Merge branch 'release' of git://git.kernel.org/pub/scm/linux/kernel/git/lenb/linux...
[linux-2.6] / 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 __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 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  * End-of-IO handler helper function which does not touch the bh after
114  * unlocking it.
115  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
116  * a race there is benign: unlock_buffer() only use the bh's address for
117  * hashing after unlocking the buffer, so it doesn't actually touch the bh
118  * itself.
119  */
120 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
121 {
122         if (uptodate) {
123                 set_buffer_uptodate(bh);
124         } else {
125                 /* This happens, due to failed READA attempts. */
126                 clear_buffer_uptodate(bh);
127         }
128         unlock_buffer(bh);
129 }
130
131 /*
132  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
133  * unlock the buffer. This is what ll_rw_block uses too.
134  */
135 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
136 {
137         __end_buffer_read_notouch(bh, uptodate);
138         put_bh(bh);
139 }
140
141 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
142 {
143         char b[BDEVNAME_SIZE];
144
145         if (uptodate) {
146                 set_buffer_uptodate(bh);
147         } else {
148                 if (!buffer_eopnotsupp(bh) && printk_ratelimit()) {
149                         buffer_io_error(bh);
150                         printk(KERN_WARNING "lost page write due to "
151                                         "I/O error on %s\n",
152                                        bdevname(bh->b_bdev, b));
153                 }
154                 set_buffer_write_io_error(bh);
155                 clear_buffer_uptodate(bh);
156         }
157         unlock_buffer(bh);
158         put_bh(bh);
159 }
160
161 /*
162  * Write out and wait upon all the dirty data associated with a block
163  * device via its mapping.  Does not take the superblock lock.
164  */
165 int sync_blockdev(struct block_device *bdev)
166 {
167         int ret = 0;
168
169         if (bdev)
170                 ret = filemap_write_and_wait(bdev->bd_inode->i_mapping);
171         return ret;
172 }
173 EXPORT_SYMBOL(sync_blockdev);
174
175 /*
176  * Write out and wait upon all dirty data associated with this
177  * device.   Filesystem data as well as the underlying block
178  * device.  Takes the superblock lock.
179  */
180 int fsync_bdev(struct block_device *bdev)
181 {
182         struct super_block *sb = get_super(bdev);
183         if (sb) {
184                 int res = fsync_super(sb);
185                 drop_super(sb);
186                 return res;
187         }
188         return sync_blockdev(bdev);
189 }
190
191 /**
192  * freeze_bdev  --  lock a filesystem and force it into a consistent state
193  * @bdev:       blockdevice to lock
194  *
195  * This takes the block device bd_mount_sem to make sure no new mounts
196  * happen on bdev until thaw_bdev() is called.
197  * If a superblock is found on this device, we take the s_umount semaphore
198  * on it to make sure nobody unmounts until the snapshot creation is done.
199  */
200 struct super_block *freeze_bdev(struct block_device *bdev)
201 {
202         struct super_block *sb;
203
204         down(&bdev->bd_mount_sem);
205         sb = get_super(bdev);
206         if (sb && !(sb->s_flags & MS_RDONLY)) {
207                 sb->s_frozen = SB_FREEZE_WRITE;
208                 smp_wmb();
209
210                 __fsync_super(sb);
211
212                 sb->s_frozen = SB_FREEZE_TRANS;
213                 smp_wmb();
214
215                 sync_blockdev(sb->s_bdev);
216
217                 if (sb->s_op->write_super_lockfs)
218                         sb->s_op->write_super_lockfs(sb);
219         }
220
221         sync_blockdev(bdev);
222         return sb;      /* thaw_bdev releases s->s_umount and bd_mount_sem */
223 }
224 EXPORT_SYMBOL(freeze_bdev);
225
226 /**
227  * thaw_bdev  -- unlock filesystem
228  * @bdev:       blockdevice to unlock
229  * @sb:         associated superblock
230  *
231  * Unlocks the filesystem and marks it writeable again after freeze_bdev().
232  */
233 void thaw_bdev(struct block_device *bdev, struct super_block *sb)
234 {
235         if (sb) {
236                 BUG_ON(sb->s_bdev != bdev);
237
238                 if (sb->s_op->unlockfs)
239                         sb->s_op->unlockfs(sb);
240                 sb->s_frozen = SB_UNFROZEN;
241                 smp_wmb();
242                 wake_up(&sb->s_wait_unfrozen);
243                 drop_super(sb);
244         }
245
246         up(&bdev->bd_mount_sem);
247 }
248 EXPORT_SYMBOL(thaw_bdev);
249
250 /*
251  * Various filesystems appear to want __find_get_block to be non-blocking.
252  * But it's the page lock which protects the buffers.  To get around this,
253  * we get exclusion from try_to_free_buffers with the blockdev mapping's
254  * private_lock.
255  *
256  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
257  * may be quite high.  This code could TryLock the page, and if that
258  * succeeds, there is no need to take private_lock. (But if
259  * private_lock is contended then so is mapping->tree_lock).
260  */
261 static struct buffer_head *
262 __find_get_block_slow(struct block_device *bdev, sector_t block)
263 {
264         struct inode *bd_inode = bdev->bd_inode;
265         struct address_space *bd_mapping = bd_inode->i_mapping;
266         struct buffer_head *ret = NULL;
267         pgoff_t index;
268         struct buffer_head *bh;
269         struct buffer_head *head;
270         struct page *page;
271         int all_mapped = 1;
272
273         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
274         page = find_get_page(bd_mapping, index);
275         if (!page)
276                 goto out;
277
278         spin_lock(&bd_mapping->private_lock);
279         if (!page_has_buffers(page))
280                 goto out_unlock;
281         head = page_buffers(page);
282         bh = head;
283         do {
284                 if (bh->b_blocknr == block) {
285                         ret = bh;
286                         get_bh(bh);
287                         goto out_unlock;
288                 }
289                 if (!buffer_mapped(bh))
290                         all_mapped = 0;
291                 bh = bh->b_this_page;
292         } while (bh != head);
293
294         /* we might be here because some of the buffers on this page are
295          * not mapped.  This is due to various races between
296          * file io on the block device and getblk.  It gets dealt with
297          * elsewhere, don't buffer_error if we had some unmapped buffers
298          */
299         if (all_mapped) {
300                 printk("__find_get_block_slow() failed. "
301                         "block=%llu, b_blocknr=%llu\n",
302                         (unsigned long long)block,
303                         (unsigned long long)bh->b_blocknr);
304                 printk("b_state=0x%08lx, b_size=%zu\n",
305                         bh->b_state, bh->b_size);
306                 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
307         }
308 out_unlock:
309         spin_unlock(&bd_mapping->private_lock);
310         page_cache_release(page);
311 out:
312         return ret;
313 }
314
315 /* If invalidate_buffers() will trash dirty buffers, it means some kind
316    of fs corruption is going on. Trashing dirty data always imply losing
317    information that was supposed to be just stored on the physical layer
318    by the user.
319
320    Thus invalidate_buffers in general usage is not allwowed to trash
321    dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
322    be preserved.  These buffers are simply skipped.
323   
324    We also skip buffers which are still in use.  For example this can
325    happen if a userspace program is reading the block device.
326
327    NOTE: In the case where the user removed a removable-media-disk even if
328    there's still dirty data not synced on disk (due a bug in the device driver
329    or due an error of the user), by not destroying the dirty buffers we could
330    generate corruption also on the next media inserted, thus a parameter is
331    necessary to handle this case in the most safe way possible (trying
332    to not corrupt also the new disk inserted with the data belonging to
333    the old now corrupted disk). Also for the ramdisk the natural thing
334    to do in order to release the ramdisk memory is to destroy dirty buffers.
335
336    These are two special cases. Normal usage imply the device driver
337    to issue a sync on the device (without waiting I/O completion) and
338    then an invalidate_buffers call that doesn't trash dirty buffers.
339
340    For handling cache coherency with the blkdev pagecache the 'update' case
341    is been introduced. It is needed to re-read from disk any pinned
342    buffer. NOTE: re-reading from disk is destructive so we can do it only
343    when we assume nobody is changing the buffercache under our I/O and when
344    we think the disk contains more recent information than the buffercache.
345    The update == 1 pass marks the buffers we need to update, the update == 2
346    pass does the actual I/O. */
347 void invalidate_bdev(struct block_device *bdev)
348 {
349         struct address_space *mapping = bdev->bd_inode->i_mapping;
350
351         if (mapping->nrpages == 0)
352                 return;
353
354         invalidate_bh_lrus();
355         invalidate_mapping_pages(mapping, 0, -1);
356 }
357
358 /*
359  * Kick pdflush then try to free up some ZONE_NORMAL memory.
360  */
361 static void free_more_memory(void)
362 {
363         struct zone **zones;
364         pg_data_t *pgdat;
365
366         wakeup_pdflush(1024);
367         yield();
368
369         for_each_online_pgdat(pgdat) {
370                 zones = pgdat->node_zonelists[gfp_zone(GFP_NOFS)].zones;
371                 if (*zones)
372                         try_to_free_pages(zones, 0, GFP_NOFS);
373         }
374 }
375
376 /*
377  * I/O completion handler for block_read_full_page() - pages
378  * which come unlocked at the end of I/O.
379  */
380 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
381 {
382         unsigned long flags;
383         struct buffer_head *first;
384         struct buffer_head *tmp;
385         struct page *page;
386         int page_uptodate = 1;
387
388         BUG_ON(!buffer_async_read(bh));
389
390         page = bh->b_page;
391         if (uptodate) {
392                 set_buffer_uptodate(bh);
393         } else {
394                 clear_buffer_uptodate(bh);
395                 if (printk_ratelimit())
396                         buffer_io_error(bh);
397                 SetPageError(page);
398         }
399
400         /*
401          * Be _very_ careful from here on. Bad things can happen if
402          * two buffer heads end IO at almost the same time and both
403          * decide that the page is now completely done.
404          */
405         first = page_buffers(page);
406         local_irq_save(flags);
407         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
408         clear_buffer_async_read(bh);
409         unlock_buffer(bh);
410         tmp = bh;
411         do {
412                 if (!buffer_uptodate(tmp))
413                         page_uptodate = 0;
414                 if (buffer_async_read(tmp)) {
415                         BUG_ON(!buffer_locked(tmp));
416                         goto still_busy;
417                 }
418                 tmp = tmp->b_this_page;
419         } while (tmp != bh);
420         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
421         local_irq_restore(flags);
422
423         /*
424          * If none of the buffers had errors and they are all
425          * uptodate then we can set the page uptodate.
426          */
427         if (page_uptodate && !PageError(page))
428                 SetPageUptodate(page);
429         unlock_page(page);
430         return;
431
432 still_busy:
433         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
434         local_irq_restore(flags);
435         return;
436 }
437
438 /*
439  * Completion handler for block_write_full_page() - pages which are unlocked
440  * during I/O, and which have PageWriteback cleared upon I/O completion.
441  */
442 static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
443 {
444         char b[BDEVNAME_SIZE];
445         unsigned long flags;
446         struct buffer_head *first;
447         struct buffer_head *tmp;
448         struct page *page;
449
450         BUG_ON(!buffer_async_write(bh));
451
452         page = bh->b_page;
453         if (uptodate) {
454                 set_buffer_uptodate(bh);
455         } else {
456                 if (printk_ratelimit()) {
457                         buffer_io_error(bh);
458                         printk(KERN_WARNING "lost page write due to "
459                                         "I/O error on %s\n",
460                                bdevname(bh->b_bdev, b));
461                 }
462                 set_bit(AS_EIO, &page->mapping->flags);
463                 set_buffer_write_io_error(bh);
464                 clear_buffer_uptodate(bh);
465                 SetPageError(page);
466         }
467
468         first = page_buffers(page);
469         local_irq_save(flags);
470         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
471
472         clear_buffer_async_write(bh);
473         unlock_buffer(bh);
474         tmp = bh->b_this_page;
475         while (tmp != bh) {
476                 if (buffer_async_write(tmp)) {
477                         BUG_ON(!buffer_locked(tmp));
478                         goto still_busy;
479                 }
480                 tmp = tmp->b_this_page;
481         }
482         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
483         local_irq_restore(flags);
484         end_page_writeback(page);
485         return;
486
487 still_busy:
488         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
489         local_irq_restore(flags);
490         return;
491 }
492
493 /*
494  * If a page's buffers are under async readin (end_buffer_async_read
495  * completion) then there is a possibility that another thread of
496  * control could lock one of the buffers after it has completed
497  * but while some of the other buffers have not completed.  This
498  * locked buffer would confuse end_buffer_async_read() into not unlocking
499  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
500  * that this buffer is not under async I/O.
501  *
502  * The page comes unlocked when it has no locked buffer_async buffers
503  * left.
504  *
505  * PageLocked prevents anyone starting new async I/O reads any of
506  * the buffers.
507  *
508  * PageWriteback is used to prevent simultaneous writeout of the same
509  * page.
510  *
511  * PageLocked prevents anyone from starting writeback of a page which is
512  * under read I/O (PageWriteback is only ever set against a locked page).
