 /*
1975                  * All buffers are uptodate - we can set the page uptodate
1976                  * as well. But not if get_block() returned an error.
1977                  */
1978                 if (!PageError(page))
1979                         SetPageUptodate(page);
1980                 unlock_page(page);
1981                 return 0;
1982         }
1983
1984         /* Stage two: lock the buffers */
1985         for (i = 0; i < nr; i++) {
1986                 bh = arr[i];
1987                 lock_buffer(bh);
1988                 mark_buffer_async_read(bh);
1989         }
1990
1991         /*
1992          * Stage 3: start the IO.  Check for uptodateness
1993          * inside the buffer lock in case another process reading
1994          * the underlying blockdev brought it uptodate (the sct fix).
1995          */
1996         for (i = 0; i < nr; i++) {
1997                 bh = arr[i];
1998                 if (buffer_uptodate(bh))
1999                         end_buffer_async_read(bh, 1);
2000                 else
2001                         submit_bh(READ, bh);
2002         }
2003         return 0;
2004 }
2005
2006 /* utility function for filesystems that need to do work on expanding
2007  * truncates.  Uses prepare/commit_write to allow the filesystem to
2008  * deal with the hole.  
2009  */
2010 static int __generic_cont_expand(struct inode *inode, loff_t size,
2011                                  pgoff_t index, unsigned int offset)
2012 {
2013         struct address_space *mapping = inode->i_mapping;
2014         struct page *page;
2015         unsigned long limit;
2016         int err;
2017
2018         err = -EFBIG;
2019         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2020         if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2021                 send_sig(SIGXFSZ, current, 0);
2022                 goto out;
2023         }
2024         if (size > inode->i_sb->s_maxbytes)
2025                 goto out;
2026
2027         err = -ENOMEM;
2028         page = grab_cache_page(mapping, index);
2029         if (!page)
2030                 goto out;
2031         err = mapping->a_ops->prepare_write(NULL, page, offset, offset);
2032         if (err) {
2033                 /*
2034                  * ->prepare_write() may have instantiated a few blocks
2035                  * outside i_size.  Trim these off again.
2036                  */
2037                 unlock_page(page);
2038                 page_cache_release(page);
2039                 vmtruncate(inode, inode->i_size);
2040                 goto out;
2041         }
2042
2043         err = mapping->a_ops->commit_write(NULL, page, offset, offset);
2044
2045         unlock_page(page);
2046         page_cache_release(page);
2047         if (err > 0)
2048                 err = 0;
2049 out:
2050         return err;
2051 }
2052
2053 int generic_cont_expand(struct inode *inode, loff_t size)
2054 {
2055         pgoff_t index;
2056         unsigned int offset;
2057
2058         offset = (size & (PAGE_CACHE_SIZE - 1)); /* Within page */
2059
2060         /* ugh.  in prepare/commit_write, if from==to==start of block, we
2061         ** skip the prepare.  make sure we never send an offset for the start
2062         ** of a block
2063         */
2064         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
2065                 /* caller must handle this extra byte. */
2066                 offset++;
2067         }
2068         index = size >> PAGE_CACHE_SHIFT;
2069
2070         return __generic_cont_expand(inode, size, index, offset);
2071 }
2072
2073 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2074 {
2075         loff_t pos = size - 1;
2076         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
2077         unsigned int offset = (pos & (PAGE_CACHE_SIZE - 1)) + 1;
2078
2079         /* prepare/commit_write can handle even if from==to==start of block. */
2080         return __generic_cont_expand(inode, size, index, offset);
2081 }
2082
2083 /*
2084  * For moronic filesystems that do not allow holes in file.
2085  * We may have to extend the file.
