Merge branch 'master' of master.kernel.org:/pub/scm/linux/kernel/git/davem/net-2.6
[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         clear_bit_unlock(BH_Lock, &bh->b_state);
80         smp_mb__after_clear_bit();
81         wake_up_bit(&bh->b_state, BH_Lock);
82 }
83
84 /*
85  * Block until a buffer comes unlocked.  This doesn't stop it
86  * from becoming locked again - you have to lock it yourself
87  * if you want to preserve its state.
88  */
89 void __wait_on_buffer(struct buffer_head * bh)
90 {
91         wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
92 }
93
94 static void
95 __clear_page_buffers(struct page *page)
96 {
97         ClearPagePrivate(page);
98         set_page_private(page, 0);
99         page_cache_release(page);
100 }
101
102
103 static int quiet_error(struct buffer_head *bh)
104 {
105         if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
106                 return 0;
107         return 1;
108 }
109
110
111 static void buffer_io_error(struct buffer_head *bh)
112 {
113         char b[BDEVNAME_SIZE];
114         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
115                         bdevname(bh->b_bdev, b),
116                         (unsigned long long)bh->b_blocknr);
117 }
118
119 /*
120  * End-of-IO handler helper function which does not touch the bh after
121  * unlocking it.
122  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
123  * a race there is benign: unlock_buffer() only use the bh's address for
124  * hashing after unlocking the buffer, so it doesn't actually touch the bh
125  * itself.
126  */
127 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
128 {
129         if (uptodate) {
130                 set_buffer_uptodate(bh);
131         } else {
132                 /* This happens, due to failed READA attempts. */
133                 clear_buffer_uptodate(bh);
134         }
135         unlock_buffer(bh);
136 }
137
138 /*
139  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
140  * unlock the buffer. This is what ll_rw_block uses too.
141  */
142 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
143 {
144         __end_buffer_read_notouch(bh, uptodate);
145         put_bh(bh);
146 }
147
148 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
149 {
150         char b[BDEVNAME_SIZE];
151
152         if (uptodate) {
153                 set_buffer_uptodate(bh);
154         } else {
155                 if (!buffer_eopnotsupp(bh) && !quiet_error(bh)) {
156                         buffer_io_error(bh);
157                         printk(KERN_WARNING "lost page write due to "
158                                         "I/O error on %s\n",
159                                        bdevname(bh->b_bdev, b));
160                 }
161                 set_buffer_write_io_error(bh);
162                 clear_buffer_uptodate(bh);
163         }
164         unlock_buffer(bh);
165         put_bh(bh);
166 }
167
168 /*
169  * Write out and wait upon all the dirty data associated with a block
170  * device via its mapping.  Does not take the superblock lock.
171  */
172 int sync_blockdev(struct block_device *bdev)
173 {
174         int ret = 0;
175
176         if (bdev)
177                 ret = filemap_write_and_wait(bdev->bd_inode->i_mapping);
178         return ret;
179 }
180 EXPORT_SYMBOL(sync_blockdev);
181
182 /*
183  * Write out and wait upon all dirty data associated with this
184  * device.   Filesystem data as well as the underlying block
185  * device.  Takes the superblock lock.
186  */
187 int fsync_bdev(struct block_device *bdev)
188 {
189         struct super_block *sb = get_super(bdev);
190         if (sb) {
191                 int res = fsync_super(sb);
192                 drop_super(sb);
193                 return res;
194         }
195         return sync_blockdev(bdev);
196 }
197
198 /**
199  * freeze_bdev  --  lock a filesystem and force it into a consistent state
200  * @bdev:       blockdevice to lock
201  *
202  * This takes the block device bd_mount_sem to make sure no new mounts
203  * happen on bdev until thaw_bdev() is called.
204  * If a superblock is found on this device, we take the s_umount semaphore
205  * on it to make sure nobody unmounts until the snapshot creation is done.
206  * The reference counter (bd_fsfreeze_count) guarantees that only the last
207  * unfreeze process can unfreeze the frozen filesystem actually when multiple
208  * freeze requests arrive simultaneously. It counts up in freeze_bdev() and
209  * count down in thaw_bdev(). When it becomes 0, thaw_bdev() will unfreeze
210  * actually.
211  */
212 struct super_block *freeze_bdev(struct block_device *bdev)
213 {
214         struct super_block *sb;
215         int error = 0;
216
217         mutex_lock(&bdev->bd_fsfreeze_mutex);
218         if (bdev->bd_fsfreeze_count > 0) {
219                 bdev->bd_fsfreeze_count++;
220                 sb = get_super(bdev);
221                 mutex_unlock(&bdev->bd_fsfreeze_mutex);
222                 return sb;
223         }
224         bdev->bd_fsfreeze_count++;
225
226         down(&bdev->bd_mount_sem);
227         sb = get_super(bdev);
228         if (sb && !(sb->s_flags & MS_RDONLY)) {
229                 sb->s_frozen = SB_FREEZE_WRITE;
230                 smp_wmb();
231
232                 __fsync_super(sb);
233
234                 sb->s_frozen = SB_FREEZE_TRANS;
235                 smp_wmb();
236
237                 sync_blockdev(sb->s_bdev);
238
239                 if (sb->s_op->freeze_fs) {
240                         error = sb->s_op->freeze_fs(sb);
241                         if (error) {
242                                 printk(KERN_ERR
243                                         "VFS:Filesystem freeze failed\n");
244                                 sb->s_frozen = SB_UNFROZEN;
245                                 drop_super(sb);
246                                 up(&bdev->bd_mount_sem);
247                                 bdev->bd_fsfreeze_count--;
248                                 mutex_unlock(&bdev->bd_fsfreeze_mutex);
249                                 return ERR_PTR(error);
250                         }
251                 }
252         }
253
254         sync_blockdev(bdev);
255         mutex_unlock(&bdev->bd_fsfreeze_mutex);
256
257         return sb;      /* thaw_bdev releases s->s_umount and bd_mount_sem */
258 }
259 EXPORT_SYMBOL(freeze_bdev);
260
261 /**
262  * thaw_bdev  -- unlock filesystem
263  * @bdev:       blockdevice to unlock
264  * @sb:         associated superblock
265  *
266  * Unlocks the filesystem and marks it writeable again after freeze_bdev().
267  */
268 int thaw_bdev(struct block_device *bdev, struct super_block *sb)
269 {
270         int error = 0;
271
272         mutex_lock(&bdev->bd_fsfreeze_mutex);
273         if (!bdev->bd_fsfreeze_count) {
274                 mutex_unlock(&bdev->bd_fsfreeze_mutex);
275                 return -EINVAL;
276         }
277
278         bdev->bd_fsfreeze_count--;
279         if (bdev->bd_fsfreeze_count > 0) {
280                 if (sb)
281                         drop_super(sb);
282                 mutex_unlock(&bdev->bd_fsfreeze_mutex);
283                 return 0;
284         }
285
286         if (sb) {
287                 BUG_ON(sb->s_bdev != bdev);
288                 if (!(sb->s_flags & MS_RDONLY)) {
289                         if (sb->s_op->unfreeze_fs) {
290                                 error = sb->s_op->unfreeze_fs(sb);
291                                 if (error) {
292                                         printk(KERN_ERR
293                                                 "VFS:Filesystem thaw failed\n");
294                                         sb->s_frozen = SB_FREEZE_TRANS;
295                                         bdev->bd_fsfreeze_count++;
296                                         mutex_unlock(&bdev->bd_fsfreeze_mutex);
297                                         return error;
298                                 }
299                         }
300                         sb->s_frozen = SB_UNFROZEN;
301                         smp_wmb();
302                         wake_up(&sb->s_wait_unfrozen);
303                 }
304                 drop_super(sb);
305         }
306
307         up(&bdev->bd_mount_sem);
308         mutex_unlock(&bdev->bd_fsfreeze_mutex);
309         return 0;
310 }
311 EXPORT_SYMBOL(thaw_bdev);
312
313 /*
314  * Various filesystems appear to want __find_get_block to be non-blocking.
315  * But it's the page lock which protects the buffers.  To get around this,
316  * we get exclusion from try_to_free_buffers with the blockdev mapping's
317  * private_lock.
318  *
319  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
320  * may be quite high.  This code could TryLock the page, and if that
321  * succeeds, there is no need to take private_lock. (But if
322  * private_lock is contended then so is mapping->tree_lock).