513  */
514 static void mark_buffer_async_read(struct buffer_head *bh)
515 {
516         bh->b_end_io = end_buffer_async_read;
517         set_buffer_async_read(bh);
518 }
519
520 void mark_buffer_async_write(struct buffer_head *bh)
521 {
522         bh->b_end_io = end_buffer_async_write;
523         set_buffer_async_write(bh);
524 }
525 EXPORT_SYMBOL(mark_buffer_async_write);
526
527
528 /*
529  * fs/buffer.c contains helper functions for buffer-backed address space's
530  * fsync functions.  A common requirement for buffer-based filesystems is
531  * that certain data from the backing blockdev needs to be written out for
532  * a successful fsync().  For example, ext2 indirect blocks need to be
533  * written back and waited upon before fsync() returns.
534  *
535  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
536  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
537  * management of a list of dependent buffers at ->i_mapping->private_list.
538  *
539  * Locking is a little subtle: try_to_free_buffers() will remove buffers
540  * from their controlling inode's queue when they are being freed.  But
541  * try_to_free_buffers() will be operating against the *blockdev* mapping
542  * at the time, not against the S_ISREG file which depends on those buffers.
543  * So the locking for private_list is via the private_lock in the address_space
544  * which backs the buffers.  Which is different from the address_space 
545  * against which the buffers are listed.  So for a particular address_space,
546  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
547  * mapping->private_list will always be protected by the backing blockdev's
548  * ->private_lock.
549  *
550  * Which introduces a requirement: all buffers on an address_space's
551  * ->private_list must be from the same address_space: the blockdev's.
552  *
553  * address_spaces which do not place buffers at ->private_list via these
554  * utility functions are free to use private_lock and private_list for
555  * whatever they want.  The only requirement is that list_empty(private_list)
556  * be true at clear_inode() time.
557  *
558  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
559  * filesystems should do that.  invalidate_inode_buffers() should just go
560  * BUG_ON(!list_empty).
561  *
562  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
563  * take an address_space, not an inode.  And it should be called
564  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
565  * queued up.
566  *
567  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
568  * list if it is already on a list.  Because if the buffer is on a list,
569  * it *must* already be on the right one.  If not, the filesystem is being
570  * silly.  This will save a ton of locking.  But first we have to ensure
571  * that buffers are taken *off* the old inode's list when they are freed
572  * (presumably in truncate).  That requires careful auditing of all
573  * filesystems (do it inside bforget()).  It could also be done by bringing
574  * b_inode back.
575  */
576
577 /*
578  * The buffer's backing address_space's private_lock must be held
579  */
580 static inline void __remove_assoc_queue(struct buffer_head *bh)
581 {
582         list_del_init(&bh->b_assoc_buffers);
583         WARN_ON(!bh->b_assoc_map);
584         if (buffer_write_io_error(bh))
585                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
586         bh->b_assoc_map = NULL;
587 }
588
589 int inode_has_buffers(struct inode *inode)
590 {
591         return !list_empty(&inode->i_data.private_list);
592 }
593
594 /*
595  * osync is designed to support O_SYNC io.  It waits synchronously for
596  * all already-submitted IO to complete, but does not queue any new
597  * writes to the disk.
598  *
599  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
600  * you dirty the buffers, and then use osync_inode_buffers to wait for
601  * completion.  Any other dirty buffers which are not yet queued for
602  * write will not be flushed to disk by the osync.
603  */
604 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
605 {
606         struct buffer_head *bh;
607         struct list_head *p;
608         int err = 0;
609
610         spin_lock(lock);
611 repeat:
612         list_for_each_prev(p, list) {
613                 bh = BH_ENTRY(p);
614                 if (buffer_locked(bh)) {
615                         get_bh(bh);
616                         spin_unlock(lock);
617                         wait_on_buffer(bh);
618                         if (!buffer_uptodate(bh))
619                                 err = -EIO;
620                         brelse(bh);
621                         spin_lock(lock);
622                         goto repeat;
623                 }
624         }
625         spin_unlock(lock);
626         return err;
627 }
628
629 /**
630  * sync_mapping_buffers - write out and wait upon a mapping's "associated"
631  *                        buffers
632  * @mapping: the mapping which wants those buffers written
633  *
634  * Starts I/O against the buffers at mapping->private_list, and waits upon
635  * that I/O.
636  *
637  * Basically, this is a convenience function for fsync().
638  * @mapping is a file or directory which needs those buffers to be written for
639  * a successful fsync().
640  */
641 int sync_mapping_buffers(struct address_space *mapping)
642 {
643         struct address_space *buffer_mapping = mapping->assoc_mapping;
644
645         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
646                 return 0;
647
648         return fsync_buffers_list(&buffer_mapping->private_lock,
649                                         &mapping->private_list);
650 }
651 EXPORT_SYMBOL(sync_mapping_buffers);
652
653 /*
654  * Called when we've recently written block `bblock', and it is known that
655  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
656  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
657  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
658  */
659 void write_boundary_block(struct block_device *bdev,
660                         sector_t bblock, unsigned blocksize)
661 {
662         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
663         if (bh) {
664                 if (buffer_dirty(bh))
665                         ll_rw_block(WRITE, 1, &bh);
666                 put_bh(bh);
667         }
668 }
669
670 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
671 {
672         struct address_space *mapping = inode->i_mapping;
673         struct address_space *buffer_mapping = bh->b_page->mapping;
674
675         mark_buffer_dirty(bh);
676         if (!mapping->assoc_mapping) {
677                 mapping->assoc_mapping = buffer_mapping;
678         } else {
679                 BUG_ON(mapping->assoc_mapping != buffer_mapping);
680         }
681         if (!bh->b_assoc_map) {
682                 spin_lock(&buffer_mapping->private_lock);
683                 list_move_tail(&bh->b_assoc_buffers,
684                                 &mapping->private_list);
685                 bh->b_assoc_map = mapping;
686                 spin_unlock(&buffer_mapping->private_lock);
687         }
688 }
689 EXPORT_SYMBOL(mark_buffer_dirty_inode);
690
691 /*
692  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
693  * dirty.
694  *
695  * If warn is true, then emit a warning if the page is not uptodate and has
696  * not been truncated.
697  */
698 static int __set_page_dirty(struct page *page,
699                 struct address_space *mapping, int warn)
700 {
701         if (unlikely(!mapping))
702                 return !TestSetPageDirty(page);
703
704         if (TestSetPageDirty(page))
705                 return 0;
706
707         write_lock_irq(&mapping->tree_lock);
708         if (page->mapping) {    /* Race with truncate? */
709                 WARN_ON_ONCE(warn && !PageUptodate(page));
710
711                 if (mapping_cap_account_dirty(mapping)) {
712                         __inc_zone_page_state(page, NR_FILE_DIRTY);
713                         __inc_bdi_stat(mapping->backing_dev_info,
714                                         BDI_RECLAIMABLE);
715                         task_io_account_write(PAGE_CACHE_SIZE);
716                 }
717                 radix_tree_tag_set(&mapping->page_tree,
718                                 page_index(page), PAGECACHE_TAG_DIRTY);
719         }
720         write_unlock_irq(&mapping->tree_lock);
721         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
722
723         return 1;
724 }
725
726 /*
727  * Add a page to the dirty page list.
728  *
729  * It is a sad fact of life that this function is called from several places
730  * deeply under spinlocking.  It may not sleep.
731  *
732  * If the page has buffers, the uptodate buffers are set dirty, to preserve
733  * dirty-state coherency between the page and the buffers.  It the page does
734  * not have buffers then when they are later attached they will all be set
735  * dirty.
736  *
737  * The buffers are dirtied before the page is dirtied.  There's a small race
738  * window in which a writepage caller may see the page cleanness but not the
739  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
740  * before the buffers, a concurrent writepage caller could clear the page dirty
741  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
742  * page on the dirty page list.
743  *
744  * We use private_lock to lock against try_to_free_buffers while using the
745  * page's buffer list.  Also use this to protect against clean buffers being
746  * added to the page after it was set dirty.
747  *
748  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
749  * address_space though.
750  */
751 int __set_page_dirty_buffers(struct page *page)
752 {
753         struct address_space *mapping = page_mapping(page);
754
755         if (unlikely(!mapping))
756                 return !TestSetPageDirty(page);
757
758         spin_lock(&mapping->private_lock);
759         if (page_has_buffers(page)) {
760                 struct buffer_head *head = page_buffers(page);
761                 struct buffer_head *bh = head;
762
763                 do {
764                         set_buffer_dirty(bh);
765                         bh = bh->b_this_page;
766                 } while (bh != head);
767         }
768         spin_unlock(&mapping->private_lock);
769
770         return __set_page_dirty(page, mapping, 1);
771 }
772 EXPORT_SYMBOL(__set_page_dirty_buffers);
773
774 /*
775  * Write out and wait upon a list of buffers.
776  *
777  * We have conflicting pressures: we want to make sure that all
778  * initially dirty buffers get waited on, but that any subsequently
779  * dirtied buffers don't.  After all, we don't want fsync to last
780  * forever if somebody is actively writing to the file.
781  *
782  * Do this in two main stages: first we copy dirty buffers to a
783  * temporary inode list, queueing the writes as we go.  Then we clean
784  * up, waiting for those writes to complete.
785  * 
786  * During this second stage, any subsequent updates to the file may end
787  * up refiling the buffer on the original inode's dirty list again, so
788  * there is a chance we will end up with a buffer queued for write but
789  * not yet completed on that list.  So, as a final cleanup we go through
790  * the osync code to catch these locked, dirty buffers without requeuing
791  * any newly dirty buffers for write.
792  */
793 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
794 {
795         struct buffer_head *bh;
796         struct list_head tmp;
797         struct address_space *mapping;
798         int err = 0, err2;
799
800         INIT_LIST_HEAD(&tmp);
801
802         spin_lock(lock);
803         while (!list_empty(list)) {
804                 bh = BH_ENTRY(list->next);
805                 mapping = bh->b_assoc_map;
806                 __remove_assoc_queue(bh);
807                 /* Avoid race with mark_buffer_dirty_inode() which does
808                  * a lockless check and we rely on seeing the dirty bit */
809                 smp_mb();
810                 if (buffer_dirty(bh) || buffer_locked(bh)) {
811                         list_add(&bh->b_assoc_buffers, &tmp);
812                         bh->b_assoc_map = mapping;
813                         if (buffer_dirty(bh)) {
814                                 get_bh(bh);
815                                 spin_unlock(lock);
816                                 /*
817                                  * Ensure any pending I/O completes so that
818                                  * ll_rw_block() actually writes the current
819                                  * contents - it is a noop if I/O is still in
820                                  * flight on potentially older contents.
821                                  */
822                                 ll_rw_block(SWRITE, 1, &bh);
823                                 brelse(bh);
824                                 spin_lock(lock);
825                         }
826                 }
827         }
828
829         while (!list_empty(&tmp)) {
830                 bh = BH_ENTRY(tmp.prev);
831                 get_bh(bh);
832                 mapping = bh->b_assoc_map;
833                 __remove_assoc_queue(bh);
834                 /* Avoid race with mark_buffer_dirty_inode() which does
835                  * a lockless check and we rely on seeing the dirty bit */
836                 smp_mb();
837                 if (buffer_dirty(bh)) {
838                         list_add(&bh->b_assoc_buffers,
839                                  &bh->b_assoc_map->private_list);
840                         bh->b_assoc_map = mapping;
841                 }
842                 spin_unlock(lock);
843                 wait_on_buffer(bh);
844                 if (!buffer_uptodate(bh))
845                         err = -EIO;
846                 brelse(bh);
847                 spin_lock(lock);
848         }
849         
850         spin_unlock(lock);
851         err2 = osync_buffers_list(lock, list);
852         if (err)
853                 return err;
854         else
855                 return err2;
856 }
857
858 /*
859  * Invalidate any and all dirty buffers on a given inode.  We are
860  * probably unmounting the fs, but that doesn't mean we have already
861  * done a sync().  Just drop the buffers from the inode list.
862  *
863  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
864  * assumes that all the buffers are against the blockdev.  Not true
865  * for reiserfs.
866  */
867 void invalidate_inode_buffers(struct inode *inode)
868 {
869         if (inode_has_buffers(inode)) {
870                 struct address_space *mapping = &inode->i_data;
871                 struct list_head *list = &mapping->private_list;
872                 struct address_space *buffer_mapping = mapping->assoc_mapping;
873
874                 spin_lock(&buffer_mapping->private_lock);
875                 while (!list_empty(list))
876                         __remove_assoc_queue(BH_ENTRY(list->next));
877                 spin_unlock(&buffer_mapping->private_lock);
878         }
879 }
880
881 /*
882  * Remove any clean buffers from the inode's buffer list.  This is called
883  * when we're trying to free the inode itself.  Those buffers can pin it.
884  *
885  * Returns true if all buffers were removed.