2086  */
2087
2088 int cont_prepare_write(struct page *page, unsigned offset,
2089                 unsigned to, get_block_t *get_block, loff_t *bytes)
2090 {
2091         struct address_space *mapping = page->mapping;
2092         struct inode *inode = mapping->host;
2093         struct page *new_page;
2094         pgoff_t pgpos;
2095         long status;
2096         unsigned zerofrom;
2097         unsigned blocksize = 1 << inode->i_blkbits;
2098
2099         while(page->index > (pgpos = *bytes>>PAGE_CACHE_SHIFT)) {
2100                 status = -ENOMEM;
2101                 new_page = grab_cache_page(mapping, pgpos);
2102                 if (!new_page)
2103                         goto out;
2104                 /* we might sleep */
2105                 if (*bytes>>PAGE_CACHE_SHIFT != pgpos) {
2106                         unlock_page(new_page);
2107                         page_cache_release(new_page);
2108                         continue;
2109                 }
2110                 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2111                 if (zerofrom & (blocksize-1)) {
2112                         *bytes |= (blocksize-1);
2113                         (*bytes)++;
2114                 }
2115                 status = __block_prepare_write(inode, new_page, zerofrom,
2116                                                 PAGE_CACHE_SIZE, get_block);
2117                 if (status)
2118                         goto out_unmap;
2119                 zero_user_page(new_page, zerofrom, PAGE_CACHE_SIZE - zerofrom,
2120                                 KM_USER0);
2121                 generic_commit_write(NULL, new_page, zerofrom, PAGE_CACHE_SIZE);
2122                 unlock_page(new_page);
2123                 page_cache_release(new_page);
2124         }
2125
2126         if (page->index < pgpos) {
2127                 /* completely inside the area */
2128                 zerofrom = offset;
2129         } else {
2130                 /* page covers the boundary, find the boundary offset */
2131                 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2132
2133                 /* if we will expand the thing last block will be filled */
2134                 if (to > zerofrom && (zerofrom & (blocksize-1))) {
2135                         *bytes |= (blocksize-1);
2136                         (*bytes)++;
2137                 }
2138
2139                 /* starting below the boundary? Nothing to zero out */
2140                 if (offset <= zerofrom)
2141                         zerofrom = offset;
2142         }
2143         status = __block_prepare_write(inode, page, zerofrom, to, get_block);
2144         if (status)
2145                 goto out1;
2146         if (zerofrom < offset) {
2147                 zero_user_page(page, zerofrom, offset - zerofrom, KM_USER0);
2148                 __block_commit_write(inode, page, zerofrom, offset);
2149         }
2150         return 0;
2151 out1:
2152         ClearPageUptodate(page);
2153         return status;
2154
2155 out_unmap:
2156         ClearPageUptodate(new_page);
2157         unlock_page(new_page);
2158         page_cache_release(new_page);
2159 out:
2160         return status;
2161 }
2162
2163 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2164                         get_block_t *get_block)
2165 {
2166         struct inode *inode = page->mapping->host;
2167         int err = __block_prepare_write(inode, page, from, to, get_block);
2168         if (err)
2169                 ClearPageUptodate(page);
2170         return err;
2171 }
2172
2173 int block_commit_write(struct page *page, unsigned from, unsigned to)
2174 {
2175         struct inode *inode = page->mapping->host;
2176         __block_commit_write(inode,page,from,to);
2177         return 0;
2178 }
2179
2180 int generic_commit_write(struct file *file, struct page *page,
2181                 unsigned from, unsigned to)
2182 {
2183         struct inode *inode = page->mapping->host;
2184         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2185         __block_commit_write(inode,page,from,to);
2186         /*
2187          * No need to use i_size_read() here, the i_size
2188          * cannot change under us because we hold i_mutex.
2189          */
2190         if (pos > inode->i_size) {
2191                 i_size_write(inode, pos);
2192                 mark_inode_dirty(inode);
2193         }
2194         return 0;
2195 }
2196
2197
2198 /*
2199  * nobh_prepare_write()'s prereads are special: the buffer_heads are freed
2200  * immediately, while under the page lock.  So it needs a special end_io
2201  * handler which does not touch the bh after unlocking it.
2202  *
2203  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
2204  * a race there is benign: unlock_buffer() only use the bh's address for
2205  * hashing after unlocking the buffer, so it doesn't actually touch the bh
2206  * itself.
2207  */
2208 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2209 {
2210         if (uptodate) {
2211                 set_buffer_uptodate(bh);
2212         } else {
2213                 /* This happens, due to failed READA attempts. */
2214                 clear_buffer_uptodate(bh);
2215         }
2216         unlock_buffer(bh);
2217 }
2218
2219 /*
2220  * On entry, the page is fully not uptodate.
2221  * On exit the page is fully uptodate in the areas outside (from,to)
2222  */
2223 int nobh_prepare_write(struct page *page, unsigned from, unsigned to,
2224                         get_block_t *get_block)
2225 {
2226         struct inode *inode = page->mapping->host;
2227         const unsigned blkbits = inode->i_blkbits;
2228         const unsigned blocksize = 1 << blkbits;
2229         struct buffer_head map_bh;
2230         struct buffer_head *read_bh[MAX_BUF_PER_PAGE];
2231         unsigned block_in_page;
2232         unsigned block_start;
2233         sector_t block_in_file;
2234         char *kaddr;
2235         int nr_reads = 0;
2236         int i;
2237         int ret = 0;
2238         int is_mapped_to_disk = 1;
2239
2240         if (PageMappedToDisk(page))
2241                 return 0;
2242
2243         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2244         map_bh.b_page = page;
2245
2246         /*
2247          * We loop across all blocks in the page, whether or not they are
2248          * part of the affected region.  This is so we can discover if the
2249          * page is fully mapped-to-disk.