323  */
324 static struct buffer_head *
325 __find_get_block_slow(struct block_device *bdev, sector_t block)
326 {
327         struct inode *bd_inode = bdev->bd_inode;
328         struct address_space *bd_mapping = bd_inode->i_mapping;
329         struct buffer_head *ret = NULL;
330         pgoff_t index;
331         struct buffer_head *bh;
332         struct buffer_head *head;
333         struct page *page;
334         int all_mapped = 1;
335
336         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
337         page = find_get_page(bd_mapping, index);
338         if (!page)
339                 goto out;
340
341         spin_lock(&bd_mapping->private_lock);
342         if (!page_has_buffers(page))
343                 goto out_unlock;
344         head = page_buffers(page);
345         bh = head;
346         do {
347                 if (bh->b_blocknr == block) {
348                         ret = bh;
349                         get_bh(bh);
350                         goto out_unlock;
351                 }
352                 if (!buffer_mapped(bh))
353                         all_mapped = 0;
354                 bh = bh->b_this_page;
355         } while (bh != head);
356
357         /* we might be here because some of the buffers on this page are
358          * not mapped.  This is due to various races between
359          * file io on the block device and getblk.  It gets dealt with
360          * elsewhere, don't buffer_error if we had some unmapped buffers
361          */
362         if (all_mapped) {
363                 printk("__find_get_block_slow() failed. "
364                         "block=%llu, b_blocknr=%llu\n",
365                         (unsigned long long)block,
366                         (unsigned long long)bh->b_blocknr);
367                 printk("b_state=0x%08lx, b_size=%zu\n",
368                         bh->b_state, bh->b_size);
369                 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
370         }
371 out_unlock:
372         spin_unlock(&bd_mapping->private_lock);
373         page_cache_release(page);
374 out:
375         return ret;
376 }
377
378 /* If invalidate_buffers() will trash dirty buffers, it means some kind
379    of fs corruption is going on. Trashing dirty data always imply losing
380    information that was supposed to be just stored on the physical layer
381    by the user.
382
383    Thus invalidate_buffers in general usage is not allwowed to trash
384    dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
385    be preserved.  These buffers are simply skipped.
386   
387    We also skip buffers which are still in use.  For example this can
388    happen if a userspace program is reading the block device.
389
390    NOTE: In the case where the user removed a removable-media-disk even if
391    there's still dirty data not synced on disk (due a bug in the device driver
392    or due an error of the user), by not destroying the dirty buffers we could
393    generate corruption also on the next media inserted, thus a parameter is
394    necessary to handle this case in the most safe way possible (trying
395    to not corrupt also the new disk inserted with the data belonging to
396    the old now corrupted disk). Also for the ramdisk the natural thing
397    to do in order to release the ramdisk memory is to destroy dirty buffers.
398
399    These are two special cases. Normal usage imply the device driver
400    to issue a sync on the device (without waiting I/O completion) and
401    then an invalidate_buffers call that doesn't trash dirty buffers.
402
403    For handling cache coherency with the blkdev pagecache the 'update' case
404    is been introduced. It is needed to re-read from disk any pinned
405    buffer. NOTE: re-reading from disk is destructive so we can do it only
406    when we assume nobody is changing the buffercache under our I/O and when
407    we think the disk contains more recent information than the buffercache.
408    The update == 1 pass marks the buffers we need to update, the update == 2
409    pass does the actual I/O. */
410 void invalidate_bdev(struct block_device *bdev)
411 {
412         struct address_space *mapping = bdev->bd_inode->i_mapping;
413
414         if (mapping->nrpages == 0)
415                 return;
416
417         invalidate_bh_lrus();
418         invalidate_mapping_pages(mapping, 0, -1);
419 }
420
421 /*
422  * Kick pdflush then try to free up some ZONE_NORMAL memory.
423  */
424 static void free_more_memory(void)
425 {
426         struct zone *zone;
427         int nid;
428
429         wakeup_pdflush(1024);
430         yield();
431
432         for_each_online_node(nid) {
433                 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
434                                                 gfp_zone(GFP_NOFS), NULL,
435                                                 &zone);
436                 if (zone)
437                         try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
438                                                 GFP_NOFS);
439         }
440 }
441
442 /*
443  * I/O completion handler for block_read_full_page() - pages
444  * which come unlocked at the end of I/O.
445  */
446 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
447 {
448         unsigned long flags;
449         struct buffer_head *first;
450         struct buffer_head *tmp;
451         struct page *page;
452         int page_uptodate = 1;
453
454         BUG_ON(!buffer_async_read(bh));
455
456         page = bh->b_page;
457         if (uptodate) {
458                 set_buffer_uptodate(bh);
459         } else {
460                 clear_buffer_uptodate(bh);
461                 if (!quiet_error(bh))
462                         buffer_io_error(bh);
463                 SetPageError(page);
464         }
465
466         /*
467          * Be _very_ careful from here on. Bad things can happen if
468          * two buffer heads end IO at almost the same time and both
469          * decide that the page is now completely done.
470          */
471         first = page_buffers(page);
472         local_irq_save(flags);
473         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
474         clear_buffer_async_read(bh);
475         unlock_buffer(bh);
476         tmp = bh;
477         do {
478                 if (!buffer_uptodate(tmp))
479                         page_uptodate = 0;
480                 if (buffer_async_read(tmp)) {
481                         BUG_ON(!buffer_locked(tmp));
482                         goto still_busy;
483                 }
484                 tmp = tmp->b_this_page;
485         } while (tmp != bh);
486         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
487         local_irq_restore(flags);
488
489         /*
490          * If none of the buffers had errors and they are all
491          * uptodate then we can set the page uptodate.
492          */
493         if (page_uptodate && !PageError(page))
494                 SetPageUptodate(page);
495         unlock_page(page);
496         return;
497
498 still_busy:
499         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
500         local_irq_restore(flags);
501         return;
502 }
503
504 /*
505  * Completion handler for block_write_full_page() - pages which are unlocked
506  * during I/O, and which have PageWriteback cleared upon I/O completion.
507  */
508 static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
509 {
510         char b[BDEVNAME_SIZE];
511         unsigned long flags;
512         struct buffer_head *first;
513         struct buffer_head *tmp;
514         struct page *page;
515
516         BUG_ON(!buffer_async_write(bh));
517
518         page = bh->b_page;
519         if (uptodate) {
520                 set_buffer_uptodate(bh);
521         } else {
522                 if (!quiet_error(bh)) {
523                         buffer_io_error(bh);
524                         printk(KERN_WARNING "lost page write due to "
525                                         "I/O error on %s\n",
526                                bdevname(bh->b_bdev, b));
527                 }
528                 set_bit(AS_EIO, &page->mapping->flags);
529                 set_buffer_write_io_error(bh);
530                 clear_buffer_uptodate(bh);
531                 SetPageError(page);
532         }
533
534         first = page_buffers(page);
535         local_irq_save(flags);
536         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
537
538         clear_buffer_async_write(bh);
539         unlock_buffer(bh);
540         tmp = bh->b_this_page;
541         while (tmp != bh) {
542                 if (buffer_async_write(tmp)) {
543                         BUG_ON(!buffer_locked(tmp));
544                         goto still_busy;
545                 }
546                 tmp = tmp->b_this_page;
547         }
548         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
549         local_irq_restore(flags);
550         end_page_writeback(page);
551         return;
552
553 still_busy:
554         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
555         local_irq_restore(flags);
556         return;
557 }
558
559 /*
560  * If a page's buffers are under async readin (end_buffer_async_read
561  * completion) then there is a possibility that another thread of
562  * control could lock one of the buffers after it has completed
563  * but while some of the other buffers have not completed.  This
564  * locked buffer would confuse end_buffer_async_read() into not unlocking
565  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
566  * that this buffer is not under async I/O.
567  *
568  * The page comes unlocked when it has no locked buffer_async buffers
569  * left.
570  *
571  * PageLocked prevents anyone starting new async I/O reads any of
572  * the buffers.
573  *
574  * PageWriteback is used to prevent simultaneous writeout of the same
575  * page.
576  *
577  * PageLocked prevents anyone from starting writeback of a page which is
578  * under read I/O (PageWriteback is only ever set against a locked page).
579  */
580 static void mark_buffer_async_read(struct buffer_head *bh)
581 {
582         bh->b_end_io = end_buffer_async_read;
583         set_buffer_async_read(bh);
584 }
585
586 void mark_buffer_async_write(struct buffer_head *bh)
587 {
588         bh->b_end_io = end_buffer_async_write;
589         set_buffer_async_write(bh);
590 }
591 EXPORT_SYMBOL(mark_buffer_async_write);
592
593
594 /*
595  * fs/buffer.c contains helper functions for buffer-backed address space's
596  * fsync functions.  A common requirement for buffer-based filesystems is
597  * that certain data from the backing blockdev needs to be written out for
598  * a successful fsync().  For example, ext2 indirect blocks need to be
599  * written back and waited upon before fsync() returns.