886  */
887 int remove_inode_buffers(struct inode *inode)
888 {
889         int ret = 1;
890
891         if (inode_has_buffers(inode)) {
892                 struct address_space *mapping = &inode->i_data;
893                 struct list_head *list = &mapping->private_list;
894                 struct address_space *buffer_mapping = mapping->assoc_mapping;
895
896                 spin_lock(&buffer_mapping->private_lock);
897                 while (!list_empty(list)) {
898                         struct buffer_head *bh = BH_ENTRY(list->next);
899                         if (buffer_dirty(bh)) {
900                                 ret = 0;
901                                 break;
902                         }
903                         __remove_assoc_queue(bh);
904                 }
905                 spin_unlock(&buffer_mapping->private_lock);
906         }
907         return ret;
908 }
909
910 /*
911  * Create the appropriate buffers when given a page for data area and
912  * the size of each buffer.. Use the bh->b_this_page linked list to
913  * follow the buffers created.  Return NULL if unable to create more
914  * buffers.
915  *
916  * The retry flag is used to differentiate async IO (paging, swapping)
917  * which may not fail from ordinary buffer allocations.
918  */
919 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
920                 int retry)
921 {
922         struct buffer_head *bh, *head;
923         long offset;
924
925 try_again:
926         head = NULL;
927         offset = PAGE_SIZE;
928         while ((offset -= size) >= 0) {
929                 bh = alloc_buffer_head(GFP_NOFS);
930                 if (!bh)
931                         goto no_grow;
932
933                 bh->b_bdev = NULL;
934                 bh->b_this_page = head;
935                 bh->b_blocknr = -1;
936                 head = bh;
937
938                 bh->b_state = 0;
939                 atomic_set(&bh->b_count, 0);
940                 bh->b_private = NULL;
941                 bh->b_size = size;
942
943                 /* Link the buffer to its page */
944                 set_bh_page(bh, page, offset);
945
946                 init_buffer(bh, NULL, NULL);
947         }
948         return head;
949 /*
950  * In case anything failed, we just free everything we got.
951  */
952 no_grow:
953         if (head) {
954                 do {
955                         bh = head;
956                         head = head->b_this_page;
957                         free_buffer_head(bh);
958                 } while (head);
959         }
960
961         /*
962          * Return failure for non-async IO requests.  Async IO requests
963          * are not allowed to fail, so we have to wait until buffer heads
964          * become available.  But we don't want tasks sleeping with 
965          * partially complete buffers, so all were released above.
966          */
967         if (!retry)
968                 return NULL;
969
970         /* We're _really_ low on memory. Now we just
971          * wait for old buffer heads to become free due to
972          * finishing IO.  Since this is an async request and
973          * the reserve list is empty, we're sure there are 
974          * async buffer heads in use.
975          */
976         free_more_memory();
977         goto try_again;
978 }
979 EXPORT_SYMBOL_GPL(alloc_page_buffers);
980
981 static inline void
982 link_dev_buffers(struct page *page, struct buffer_head *head)
983 {
984         struct buffer_head *bh, *tail;
985
986         bh = head;
987         do {
988                 tail = bh;
989                 bh = bh->b_this_page;
990         } while (bh);
991         tail->b_this_page = head;
992         attach_page_buffers(page, head);
993 }
994
995 /*
996  * Initialise the state of a blockdev page's buffers.
997  */ 
998 static void
999 init_page_buffers(struct page *page, struct block_device *bdev,
1000                         sector_t block, int size)
1001 {
1002         struct buffer_head *head = page_buffers(page);
1003         struct buffer_head *bh = head;
1004         int uptodate = PageUptodate(page);
1005
1006         do {
1007                 if (!buffer_mapped(bh)) {
1008                         init_buffer(bh, NULL, NULL);
1009                         bh->b_bdev = bdev;
1010                         bh->b_blocknr = block;
1011                         if (uptodate)
1012                                 set_buffer_uptodate(bh);
1013                         set_buffer_mapped(bh);
1014                 }
1015                 block++;
1016                 bh = bh->b_this_page;
1017         } while (bh != head);
1018 }
1019
1020 /*
1021  * Create the page-cache page that contains the requested block.
1022  *
1023  * This is user purely for blockdev mappings.
1024  */
1025 static struct page *
1026 grow_dev_page(struct block_device *bdev, sector_t block,
1027                 pgoff_t index, int size)
1028 {
1029         struct inode *inode = bdev->bd_inode;
1030         struct page *page;
1031         struct buffer_head *bh;
1032
1033         page = find_or_create_page(inode->i_mapping, index,
1034                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
1035         if (!page)
1036                 return NULL;
1037
1038         BUG_ON(!PageLocked(page));
1039
1040         if (page_has_buffers(page)) {
1041                 bh = page_buffers(page);
1042                 if (bh->b_size == size) {
1043                         init_page_buffers(page, bdev, block, size);
1044                         return page;
1045                 }
1046                 if (!try_to_free_buffers(page))
1047                         goto failed;
1048         }
1049
1050         /*
1051          * Allocate some buffers for this page
1052          */
1053         bh = alloc_page_buffers(page, size, 0);
1054         if (!bh)
1055                 goto failed;
1056
1057         /*
1058          * Link the page to the buffers and initialise them.  Take the
1059          * lock to be atomic wrt __find_get_block(), which does not
1060          * run under the page lock.
1061          */
1062         spin_lock(&inode->i_mapping->private_lock);
1063         link_dev_buffers(page, bh);
1064         init_page_buffers(page, bdev, block, size);
1065         spin_unlock(&inode->i_mapping->private_lock);
1066         return page;
1067
1068 failed:
1069         BUG();
1070         unlock_page(page);
1071         page_cache_release(page);
1072         return NULL;
1073 }
1074
1075 /*
1076  * Create buffers for the specified block device block's page.  If
1077  * that page was dirty, the buffers are set dirty also.
1078  */
1079 static int
1080 grow_buffers(struct block_device *bdev, sector_t block, int size)
1081 {
1082         struct page *page;
1083         pgoff_t index;
1084         int sizebits;
1085
1086         sizebits = -1;
1087         do {
1088                 sizebits++;
1089         } while ((size << sizebits) < PAGE_SIZE);
1090
1091         index = block >> sizebits;
1092
1093         /*
1094          * Check for a block which wants to lie outside our maximum possible
1095          * pagecache index.  (this comparison is done using sector_t types).
1096          */
1097         if (unlikely(index != block >> sizebits)) {
1098                 char b[BDEVNAME_SIZE];
1099
1100                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1101                         "device %s\n",
1102                         __FUNCTION__, (unsigned long long)block,
1103                         bdevname(bdev, b));
1104                 return -EIO;
1105         }
1106         block = index << sizebits;
1107         /* Create a page with the proper size buffers.. */
1108         page = grow_dev_page(bdev, block, index, size);
1109         if (!page)
1110                 return 0;
1111         unlock_page(page);
1112         page_cache_release(page);
1113         return 1;
1114 }
1115
1116 static struct buffer_head *
1117 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1118 {
1119         /* Size must be multiple of hard sectorsize */
1120         if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1121                         (size < 512 || size > PAGE_SIZE))) {
1122                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1123                                         size);
1124                 printk(KERN_ERR "hardsect size: %d\n",
1125                                         bdev_hardsect_size(bdev));
1126
1127                 dump_stack();
1128                 return NULL;
1129         }
1130
1131         for (;;) {
1132                 struct buffer_head * bh;
1133                 int ret;
1134
1135                 bh = __find_get_block(bdev, block, size);
1136                 if (bh)
1137                         return bh;
1138
1139                 ret = grow_buffers(bdev, block, size);
1140                 if (ret < 0)
1141                         return NULL;
1142                 if (ret == 0)
1143                         free_more_memory();
1144         }
1145 }
1146
1147 /*
1148  * The relationship between dirty buffers and dirty pages:
1149  *
1150  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1151  * the page is tagged dirty in its radix tree.
1152  *
1153  * At all times, the dirtiness of the buffers represents the dirtiness of
1154  * subsections of the page.  If the page has buffers, the page dirty bit is
1155  * merely a hint about the true dirty state.
1156  *
1157  * When a page is set dirty in its entirety, all its buffers are marked dirty
1158  * (if the page has buffers).
1159  *
1160  * When a buffer is marked dirty, its page is dirtied, but the page's other
1161  * buffers are not.
1162  *
1163  * Also.  When blockdev buffers are explicitly read with bread(), they
1164  * individually become uptodate.  But their backing page remains not
1165  * uptodate - even if all of its buffers are uptodate.  A subsequent
1166  * block_read_full_page() against that page will discover all the uptodate
1167  * buffers, will set the page uptodate and will perform no I/O.
1168  */
1169
1170 /**
1171  * mark_buffer_dirty - mark a buffer_head as needing writeout
1172  * @bh: the buffer_head to mark dirty
1173  *
1174  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1175  * backing page dirty, then tag the page as dirty in its address_space's radix
1176  * tree and then attach the address_space's inode to its superblock's dirty
1177  * inode list.
1178  *
1179  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1180  * mapping->tree_lock and the global inode_lock.
1181  */
1182 void mark_buffer_dirty(struct buffer_head *bh)
1183 {
1184         WARN_ON_ONCE(!buffer_uptodate(bh));
1185         if (!buffer_dirty(bh) && !test_set_buffer_dirty(bh))
1186                 __set_page_dirty(bh->b_page, page_mapping(bh->b_page), 0);
1187 }
1188
1189 /*
1190  * Decrement a buffer_head's reference count.  If all buffers against a page
1191  * have zero reference count, are clean and unlocked, and if the page is clean
1192  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1193  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1194  * a page but it ends up not being freed, and buffers may later be reattached).
1195  */
1196 void __brelse(struct buffer_head * buf)
1197 {
1198         if (atomic_read(&buf->b_count)) {
1199                 put_bh(buf);
1200                 return;
1201         }
1202         printk(KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1203         WARN_ON(1);
1204 }
1205
1206 /*
1207  * bforget() is like brelse(), except it discards any
1208  * potentially dirty data.
1209  */
1210 void __bforget(struct buffer_head *bh)
1211 {
1212         clear_buffer_dirty(bh);
1213         if (bh->b_assoc_map) {
1214                 struct address_space *buffer_mapping = bh->b_page->mapping;
1215
1216                 spin_lock(&buffer_mapping->private_lock);
1217                 list_del_init(&bh->b_assoc_buffers);
1218                 bh->b_assoc_map = NULL;
1219                 spin_unlock(&buffer_mapping->private_lock);
1220         }
1221         __brelse(bh);
1222 }
1223
1224 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1225 {
1226         lock_buffer(bh);
1227         if (buffer_uptodate(bh)) {
1228                 unlock_buffer(bh);
1229                 return bh;
1230         } else {
1231                 get_bh(bh);
1232                 bh->b_end_io = end_buffer_read_sync;
1233                 submit_bh(READ, bh);
1234                 wait_on_buffer(bh);
1235                 if (buffer_uptodate(bh))
1236                         return bh;
1237         }
1238         brelse(bh);
1239         return NULL;
1240 }
1241
1242 /*
1243  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1244  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1245  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1246  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1247  * CPU's LRUs at the same time.
1248  *
1249  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1250  * sb_find_get_block().
1251  *
1252  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1253  * a local interrupt disable for that.
1254  */
1255
1256 #define BH_LRU_SIZE     8
1257
1258 struct bh_lru {
1259         struct buffer_head *bhs[BH_LRU_SIZE];
1260 };
1261
1262 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1263
1264 #ifdef CONFIG_SMP
1265 #define bh_lru_lock()   local_irq_disable()
1266 #define bh_lru_unlock() local_irq_enable()
1267 #else
1268 #define bh_lru_lock()   preempt_disable()
1269 #define bh_lru_unlock() preempt_enable()
1270 #endif
1271
1272 static inline void check_irqs_on(void)
1273 {
1274 #ifdef irqs_disabled
1275         BUG_ON(irqs_disabled());
1276 #endif
1277 }
1278
1279 /*
1280  * The LRU management algorithm is dopey-but-simple.  Sorry.
1281  */
1282 static void bh_lru_install(struct buffer_head *bh)
1283 {
1284         struct buffer_head *evictee = NULL;
1285         struct bh_lru *lru;
1286
1287         check_irqs_on();
1288         bh_lru_lock();
1289         lru = &__get_cpu_var(bh_lrus);
1290         if (lru->bhs[0] != bh) {
1291                 struct buffer_head *bhs[BH_LRU_SIZE];
1292                 int in;
1293                 int out = 0;
1294
1295                 get_bh(bh);
1296                 bhs[out++] = bh;
1297                 for (in = 0; in < BH_LRU_SIZE; in++) {
1298                         struct buffer_head *bh2 = lru->bhs[in];
1299
1300                         if (bh2 == bh) {
1301                                 __brelse(bh2);
1302                         } else {
1303                                 if (out >= BH_LRU_SIZE) {
1304                                         BUG_ON(evictee != NULL);
1305                                         evictee = bh2;
1306                                 } else {
1307                                         bhs[out++] = bh2;
1308                                 }
1309                         }
1310                 }
1311                 while (out < BH_LRU_SIZE)
1312                         bhs[out++] = NULL;
1313                 memcpy(lru->bhs, bhs, sizeof(bhs));
1314         }
1315         bh_lru_unlock();
1316
1317         if (evictee)
1318                 __brelse(evictee);
1319 }
1320
1321 /*
1322  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1323  */
1324 static struct buffer_head *
1325 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1326 {
1327         struct buffer_head *ret = NULL;
1328         struct bh_lru *lru;
1329         unsigned int i;
1330
1331         check_irqs_on();
1332         bh_lru_lock();
1333         lru = &__get_cpu_var(bh_lrus);
1334         for (i = 0; i < BH_LRU_SIZE; i++) {
1335                 struct buffer_head *bh = lru->bhs[i];
1336
1337                 if (bh && bh->b_bdev == bdev &&
1338                                 bh->b_blocknr == block && bh->b_size == size) {
1339                         if (i) {
1340                                 while (i) {
1341                                         lru->bhs[i] = lru->bhs[i - 1];
1342                                         i--;
1343                                 }
1344                                 lru->bhs[0] = bh;
1345                         }
1346                         get_bh(bh);
1347                         ret = bh;
1348                         break;
1349                 }
1350         }
1351         bh_lru_unlock();
1352         return ret;
1353 }
1354
1355 /*
1356  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1357  * it in the LRU and mark it as accessed.  If it is not present then return
1358  * NULL
1359  */
1360 struct buffer_head *
1361 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1362 {
1363         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1364
1365         if (bh == NULL) {
1366                 bh = __find_get_block_slow(bdev, block);
1367                 if (bh)
1368                         bh_lru_install(bh);
1369         }
1370         if (bh)
1371                 touch_buffer(bh);
1372         return bh;
1373 }
1374 EXPORT_SYMBOL(__find_get_block);
1375
1376 /*
1377  * __getblk will locate (and, if necessary, create) the buffer_head
1378  * which corresponds to the passed block_device, block and size. The
1379  * returned buffer has its reference count incremented.