2250          */
2251         for (block_start = 0, block_in_page = 0;
2252                   block_start < PAGE_CACHE_SIZE;
2253                   block_in_page++, block_start += blocksize) {
2254                 unsigned block_end = block_start + blocksize;
2255                 int create;
2256
2257                 map_bh.b_state = 0;
2258                 create = 1;
2259                 if (block_start >= to)
2260                         create = 0;
2261                 map_bh.b_size = blocksize;
2262                 ret = get_block(inode, block_in_file + block_in_page,
2263                                         &map_bh, create);
2264                 if (ret)
2265                         goto failed;
2266                 if (!buffer_mapped(&map_bh))
2267                         is_mapped_to_disk = 0;
2268                 if (buffer_new(&map_bh))
2269                         unmap_underlying_metadata(map_bh.b_bdev,
2270                                                         map_bh.b_blocknr);
2271                 if (PageUptodate(page))
2272                         continue;
2273                 if (buffer_new(&map_bh) || !buffer_mapped(&map_bh)) {
2274                         kaddr = kmap_atomic(page, KM_USER0);
2275                         if (block_start < from)
2276                                 memset(kaddr+block_start, 0, from-block_start);
2277                         if (block_end > to)
2278                                 memset(kaddr + to, 0, block_end - to);
2279                         flush_dcache_page(page);
2280                         kunmap_atomic(kaddr, KM_USER0);
2281                         continue;
2282                 }
2283                 if (buffer_uptodate(&map_bh))
2284                         continue;       /* reiserfs does this */
2285                 if (block_start < from || block_end > to) {
2286                         struct buffer_head *bh = alloc_buffer_head(GFP_NOFS);
2287
2288                         if (!bh) {
2289                                 ret = -ENOMEM;
2290                                 goto failed;
2291                         }
2292                         bh->b_state = map_bh.b_state;
2293                         atomic_set(&bh->b_count, 0);
2294                         bh->b_this_page = NULL;
2295                         bh->b_page = page;
2296                         bh->b_blocknr = map_bh.b_blocknr;
2297                         bh->b_size = blocksize;
2298                         bh->b_data = (char *)(long)block_start;
2299                         bh->b_bdev = map_bh.b_bdev;
2300                         bh->b_private = NULL;
2301                         read_bh[nr_reads++] = bh;
2302                 }
2303         }
2304
2305         if (nr_reads) {
2306                 struct buffer_head *bh;
2307
2308                 /*
2309                  * The page is locked, so these buffers are protected from
2310                  * any VM or truncate activity.  Hence we don't need to care
2311                  * for the buffer_head refcounts.
2312                  */
2313                 for (i = 0; i < nr_reads; i++) {
2314                         bh = read_bh[i];
2315                         lock_buffer(bh);
2316                         bh->b_end_io = end_buffer_read_nobh;
2317                         submit_bh(READ, bh);
2318                 }
2319                 for (i = 0; i < nr_reads; i++) {
2320                         bh = read_bh[i];
2321                         wait_on_buffer(bh);
2322                         if (!buffer_uptodate(bh))
2323                                 ret = -EIO;
2324                         free_buffer_head(bh);
2325                         read_bh[i] = NULL;
2326                 }
2327                 if (ret)
2328                         goto failed;
2329         }
2330
2331         if (is_mapped_to_disk)
2332                 SetPageMappedToDisk(page);
2333
2334         return 0;
2335
2336 failed:
2337         for (i = 0; i < nr_reads; i++) {
2338                 if (read_bh[i])
2339                         free_buffer_head(read_bh[i]);
2340         }
2341
2342         /*
2343          * Error recovery is pretty slack.  Clear the page and mark it dirty
2344          * so we'll later zero out any blocks which _were_ allocated.
2345          */
2346         zero_user_page(page, 0, PAGE_CACHE_SIZE, KM_USER0);
2347         SetPageUptodate(page);
2348         set_page_dirty(page);
2349         return ret;
2350 }
2351 EXPORT_SYMBOL(nobh_prepare_write);
2352
2353 /*
2354  * Make sure any changes to nobh_commit_write() are reflected in
2355  * nobh_truncate_page(), since it doesn't call commit_write().