600  *
601  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
602  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
603  * management of a list of dependent buffers at ->i_mapping->private_list.
604  *
605  * Locking is a little subtle: try_to_free_buffers() will remove buffers
606  * from their controlling inode's queue when they are being freed.  But
607  * try_to_free_buffers() will be operating against the *blockdev* mapping
608  * at the time, not against the S_ISREG file which depends on those buffers.
609  * So the locking for private_list is via the private_lock in the address_space
610  * which backs the buffers.  Which is different from the address_space 
611  * against which the buffers are listed.  So for a particular address_space,
612  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
613  * mapping->private_list will always be protected by the backing blockdev's
614  * ->private_lock.
615  *
616  * Which introduces a requirement: all buffers on an address_space's
617  * ->private_list must be from the same address_space: the blockdev's.
618  *
619  * address_spaces which do not place buffers at ->private_list via these
620  * utility functions are free to use private_lock and private_list for
621  * whatever they want.  The only requirement is that list_empty(private_list)
622  * be true at clear_inode() time.
623  *
624  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
625  * filesystems should do that.  invalidate_inode_buffers() should just go
626  * BUG_ON(!list_empty).
627  *
628  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
629  * take an address_space, not an inode.  And it should be called
630  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
631  * queued up.
632  *
633  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
634  * list if it is already on a list.  Because if the buffer is on a list,
635  * it *must* already be on the right one.  If not, the filesystem is being
636  * silly.  This will save a ton of locking.  But first we have to ensure
637  * that buffers are taken *off* the old inode's list when they are freed
638  * (presumably in truncate).  That requires careful auditing of all
639  * filesystems (do it inside bforget()).  It could also be done by bringing
640  * b_inode back.
641  */
642
643 /*
644  * The buffer's backing address_space's private_lock must be held
645  */
646 static void __remove_assoc_queue(struct buffer_head *bh)
647 {
648         list_del_init(&bh->b_assoc_buffers);
649         WARN_ON(!bh->b_assoc_map);
650         if (buffer_write_io_error(bh))
651                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
652         bh->b_assoc_map = NULL;
653 }
654
655 int inode_has_buffers(struct inode *inode)
656 {
657         return !list_empty(&inode->i_data.private_list);
658 }
659
660 /*
661  * osync is designed to support O_SYNC io.  It waits synchronously for
662  * all already-submitted IO to complete, but does not queue any new
663  * writes to the disk.
664  *
665  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
666  * you dirty the buffers, and then use osync_inode_buffers to wait for
667  * completion.  Any other dirty buffers which are not yet queued for
668  * write will not be flushed to disk by the osync.
669  */
670 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
671 {
672         struct buffer_head *bh;
673         struct list_head *p;
674         int err = 0;
675
676         spin_lock(lock);
677 repeat:
678         list_for_each_prev(p, list) {
679                 bh = BH_ENTRY(p);
680                 if (buffer_locked(bh)) {
681                         get_bh(bh);
682                         spin_unlock(lock);
683                         wait_on_buffer(bh);
684                         if (!buffer_uptodate(bh))
685                                 err = -EIO;
686                         brelse(bh);
687                         spin_lock(lock);
688                         goto repeat;
689                 }
690         }
691         spin_unlock(lock);
692         return err;
693 }
694
695 /**
696  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
697  * @mapping: the mapping which wants those buffers written
698  *
699  * Starts I/O against the buffers at mapping->private_list, and waits upon
700  * that I/O.
701  *
702  * Basically, this is a convenience function for fsync().
703  * @mapping is a file or directory which needs those buffers to be written for
704  * a successful fsync().
705  */
706 int sync_mapping_buffers(struct address_space *mapping)
707 {
708         struct address_space *buffer_mapping = mapping->assoc_mapping;
709
710         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
711                 return 0;
712
713         return fsync_buffers_list(&buffer_mapping->private_lock,
714                                         &mapping->private_list);
715 }
716 EXPORT_SYMBOL(sync_mapping_buffers);
717
718 /*
719  * Called when we've recently written block `bblock', and it is known that
720  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
721  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
722  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
723  */
724 void write_boundary_block(struct block_device *bdev,
725                         sector_t bblock, unsigned blocksize)
726 {
727         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
728         if (bh) {
729                 if (buffer_dirty(bh))
730                         ll_rw_block(WRITE, 1, &bh);
731                 put_bh(bh);
732         }
733 }
734
735 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
736 {
737         struct address_space *mapping = inode->i_mapping;
738         struct address_space *buffer_mapping = bh->b_page->mapping;
739
740         mark_buffer_dirty(bh);
741         if (!mapping->assoc_mapping) {
742                 mapping->assoc_mapping = buffer_mapping;
743         } else {
744                 BUG_ON(mapping->assoc_mapping != buffer_mapping);
745         }
746         if (!bh->b_assoc_map) {
747                 spin_lock(&buffer_mapping->private_lock);
748                 list_move_tail(&bh->b_assoc_buffers,
749                                 &mapping->private_list);
750                 bh->b_assoc_map = mapping;
751                 spin_unlock(&buffer_mapping->private_lock);
752         }
753 }
754 EXPORT_SYMBOL(mark_buffer_dirty_inode);
755
756 /*
757  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
758  * dirty.
759  *
760  * If warn is true, then emit a warning if the page is not uptodate and has
761  * not been truncated.
762  */
763 static int __set_page_dirty(struct page *page,
764                 struct address_space *mapping, int warn)
765 {
766         if (unlikely(!mapping))
767                 return !TestSetPageDirty(page);
768
769         if (TestSetPageDirty(page))
770                 return 0;
771
772         spin_lock_irq(&mapping->tree_lock);
773         if (page->mapping) {    /* Race with truncate? */
774                 WARN_ON_ONCE(warn && !PageUptodate(page));
775
776                 if (mapping_cap_account_dirty(mapping)) {
777                         __inc_zone_page_state(page, NR_FILE_DIRTY);
778                         __inc_bdi_stat(mapping->backing_dev_info,
779                                         BDI_RECLAIMABLE);
780                         task_dirty_inc(current);
781                         task_io_account_write(PAGE_CACHE_SIZE);
782                 }
783                 radix_tree_tag_set(&mapping->page_tree,
784                                 page_index(page), PAGECACHE_TAG_DIRTY);
785         }
786         spin_unlock_irq(&mapping->tree_lock);
787         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
788
789         return 1;
790 }
791
792 /*
793  * Add a page to the dirty page list.
794  *
795  * It is a sad fact of life that this function is called from several places
796  * deeply under spinlocking.  It may not sleep.
797  *
798  * If the page has buffers, the uptodate buffers are set dirty, to preserve
799  * dirty-state coherency between the page and the buffers.  It the page does
800  * not have buffers then when they are later attached they will all be set
801  * dirty.
802  *
803  * The buffers are dirtied before the page is dirtied.  There's a small race
804  * window in which a writepage caller may see the page cleanness but not the
805  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
806  * before the buffers, a concurrent writepage caller could clear the page dirty
807  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
808  * page on the dirty page list.
809  *
810  * We use private_lock to lock against try_to_free_buffers while using the
811  * page's buffer list.  Also use this to protect against clean buffers being
812  * added to the page after it was set dirty.
813  *
814  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
815  * address_space though.
816  */
817 int __set_page_dirty_buffers(struct page *page)
818 {
819         struct address_space *mapping = page_mapping(page);
820
821         if (unlikely(!mapping))
822                 return !TestSetPageDirty(page);
823
824         spin_lock(&mapping->private_lock);
825         if (page_has_buffers(page)) {
826                 struct buffer_head *head = page_buffers(page);
827                 struct buffer_head *bh = head;
828
829                 do {
830                         set_buffer_dirty(bh);
831                         bh = bh->b_this_page;
832                 } while (bh != head);
833         }
834         spin_unlock(&mapping->private_lock);
835
836         return __set_page_dirty(page, mapping, 1);
837 }
838 EXPORT_SYMBOL(__set_page_dirty_buffers);
839
840 /*
841  * Write out and wait upon a list of buffers.
842  *
843  * We have conflicting pressures: we want to make sure that all
844  * initially dirty buffers get waited on, but that any subsequently
845  * dirtied buffers don't.  After all, we don't want fsync to last
846  * forever if somebody is actively writing to the file.
847  *
848  * Do this in two main stages: first we copy dirty buffers to a
849  * temporary inode list, queueing the writes as we go.  Then we clean
850  * up, waiting for those writes to complete.