1380  *
1381  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1382  * illegal block number, __getblk() will happily return a buffer_head
1383  * which represents the non-existent block.  Very weird.
1384  *
1385  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1386  * attempt is failing.  FIXME, perhaps?
1387  */
1388 struct buffer_head *
1389 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1390 {
1391         struct buffer_head *bh = __find_get_block(bdev, block, size);
1392
1393         might_sleep();
1394         if (bh == NULL)
1395                 bh = __getblk_slow(bdev, block, size);
1396         return bh;
1397 }
1398 EXPORT_SYMBOL(__getblk);
1399
1400 /*
1401  * Do async read-ahead on a buffer..
1402  */
1403 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1404 {
1405         struct buffer_head *bh = __getblk(bdev, block, size);
1406         if (likely(bh)) {
1407                 ll_rw_block(READA, 1, &bh);
1408                 brelse(bh);
1409         }
1410 }
1411 EXPORT_SYMBOL(__breadahead);
1412
1413 /**
1414  *  __bread() - reads a specified block and returns the bh
1415  *  @bdev: the block_device to read from
1416  *  @block: number of block
1417  *  @size: size (in bytes) to read
1418  * 
1419  *  Reads a specified block, and returns buffer head that contains it.
1420  *  It returns NULL if the block was unreadable.
1421  */
1422 struct buffer_head *
1423 __bread(struct block_device *bdev, sector_t block, unsigned size)
1424 {
1425         struct buffer_head *bh = __getblk(bdev, block, size);
1426
1427         if (likely(bh) && !buffer_uptodate(bh))
1428                 bh = __bread_slow(bh);
1429         return bh;
1430 }
1431 EXPORT_SYMBOL(__bread);
1432
1433 /*
1434  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1435  * This doesn't race because it runs in each cpu either in irq
1436  * or with preempt disabled.
1437  */
1438 static void invalidate_bh_lru(void *arg)
1439 {
1440         struct bh_lru *b = &get_cpu_var(bh_lrus);
1441         int i;
1442
1443         for (i = 0; i < BH_LRU_SIZE; i++) {
1444                 brelse(b->bhs[i]);
1445                 b->bhs[i] = NULL;
1446         }
1447         put_cpu_var(bh_lrus);
1448 }
1449         
1450 void invalidate_bh_lrus(void)
1451 {
1452         on_each_cpu(invalidate_bh_lru, NULL, 1, 1);
1453 }
1454 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1455
1456 void set_bh_page(struct buffer_head *bh,
1457                 struct page *page, unsigned long offset)
1458 {
1459         bh->b_page = page;
1460         BUG_ON(offset >= PAGE_SIZE);
1461         if (PageHighMem(page))
1462                 /*
1463                  * This catches illegal uses and preserves the offset:
1464                  */
1465                 bh->b_data = (char *)(0 + offset);
1466         else
1467                 bh->b_data = page_address(page) + offset;
1468 }
1469 EXPORT_SYMBOL(set_bh_page);
1470
1471 /*
1472  * Called when truncating a buffer on a page completely.
1473  */
1474 static void discard_buffer(struct buffer_head * bh)
1475 {
1476         lock_buffer(bh);
1477         clear_buffer_dirty(bh);
1478         bh->b_bdev = NULL;
1479         clear_buffer_mapped(bh);
1480         clear_buffer_req(bh);
1481         clear_buffer_new(bh);
1482         clear_buffer_delay(bh);
1483         clear_buffer_unwritten(bh);
1484         unlock_buffer(bh);
1485 }
1486
1487 /**
1488  * block_invalidatepage - invalidate part of all of a buffer-backed page
1489  *
1490  * @page: the page which is affected
1491  * @offset: the index of the truncation point
1492  *
1493  * block_invalidatepage() is called when all or part of the page has become
1494  * invalidatedby a truncate operation.
1495  *
1496  * block_invalidatepage() does not have to release all buffers, but it must
1497  * ensure that no dirty buffer is left outside @offset and that no I/O
1498  * is underway against any of the blocks which are outside the truncation
1499  * point.  Because the caller is about to free (and possibly reuse) those
1500  * blocks on-disk.
1501  */
1502 void block_invalidatepage(struct page *page, unsigned long offset)
1503 {
1504         struct buffer_head *head, *bh, *next;
1505         unsigned int curr_off = 0;
1506
1507         BUG_ON(!PageLocked(page));
1508         if (!page_has_buffers(page))
1509                 goto out;
1510
1511         head = page_buffers(page);
1512         bh = head;
1513         do {
1514                 unsigned int next_off = curr_off + bh->b_size;
1515                 next = bh->b_this_page;
1516
1517                 /*
1518                  * is this block fully invalidated?
1519                  */
1520                 if (offset <= curr_off)
1521                         discard_buffer(bh);
1522                 curr_off = next_off;
1523                 bh = next;
1524         } while (bh != head);
1525
1526         /*
1527          * We release buffers only if the entire page is being invalidated.
1528          * The get_block cached value has been unconditionally invalidated,
1529          * so real IO is not possible anymore.
1530          */
1531         if (offset == 0)
1532                 try_to_release_page(page, 0);
1533 out:
1534         return;
1535 }
1536 EXPORT_SYMBOL(block_invalidatepage);
1537
1538 /*
1539  * We attach and possibly dirty the buffers atomically wrt
1540  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1541  * is already excluded via the page lock.
1542  */
1543 void create_empty_buffers(struct page *page,
1544                         unsigned long blocksize, unsigned long b_state)
1545 {
1546         struct buffer_head *bh, *head, *tail;
1547
1548         head = alloc_page_buffers(page, blocksize, 1);
1549         bh = head;
1550         do {
1551                 bh->b_state |= b_state;
1552                 tail = bh;
1553                 bh = bh->b_this_page;
1554         } while (bh);
1555         tail->b_this_page = head;
1556
1557         spin_lock(&page->mapping->private_lock);
1558         if (PageUptodate(page) || PageDirty(page)) {
1559                 bh = head;
1560                 do {
1561                         if (PageDirty(page))
1562                                 set_buffer_dirty(bh);
1563                         if (PageUptodate(page))
1564                                 set_buffer_uptodate(bh);
1565                         bh = bh->b_this_page;
1566                 } while (bh != head);
1567         }
1568         attach_page_buffers(page, head);
1569         spin_unlock(&page->mapping->private_lock);
1570 }
1571 EXPORT_SYMBOL(create_empty_buffers);
1572
1573 /*
1574  * We are taking a block for data and we don't want any output from any
1575  * buffer-cache aliases starting from return from that function and
1576  * until the moment when something will explicitly mark the buffer
1577  * dirty (hopefully that will not happen until we will free that block ;-)
1578  * We don't even need to mark it not-uptodate - nobody can expect
1579  * anything from a newly allocated buffer anyway. We used to used
1580  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1581  * don't want to mark the alias unmapped, for example - it would confuse
1582  * anyone who might pick it with bread() afterwards...
1583  *
1584  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1585  * be writeout I/O going on against recently-freed buffers.  We don't
1586  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1587  * only if we really need to.  That happens here.
1588  */
1589 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1590 {
1591         struct buffer_head *old_bh;
1592
1593         might_sleep();
1594
1595         old_bh = __find_get_block_slow(bdev, block);
1596         if (old_bh) {
1597                 clear_buffer_dirty(old_bh);
1598                 wait_on_buffer(old_bh);
1599                 clear_buffer_req(old_bh);
1600                 __brelse(old_bh);
1601         }
1602 }
1603 EXPORT_SYMBOL(unmap_underlying_metadata);
1604
1605 /*
1606  * NOTE! All mapped/uptodate combinations are valid:
1607  *
1608  *      Mapped  Uptodate        Meaning
1609  *
1610  *      No      No              "unknown" - must do get_block()
1611  *      No      Yes             "hole" - zero-filled
1612  *      Yes     No              "allocated" - allocated on disk, not read in
1613  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1614  *
1615  * "Dirty" is valid only with the last case (mapped+uptodate).
1616  */
1617
1618 /*
1619  * While block_write_full_page is writing back the dirty buffers under
1620  * the page lock, whoever dirtied the buffers may decide to clean them
1621  * again at any time.  We handle that by only looking at the buffer
1622  * state inside lock_buffer().
1623  *
1624  * If block_write_full_page() is called for regular writeback
1625  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1626  * locked buffer.   This only can happen if someone has written the buffer
1627  * directly, with submit_bh().  At the address_space level PageWriteback
1628  * prevents this contention from occurring.
1629  */
1630 static int __block_write_full_page(struct inode *inode, struct page *page,
1631                         get_block_t *get_block, struct writeback_control *wbc)
1632 {
1633         int err;
1634         sector_t block;
1635         sector_t last_block;
1636         struct buffer_head *bh, *head;
1637         const unsigned blocksize = 1 << inode->i_blkbits;
1638         int nr_underway = 0;
1639
1640         BUG_ON(!PageLocked(page));
1641
1642         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1643
1644         if (!page_has_buffers(page)) {
1645                 create_empty_buffers(page, blocksize,
1646                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1647         }
1648
1649         /*
1650          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1651          * here, and the (potentially unmapped) buffers may become dirty at
1652          * any time.  If a buffer becomes dirty here after we've inspected it
1653          * then we just miss that fact, and the page stays dirty.
1654          *
1655          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1656          * handle that here by just cleaning them.
1657          */
1658
1659         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1660         head = page_buffers(page);
1661         bh = head;
1662
1663         /*
1664          * Get all the dirty buffers mapped to disk addresses and
1665          * handle any aliases from the underlying blockdev's mapping.
1666          */
1667         do {
1668                 if (block > last_block) {
1669                         /*
1670                          * mapped buffers outside i_size will occur, because
1671                          * this page can be outside i_size when there is a
1672                          * truncate in progress.
1673                          */
1674                         /*
1675                          * The buffer was zeroed by block_write_full_page()
1676                          */
1677                         clear_buffer_dirty(bh);
1678                         set_buffer_uptodate(bh);
1679                 } else if (!buffer_mapped(bh) && buffer_dirty(bh)) {
1680                         WARN_ON(bh->b_size != blocksize);
1681                         err = get_block(inode, block, bh, 1);
1682                         if (err)
1683                                 goto recover;
1684                         if (buffer_new(bh)) {
1685                                 /* blockdev mappings never come here */
1686                                 clear_buffer_new(bh);
1687                                 unmap_underlying_metadata(bh->b_bdev,
1688                                                         bh->b_blocknr);
1689                         }
1690                 }
1691                 bh = bh->b_this_page;
1692                 block++;
1693         } while (bh != head);
1694
1695         do {
1696                 if (!buffer_mapped(bh))
1697                         continue;
1698                 /*
1699                  * If it's a fully non-blocking write attempt and we cannot
1700                  * lock the buffer then redirty the page.  Note that this can
1701                  * potentially cause a busy-wait loop from pdflush and kswapd
1702                  * activity, but those code paths have their own higher-level
1703                  * throttling.
1704                  */
1705                 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1706                         lock_buffer(bh);
1707                 } else if (test_set_buffer_locked(bh)) {
1708                         redirty_page_for_writepage(wbc, page);
1709                         continue;
1710                 }
1711                 if (test_clear_buffer_dirty(bh)) {
1712                         mark_buffer_async_write(bh);
1713                 } else {
1714                         unlock_buffer(bh);
1715                 }
1716         } while ((bh = bh->b_this_page) != head);
1717
1718         /*
1719          * The page and its buffers are protected by PageWriteback(), so we can
1720          * drop the bh refcounts early.