2356  */
2357 int nobh_commit_write(struct file *file, struct page *page,
2358                 unsigned from, unsigned to)
2359 {
2360         struct inode *inode = page->mapping->host;
2361         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2362
2363         SetPageUptodate(page);
2364         set_page_dirty(page);
2365         if (pos > inode->i_size) {
2366                 i_size_write(inode, pos);
2367                 mark_inode_dirty(inode);
2368         }
2369         return 0;
2370 }
2371 EXPORT_SYMBOL(nobh_commit_write);
2372
2373 /*
2374  * nobh_writepage() - based on block_full_write_page() except
2375  * that it tries to operate without attaching bufferheads to
2376  * the page.
2377  */
2378 int nobh_writepage(struct page *page, get_block_t *get_block,
2379                         struct writeback_control *wbc)
2380 {
2381         struct inode * const inode = page->mapping->host;
2382         loff_t i_size = i_size_read(inode);
2383         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2384         unsigned offset;
2385         int ret;
2386
2387         /* Is the page fully inside i_size? */
2388         if (page->index < end_index)
2389                 goto out;
2390
2391         /* Is the page fully outside i_size? (truncate in progress) */
2392         offset = i_size & (PAGE_CACHE_SIZE-1);
2393         if (page->index >= end_index+1 || !offset) {
2394                 /*
2395                  * The page may have dirty, unmapped buffers.  For example,
2396                  * they may have been added in ext3_writepage().  Make them
2397                  * freeable here, so the page does not leak.
2398                  */
2399 #if 0
2400                 /* Not really sure about this  - do we need this ? */
2401                 if (page->mapping->a_ops->invalidatepage)
2402                         page->mapping->a_ops->invalidatepage(page, offset);
2403 #endif
2404                 unlock_page(page);
2405                 return 0; /* don't care */
2406         }
2407
2408         /*
2409          * The page straddles i_size.  It must be zeroed out on each and every
2410          * writepage invocation because it may be mmapped.  "A file is mapped
2411          * in multiples of the page size.  For a file that is not a multiple of
2412          * the  page size, the remaining memory is zeroed when mapped, and
2413          * writes to that region are not written out to the file."
2414          */
2415         zero_user_page(page, offset, PAGE_CACHE_SIZE - offset, KM_USER0);
2416 out:
2417         ret = mpage_writepage(page, get_block, wbc);
2418         if (ret == -EAGAIN)
2419                 ret = __block_write_full_page(inode, page, get_block, wbc);
2420         return ret;
2421 }
2422 EXPORT_SYMBOL(nobh_writepage);
2423
2424 /*
2425  * This function assumes that ->prepare_write() uses nobh_prepare_write().
2426  */
2427 int nobh_truncate_page(struct address_space *mapping, loff_t from)
2428 {
2429         struct inode *inode = mapping->host;
2430         unsigned blocksize = 1 << inode->i_blkbits;
2431         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2432         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2433         unsigned to;
2434         struct page *page;
2435         const struct address_space_operations *a_ops = mapping->a_ops;
2436         int ret = 0;
2437
2438         if ((offset & (blocksize - 1)) == 0)
2439                 goto out;
2440
2441         ret = -ENOMEM;
2442         page = grab_cache_page(mapping, index);
2443         if (!page)
2444                 goto out;
2445
2446         to = (offset + blocksize) & ~(blocksize - 1);
2447         ret = a_ops->prepare_write(NULL, page, offset, to);
2448         if (ret == 0) {
2449                 zero_user_page(page, offset, PAGE_CACHE_SIZE - offset,
2450                                 KM_USER0);
2451                 /*
2452                  * It would be more correct to call aops->commit_write()
2453                  * here, but this is more efficient.