851  * 
852  * During this second stage, any subsequent updates to the file may end
853  * up refiling the buffer on the original inode's dirty list again, so
854  * there is a chance we will end up with a buffer queued for write but
855  * not yet completed on that list.  So, as a final cleanup we go through
856  * the osync code to catch these locked, dirty buffers without requeuing
857  * any newly dirty buffers for write.
858  */
859 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
860 {
861         struct buffer_head *bh;
862         struct list_head tmp;
863         struct address_space *mapping;
864         int err = 0, err2;
865
866         INIT_LIST_HEAD(&tmp);
867
868         spin_lock(lock);
869         while (!list_empty(list)) {
870                 bh = BH_ENTRY(list->next);
871                 mapping = bh->b_assoc_map;
872                 __remove_assoc_queue(bh);
873                 /* Avoid race with mark_buffer_dirty_inode() which does
874                  * a lockless check and we rely on seeing the dirty bit */
875                 smp_mb();
876                 if (buffer_dirty(bh) || buffer_locked(bh)) {
877                         list_add(&bh->b_assoc_buffers, &tmp);
878                         bh->b_assoc_map = mapping;
879                         if (buffer_dirty(bh)) {
880                                 get_bh(bh);
881                                 spin_unlock(lock);
882                                 /*
883                                  * Ensure any pending I/O completes so that
884                                  * ll_rw_block() actually writes the current
885                                  * contents - it is a noop if I/O is still in
886                                  * flight on potentially older contents.
887                                  */
888                                 ll_rw_block(SWRITE_SYNC, 1, &bh);
889                                 brelse(bh);
890                                 spin_lock(lock);
891                         }
892                 }
893         }
894
895         while (!list_empty(&tmp)) {
896                 bh = BH_ENTRY(tmp.prev);
897                 get_bh(bh);
898                 mapping = bh->b_assoc_map;
899                 __remove_assoc_queue(bh);
900                 /* Avoid race with mark_buffer_dirty_inode() which does
901                  * a lockless check and we rely on seeing the dirty bit */
902                 smp_mb();
903                 if (buffer_dirty(bh)) {
904                         list_add(&bh->b_assoc_buffers,
905                                  &mapping->private_list);
906                         bh->b_assoc_map = mapping;
907                 }
908                 spin_unlock(lock);
909                 wait_on_buffer(bh);
910                 if (!buffer_uptodate(bh))
911                         err = -EIO;
912                 brelse(bh);
913                 spin_lock(lock);
914         }
915         
916         spin_unlock(lock);
917         err2 = osync_buffers_list(lock, list);
918         if (err)
919                 return err;
920         else
921                 return err2;
922 }
923
924 /*
925  * Invalidate any and all dirty buffers on a given inode.  We are
926  * probably unmounting the fs, but that doesn't mean we have already
927  * done a sync().  Just drop the buffers from the inode list.
928  *
929  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
930  * assumes that all the buffers are against the blockdev.  Not true
931  * for reiserfs.
932  */
933 void invalidate_inode_buffers(struct inode *inode)
934 {
935         if (inode_has_buffers(inode)) {
936                 struct address_space *mapping = &inode->i_data;
937                 struct list_head *list = &mapping->private_list;
938                 struct address_space *buffer_mapping = mapping->assoc_mapping;
939
940                 spin_lock(&buffer_mapping->private_lock);
941                 while (!list_empty(list))
942                         __remove_assoc_queue(BH_ENTRY(list->next));
943                 spin_unlock(&buffer_mapping->private_lock);
944         }
945 }
946 EXPORT_SYMBOL(invalidate_inode_buffers);
947
948 /*
949  * Remove any clean buffers from the inode's buffer list.  This is called
950  * when we're trying to free the inode itself.  Those buffers can pin it.
951  *
952  * Returns true if all buffers were removed.
953  */
954 int remove_inode_buffers(struct inode *inode)
955 {
956         int ret = 1;
957
958         if (inode_has_buffers(inode)) {
959                 struct address_space *mapping = &inode->i_data;
960                 struct list_head *list = &mapping->private_list;
961                 struct address_space *buffer_mapping = mapping->assoc_mapping;
962
963                 spin_lock(&buffer_mapping->private_lock);
964                 while (!list_empty(list)) {
965                         struct buffer_head *bh = BH_ENTRY(list->next);
966                         if (buffer_dirty(bh)) {
967                                 ret = 0;
968                                 break;
969                         }
970                         __remove_assoc_queue(bh);
971                 }
972                 spin_unlock(&buffer_mapping->private_lock);
973         }
974         return ret;
975 }
976
977 /*
978  * Create the appropriate buffers when given a page for data area and
979  * the size of each buffer.. Use the bh->b_this_page linked list to
980  * follow the buffers created.  Return NULL if unable to create more
981  * buffers.
982  *
983  * The retry flag is used to differentiate async IO (paging, swapping)
984  * which may not fail from ordinary buffer allocations.
985  */
986 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
987                 int retry)
988 {
989         struct buffer_head *bh, *head;
990         long offset;
991
992 try_again:
993         head = NULL;
994         offset = PAGE_SIZE;
995         while ((offset -= size) >= 0) {
996                 bh = alloc_buffer_head(GFP_NOFS);
997                 if (!bh)
998                         goto no_grow;
999
1000                 bh->b_bdev = NULL;
1001                 bh->b_this_page = head;
1002                 bh->b_blocknr = -1;
1003                 head = bh;
1004
1005                 bh->b_state = 0;
1006                 atomic_set(&bh->b_count, 0);
1007                 bh->b_private = NULL;
1008                 bh->b_size = size;
1009
1010                 /* Link the buffer to its page */
1011                 set_bh_page(bh, page, offset);
1012
1013                 init_buffer(bh, NULL, NULL);
1014         }
1015         return head;
1016 /*
1017  * In case anything failed, we just free everything we got.
1018  */
1019 no_grow:
1020         if (head) {
1021                 do {
1022                         bh = head;
1023                         head = head->b_this_page;
1024                         free_buffer_head(bh);
1025                 } while (head);
1026         }
1027
1028         /*
1029          * Return failure for non-async IO requests.  Async IO requests
1030          * are not allowed to fail, so we have to wait until buffer heads
1031          * become available.  But we don't want tasks sleeping with 
1032          * partially complete buffers, so all were released above.
1033          */
1034         if (!retry)
1035                 return NULL;
1036
1037         /* We're _really_ low on memory. Now we just
1038          * wait for old buffer heads to become free due to
1039          * finishing IO.  Since this is an async request and
1040          * the reserve list is empty, we're sure there are 
1041          * async buffer heads in use.
1042          */
1043         free_more_memory();
1044         goto try_again;
1045 }
1046 EXPORT_SYMBOL_GPL(alloc_page_buffers);
1047
1048 static inline void
1049 link_dev_buffers(struct page *page, struct buffer_head *head)
1050 {
1051         struct buffer_head *bh, *tail;
1052
1053         bh = head;
1054         do {
1055                 tail = bh;
1056                 bh = bh->b_this_page;
1057         } while (bh);
1058         tail->b_this_page = head;
1059         attach_page_buffers(page, head);
1060 }
1061
1062 /*
1063  * Initialise the state of a blockdev page's buffers.
1064  */ 
1065 static void
1066 init_page_buffers(struct page *page, struct block_device *bdev,
1067                         sector_t block, int size)
1068 {
1069         struct buffer_head *head = page_buffers(page);
1070         struct buffer_head *bh = head;
1071         int uptodate = PageUptodate(page);
1072
1073         do {
1074                 if (!buffer_mapped(bh)) {
1075                         init_buffer(bh, NULL, NULL);
1076                         bh->b_bdev = bdev;
1077                         bh->b_blocknr = block;
1078                         if (uptodate)
1079                                 set_buffer_uptodate(bh);
1080                         set_buffer_mapped(bh);
1081                 }
1082                 block++;
1083                 bh = bh->b_this_page;
1084         } while (bh != head);
1085 }
1086
1087 /*
1088  * Create the page-cache page that contains the requested block.
1089  *
1090  * This is user purely for blockdev mappings.