1721          */
1722         BUG_ON(PageWriteback(page));
1723         set_page_writeback(page);
1724
1725         do {
1726                 struct buffer_head *next = bh->b_this_page;
1727                 if (buffer_async_write(bh)) {
1728                         submit_bh(WRITE, bh);
1729                         nr_underway++;
1730                 }
1731                 bh = next;
1732         } while (bh != head);
1733         unlock_page(page);
1734
1735         err = 0;
1736 done:
1737         if (nr_underway == 0) {
1738                 /*
1739                  * The page was marked dirty, but the buffers were
1740                  * clean.  Someone wrote them back by hand with
1741                  * ll_rw_block/submit_bh.  A rare case.
1742                  */
1743                 end_page_writeback(page);
1744
1745                 /*
1746                  * The page and buffer_heads can be released at any time from
1747                  * here on.
1748                  */
1749         }
1750         return err;
1751
1752 recover:
1753         /*
1754          * ENOSPC, or some other error.  We may already have added some
1755          * blocks to the file, so we need to write these out to avoid
1756          * exposing stale data.
1757          * The page is currently locked and not marked for writeback
1758          */
1759         bh = head;
1760         /* Recovery: lock and submit the mapped buffers */
1761         do {
1762                 if (buffer_mapped(bh) && buffer_dirty(bh)) {
1763                         lock_buffer(bh);
1764                         mark_buffer_async_write(bh);
1765                 } else {
1766                         /*
1767                          * The buffer may have been set dirty during
1768                          * attachment to a dirty page.
1769                          */
1770                         clear_buffer_dirty(bh);
1771                 }
1772         } while ((bh = bh->b_this_page) != head);
1773         SetPageError(page);
1774         BUG_ON(PageWriteback(page));
1775         mapping_set_error(page->mapping, err);
1776         set_page_writeback(page);
1777         do {
1778                 struct buffer_head *next = bh->b_this_page;
1779                 if (buffer_async_write(bh)) {
1780                         clear_buffer_dirty(bh);
1781                         submit_bh(WRITE, bh);
1782                         nr_underway++;
1783                 }
1784                 bh = next;
1785         } while (bh != head);
1786         unlock_page(page);
1787         goto done;
1788 }
1789
1790 /*
1791  * If a page has any new buffers, zero them out here, and mark them uptodate
1792  * and dirty so they'll be written out (in order to prevent uninitialised
1793  * block data from leaking). And clear the new bit.
1794  */
1795 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1796 {
1797         unsigned int block_start, block_end;
1798         struct buffer_head *head, *bh;
1799
1800         BUG_ON(!PageLocked(page));
1801         if (!page_has_buffers(page))
1802                 return;
1803
1804         bh = head = page_buffers(page);
1805         block_start = 0;
1806         do {
1807                 block_end = block_start + bh->b_size;
1808
1809                 if (buffer_new(bh)) {
1810                         if (block_end > from && block_start < to) {
1811                                 if (!PageUptodate(page)) {
1812                                         unsigned start, size;
1813
1814                                         start = max(from, block_start);
1815                                         size = min(to, block_end) - start;
1816
1817                                         zero_user(page, start, size);
1818                                         set_buffer_uptodate(bh);
1819                                 }
1820
1821                                 clear_buffer_new(bh);
1822                                 mark_buffer_dirty(bh);
1823                         }
1824                 }
1825
1826                 block_start = block_end;
1827                 bh = bh->b_this_page;
1828         } while (bh != head);
1829 }
1830 EXPORT_SYMBOL(page_zero_new_buffers);
1831
1832 static int __block_prepare_write(struct inode *inode, struct page *page,
1833                 unsigned from, unsigned to, get_block_t *get_block)
1834 {
1835         unsigned block_start, block_end;
1836         sector_t block;
1837         int err = 0;
1838         unsigned blocksize, bbits;
1839         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1840
1841         BUG_ON(!PageLocked(page));
1842         BUG_ON(from > PAGE_CACHE_SIZE);
1843         BUG_ON(to > PAGE_CACHE_SIZE);
1844         BUG_ON(from > to);
1845
1846         blocksize = 1 << inode->i_blkbits;
1847         if (!page_has_buffers(page))
1848                 create_empty_buffers(page, blocksize, 0);
1849         head = page_buffers(page);
1850
1851         bbits = inode->i_blkbits;
1852         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1853
1854         for(bh = head, block_start = 0; bh != head || !block_start;
1855             block++, block_start=block_end, bh = bh->b_this_page) {
1856                 block_end = block_start + blocksize;
1857                 if (block_end <= from || block_start >= to) {
1858                         if (PageUptodate(page)) {
1859                                 if (!buffer_uptodate(bh))
1860                                         set_buffer_uptodate(bh);
1861                         }
1862                         continue;
1863                 }
1864                 if (buffer_new(bh))
1865                         clear_buffer_new(bh);
1866                 if (!buffer_mapped(bh)) {
1867                         WARN_ON(bh->b_size != blocksize);
1868                         err = get_block(inode, block, bh, 1);
1869                         if (err)
1870                                 break;
1871                         if (buffer_new(bh)) {
1872                                 unmap_underlying_metadata(bh->b_bdev,
1873                                                         bh->b_blocknr);
1874                                 if (PageUptodate(page)) {
1875                                         clear_buffer_new(bh);
1876                                         set_buffer_uptodate(bh);
1877                                         mark_buffer_dirty(bh);
1878                                         continue;
1879                                 }
1880                                 if (block_end > to || block_start < from)
1881                                         zero_user_segments(page,
1882                                                 to, block_end,
1883                                                 block_start, from);
1884                                 continue;
1885                         }
1886                 }
1887                 if (PageUptodate(page)) {
1888                         if (!buffer_uptodate(bh))
1889                                 set_buffer_uptodate(bh);
1890                         continue; 
1891                 }
1892                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1893                     !buffer_unwritten(bh) &&
1894                      (block_start < from || block_end > to)) {
1895                         ll_rw_block(READ, 1, &bh);
1896                         *wait_bh++=bh;
1897                 }
1898         }
1899         /*
1900          * If we issued read requests - let them complete.
1901          */
1902         while(wait_bh > wait) {
1903                 wait_on_buffer(*--wait_bh);
1904                 if (!buffer_uptodate(*wait_bh))
1905                         err = -EIO;
1906         }
1907         if (unlikely(err))
1908                 page_zero_new_buffers(page, from, to);
1909         return err;
1910 }
1911
1912 static int __block_commit_write(struct inode *inode, struct page *page,
1913                 unsigned from, unsigned to)
1914 {
1915         unsigned block_start, block_end;
1916         int partial = 0;
1917         unsigned blocksize;
1918         struct buffer_head *bh, *head;
1919
1920         blocksize = 1 << inode->i_blkbits;
1921
1922         for(bh = head = page_buffers(page), block_start = 0;
1923             bh != head || !block_start;
1924             block_start=block_end, bh = bh->b_this_page) {
1925                 block_end = block_start + blocksize;
1926                 if (block_end <= from || block_start >= to) {
1927                         if (!buffer_uptodate(bh))
1928                                 partial = 1;
1929                 } else {
1930                         set_buffer_uptodate(bh);
1931                         mark_buffer_dirty(bh);
1932                 }
1933                 clear_buffer_new(bh);
1934         }
1935
1936         /*
1937          * If this is a partial write which happened to make all buffers
1938          * uptodate then we can optimize away a bogus readpage() for
1939          * the next read(). Here we 'discover' whether the page went
1940          * uptodate as a result of this (potentially partial) write.
1941          */
1942         if (!partial)
1943                 SetPageUptodate(page);
1944         return 0;
1945 }
1946
1947 /*
1948  * block_write_begin takes care of the basic task of block allocation and
1949  * bringing partial write blocks uptodate first.
1950  *
1951  * If *pagep is not NULL, then block_write_begin uses the locked page
1952  * at *pagep rather than allocating its own. In this case, the page will
1953  * not be unlocked or deallocated on failure.
1954  */
1955 int block_write_begin(struct file *file, struct address_space *mapping,
1956                         loff_t pos, unsigned len, unsigned flags,
1957                         struct page **pagep, void **fsdata,
1958                         get_block_t *get_block)
1959 {
1960         struct inode *inode = mapping->host;
1961         int status = 0;
1962         struct page *page;
1963         pgoff_t index;
1964         unsigned start, end;
1965         int ownpage = 0;
1966
1967         index = pos >> PAGE_CACHE_SHIFT;
1968         start = pos & (PAGE_CACHE_SIZE - 1);
1969         end = start + len;
1970
1971         page = *pagep;
1972         if (page == NULL) {
1973                 ownpage = 1;
1974                 page = __grab_cache_page(mapping, index);
1975                 if (!page) {
1976                         status = -ENOMEM;
1977                         goto out;
1978                 }
1979                 *pagep = page;
1980         } else
1981                 BUG_ON(!PageLocked(page));
1982
1983         status = __block_prepare_write(inode, page, start, end, get_block);
1984         if (unlikely(status)) {
1985                 ClearPageUptodate(page);
1986
1987                 if (ownpage) {
1988                         unlock_page(page);
1989                         page_cache_release(page);
1990                         *pagep = NULL;
1991
1992                         /*
1993                          * prepare_write() may have instantiated a few blocks
1994                          * outside i_size.  Trim these off again. Don't need
1995                          * i_size_read because we hold i_mutex.
1996                          */
1997                         if (pos + len > inode->i_size)
1998                                 vmtruncate(inode, inode->i_size);
1999                 }
2000                 goto out;
2001         }
2002
2003 out:
2004         return status;
2005 }
2006 EXPORT_SYMBOL(block_write_begin);
2007
2008 int block_write_end(struct file *file, struct address_space *mapping,
2009                         loff_t pos, unsigned len, unsigned copied,
2010                         struct page *page, void *fsdata)
2011 {
2012         struct inode *inode = mapping->host;
2013         unsigned start;
2014
2015         start = pos & (PAGE_CACHE_SIZE - 1);
2016
2017         if (unlikely(copied < len)) {
2018                 /*
2019                  * The buffers that were written will now be uptodate, so we
2020                  * don't have to worry about a readpage reading them and
2021                  * overwriting a partial write. However if we have encountered
2022                  * a short write and only partially written into a buffer, it
2023                  * will not be marked uptodate, so a readpage might come in and
2024                  * destroy our partial write.
2025                  *
2026                  * Do the simplest thing, and just treat any short write to a
2027                  * non uptodate page as a zero-length write, and force the
2028                  * caller to redo the whole thing.
2029                  */
2030                 if (!PageUptodate(page))
2031                         copied = 0;
2032
2033                 page_zero_new_buffers(page, start+copied, start+len);
2034         }
2035         flush_dcache_page(page);
2036
2037         /* This could be a short (even 0-length) commit */
2038         __block_commit_write(inode, page, start, start+copied);
2039
2040         return copied;
2041 }
2042 EXPORT_SYMBOL(block_write_end);
2043
2044 int generic_write_end(struct file *file, struct address_space *mapping,
2045                         loff_t pos, unsigned len, unsigned copied,
2046                         struct page *page, void *fsdata)
2047 {
2048         struct inode *inode = mapping->host;
2049
2050         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2051
2052         /*
2053          * No need to use i_size_read() here, the i_size
2054          * cannot change under us because we hold i_mutex.
2055          *
2056          * But it's important to update i_size while still holding page lock:
2057          * page writeout could otherwise come in and zero beyond i_size.
2058          */
2059         if (pos+copied > inode->i_size) {
2060                 i_size_write(inode, pos+copied);
2061                 mark_inode_dirty(inode);
2062         }
2063
2064         unlock_page(page);
2065         page_cache_release(page);
2066
2067         return copied;
2068 }
2069 EXPORT_SYMBOL(generic_write_end);
2070
2071 /*
2072  * Generic "read page" function for block devices that have the normal
2073  * get_block functionality. This is most of the block device filesystems.
2074  * Reads the page asynchronously --- the unlock_buffer() and
2075  * set/clear_buffer_uptodate() functions propagate buffer state into the
2076  * page struct once IO has completed.
2077  */
2078 int block_read_full_page(struct page *page, get_block_t *get_block)
2079 {
2080         struct inode *inode = page->mapping->host;
2081         sector_t iblock, lblock;
2082         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2083         unsigned int blocksize;
2084         int nr, i;
2085         int fully_mapped = 1;
2086
2087         BUG_ON(!PageLocked(page));
2088         blocksize = 1 << inode->i_blkbits;
2089         if (!page_has_buffers(page))
2090                 create_empty_buffers(page, blocksize, 0);
2091         head = page_buffers(page);
2092
2093         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2094         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2095         bh = head;
2096         nr = 0;
2097         i = 0;
2098
2099         do {
2100                 if (buffer_uptodate(bh))
2101                         continue;
2102
2103                 if (!buffer_mapped(bh)) {
2104                         int err = 0;
2105
2106                         fully_mapped = 0;
2107                         if (iblock < lblock) {
2108                                 WARN_ON(bh->b_size != blocksize);
2109                                 err = get_block(inode, iblock, bh, 0);
2110                                 if (err)
2111                                         SetPageError(page);
2112                         }
2113                         if (!buffer_mapped(bh)) {
2114                                 zero_user(page, i * blocksize, blocksize);
2115                                 if (!err)
2116                                         set_buffer_uptodate(bh);
2117                                 continue;
2118                         }
2119                         /*
2120                          * get_block() might have updated the buffer
2121                          * synchronously
2122                          */
2123                         if (buffer_uptodate(bh))
2124                                 continue;
2125                 }
2126                 arr[nr++] = bh;
2127         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2128
2129         if (fully_mapped)
2130                 SetPageMappedToDisk(page);
2131
2132         if (!nr) {
2133                 /*
2134                  * All buffers are uptodate - we can set the page uptodate
2135                  * as well. But not if get_block() returned an error.