2454                  */
2455                 SetPageUptodate(page);
2456                 set_page_dirty(page);
2457         }
2458         unlock_page(page);
2459         page_cache_release(page);
2460 out:
2461         return ret;
2462 }
2463 EXPORT_SYMBOL(nobh_truncate_page);
2464
2465 int block_truncate_page(struct address_space *mapping,
2466                         loff_t from, get_block_t *get_block)
2467 {
2468         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2469         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2470         unsigned blocksize;
2471         sector_t iblock;
2472         unsigned length, pos;
2473         struct inode *inode = mapping->host;
2474         struct page *page;
2475         struct buffer_head *bh;
2476         int err;
2477
2478         blocksize = 1 << inode->i_blkbits;
2479         length = offset & (blocksize - 1);
2480
2481         /* Block boundary? Nothing to do */
2482         if (!length)
2483                 return 0;
2484
2485         length = blocksize - length;
2486         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2487         
2488         page = grab_cache_page(mapping, index);
2489         err = -ENOMEM;
2490         if (!page)
2491                 goto out;
2492
2493         if (!page_has_buffers(page))
2494                 create_empty_buffers(page, blocksize, 0);
2495
2496         /* Find the buffer that contains "offset" */
2497         bh = page_buffers(page);
2498         pos = blocksize;
2499         while (offset >= pos) {
2500                 bh = bh->b_this_page;
2501                 iblock++;
2502                 pos += blocksize;
2503         }
2504
2505         err = 0;
2506         if (!buffer_mapped(bh)) {
2507                 WARN_ON(bh->b_size != blocksize);
2508                 err = get_block(inode, iblock, bh, 0);
2509                 if (err)
2510                         goto unlock;
2511                 /* unmapped? It's a hole - nothing to do */
2512                 if (!buffer_mapped(bh))
2513                         goto unlock;
2514         }
2515
2516         /* Ok, it's mapped. Make sure it's up-to-date */
2517         if (PageUptodate(page))
2518                 set_buffer_uptodate(bh);
2519
2520         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2521                 err = -EIO;
2522                 ll_rw_block(READ, 1, &bh);
2523                 wait_on_buffer(bh);
2524                 /* Uhhuh. Read error. Complain and punt. */
2525                 if (!buffer_uptodate(bh))
2526                         goto unlock;
2527         }
2528
2529         zero_user_page(page, offset, length, KM_USER0);
2530         mark_buffer_dirty(bh);
2531         err = 0;
2532
2533 unlock:
2534         unlock_page(page);
2535         page_cache_release(page);
2536 out:
2537         return err;
2538 }
2539
2540 /*
2541  * The generic ->writepage function for buffer-backed address_spaces
2542  */
2543 int block_write_full_page(struct page *page, get_block_t *get_block,
2544                         struct writeback_control *wbc)
2545 {
2546         struct inode * const inode = page->mapping->host;
2547         loff_t i_size = i_size_read(inode);
2548         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2549         unsigned offset;
2550
2551         /* Is the page fully inside i_size? */
2552         if (page->index < end_index)
2553                 return __block_write_full_page(inode, page, get_block, wbc);
2554
2555         /* Is the page fully outside i_size? (truncate in progress) */
2556         offset = i_size & (PAGE_CACHE_SIZE-1);
2557         if (page->index >= end_index+1 || !offset) {
2558                 /*
2559                  * The page may have dirty, unmapped buffers.  For example,
2560                  * they may have been added in ext3_writepage().  Make them
2561                  * freeable here, so the page does not leak.
2562                  */
2563                 do_invalidatepage(page, 0);
2564                 unlock_page(page);
2565                 return 0; /* don't care */
2566         }
2567
2568         /*
2569          * The page straddles i_size.  It must be zeroed out on each and every
2570          * writepage invokation because it may be mmapped.  "A file is mapped
2571          * in multiples of the page size.  For a file that is not a multiple of
2572          * the  page size, the remaining memory is zeroed when mapped, and
2573          * writes to that region are not written out to the file."
2574          */
2575         zero_user_page(page, offset, PAGE_CACHE_SIZE - offset, KM_USER0);
2576         return __block_write_full_page(inode, page, get_block, wbc);
2577 }
2578
2579 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2580                             get_block_t *get_block)
2581 {
2582         struct buffer_head tmp;
2583         struct inode *inode = mapping->host;
2584         tmp.b_state = 0;
2585         tmp.b_blocknr = 0;
2586         tmp.b_size = 1 << inode->i_blkbits;
2587         get_block(inode, block, &tmp, 0);
2588         return tmp.b_blocknr;
2589 }
2590
2591 static int end_bio_bh_io_sync(struct bio *bio, unsigned int bytes_done, int err)
2592 {
2593         struct buffer_head *bh = bio->bi_private;
2594
2595         if (bio->bi_size)
2596                 return 1;
2597
2598         if (err == -EOPNOTSUPP) {
2599                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2600                 set_bit(BH_Eopnotsupp, &bh->b_state);
2601         }
2602
2603         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2604         bio_put(bio);
2605         return 0;
2606 }
2607
2608 int submit_bh(int rw, struct buffer_head * bh)
2609 {
2610         struct bio *bio;
2611         int ret = 0;
2612
2613         BUG_ON(!buffer_locked(bh));
2614         BUG_ON(!buffer_mapped(bh));
2615         BUG_ON(!bh->b_end_io);
2616
2617         if (buffer_ordered(bh) && (rw == WRITE))
2618                 rw = WRITE_BARRIER;
2619
2620         /*
2621          * Only clear out a write error when rewriting, should this
2622          * include WRITE_SYNC as well?