1091  */
1092 static struct page *
1093 grow_dev_page(struct block_device *bdev, sector_t block,
1094                 pgoff_t index, int size)
1095 {
1096         struct inode *inode = bdev->bd_inode;
1097         struct page *page;
1098         struct buffer_head *bh;
1099
1100         page = find_or_create_page(inode->i_mapping, index,
1101                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
1102         if (!page)
1103                 return NULL;
1104
1105         BUG_ON(!PageLocked(page));
1106
1107         if (page_has_buffers(page)) {
1108                 bh = page_buffers(page);
1109                 if (bh->b_size == size) {
1110                         init_page_buffers(page, bdev, block, size);
1111                         return page;
1112                 }
1113                 if (!try_to_free_buffers(page))
1114                         goto failed;
1115         }
1116
1117         /*
1118          * Allocate some buffers for this page
1119          */
1120         bh = alloc_page_buffers(page, size, 0);
1121         if (!bh)
1122                 goto failed;
1123
1124         /*
1125          * Link the page to the buffers and initialise them.  Take the
1126          * lock to be atomic wrt __find_get_block(), which does not
1127          * run under the page lock.
1128          */
1129         spin_lock(&inode->i_mapping->private_lock);
1130         link_dev_buffers(page, bh);
1131         init_page_buffers(page, bdev, block, size);
1132         spin_unlock(&inode->i_mapping->private_lock);
1133         return page;
1134
1135 failed:
1136         BUG();
1137         unlock_page(page);
1138         page_cache_release(page);
1139         return NULL;
1140 }
1141
1142 /*
1143  * Create buffers for the specified block device block's page.  If
1144  * that page was dirty, the buffers are set dirty also.
1145  */
1146 static int
1147 grow_buffers(struct block_device *bdev, sector_t block, int size)
1148 {
1149         struct page *page;
1150         pgoff_t index;
1151         int sizebits;
1152
1153         sizebits = -1;
1154         do {
1155                 sizebits++;
1156         } while ((size << sizebits) < PAGE_SIZE);
1157
1158         index = block >> sizebits;
1159
1160         /*
1161          * Check for a block which wants to lie outside our maximum possible
1162          * pagecache index.  (this comparison is done using sector_t types).
1163          */
1164         if (unlikely(index != block >> sizebits)) {
1165                 char b[BDEVNAME_SIZE];
1166
1167                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1168                         "device %s\n",
1169                         __func__, (unsigned long long)block,
1170                         bdevname(bdev, b));
1171                 return -EIO;
1172         }
1173         block = index << sizebits;
1174         /* Create a page with the proper size buffers.. */
1175         page = grow_dev_page(bdev, block, index, size);
1176         if (!page)
1177                 return 0;
1178         unlock_page(page);
1179         page_cache_release(page);
1180         return 1;
1181 }
1182
1183 static struct buffer_head *
1184 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1185 {
1186         /* Size must be multiple of hard sectorsize */
1187         if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1188                         (size < 512 || size > PAGE_SIZE))) {
1189                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1190                                         size);
1191                 printk(KERN_ERR "hardsect size: %d\n",
1192                                         bdev_hardsect_size(bdev));
1193
1194                 dump_stack();
1195                 return NULL;
1196         }
1197
1198         for (;;) {
1199                 struct buffer_head * bh;
1200                 int ret;
1201
1202                 bh = __find_get_block(bdev, block, size);
1203                 if (bh)
1204                         return bh;
1205
1206                 ret = grow_buffers(bdev, block, size);
1207                 if (ret < 0)
1208                         return NULL;
1209                 if (ret == 0)
1210                         free_more_memory();
1211         }
1212 }
1213
1214 /*
1215  * The relationship between dirty buffers and dirty pages:
1216  *
1217  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1218  * the page is tagged dirty in its radix tree.
1219  *
1220  * At all times, the dirtiness of the buffers represents the dirtiness of
1221  * subsections of the page.  If the page has buffers, the page dirty bit is
1222  * merely a hint about the true dirty state.
1223  *
1224  * When a page is set dirty in its entirety, all its buffers are marked dirty
1225  * (if the page has buffers).
1226  *
1227  * When a buffer is marked dirty, its page is dirtied, but the page's other
1228  * buffers are not.
1229  *
1230  * Also.  When blockdev buffers are explicitly read with bread(), they
1231  * individually become uptodate.  But their backing page remains not
1232  * uptodate - even if all of its buffers are uptodate.  A subsequent
1233  * block_read_full_page() against that page will discover all the uptodate
1234  * buffers, will set the page uptodate and will perform no I/O.
1235  */
1236
1237 /**
1238  * mark_buffer_dirty - mark a buffer_head as needing writeout
1239  * @bh: the buffer_head to mark dirty
1240  *
1241  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1242  * backing page dirty, then tag the page as dirty in its address_space's radix
1243  * tree and then attach the address_space's inode to its superblock's dirty
1244  * inode list.
1245  *
1246  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1247  * mapping->tree_lock and the global inode_lock.
1248  */
1249 void mark_buffer_dirty(struct buffer_head *bh)
1250 {
1251         WARN_ON_ONCE(!buffer_uptodate(bh));
1252
1253         /*
1254          * Very *carefully* optimize the it-is-already-dirty case.
1255          *
1256          * Don't let the final "is it dirty" escape to before we
1257          * perhaps modified the buffer.
1258          */
1259         if (buffer_dirty(bh)) {
1260                 smp_mb();
1261                 if (buffer_dirty(bh))
1262                         return;
1263         }
1264
1265         if (!test_set_buffer_dirty(bh))
1266                 __set_page_dirty(bh->b_page, page_mapping(bh->b_page), 0);
1267 }
1268
1269 /*
1270  * Decrement a buffer_head's reference count.  If all buffers against a page
1271  * have zero reference count, are clean and unlocked, and if the page is clean
1272  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1273  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1274  * a page but it ends up not being freed, and buffers may later be reattached).
1275  */
1276 void __brelse(struct buffer_head * buf)
1277 {
1278         if (atomic_read(&buf->b_count)) {
1279                 put_bh(buf);
1280                 return;
1281         }
1282         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1283 }
1284
1285 /*
1286  * bforget() is like brelse(), except it discards any
1287  * potentially dirty data.
1288  */
1289 void __bforget(struct buffer_head *bh)
1290 {
1291         clear_buffer_dirty(bh);
1292         if (bh->b_assoc_map) {
1293                 struct address_space *buffer_mapping = bh->b_page->mapping;
1294
1295                 spin_lock(&buffer_mapping->private_lock);
1296                 list_del_init(&bh->b_assoc_buffers);
1297                 bh->b_assoc_map = NULL;
1298                 spin_unlock(&buffer_mapping->private_lock);
1299         }
1300         __brelse(bh);
1301 }
1302
1303 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1304 {
1305         lock_buffer(bh);
1306         if (buffer_uptodate(bh)) {
1307                 unlock_buffer(bh);
1308                 return bh;
1309         } else {
1310                 get_bh(bh);
1311                 bh->b_end_io = end_buffer_read_sync;
1312                 submit_bh(READ, bh);
1313                 wait_on_buffer(bh);
1314                 if (buffer_uptodate(bh))
1315                         return bh;
1316         }
1317         brelse(bh);
1318         return NULL;
1319 }
1320
1321 /*
1322  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1323  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1324  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1325  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1326  * CPU's LRUs at the same time.
1327  *
1328  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1329  * sb_find_get_block().
1330  *
1331  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1332  * a local interrupt disable for that.
1333  */
1334
1335 #define BH_LRU_SIZE     8
1336
1337 struct bh_lru {
1338         struct buffer_head *bhs[BH_LRU_SIZE];
1339 };
1340
1341 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1342
1343 #ifdef CONFIG_SMP
1344 #define bh_lru_lock()   local_irq_disable()
1345 #define bh_lru_unlock() local_irq_enable()
1346 #else
1347 #define bh_lru_lock()   preempt_disable()
1348 #define bh_lru_unlock() preempt_enable()
1349 #endif
1350
1351 static inline void check_irqs_on(void)
1352 {
1353 #ifdef irqs_disabled
1354         BUG_ON(irqs_disabled());
1355 #endif
1356 }
1357
1358 /*
1359  * The LRU management algorithm is dopey-but-simple.  Sorry.