2136                  */
2137                 if (!PageError(page))
2138                         SetPageUptodate(page);
2139                 unlock_page(page);
2140                 return 0;
2141         }
2142
2143         /* Stage two: lock the buffers */
2144         for (i = 0; i < nr; i++) {
2145                 bh = arr[i];
2146                 lock_buffer(bh);
2147                 mark_buffer_async_read(bh);
2148         }
2149
2150         /*
2151          * Stage 3: start the IO.  Check for uptodateness
2152          * inside the buffer lock in case another process reading
2153          * the underlying blockdev brought it uptodate (the sct fix).
2154          */
2155         for (i = 0; i < nr; i++) {
2156                 bh = arr[i];
2157                 if (buffer_uptodate(bh))
2158                         end_buffer_async_read(bh, 1);
2159                 else
2160                         submit_bh(READ, bh);
2161         }
2162         return 0;
2163 }
2164
2165 /* utility function for filesystems that need to do work on expanding
2166  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2167  * deal with the hole.  
2168  */
2169 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2170 {
2171         struct address_space *mapping = inode->i_mapping;
2172         struct page *page;
2173         void *fsdata;
2174         unsigned long limit;
2175         int err;
2176
2177         err = -EFBIG;
2178         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2179         if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2180                 send_sig(SIGXFSZ, current, 0);
2181                 goto out;
2182         }
2183         if (size > inode->i_sb->s_maxbytes)
2184                 goto out;
2185
2186         err = pagecache_write_begin(NULL, mapping, size, 0,
2187                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2188                                 &page, &fsdata);
2189         if (err)
2190                 goto out;
2191
2192         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2193         BUG_ON(err > 0);
2194
2195 out:
2196         return err;
2197 }
2198
2199 int cont_expand_zero(struct file *file, struct address_space *mapping,
2200                         loff_t pos, loff_t *bytes)
2201 {
2202         struct inode *inode = mapping->host;
2203         unsigned blocksize = 1 << inode->i_blkbits;
2204         struct page *page;
2205         void *fsdata;
2206         pgoff_t index, curidx;
2207         loff_t curpos;
2208         unsigned zerofrom, offset, len;
2209         int err = 0;
2210
2211         index = pos >> PAGE_CACHE_SHIFT;
2212         offset = pos & ~PAGE_CACHE_MASK;
2213
2214         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2215                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2216                 if (zerofrom & (blocksize-1)) {
2217                         *bytes |= (blocksize-1);
2218                         (*bytes)++;
2219                 }
2220                 len = PAGE_CACHE_SIZE - zerofrom;
2221
2222                 err = pagecache_write_begin(file, mapping, curpos, len,
2223                                                 AOP_FLAG_UNINTERRUPTIBLE,
2224                                                 &page, &fsdata);
2225                 if (err)
2226                         goto out;
2227                 zero_user(page, zerofrom, len);
2228                 err = pagecache_write_end(file, mapping, curpos, len, len,
2229                                                 page, fsdata);
2230                 if (err < 0)
2231                         goto out;
2232                 BUG_ON(err != len);
2233                 err = 0;
2234         }
2235
2236         /* page covers the boundary, find the boundary offset */
2237         if (index == curidx) {
2238                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2239                 /* if we will expand the thing last block will be filled */
2240                 if (offset <= zerofrom) {
2241                         goto out;
2242                 }
2243                 if (zerofrom & (blocksize-1)) {
2244                         *bytes |= (blocksize-1);
2245                         (*bytes)++;
2246                 }
2247                 len = offset - zerofrom;
2248
2249                 err = pagecache_write_begin(file, mapping, curpos, len,
2250                                                 AOP_FLAG_UNINTERRUPTIBLE,
2251                                                 &page, &fsdata);
2252                 if (err)
2253                         goto out;
2254                 zero_user(page, zerofrom, len);
2255                 err = pagecache_write_end(file, mapping, curpos, len, len,
2256                                                 page, fsdata);
2257                 if (err < 0)
2258                         goto out;
2259                 BUG_ON(err != len);
2260                 err = 0;
2261         }
2262 out:
2263         return err;
2264 }
2265
2266 /*
2267  * For moronic filesystems that do not allow holes in file.
2268  * We may have to extend the file.
2269  */
2270 int cont_write_begin(struct file *file, struct address_space *mapping,
2271                         loff_t pos, unsigned len, unsigned flags,
2272                         struct page **pagep, void **fsdata,
2273                         get_block_t *get_block, loff_t *bytes)
2274 {
2275         struct inode *inode = mapping->host;
2276         unsigned blocksize = 1 << inode->i_blkbits;
2277         unsigned zerofrom;
2278         int err;
2279
2280         err = cont_expand_zero(file, mapping, pos, bytes);
2281         if (err)
2282                 goto out;
2283
2284         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2285         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2286                 *bytes |= (blocksize-1);
2287                 (*bytes)++;
2288         }
2289
2290         *pagep = NULL;
2291         err = block_write_begin(file, mapping, pos, len,
2292                                 flags, pagep, fsdata, get_block);
2293 out:
2294         return err;
2295 }
2296
2297 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2298                         get_block_t *get_block)
2299 {
2300         struct inode *inode = page->mapping->host;
2301         int err = __block_prepare_write(inode, page, from, to, get_block);
2302         if (err)
2303                 ClearPageUptodate(page);
2304         return err;
2305 }
2306
2307 int block_commit_write(struct page *page, unsigned from, unsigned to)
2308 {
2309         struct inode *inode = page->mapping->host;
2310         __block_commit_write(inode,page,from,to);
2311         return 0;
2312 }
2313
2314 int generic_commit_write(struct file *file, struct page *page,
2315                 unsigned from, unsigned to)
2316 {
2317         struct inode *inode = page->mapping->host;
2318         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2319         __block_commit_write(inode,page,from,to);
2320         /*
2321          * No need to use i_size_read() here, the i_size
2322          * cannot change under us because we hold i_mutex.
2323          */
2324         if (pos > inode->i_size) {
2325                 i_size_write(inode, pos);
2326                 mark_inode_dirty(inode);
2327         }
2328         return 0;
2329 }
2330
2331 /*
2332  * block_page_mkwrite() is not allowed to change the file size as it gets
2333  * called from a page fault handler when a page is first dirtied. Hence we must
2334  * be careful to check for EOF conditions here. We set the page up correctly
2335  * for a written page which means we get ENOSPC checking when writing into
2336  * holes and correct delalloc and unwritten extent mapping on filesystems that
2337  * support these features.
2338  *
2339  * We are not allowed to take the i_mutex here so we have to play games to
2340  * protect against truncate races as the page could now be beyond EOF.  Because
2341  * vmtruncate() writes the inode size before removing pages, once we have the
2342  * page lock we can determine safely if the page is beyond EOF. If it is not
2343  * beyond EOF, then the page is guaranteed safe against truncation until we
2344  * unlock the page.
2345  */
2346 int
2347 block_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2348                    get_block_t get_block)
2349 {
2350         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2351         unsigned long end;
2352         loff_t size;
2353         int ret = -EINVAL;
2354
2355         lock_page(page);
2356         size = i_size_read(inode);
2357         if ((page->mapping != inode->i_mapping) ||
2358             (page_offset(page) > size)) {
2359                 /* page got truncated out from underneath us */
2360                 goto out_unlock;
2361         }
2362
2363         /* page is wholly or partially inside EOF */
2364         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2365                 end = size & ~PAGE_CACHE_MASK;
2366         else
2367                 end = PAGE_CACHE_SIZE;
2368
2369         ret = block_prepare_write(page, 0, end, get_block);
2370         if (!ret)
2371                 ret = block_commit_write(page, 0, end);
2372
2373 out_unlock:
2374         unlock_page(page);
2375         return ret;
2376 }
2377
2378 /*
2379  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2380  * immediately, while under the page lock.  So it needs a special end_io
2381  * handler which does not touch the bh after unlocking it.
2382  */
2383 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2384 {
2385         __end_buffer_read_notouch(bh, uptodate);
2386 }
2387
2388 /*
2389  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2390  * the page (converting it to circular linked list and taking care of page
2391  * dirty races).
2392  */
2393 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2394 {
2395         struct buffer_head *bh;
2396
2397         BUG_ON(!PageLocked(page));
2398
2399         spin_lock(&page->mapping->private_lock);
2400         bh = head;
2401         do {
2402                 if (PageDirty(page))
2403                         set_buffer_dirty(bh);
2404                 if (!bh->b_this_page)
2405                         bh->b_this_page = head;
2406                 bh = bh->b_this_page;
2407         } while (bh != head);
2408         attach_page_buffers(page, head);
2409         spin_unlock(&page->mapping->private_lock);
2410 }
2411
2412 /*
2413  * On entry, the page is fully not uptodate.
2414  * On exit the page is fully uptodate in the areas outside (from,to)
2415  */
2416 int nobh_write_begin(struct file *file, struct address_space *mapping,
2417                         loff_t pos, unsigned len, unsigned flags,
2418                         struct page **pagep, void **fsdata,
2419                         get_block_t *get_block)
2420 {
2421         struct inode *inode = mapping->host;
2422         const unsigned blkbits = inode->i_blkbits;
2423         const unsigned blocksize = 1 << blkbits;
2424         struct buffer_head *head, *bh;
2425         struct page *page;
2426         pgoff_t index;
2427         unsigned from, to;
2428         unsigned block_in_page;
2429         unsigned block_start, block_end;
2430         sector_t block_in_file;
2431         int nr_reads = 0;
2432         int ret = 0;
2433         int is_mapped_to_disk = 1;
2434
2435         index = pos >> PAGE_CACHE_SHIFT;
2436         from = pos & (PAGE_CACHE_SIZE - 1);
2437         to = from + len;
2438
2439         page = __grab_cache_page(mapping, index);
2440         if (!page)
2441                 return -ENOMEM;
2442         *pagep = page;
2443         *fsdata = NULL;
2444
2445         if (page_has_buffers(page)) {
2446                 unlock_page(page);
2447                 page_cache_release(page);
2448                 *pagep = NULL;
2449                 return block_write_begin(file, mapping, pos, len, flags, pagep,
2450                                         fsdata, get_block);
2451         }
2452
2453         if (PageMappedToDisk(page))
2454                 return 0;
2455
2456         /*
2457          * Allocate buffers so that we can keep track of state, and potentially
2458          * attach them to the page if an error occurs. In the common case of
2459          * no error, they will just be freed again without ever being attached
2460          * to the page (which is all OK, because we're under the page lock).
2461          *
2462          * Be careful: the buffer linked list is a NULL terminated one, rather
2463          * than the circular one we're used to.
2464          */
2465         head = alloc_page_buffers(page, blocksize, 0);
2466         if (!head) {
2467                 ret = -ENOMEM;
2468                 goto out_release;
2469         }
2470
2471         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2472
2473         /*
2474          * We loop across all blocks in the page, whether or not they are
2475          * part of the affected region.  This is so we can discover if the
2476          * page is fully mapped-to-disk.
2477          */
2478         for (block_start = 0, block_in_page = 0, bh = head;
2479                   block_start < PAGE_CACHE_SIZE;
2480                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2481                 int create;
2482
2483                 block_end = block_start + blocksize;
2484                 bh->b_state = 0;
2485                 create = 1;
2486                 if (block_start >= to)
2487                         create = 0;
2488                 ret = get_block(inode, block_in_file + block_in_page,
2489                                         bh, create);
2490                 if (ret)
2491                         goto failed;
2492                 if (!buffer_mapped(bh))
2493                         is_mapped_to_disk = 0;
2494                 if (buffer_new(bh))
2495                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2496                 if (PageUptodate(page)) {
2497                         set_buffer_uptodate(bh);
2498                         continue;
2499                 }
2500                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2501                         zero_user_segments(page, block_start, from,
2502                                                         to, block_end);
2503                         continue;
2504                 }
2505                 if (buffer_uptodate(bh))
2506                         continue;       /* reiserfs does this */
2507                 if (block_start < from || block_end > to) {
2508                         lock_buffer(bh);
2509                         bh->b_end_io = end_buffer_read_nobh;
2510                         submit_bh(READ, bh);
2511                         nr_reads++;
2512                 }
2513         }
2514
2515         if (nr_reads) {
2516                 /*
2517                  * The page is locked, so these buffers are protected from
2518                  * any VM or truncate activity.  Hence we don't need to care
2519                  * for the buffer_head refcounts.