2623          */
2624         if (test_set_buffer_req(bh) && (rw == WRITE || rw == WRITE_BARRIER))
2625                 clear_buffer_write_io_error(bh);
2626
2627         /*
2628          * from here on down, it's all bio -- do the initial mapping,
2629          * submit_bio -> generic_make_request may further map this bio around
2630          */
2631         bio = bio_alloc(GFP_NOIO, 1);
2632
2633         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2634         bio->bi_bdev = bh->b_bdev;
2635         bio->bi_io_vec[0].bv_page = bh->b_page;
2636         bio->bi_io_vec[0].bv_len = bh->b_size;
2637         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2638
2639         bio->bi_vcnt = 1;
2640         bio->bi_idx = 0;
2641         bio->bi_size = bh->b_size;
2642
2643         bio->bi_end_io = end_bio_bh_io_sync;
2644         bio->bi_private = bh;
2645
2646         bio_get(bio);
2647         submit_bio(rw, bio);
2648
2649         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2650                 ret = -EOPNOTSUPP;
2651
2652         bio_put(bio);
2653         return ret;
2654 }
2655
2656 /**
2657  * ll_rw_block: low-level access to block devices (DEPRECATED)
2658  * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
2659  * @nr: number of &struct buffer_heads in the array
2660  * @bhs: array of pointers to &struct buffer_head
2661  *
2662  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2663  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2664  * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2665  * are sent to disk. The fourth %READA option is described in the documentation
2666  * for generic_make_request() which ll_rw_block() calls.
2667  *
2668  * This function drops any buffer that it cannot get a lock on (with the
2669  * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
2670  * clean when doing a write request, and any buffer that appears to be
2671  * up-to-date when doing read request.  Further it marks as clean buffers that
2672  * are processed for writing (the buffer cache won't assume that they are
2673  * actually clean until the buffer gets unlocked).
2674  *
2675  * ll_rw_block sets b_end_io to simple completion handler that marks
2676  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2677  * any waiters. 
2678  *
2679  * All of the buffers must be for the same device, and must also be a
2680  * multiple of the current approved size for the device.
2681  */
2682 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2683 {
2684         int i;
2685
2686         for (i = 0; i < nr; i++) {
2687                 struct buffer_head *bh = bhs[i];
2688
2689                 if (rw == SWRITE)
2690                         lock_buffer(bh);
2691                 else if (test_set_buffer_locked(bh))
2692                         continue;
2693
2694                 if (rw == WRITE || rw == SWRITE) {
2695                         if (test_clear_buffer_dirty(bh)) {
2696                                 bh->b_end_io = end_buffer_write_sync;
2697                                 get_bh(bh);
2698                                 submit_bh(WRITE, bh);
2699                                 continue;
2700                         }
2701                 } else {
2702                         if (!buffer_uptodate(bh)) {
2703                                 bh->b_end_io = end_buffer_read_sync;
2704                                 get_bh(bh);
2705                                 submit_bh(rw, bh);
2706                                 continue;
2707                         }
2708                 }
2709                 unlock_buffer(bh);
2710         }
2711 }
2712
2713 /*
2714  * For a data-integrity writeout, we need to wait upon any in-progress I/O
2715  * and then start new I/O and then wait upon it.  The caller must have a ref on
2716  * the buffer_head.
2717  */
2718 int sync_dirty_buffer(struct buffer_head *bh)
2719 {
2720         int ret = 0;
2721
2722         WARN_ON(atomic_read(&bh->b_count) < 1);
2723         lock_buffer(bh);
2724         if (test_clear_buffer_dirty(bh)) {
2725                 get_bh(bh);
2726                 bh->b_end_io = end_buffer_write_sync;
2727                 ret = submit_bh(WRITE, bh);
2728                 wait_on_buffer(bh);
2729                 if (buffer_eopnotsupp(bh)) {
2730                         clear_buffer_eopnotsupp(bh);
2731                         ret = -EOPNOTSUPP;
2732                 }
2733                 if (!ret && !buffer_uptodate(bh))
2734                         ret = -EIO;
2735         } else {
2736                 unlock_buffer(bh);
2737         }
2738         return ret;
2739 }
2740
2741 /*
2742  * try_to_free_buffers() checks if all the buffers on this particular page
2743  * are unused, and releases them if so.
2744  *
2745  * Exclusion against try_to_free_buffers may be obtained by either
2746  * locking the page or by holding its mapping's private_lock.