1360  */
1361 static void bh_lru_install(struct buffer_head *bh)
1362 {
1363         struct buffer_head *evictee = NULL;
1364         struct bh_lru *lru;
1365
1366         check_irqs_on();
1367         bh_lru_lock();
1368         lru = &__get_cpu_var(bh_lrus);
1369         if (lru->bhs[0] != bh) {
1370                 struct buffer_head *bhs[BH_LRU_SIZE];
1371                 int in;
1372                 int out = 0;
1373
1374                 get_bh(bh);
1375                 bhs[out++] = bh;
1376                 for (in = 0; in < BH_LRU_SIZE; in++) {
1377                         struct buffer_head *bh2 = lru->bhs[in];
1378
1379                         if (bh2 == bh) {
1380                                 __brelse(bh2);
1381                         } else {
1382                                 if (out >= BH_LRU_SIZE) {
1383                                         BUG_ON(evictee != NULL);
1384                                         evictee = bh2;
1385                                 } else {
1386                                         bhs[out++] = bh2;
1387                                 }
1388                         }
1389                 }
1390                 while (out < BH_LRU_SIZE)
1391                         bhs[out++] = NULL;
1392                 memcpy(lru->bhs, bhs, sizeof(bhs));
1393         }
1394         bh_lru_unlock();
1395
1396         if (evictee)
1397                 __brelse(evictee);
1398 }
1399
1400 /*
1401  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1402  */
1403 static struct buffer_head *
1404 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1405 {
1406         struct buffer_head *ret = NULL;
1407         struct bh_lru *lru;
1408         unsigned int i;
1409
1410         check_irqs_on();
1411         bh_lru_lock();
1412         lru = &__get_cpu_var(bh_lrus);
1413         for (i = 0; i < BH_LRU_SIZE; i++) {
1414                 struct buffer_head *bh = lru->bhs[i];
1415
1416                 if (bh && bh->b_bdev == bdev &&
1417                                 bh->b_blocknr == block && bh->b_size == size) {
1418                         if (i) {
1419                                 while (i) {
1420                                         lru->bhs[i] = lru->bhs[i - 1];
1421                                         i--;
1422                                 }
1423                                 lru->bhs[0] = bh;
1424                         }
1425                         get_bh(bh);
1426                         ret = bh;
1427                         break;
1428                 }
1429         }
1430         bh_lru_unlock();
1431         return ret;
1432 }
1433
1434 /*
1435  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1436  * it in the LRU and mark it as accessed.  If it is not present then return
1437  * NULL
1438  */
1439 struct buffer_head *
1440 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1441 {
1442         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1443
1444         if (bh == NULL) {
1445                 bh = __find_get_block_slow(bdev, block);
1446                 if (bh)
1447                         bh_lru_install(bh);
1448         }
1449         if (bh)
1450                 touch_buffer(bh);
1451         return bh;
1452 }
1453 EXPORT_SYMBOL(__find_get_block);
1454
1455 /*
1456  * __getblk will locate (and, if necessary, create) the buffer_head
1457  * which corresponds to the passed block_device, block and size. The
1458  * returned buffer has its reference count incremented.
1459  *
1460  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1461  * illegal block number, __getblk() will happily return a buffer_head
1462  * which represents the non-existent block.  Very weird.
1463  *
1464  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1465  * attempt is failing.  FIXME, perhaps?
1466  */
1467 struct buffer_head *
1468 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1469 {
1470         struct buffer_head *bh = __find_get_block(bdev, block, size);
1471
1472         might_sleep();
1473         if (bh == NULL)
1474                 bh = __getblk_slow(bdev, block, size);
1475         return bh;
1476 }
1477 EXPORT_SYMBOL(__getblk);
1478
1479 /*
1480  * Do async read-ahead on a buffer..
1481  */
1482 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1483 {
1484         struct buffer_head *bh = __getblk(bdev, block, size);
1485         if (likely(bh)) {
1486                 ll_rw_block(READA, 1, &bh);
1487                 brelse(bh);
1488         }
1489 }
1490 EXPORT_SYMBOL(__breadahead);
1491
1492 /**
1493  *  __bread() - reads a specified block and returns the bh
1494  *  @bdev: the block_device to read from
1495  *  @block: number of block
1496  *  @size: size (in bytes) to read
1497  * 
1498  *  Reads a specified block, and returns buffer head that contains it.
1499  *  It returns NULL if the block was unreadable.
1500  */
1501 struct buffer_head *
1502 __bread(struct block_device *bdev, sector_t block, unsigned size)
1503 {
1504         struct buffer_head *bh = __getblk(bdev, block, size);
1505
1506         if (likely(bh) && !buffer_uptodate(bh))
1507                 bh = __bread_slow(bh);
1508         return bh;
1509 }
1510 EXPORT_SYMBOL(__bread);
1511
1512 /*
1513  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1514  * This doesn't race because it runs in each cpu either in irq
1515  * or with preempt disabled.
1516  */
1517 static void invalidate_bh_lru(void *arg)
1518 {
1519         struct bh_lru *b = &get_cpu_var(bh_lrus);
1520         int i;
1521
1522         for (i = 0; i < BH_LRU_SIZE; i++) {
1523                 brelse(b->bhs[i]);
1524                 b->bhs[i] = NULL;
1525         }
1526         put_cpu_var(bh_lrus);
1527 }
1528         
1529 void invalidate_bh_lrus(void)
1530 {
1531         on_each_cpu(invalidate_bh_lru, NULL, 1);
1532 }
1533 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1534
1535 void set_bh_page(struct buffer_head *bh,
1536                 struct page *page, unsigned long offset)
1537 {
1538         bh->b_page = page;
1539         BUG_ON(offset >= PAGE_SIZE);
1540         if (PageHighMem(page))
1541                 /*
1542                  * This catches illegal uses and preserves the offset:
1543                  */
1544                 bh->b_data = (char *)(0 + offset);
1545         else
1546                 bh->b_data = page_address(page) + offset;
1547 }
1548 EXPORT_SYMBOL(set_bh_page);
1549
1550 /*
1551  * Called when truncating a buffer on a page completely.
1552  */
1553 static void discard_buffer(struct buffer_head * bh)
1554 {
1555         lock_buffer(bh);
1556         clear_buffer_dirty(bh);
1557         bh->b_bdev = NULL;
1558         clear_buffer_mapped(bh);
1559         clear_buffer_req(bh);
1560         clear_buffer_new(bh);
1561         clear_buffer_delay(bh);
1562         clear_buffer_unwritten(bh);
1563         unlock_buffer(bh);
1564 }
1565
1566 /**
1567  * block_invalidatepage - invalidate part of all of a buffer-backed page
1568  *
1569  * @page: the page which is affected
1570  * @offset: the index of the truncation point
1571  *
1572  * block_invalidatepage() is called when all or part of the page has become
1573  * invalidatedby a truncate operation.
1574  *
1575  * block_invalidatepage() does not have to release all buffers, but it must
1576  * ensure that no dirty buffer is left outside @offset and that no I/O
1577  * is underway against any of the blocks which are outside the truncation
1578  * point.  Because the caller is about to free (and possibly reuse) those
1579  * blocks on-disk.
1580  */
1581 void block_invalidatepage(struct page *page, unsigned long offset)
1582 {
1583         struct buffer_head *head, *bh, *next;
1584         unsigned int curr_off = 0;
1585
1586         BUG_ON(!PageLocked(page));
1587         if (!page_has_buffers(page))
1588                 goto out;
1589
1590         head = page_buffers(page);
1591         bh = head;
1592         do {
1593                 unsigned int next_off = curr_off + bh->b_size;
1594                 next = bh->b_this_page;
1595
1596                 /*
1597                  * is this block fully invalidated?
1598                  */
1599                 if (offset <= curr_off)
1600                         discard_buffer(bh);
1601                 curr_off = next_off;
1602                 bh = next;
1603         } while (bh != head);
1604
1605         /*
1606          * We release buffers only if the entire page is being invalidated.
1607          * The get_block cached value has been unconditionally invalidated,
1608          * so real IO is not possible anymore.
1609          */
1610         if (offset == 0)
1611                 try_to_release_page(page, 0);
1612 out:
1613         return;
1614 }
1615 EXPORT_SYMBOL(block_invalidatepage);
1616
1617 /*
1618  * We attach and possibly dirty the buffers atomically wrt
1619  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1620  * is already excluded via the page lock.
1621  */
1622 void create_empty_buffers(struct page *page,
1623                         unsigned long blocksize, unsigned long b_state)
1624 {
1625         struct buffer_head *bh, *head, *tail;
1626
1627         head = alloc_page_buffers(page, blocksize, 1);
1628         bh = head;
1629         do {
1630                 bh->b_state |= b_state;
1631                 tail = bh;
1632                 bh = bh->b_this_page;
1633         } while (bh);
1634         tail->b_this_page = head;
1635
1636         spin_lock(&page->mapping->private_lock);
1637         if (PageUptodate(page) || PageDirty(page)) {
1638                 bh = head;
1639                 do {
1640                         if (PageDirty(page))
1641                                 set_buffer_dirty(bh);
1642                         if (PageUptodate(page))
1643                                 set_buffer_uptodate(bh);
1644                         bh = bh->b_this_page;
1645                 } while (bh != head);
1646         }
1647         attach_page_buffers(page, head);
1648         spin_unlock(&page->mapping->private_lock);
1649 }
1650 EXPORT_SYMBOL(create_empty_buffers);
1651
1652 /*
1653  * We are taking a block for data and we don't want any output from any
1654  * buffer-cache aliases starting from return from that function and
1655  * until the moment when something will explicitly mark the buffer
1656  * dirty (hopefully that will not happen until we will free that block ;-)
1657  * We don't even need to mark it not-uptodate - nobody can expect
1658  * anything from a newly allocated buffer anyway. We used to used
1659  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1660  * don't want to mark the alias unmapped, for example - it would confuse
1661  * anyone who might pick it with bread() afterwards...