2520                  */
2521                 for (bh = head; bh; bh = bh->b_this_page) {
2522                         wait_on_buffer(bh);
2523                         if (!buffer_uptodate(bh))
2524                                 ret = -EIO;
2525                 }
2526                 if (ret)
2527                         goto failed;
2528         }
2529
2530         if (is_mapped_to_disk)
2531                 SetPageMappedToDisk(page);
2532
2533         *fsdata = head; /* to be released by nobh_write_end */
2534
2535         return 0;
2536
2537 failed:
2538         BUG_ON(!ret);
2539         /*
2540          * Error recovery is a bit difficult. We need to zero out blocks that
2541          * were newly allocated, and dirty them to ensure they get written out.
2542          * Buffers need to be attached to the page at this point, otherwise
2543          * the handling of potential IO errors during writeout would be hard
2544          * (could try doing synchronous writeout, but what if that fails too?)
2545          */
2546         attach_nobh_buffers(page, head);
2547         page_zero_new_buffers(page, from, to);
2548
2549 out_release:
2550         unlock_page(page);
2551         page_cache_release(page);
2552         *pagep = NULL;
2553
2554         if (pos + len > inode->i_size)
2555                 vmtruncate(inode, inode->i_size);
2556
2557         return ret;
2558 }
2559 EXPORT_SYMBOL(nobh_write_begin);
2560
2561 int nobh_write_end(struct file *file, struct address_space *mapping,
2562                         loff_t pos, unsigned len, unsigned copied,
2563                         struct page *page, void *fsdata)
2564 {
2565         struct inode *inode = page->mapping->host;
2566         struct buffer_head *head = fsdata;
2567         struct buffer_head *bh;
2568
2569         if (!PageMappedToDisk(page)) {
2570                 if (unlikely(copied < len) && !page_has_buffers(page))
2571                         attach_nobh_buffers(page, head);
2572                 if (page_has_buffers(page))
2573                         return generic_write_end(file, mapping, pos, len,
2574                                                 copied, page, fsdata);
2575         }
2576
2577         SetPageUptodate(page);
2578         set_page_dirty(page);
2579         if (pos+copied > inode->i_size) {
2580                 i_size_write(inode, pos+copied);
2581                 mark_inode_dirty(inode);
2582         }
2583
2584         unlock_page(page);
2585         page_cache_release(page);
2586
2587         while (head) {
2588                 bh = head;
2589                 head = head->b_this_page;
2590                 free_buffer_head(bh);
2591         }
2592
2593         return copied;
2594 }
2595 EXPORT_SYMBOL(nobh_write_end);
2596
2597 /*
2598  * nobh_writepage() - based on block_full_write_page() except
2599  * that it tries to operate without attaching bufferheads to
2600  * the page.
2601  */
2602 int nobh_writepage(struct page *page, get_block_t *get_block,
2603                         struct writeback_control *wbc)
2604 {
2605         struct inode * const inode = page->mapping->host;
2606         loff_t i_size = i_size_read(inode);
2607         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2608         unsigned offset;
2609         int ret;
2610
2611         /* Is the page fully inside i_size? */
2612         if (page->index < end_index)
2613                 goto out;
2614
2615         /* Is the page fully outside i_size? (truncate in progress) */
2616         offset = i_size & (PAGE_CACHE_SIZE-1);
2617         if (page->index >= end_index+1 || !offset) {
2618                 /*
2619                  * The page may have dirty, unmapped buffers.  For example,
2620                  * they may have been added in ext3_writepage().  Make them
2621                  * freeable here, so the page does not leak.
2622                  */
2623 #if 0
2624                 /* Not really sure about this  - do we need this ? */
2625                 if (page->mapping->a_ops->invalidatepage)
2626                         page->mapping->a_ops->invalidatepage(page, offset);
2627 #endif
2628                 unlock_page(page);
2629                 return 0; /* don't care */
2630         }
2631
2632         /*
2633          * The page straddles i_size.  It must be zeroed out on each and every
2634          * writepage invocation because it may be mmapped.  "A file is mapped
2635          * in multiples of the page size.  For a file that is not a multiple of
2636          * the  page size, the remaining memory is zeroed when mapped, and
2637          * writes to that region are not written out to the file."
2638          */
2639         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2640 out:
2641         ret = mpage_writepage(page, get_block, wbc);
2642         if (ret == -EAGAIN)
2643                 ret = __block_write_full_page(inode, page, get_block, wbc);
2644         return ret;
2645 }
2646 EXPORT_SYMBOL(nobh_writepage);
2647
2648 int nobh_truncate_page(struct address_space *mapping,
2649                         loff_t from, get_block_t *get_block)
2650 {
2651         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2652         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2653         unsigned blocksize;
2654         sector_t iblock;
2655         unsigned length, pos;
2656         struct inode *inode = mapping->host;
2657         struct page *page;
2658         struct buffer_head map_bh;
2659         int err;
2660
2661         blocksize = 1 << inode->i_blkbits;
2662         length = offset & (blocksize - 1);
2663
2664         /* Block boundary? Nothing to do */
2665         if (!length)
2666                 return 0;
2667
2668         length = blocksize - length;
2669         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2670
2671         page = grab_cache_page(mapping, index);
2672         err = -ENOMEM;
2673         if (!page)
2674                 goto out;
2675
2676         if (page_has_buffers(page)) {
2677 has_buffers:
2678                 unlock_page(page);
2679                 page_cache_release(page);
2680                 return block_truncate_page(mapping, from, get_block);
2681         }
2682
2683         /* Find the buffer that contains "offset" */
2684         pos = blocksize;
2685         while (offset >= pos) {
2686                 iblock++;
2687                 pos += blocksize;
2688         }
2689
2690         err = get_block(inode, iblock, &map_bh, 0);
2691         if (err)
2692                 goto unlock;
2693         /* unmapped? It's a hole - nothing to do */
2694         if (!buffer_mapped(&map_bh))
2695                 goto unlock;
2696
2697         /* Ok, it's mapped. Make sure it's up-to-date */
2698         if (!PageUptodate(page)) {
2699                 err = mapping->a_ops->readpage(NULL, page);
2700                 if (err) {
2701                         page_cache_release(page);
2702                         goto out;
2703                 }
2704                 lock_page(page);
2705                 if (!PageUptodate(page)) {
2706                         err = -EIO;
2707                         goto unlock;
2708                 }
2709                 if (page_has_buffers(page))
2710                         goto has_buffers;
2711         }
2712         zero_user(page, offset, length);
2713         set_page_dirty(page);
2714         err = 0;
2715
2716 unlock:
2717         unlock_page(page);
2718         page_cache_release(page);
2719 out:
2720         return err;
2721 }
2722 EXPORT_SYMBOL(nobh_truncate_page);
2723
2724 int block_truncate_page(struct address_space *mapping,
2725                         loff_t from, get_block_t *get_block)
2726 {
2727         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2728         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2729         unsigned blocksize;
2730         sector_t iblock;
2731         unsigned length, pos;
2732         struct inode *inode = mapping->host;
2733         struct page *page;
2734         struct buffer_head *bh;
2735         int err;
2736
2737         blocksize = 1 << inode->i_blkbits;
2738         length = offset & (blocksize - 1);
2739
2740         /* Block boundary? Nothing to do */
2741         if (!length)
2742                 return 0;
2743
2744         length = blocksize - length;
2745         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2746         
2747         page = grab_cache_page(mapping, index);
2748         err = -ENOMEM;
2749         if (!page)
2750                 goto out;
2751
2752         if (!page_has_buffers(page))
2753                 create_empty_buffers(page, blocksize, 0);
2754
2755         /* Find the buffer that contains "offset" */
2756         bh = page_buffers(page);
2757         pos = blocksize;
2758         while (offset >= pos) {
2759                 bh = bh->b_this_page;
2760                 iblock++;
2761                 pos += blocksize;
2762         }
2763
2764         err = 0;
2765         if (!buffer_mapped(bh)) {
2766                 WARN_ON(bh->b_size != blocksize);
2767                 err = get_block(inode, iblock, bh, 0);
2768                 if (err)
2769                         goto unlock;
2770                 /* unmapped? It's a hole - nothing to do */
2771                 if (!buffer_mapped(bh))
2772                         goto unlock;
2773         }
2774
2775         /* Ok, it's mapped. Make sure it's up-to-date */
2776         if (PageUptodate(page))
2777                 set_buffer_uptodate(bh);
2778
2779         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2780                 err = -EIO;
2781                 ll_rw_block(READ, 1, &bh);
2782                 wait_on_buffer(bh);
2783                 /* Uhhuh. Read error. Complain and punt. */
2784                 if (!buffer_uptodate(bh))
2785                         goto unlock;
2786         }
2787
2788         zero_user(page, offset, length);
2789         mark_buffer_dirty(bh);
2790         err = 0;
2791
2792 unlock:
2793         unlock_page(page);
2794         page_cache_release(page);
2795 out:
2796         return err;
2797 }
2798
2799 /*
2800  * The generic ->writepage function for buffer-backed address_spaces
2801  */
2802 int block_write_full_page(struct page *page, get_block_t *get_block,
2803                         struct writeback_control *wbc)
2804 {
2805         struct inode * const inode = page->mapping->host;
2806         loff_t i_size = i_size_read(inode);
2807         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2808         unsigned offset;
2809
2810         /* Is the page fully inside i_size? */
2811         if (page->index < end_index)
2812                 return __block_write_full_page(inode, page, get_block, wbc);
2813
2814         /* Is the page fully outside i_size? (truncate in progress) */
2815         offset = i_size & (PAGE_CACHE_SIZE-1);
2816         if (page->index >= end_index+1 || !offset) {
2817                 /*
2818                  * The page may have dirty, unmapped buffers.  For example,
2819                  * they may have been added in ext3_writepage().  Make them
2820                  * freeable here, so the page does not leak.
2821                  */
2822                 do_invalidatepage(page, 0);
2823                 unlock_page(page);
2824                 return 0; /* don't care */
2825         }
2826
2827         /*
2828          * The page straddles i_size.  It must be zeroed out on each and every
2829          * writepage invokation because it may be mmapped.  "A file is mapped
2830          * in multiples of the page size.  For a file that is not a multiple of
2831          * the  page size, the remaining memory is zeroed when mapped, and
2832          * writes to that region are not written out to the file."
2833          */
2834         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2835         return __block_write_full_page(inode, page, get_block, wbc);
2836 }
2837
2838 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2839                             get_block_t *get_block)
2840 {
2841         struct buffer_head tmp;
2842         struct inode *inode = mapping->host;
2843         tmp.b_state = 0;
2844         tmp.b_blocknr = 0;
2845         tmp.b_size = 1 << inode->i_blkbits;
2846         get_block(inode, block, &tmp, 0);
2847         return tmp.b_blocknr;
2848 }
2849
2850 static void end_bio_bh_io_sync(struct bio *bio, int err)
2851 {
2852         struct buffer_head *bh = bio->bi_private;
2853
2854         if (err == -EOPNOTSUPP) {
2855                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2856                 set_bit(BH_Eopnotsupp, &bh->b_state);
2857         }
2858
2859         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2860         bio_put(bio);
2861 }
2862
2863 int submit_bh(int rw, struct buffer_head * bh)
2864 {
2865         struct bio *bio;
2866         int ret = 0;
2867
2868         BUG_ON(!buffer_locked(bh));
2869         BUG_ON(!buffer_mapped(bh));
2870         BUG_ON(!bh->b_end_io);
2871
2872         if (buffer_ordered(bh) && (rw == WRITE))
2873                 rw = WRITE_BARRIER;
2874
2875         /*
2876          * Only clear out a write error when rewriting, should this
2877          * include WRITE_SYNC as well?
2878          */
2879         if (test_set_buffer_req(bh) && (rw == WRITE || rw == WRITE_BARRIER))
2880                 clear_buffer_write_io_error(bh);
2881
2882         /*
2883          * from here on down, it's all bio -- do the initial mapping,
2884          * submit_bio -> generic_make_request may further map this bio around
2885          */
2886         bio = bio_alloc(GFP_NOIO, 1);
2887
2888         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2889         bio->bi_bdev = bh->b_bdev;
2890         bio->bi_io_vec[0].bv_page = bh->b_page;
2891         bio->bi_io_vec[0].bv_len = bh->b_size;
2892         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2893
2894         bio->bi_vcnt = 1;
2895         bio->bi_idx = 0;
2896         bio->bi_size = bh->b_size;
2897
2898         bio->bi_end_io = end_bio_bh_io_sync;
2899         bio->bi_private = bh;
2900
2901         bio_get(bio);
2902         submit_bio(rw, bio);
2903
2904         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2905                 ret = -EOPNOTSUPP;
2906
2907         bio_put(bio);
2908         return ret;
2909 }
2910
2911 /**
2912  * ll_rw_block: low-level access to block devices (DEPRECATED)
2913  * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
2914  * @nr: number of &struct buffer_heads in the array
2915  * @bhs: array of pointers to &struct buffer_head
2916  *
2917  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2918  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2919  * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2920  * are sent to disk. The fourth %READA option is described in the documentation
2921  * for generic_make_request() which ll_rw_block() calls.
2922  *
2923  * This function drops any buffer that it cannot get a lock on (with the
2924  * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
2925  * clean when doing a write request, and any buffer that appears to be
2926  * up-to-date when doing read request.  Further it marks as clean buffers that
2927  * are processed for writing (the buffer cache won't assume that they are
2928  * actually clean until the buffer gets unlocked).