2747  *
2748  * If the page is dirty but all the buffers are clean then we need to
2749  * be sure to mark the page clean as well.  This is because the page
2750  * may be against a block device, and a later reattachment of buffers
2751  * to a dirty page will set *all* buffers dirty.  Which would corrupt
2752  * filesystem data on the same device.
2753  *
2754  * The same applies to regular filesystem pages: if all the buffers are
2755  * clean then we set the page clean and proceed.  To do that, we require
2756  * total exclusion from __set_page_dirty_buffers().  That is obtained with
2757  * private_lock.
2758  *
2759  * try_to_free_buffers() is non-blocking.
2760  */
2761 static inline int buffer_busy(struct buffer_head *bh)
2762 {
2763         return atomic_read(&bh->b_count) |
2764                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
2765 }
2766
2767 static int
2768 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
2769 {
2770         struct buffer_head *head = page_buffers(page);
2771         struct buffer_head *bh;
2772
2773         bh = head;
2774         do {
2775                 if (buffer_write_io_error(bh) && page->mapping)
2776                         set_bit(AS_EIO, &page->mapping->flags);
2777                 if (buffer_busy(bh))
2778                         goto failed;
2779                 bh = bh->b_this_page;
2780         } while (bh != head);
2781
2782         do {
2783                 struct buffer_head *next = bh->b_this_page;
2784
2785                 if (!list_empty(&bh->b_assoc_buffers))
2786                         __remove_assoc_queue(bh);
2787                 bh = next;
2788         } while (bh != head);
2789         *buffers_to_free = head;
2790         __clear_page_buffers(page);
2791         return 1;
2792 failed:
2793         return 0;
2794 }
2795
2796 int try_to_free_buffers(struct page *page)
2797 {
2798         struct address_space * const mapping = page->mapping;
2799         struct buffer_head *buffers_to_free = NULL;
2800         int ret = 0;
2801
2802         BUG_ON(!PageLocked(page));
2803         if (PageWriteback(page))
2804                 return 0;
2805
2806         if (mapping == NULL) {          /* can this still happen? */
2807                 ret = drop_buffers(page, &buffers_to_free);
2808                 goto out;
2809         }
2810
2811         spin_lock(&mapping->private_lock);
2812         ret = drop_buffers(page, &buffers_to_free);
2813
2814         /*
2815          * If the filesystem writes its buffers by hand (eg ext3)
2816          * then we can have clean buffers against a dirty page.  We
2817          * clean the page here; otherwise the VM will never notice
2818          * that the filesystem did any IO at all.
2819          *
2820          * Also, during truncate, discard_buffer will have marked all
2821          * the page's buffers clean.  We discover that here and clean
2822          * the page also.
2823          *
2824          * private_lock must be held over this entire operation in order
2825          * to synchronise against __set_page_dirty_buffers and prevent the
2826          * dirty bit from being lost.
2827          */
2828         if (ret)
2829                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
2830         spin_unlock(&mapping->private_lock);
2831 out:
2832         if (buffers_to_free) {
2833                 struct buffer_head *bh = buffers_to_free;
2834
2835                 do {
2836                         struct buffer_head *next = bh->b_this_page;
2837                         free_buffer_head(bh);
2838                         bh = next;
2839                 } while (bh != buffers_to_free);
2840         }
2841         return ret;
2842 }
2843 EXPORT_SYMBOL(try_to_free_buffers);
2844
2845 void block_sync_page(struct page *page)
2846 {
2847         struct address_space *mapping;
2848
2849         smp_mb();
2850         mapping = page_mapping(page);
2851         if (mapping)
2852                 blk_run_backing_dev(mapping->backing_dev_info, page);
2853 }
2854
2855 /*
2856  * There are no bdflush tunables left.  But distributions are
2857  * still running obsolete flush daemons, so we terminate them here.
2858  *
2859  * Use of bdflush() is deprecated and will be removed in a future kernel.
2860  * The `pdflush' kernel threads fully replace bdflush daemons and this call.
2861  */
2862 asmlinkage long sys_bdflush(int func, long data)
2863 {
2864         static int msg_count;
2865
2866         if (!capable(CAP_SYS_ADMIN))
2867                 return -EPERM;
2868
2869         if (msg_count < 5) {
2870                 msg_count++;
2871                 printk(KERN_INFO
2872                         "warning: process `%s' used the obsolete bdflush"
2873                         " system call\n", current->comm);
2874                 printk(KERN_INFO "Fix your initscripts?\n");
2875         }
2876
2877         if (func == 1)
2878                 do_exit(0);
2879         return 0;
2880 }
2881
2882 /*
2883  * Buffer-head allocation
2884  */
2885 static struct kmem_cache *bh_cachep;
2886
2887 /*
2888  * Once the number of bh's in the machine exceeds this level, we start
2889  * stripping them in writeback.