1662  *
1663  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1664  * be writeout I/O going on against recently-freed buffers.  We don't
1665  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1666  * only if we really need to.  That happens here.
1667  */
1668 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1669 {
1670         struct buffer_head *old_bh;
1671
1672         might_sleep();
1673
1674         old_bh = __find_get_block_slow(bdev, block);
1675         if (old_bh) {
1676                 clear_buffer_dirty(old_bh);
1677                 wait_on_buffer(old_bh);
1678                 clear_buffer_req(old_bh);
1679                 __brelse(old_bh);
1680         }
1681 }
1682 EXPORT_SYMBOL(unmap_underlying_metadata);
1683
1684 /*
1685  * NOTE! All mapped/uptodate combinations are valid:
1686  *
1687  *      Mapped  Uptodate        Meaning
1688  *
1689  *      No      No              "unknown" - must do get_block()
1690  *      No      Yes             "hole" - zero-filled
1691  *      Yes     No              "allocated" - allocated on disk, not read in
1692  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1693  *
1694  * "Dirty" is valid only with the last case (mapped+uptodate).
1695  */
1696
1697 /*
1698  * While block_write_full_page is writing back the dirty buffers under
1699  * the page lock, whoever dirtied the buffers may decide to clean them
1700  * again at any time.  We handle that by only looking at the buffer
1701  * state inside lock_buffer().
1702  *
1703  * If block_write_full_page() is called for regular writeback
1704  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1705  * locked buffer.   This only can happen if someone has written the buffer
1706  * directly, with submit_bh().  At the address_space level PageWriteback
1707  * prevents this contention from occurring.
1708  */
1709 static int __block_write_full_page(struct inode *inode, struct page *page,
1710                         get_block_t *get_block, struct writeback_control *wbc)
1711 {
1712         int err;
1713         sector_t block;
1714         sector_t last_block;
1715         struct buffer_head *bh, *head;
1716         const unsigned blocksize = 1 << inode->i_blkbits;
1717         int nr_underway = 0;
1718
1719         BUG_ON(!PageLocked(page));
1720
1721         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1722
1723         if (!page_has_buffers(page)) {
1724                 create_empty_buffers(page, blocksize,
1725                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1726         }
1727
1728         /*
1729          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1730          * here, and the (potentially unmapped) buffers may become dirty at
1731          * any time.  If a buffer becomes dirty here after we've inspected it
1732          * then we just miss that fact, and the page stays dirty.
1733          *
1734          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1735          * handle that here by just cleaning them.
1736          */
1737
1738         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1739         head = page_buffers(page);
1740         bh = head;
1741
1742         /*
1743          * Get all the dirty buffers mapped to disk addresses and
1744          * handle any aliases from the underlying blockdev's mapping.
1745          */
1746         do {
1747                 if (block > last_block) {
1748                         /*
1749                          * mapped buffers outside i_size will occur, because
1750                          * this page can be outside i_size when there is a
1751                          * truncate in progress.
1752                          */
1753                         /*
1754                          * The buffer was zeroed by block_write_full_page()
1755                          */
1756                         clear_buffer_dirty(bh);
1757                         set_buffer_uptodate(bh);
1758                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1759                            buffer_dirty(bh)) {
1760                         WARN_ON(bh->b_size != blocksize);
1761                         err = get_block(inode, block, bh, 1);
1762                         if (err)
1763                                 goto recover;
1764                         clear_buffer_delay(bh);
1765                         if (buffer_new(bh)) {
1766                                 /* blockdev mappings never come here */
1767                                 clear_buffer_new(bh);
1768                                 unmap_underlying_metadata(bh->b_bdev,
1769                                                         bh->b_blocknr);
1770                         }
1771                 }
1772                 bh = bh->b_this_page;
1773                 block++;
1774         } while (bh != head);
1775
1776         do {
1777                 if (!buffer_mapped(bh))
1778                         continue;
1779                 /*
1780                  * If it's a fully non-blocking write attempt and we cannot
1781                  * lock the buffer then redirty the page.  Note that this can
1782                  * potentially cause a busy-wait loop from pdflush and kswapd
1783                  * activity, but those code paths have their own higher-level
1784                  * throttling.
1785                  */
1786                 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1787                         lock_buffer(bh);
1788                 } else if (!trylock_buffer(bh)) {
1789                         redirty_page_for_writepage(wbc, page);
1790                         continue;
1791                 }
1792                 if (test_clear_buffer_dirty(bh)) {
1793                         mark_buffer_async_write(bh);
1794                 } else {
1795                         unlock_buffer(bh);
1796                 }
1797         } while ((bh = bh->b_this_page) != head);
1798
1799         /*
1800          * The page and its buffers are protected by PageWriteback(), so we can
1801          * drop the bh refcounts early.
1802          */
1803         BUG_ON(PageWriteback(page));
1804         set_page_writeback(page);
1805
1806         do {
1807                 struct buffer_head *next = bh->b_this_page;
1808                 if (buffer_async_write(bh)) {
1809                         submit_bh(WRITE, bh);
1810                         nr_underway++;
1811                 }
1812                 bh = next;
1813         } while (bh != head);
1814         unlock_page(page);
1815
1816         err = 0;
1817 done:
1818         if (nr_underway == 0) {
1819                 /*
1820                  * The page was marked dirty, but the buffers were
1821                  * clean.  Someone wrote them back by hand with
1822                  * ll_rw_block/submit_bh.  A rare case.
1823                  */
1824                 end_page_writeback(page);
1825
1826                 /*
1827                  * The page and buffer_heads can be released at any time from
1828                  * here on.
1829                  */
1830         }
1831         return err;
1832
1833 recover:
1834         /*
1835          * ENOSPC, or some other error.  We may already have added some
1836          * blocks to the file, so we need to write these out to avoid
1837          * exposing stale data.
1838          * The page is currently locked and not marked for writeback
1839          */
1840         bh = head;
1841         /* Recovery: lock and submit the mapped buffers */
1842         do {
1843                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1844                     !buffer_delay(bh)) {
1845                         lock_buffer(bh);
1846                         mark_buffer_async_write(bh);
1847                 } else {
1848                         /*
1849                          * The buffer may have been set dirty during
1850                          * attachment to a dirty page.
1851                          */
1852                         clear_buffer_dirty(bh);
1853                 }
1854         } while ((bh = bh->b_this_page) != head);
1855         SetPageError(page);
1856         BUG_ON(PageWriteback(page));
1857         mapping_set_error(page->mapping, err);
1858         set_page_writeback(page);
1859         do {
1860                 struct buffer_head *next = bh->b_this_page;
1861                 if (buffer_async_write(bh)) {
1862                         clear_buffer_dirty(bh);
1863                         submit_bh(WRITE, bh);
1864                         nr_underway++;
1865                 }
1866                 bh = next;
1867         } while (bh != head);
1868         unlock_page(page);
1869         goto done;
1870 }
1871
1872 /*
1873  * If a page has any new buffers, zero them out here, and mark them uptodate
1874  * and dirty so they'll be written out (in order to prevent uninitialised
1875  * block data from leaking). And clear the new bit.