2929  *
2930  * ll_rw_block sets b_end_io to simple completion handler that marks
2931  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2932  * any waiters. 
2933  *
2934  * All of the buffers must be for the same device, and must also be a
2935  * multiple of the current approved size for the device.
2936  */
2937 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2938 {
2939         int i;
2940
2941         for (i = 0; i < nr; i++) {
2942                 struct buffer_head *bh = bhs[i];
2943
2944                 if (rw == SWRITE)
2945                         lock_buffer(bh);
2946                 else if (test_set_buffer_locked(bh))
2947                         continue;
2948
2949                 if (rw == WRITE || rw == SWRITE) {
2950                         if (test_clear_buffer_dirty(bh)) {
2951                                 bh->b_end_io = end_buffer_write_sync;
2952                                 get_bh(bh);
2953                                 submit_bh(WRITE, bh);
2954                                 continue;
2955                         }
2956                 } else {
2957                         if (!buffer_uptodate(bh)) {
2958                                 bh->b_end_io = end_buffer_read_sync;
2959                                 get_bh(bh);
2960                                 submit_bh(rw, bh);
2961                                 continue;
2962                         }
2963                 }
2964                 unlock_buffer(bh);
2965         }
2966 }
2967
2968 /*
2969  * For a data-integrity writeout, we need to wait upon any in-progress I/O
2970  * and then start new I/O and then wait upon it.  The caller must have a ref on
2971  * the buffer_head.
2972  */
2973 int sync_dirty_buffer(struct buffer_head *bh)
2974 {
2975         int ret = 0;
2976
2977         WARN_ON(atomic_read(&bh->b_count) < 1);
2978         lock_buffer(bh);
2979         if (test_clear_buffer_dirty(bh)) {
2980                 get_bh(bh);
2981                 bh->b_end_io = end_buffer_write_sync;
2982                 ret = submit_bh(WRITE, bh);
2983                 wait_on_buffer(bh);
2984                 if (buffer_eopnotsupp(bh)) {
2985                         clear_buffer_eopnotsupp(bh);
2986                         ret = -EOPNOTSUPP;
2987                 }
2988                 if (!ret && !buffer_uptodate(bh))
2989                         ret = -EIO;
2990         } else {
2991                 unlock_buffer(bh);
2992         }
2993         return ret;
2994 }
2995
2996 /*
2997  * try_to_free_buffers() checks if all the buffers on this particular page
2998  * are unused, and releases them if so.
2999  *
3000  * Exclusion against try_to_free_buffers may be obtained by either
3001  * locking the page or by holding its mapping's private_lock.
3002  *
3003  * If the page is dirty but all the buffers are clean then we need to
3004  * be sure to mark the page clean as well.  This is because the page
3005  * may be against a block device, and a later reattachment of buffers
3006  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3007  * filesystem data on the same device.
3008  *
3009  * The same applies to regular filesystem pages: if all the buffers are
3010  * clean then we set the page clean and proceed.  To do that, we require
3011  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3012  * private_lock.
3013  *
3014  * try_to_free_buffers() is non-blocking.
3015  */
3016 static inline int buffer_busy(struct buffer_head *bh)
3017 {
3018         return atomic_read(&bh->b_count) |
3019                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3020 }
3021
3022 static int
3023 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3024 {
3025         struct buffer_head *head = page_buffers(page);
3026         struct buffer_head *bh;
3027
3028         bh = head;
3029         do {
3030                 if (buffer_write_io_error(bh) && page->mapping)
3031                         set_bit(AS_EIO, &page->mapping->flags);
3032                 if (buffer_busy(bh))
3033                         goto failed;
3034                 bh = bh->b_this_page;
3035         } while (bh != head);
3036
3037         do {
3038                 struct buffer_head *next = bh->b_this_page;
3039
3040                 if (bh->b_assoc_map)
3041                         __remove_assoc_queue(bh);
3042                 bh = next;
3043         } while (bh != head);
3044         *buffers_to_free = head;
3045         __clear_page_buffers(page);
3046         return 1;
3047 failed:
3048         return 0;
3049 }
3050
3051 int try_to_free_buffers(struct page *page)
3052 {
3053         struct address_space * const mapping = page->mapping;
3054         struct buffer_head *buffers_to_free = NULL;
3055         int ret = 0;
3056
3057         BUG_ON(!PageLocked(page));
3058         if (PageWriteback(page))
3059                 return 0;
3060
3061         if (mapping == NULL) {          /* can this still happen? */
3062                 ret = drop_buffers(page, &buffers_to_free);
3063                 goto out;
3064         }
3065
3066         spin_lock(&mapping->private_lock);
3067         ret = drop_buffers(page, &buffers_to_free);
3068
3069         /*
3070          * If the filesystem writes its buffers by hand (eg ext3)
3071          * then we can have clean buffers against a dirty page.  We
3072          * clean the page here; otherwise the VM will never notice
3073          * that the filesystem did any IO at all.
3074          *
3075          * Also, during truncate, discard_buffer will have marked all
3076          * the page's buffers clean.  We discover that here and clean
3077          * the page also.
3078          *
3079          * private_lock must be held over this entire operation in order
3080          * to synchronise against __set_page_dirty_buffers and prevent the
3081          * dirty bit from being lost.
3082          */
3083         if (ret)
3084                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3085         spin_unlock(&mapping->private_lock);
3086 out:
3087         if (buffers_to_free) {
3088                 struct buffer_head *bh = buffers_to_free;
3089
3090                 do {
3091                         struct buffer_head *next = bh->b_this_page;
3092                         free_buffer_head(bh);
3093                         bh = next;
3094                 } while (bh != buffers_to_free);
3095         }
3096         return ret;
3097 }
3098 EXPORT_SYMBOL(try_to_free_buffers);
3099
3100 void block_sync_page(struct page *page)
3101 {
3102         struct address_space *mapping;
3103
3104         smp_mb();
3105         mapping = page_mapping(page);
3106         if (mapping)
3107                 blk_run_backing_dev(mapping->backing_dev_info, page);
3108 }
3109
3110 /*
3111  * There are no bdflush tunables left.  But distributions are
3112  * still running obsolete flush daemons, so we terminate them here.
3113  *
3114  * Use of bdflush() is deprecated and will be removed in a future kernel.
3115  * The `pdflush' kernel threads fully replace bdflush daemons and this call.
3116  */
3117 asmlinkage long sys_bdflush(int func, long data)
3118 {
3119         static int msg_count;
3120
3121         if (!capable(CAP_SYS_ADMIN))
3122                 return -EPERM;
3123
3124         if (msg_count < 5) {
3125                 msg_count++;
3126                 printk(KERN_INFO
3127                         "warning: process `%s' used the obsolete bdflush"
3128                         " system call\n", current->comm);
3129                 printk(KERN_INFO "Fix your initscripts?\n");
3130         }
3131
3132         if (func == 1)
3133                 do_exit(0);
3134         return 0;
3135 }
3136
3137 /*
3138  * Buffer-head allocation
3139  */
3140 static struct kmem_cache *bh_cachep;
3141
3142 /*
3143  * Once the number of bh's in the machine exceeds this level, we start
3144  * stripping them in writeback.
3145  */
3146 static int max_buffer_heads;
3147
3148 int buffer_heads_over_limit;
3149
3150 struct bh_accounting {
3151         int nr;                 /* Number of live bh's */
3152         int ratelimit;          /* Limit cacheline bouncing */
3153 };
3154
3155 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3156
3157 static void recalc_bh_state(void)
3158 {
3159         int i;
3160         int tot = 0;
3161
3162         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3163                 return;
3164         __get_cpu_var(bh_accounting).ratelimit = 0;
3165         for_each_online_cpu(i)
3166                 tot += per_cpu(bh_accounting, i).nr;
3167         buffer_heads_over_limit = (tot > max_buffer_heads);
3168 }
3169         
3170 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3171 {
3172         struct buffer_head *ret = kmem_cache_alloc(bh_cachep,
3173                                 set_migrateflags(gfp_flags, __GFP_RECLAIMABLE));
3174         if (ret) {
3175                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3176                 get_cpu_var(bh_accounting).nr++;
3177                 recalc_bh_state();
3178                 put_cpu_var(bh_accounting);
3179         }
3180         return ret;
3181 }
3182 EXPORT_SYMBOL(alloc_buffer_head);
3183
3184 void free_buffer_head(struct buffer_head *bh)
3185 {
3186         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3187         kmem_cache_free(bh_cachep, bh);
3188         get_cpu_var(bh_accounting).nr--;
3189         recalc_bh_state();
3190         put_cpu_var(bh_accounting);
3191 }
3192 EXPORT_SYMBOL(free_buffer_head);
3193
3194 static void buffer_exit_cpu(int cpu)
3195 {
3196         int i;
3197         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3198
3199         for (i = 0; i < BH_LRU_SIZE; i++) {
3200                 brelse(b->bhs[i]);
3201                 b->bhs[i] = NULL;
3202         }
3203         get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3204         per_cpu(bh_accounting, cpu).nr = 0;
3205         put_cpu_var(bh_accounting);
3206 }
3207
3208 static int buffer_cpu_notify(struct notifier_block *self,
3209                               unsigned long action, void *hcpu)
3210 {
3211         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3212                 buffer_exit_cpu((unsigned long)hcpu);
3213         return NOTIFY_OK;
3214 }
3215
3216 /**
3217  * bh_uptodate_or_lock: Test whether the buffer is uptodate
3218  * @bh: struct buffer_head
3219  *
3220  * Return true if the buffer is up-to-date and false,
3221  * with the buffer locked, if not.
3222  */
3223 int bh_uptodate_or_lock(struct buffer_head *bh)
3224 {
3225         if (!buffer_uptodate(bh)) {
3226                 lock_buffer(bh);
3227                 if (!buffer_uptodate(bh))
3228                         return 0;
3229                 unlock_buffer(bh);
3230         }
3231         return 1;
3232 }
3233 EXPORT_SYMBOL(bh_uptodate_or_lock);
3234
3235 /**
3236  * bh_submit_read: Submit a locked buffer for reading
3237  * @bh: struct buffer_head
3238  *
3239  * Returns zero on success and -EIO on error.
3240  */
3241 int bh_submit_read(struct buffer_head *bh)
3242 {
3243         BUG_ON(!buffer_locked(bh));
3244
3245         if (buffer_uptodate(bh)) {
3246                 unlock_buffer(bh);
3247                 return 0;
3248         }
3249
3250         get_bh(bh);
3251         bh->b_end_io = end_buffer_read_sync;
3252         submit_bh(READ, bh);
3253         wait_on_buffer(bh);
3254         if (buffer_uptodate(bh))
3255                 return 0;
3256         return -EIO;
3257 }
3258 EXPORT_SYMBOL(bh_submit_read);
3259
3260 static void
3261 init_buffer_head(struct kmem_cache *cachep, void *data)
3262 {
3263         struct buffer_head *bh = data;
3264
3265         memset(bh, 0, sizeof(*bh));
3266         INIT_LIST_HEAD(&bh->b_assoc_buffers);
3267 }
3268
3269 void __init buffer_init(void)
3270 {
3271         int nrpages;
3272
3273         bh_cachep = kmem_cache_create("buffer_head",
3274                         sizeof(struct buffer_head), 0,
3275                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3276                                 SLAB_MEM_SPREAD),
3277                                 init_buffer_head);
3278
3279         /*
3280          * Limit the bh occupancy to 10% of ZONE_NORMAL
3281          */
3282         nrpages = (nr_free_buffer_pages() * 10) / 100;
3283         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3284         hotcpu_notifier(buffer_cpu_notify, 0);
3285 }
3286
3287 EXPORT_SYMBOL(__bforget);
3288 EXPORT_SYMBOL(__brelse);
3289 EXPORT_SYMBOL(__wait_on_buffer);
3290 EXPORT_SYMBOL(block_commit_write);
3291 EXPORT_SYMBOL(block_prepare_write);
3292 EXPORT_SYMBOL(block_page_mkwrite);
3293 EXPORT_SYMBOL(block_read_full_page);
3294 EXPORT_SYMBOL(block_sync_page);
3295 EXPORT_SYMBOL(block_truncate_page);
3296 EXPORT_SYMBOL(block_write_full_page);
3297 EXPORT_SYMBOL(cont_write_begin);
3298 EXPORT_SYMBOL(end_buffer_read_sync);
3299 EXPORT_SYMBOL(end_buffer_write_sync);
3300 EXPORT_SYMBOL(file_fsync);
3301 EXPORT_SYMBOL(fsync_bdev);
3302 EXPORT_SYMBOL(generic_block_bmap);
3303 EXPORT_SYMBOL(generic_commit_write);
3304 EXPORT_SYMBOL(generic_cont_expand_simple);
3305 EXPORT_SYMBOL(init_buffer);
3306 EXPORT_SYMBOL(invalidate_bdev);
3307 EXPORT_SYMBOL(ll_rw_block);
3308 EXPORT_SYMBOL(mark_buffer_dirty);
3309 EXPORT_SYMBOL(submit_bh);
3310 EXPORT_SYMBOL(sync_dirty_buffer);
3311 EXPORT_SYMBOL(unlock_buffer);