2890  */
2891 static int max_buffer_heads;
2892
2893 int buffer_heads_over_limit;
2894
2895 struct bh_accounting {
2896         int nr;                 /* Number of live bh's */
2897         int ratelimit;          /* Limit cacheline bouncing */
2898 };
2899
2900 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
2901
2902 static void recalc_bh_state(void)
2903 {
2904         int i;
2905         int tot = 0;
2906
2907         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
2908                 return;
2909         __get_cpu_var(bh_accounting).ratelimit = 0;
2910         for_each_online_cpu(i)
2911                 tot += per_cpu(bh_accounting, i).nr;
2912         buffer_heads_over_limit = (tot > max_buffer_heads);
2913 }
2914         
2915 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
2916 {
2917         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
2918         if (ret) {
2919                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
2920                 get_cpu_var(bh_accounting).nr++;
2921                 recalc_bh_state();
2922                 put_cpu_var(bh_accounting);
2923         }
2924         return ret;
2925 }
2926 EXPORT_SYMBOL(alloc_buffer_head);
2927
2928 void free_buffer_head(struct buffer_head *bh)
2929 {
2930         BUG_ON(!list_empty(&bh->b_assoc_buffers));
2931         kmem_cache_free(bh_cachep, bh);
2932         get_cpu_var(bh_accounting).nr--;
2933         recalc_bh_state();
2934         put_cpu_var(bh_accounting);
2935 }
2936 EXPORT_SYMBOL(free_buffer_head);
2937
2938 static void buffer_exit_cpu(int cpu)
2939 {
2940         int i;
2941         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
2942
2943         for (i = 0; i < BH_LRU_SIZE; i++) {
2944                 brelse(b->bhs[i]);
2945                 b->bhs[i] = NULL;
2946         }
2947         get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
2948         per_cpu(bh_accounting, cpu).nr = 0;
2949         put_cpu_var(bh_accounting);
2950 }
2951
2952 static int buffer_cpu_notify(struct notifier_block *self,
2953                               unsigned long action, void *hcpu)
2954 {
2955         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
2956                 buffer_exit_cpu((unsigned long)hcpu);
2957         return NOTIFY_OK;
2958 }
2959
2960 void __init buffer_init(void)
2961 {
2962         int nrpages;
2963
2964         bh_cachep = KMEM_CACHE(buffer_head,
2965                         SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD);
2966
2967         /*
2968          * Limit the bh occupancy to 10% of ZONE_NORMAL
2969          */
2970         nrpages = (nr_free_buffer_pages() * 10) / 100;
2971         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
2972         hotcpu_notifier(buffer_cpu_notify, 0);
2973 }
2974
2975 EXPORT_SYMBOL(__bforget);
2976 EXPORT_SYMBOL(__brelse);
2977 EXPORT_SYMBOL(__wait_on_buffer);
2978 EXPORT_SYMBOL(block_commit_write);
2979 EXPORT_SYMBOL(block_prepare_write);
2980 EXPORT_SYMBOL(block_read_full_page);
2981 EXPORT_SYMBOL(block_sync_page);
2982 EXPORT_SYMBOL(block_truncate_page);
2983 EXPORT_SYMBOL(block_write_full_page);
2984 EXPORT_SYMBOL(cont_prepare_write);
2985 EXPORT_SYMBOL(end_buffer_read_sync);
2986 EXPORT_SYMBOL(end_buffer_write_sync);
2987 EXPORT_SYMBOL(file_fsync);
2988 EXPORT_SYMBOL(fsync_bdev);
2989 EXPORT_SYMBOL(generic_block_bmap);
2990 EXPORT_SYMBOL(generic_commit_write);
2991 EXPORT_SYMBOL(generic_cont_expand);
2992 EXPORT_SYMBOL(generic_cont_expand_simple);
2993 EXPORT_SYMBOL(init_buffer);
2994 EXPORT_SYMBOL(invalidate_bdev);
2995 EXPORT_SYMBOL(ll_rw_block);
2996 EXPORT_SYMBOL(mark_buffer_dirty);
2997 EXPORT_SYMBOL(submit_bh);
2998 EXPORT_SYMBOL(sync_dirty_buffer);
2999 EXPORT_SYMBOL(unlock_buffer);