1876  */
1877 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1878 {
1879         unsigned int block_start, block_end;
1880         struct buffer_head *head, *bh;
1881
1882         BUG_ON(!PageLocked(page));
1883         if (!page_has_buffers(page))
1884                 return;
1885
1886         bh = head = page_buffers(page);
1887         block_start = 0;
1888         do {
1889                 block_end = block_start + bh->b_size;
1890
1891                 if (buffer_new(bh)) {
1892                         if (block_end > from && block_start < to) {
1893                                 if (!PageUptodate(page)) {
1894                                         unsigned start, size;
1895
1896                                         start = max(from, block_start);
1897                                         size = min(to, block_end) - start;
1898
1899                                         zero_user(page, start, size);
1900                                         set_buffer_uptodate(bh);
1901                                 }
1902
1903                                 clear_buffer_new(bh);
1904                                 mark_buffer_dirty(bh);
1905                         }
1906                 }
1907
1908                 block_start = block_end;
1909                 bh = bh->b_this_page;
1910         } while (bh != head);
1911 }
1912 EXPORT_SYMBOL(page_zero_new_buffers);
1913
1914 static int __block_prepare_write(struct inode *inode, struct page *page,
1915                 unsigned from, unsigned to, get_block_t *get_block)
1916 {
1917         unsigned block_start, block_end;
1918         sector_t block;
1919         int err = 0;
1920         unsigned blocksize, bbits;
1921         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1922
1923         BUG_ON(!PageLocked(page));
1924         BUG_ON(from > PAGE_CACHE_SIZE);
1925         BUG_ON(to > PAGE_CACHE_SIZE);
1926         BUG_ON(from > to);
1927
1928         blocksize = 1 << inode->i_blkbits;
1929         if (!page_has_buffers(page))
1930                 create_empty_buffers(page, blocksize, 0);
1931         head = page_buffers(page);
1932
1933         bbits = inode->i_blkbits;
1934         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1935
1936         for(bh = head, block_start = 0; bh != head || !block_start;
1937             block++, block_start=block_end, bh = bh->b_this_page) {
1938                 block_end = block_start + blocksize;
1939                 if (block_end <= from || block_start >= to) {
1940                         if (PageUptodate(page)) {
1941                                 if (!buffer_uptodate(bh))
1942                                         set_buffer_uptodate(bh);
1943                         }
1944                         continue;
1945                 }
1946                 if (buffer_new(bh))
1947                         clear_buffer_new(bh);
1948                 if (!buffer_mapped(bh)) {
1949                         WARN_ON(bh->b_size != blocksize);
1950                         err = get_block(inode, block, bh, 1);
1951                         if (err)
1952                                 break;
1953                         if (buffer_new(bh)) {
1954                                 unmap_underlying_metadata(bh->b_bdev,
1955                                                         bh->b_blocknr);
1956                                 if (PageUptodate(page)) {
1957                                         clear_buffer_new(bh);
1958                                         set_buffer_uptodate(bh);
1959                                         mark_buffer_dirty(bh);
1960                                         continue;
1961                                 }
1962                                 if (block_end > to || block_start < from)
1963                                         zero_user_segments(page,
1964                                                 to, block_end,
1965                                                 block_start, from);
1966                                 continue;
1967                         }
1968                 }
1969                 if (PageUptodate(page)) {
1970                         if (!buffer_uptodate(bh))
1971                                 set_buffer_uptodate(bh);
1972                         continue; 
1973                 }
1974                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1975                     !buffer_unwritten(bh) &&
1976                      (block_start < from || block_end > to)) {
1977                         ll_rw_block(READ, 1, &bh);
1978                         *wait_bh++=bh;
1979                 }
1980         }
1981         /*
1982          * If we issued read requests - let them complete.
1983          */
1984         while(wait_bh > wait) {
1985                 wait_on_buffer(*--wait_bh);
1986                 if (!buffer_uptodate(*wait_bh))
1987                         err = -EIO;
1988         }
1989         if (unlikely(err))
1990                 page_zero_new_buffers(page, from, to);
1991         return err;
1992 }
1993
1994 static int __block_commit_write(struct inode *inode, struct page *page,
1995                 unsigned from, unsigned to)
1996 {
1997         unsigned block_start, block_end;
1998         int partial = 0;
1999         unsigned blocksize;
2000         struct buffer_head *bh, *head;
2001
2002         blocksize = 1 << inode->i_blkbits;
2003
2004         for(bh = head = page_buffers(page), block_start = 0;
2005             bh != head || !block_start;
2006             block_start=block_end, bh = bh->b_this_page) {
2007                 block_end = block_start + blocksize;
2008                 if (block_end <= from || block_start >= to) {
2009                         if (!buffer_uptodate(bh))
2010                                 partial = 1;
2011                 } else {
2012                         set_buffer_uptodate(bh);
2013                         mark_buffer_dirty(bh);
2014                 }
2015                 clear_buffer_new(bh);
2016         }
2017
2018         /*
2019          * If this is a partial write which happened to make all buffers
2020          * uptodate then we can optimize away a bogus readpage() for
2021          * the next read(). Here we 'discover' whether the page went
2022          * uptodate as a result of this (potentially partial) write.
2023          */
2024         if (!partial)
2025                 SetPageUptodate(page);
2026         return 0;
2027 }
2028
2029 /*
2030  * block_write_begin takes care of the basic task of block allocation and
2031  * bringing partial write blocks uptodate first.
2032  *
2033  * If *pagep is not NULL, then block_write_begin uses the locked page
2034  * at *pagep rather than allocating its own. In this case, the page will
2035  * not be unlocked or deallocated on failure.
2036  */
2037 int block_write_begin(struct file *file, struct address_space *mapping,
2038                         loff_t pos, unsigned len, unsigned flags,
2039                         struct page **pagep, void **fsdata,
2040                         get_block_t *get_block)
2041 {
2042         struct inode *inode = mapping->host;
2043         int status = 0;
2044         struct page *page;
2045         pgoff_t index;
2046         unsigned start, end;
2047         int ownpage = 0;
2048
2049         index = pos >> PAGE_CACHE_SHIFT;
2050         start = pos & (PAGE_CACHE_SIZE - 1);
2051         end = start + len;
2052
2053         page = *pagep;
2054         if (page == NULL) {
2055                 ownpage = 1;
2056                 page = grab_cache_page_write_begin(mapping, index, flags);
2057                 if (!page) {
2058                         status = -ENOMEM;
2059                         goto out;
2060                 }
2061                 *pagep = page;
2062         } else
2063                 BUG_ON(!PageLocked(page));
2064
2065         status = __block_prepare_write(inode, page, start, end, get_block);
2066         if (unlikely(status)) {
2067                 ClearPageUptodate(page);
2068
2069                 if (ownpage) {
2070                         unlock_page(page);
2071                         page_cache_release(page);
2072                         *pagep = NULL;
2073
2074                         /*
2075                          * prepare_write() may have instantiated a few blocks
2076                          * outside i_size.  Trim these off again. Don't need
2077                          * i_size_read because we hold i_mutex.
2078                          */
2079                         if (pos + len > inode->i_size)
2080                                 vmtruncate(inode, inode->i_size);
2081                 }
2082         }
2083
2084 out:
2085         return status;
2086 }
2087 EXPORT_SYMBOL(block_write_begin);
2088
2089 int block_write_end(struct file *file, struct address_space *mapping,
2090                         loff_t pos, unsigned len, unsigned copied,
2091                         struct page *page, void *fsdata)
2092 {
2093         struct inode *inode = mapping->host;
2094         unsigned start;
2095
2096         start = pos & (PAGE_CACHE_SIZE - 1);
2097
2098         if (unlikely(copied < len)) {
2099                 /*
2100                  * The buffers that were written will now be uptodate, so we
2101                  * don't have to worry about a readpage reading them and
2102                  * overwriting a partial write. However if we have encountered
2103                  * a short write and only partially written into a buffer, it
2104                  * will not be marked uptodate, so a readpage might come in and
2105                  * destroy our partial write.
2106                  *
2107                  * Do the simplest thing, and just treat any short write to a
2108                  * non uptodate page as a zero-length write, and force the
2109                  * caller to redo the whole thing.
2110                  */
2111                 if (!PageUptodate(page))
2112                         copied = 0;
2113
2114                 page_zero_new_buffers(page, start+copied, start+len);
2115         }
2116         flush_dcache_page(page);
2117
2118         /* This could be a short (even 0-length) commit */
2119         __block_commit_write(inode, page, start, start+copied);
2120
2121         return copied;
2122 }
2123 EXPORT_SYMBOL(block_write_end);
2124
2125 int generic_write_end(struct file *file, struct address_space *mapping,
2126                         loff_t pos, unsigned len, unsigned copied,
2127                         struct page *page, void *fsdata)
2128 {
2129         struct inode *inode = mapping->host;
2130         int i_size_changed = 0;
2131
2132         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2133
2134         /*
2135          * No need to use i_size_read() here, the i_size
2136          * cannot change under us because we hold i_mutex.
2137          *
2138          * But it's important to update i_size while still holding page lock:
2139          * page writeout could otherwise come in and zero beyond i_size.
2140          */
2141         if (pos+copied > inode->i_size) {
2142                 i_size_write(inode, pos+copied);
2143                 i_size_changed = 1;
2144         }
2145
2146         unlock_page(page);
2147         page_cache_release(page);
2148
2149         /*
2150          * Don't mark the inode dirty under page lock. First, it unnecessarily
2151          * makes the holding time of page lock longer. Second, it forces lock
2152          * ordering of page lock and transaction start for journaling
2153          * filesystems.
2154          */
2155         if (i_size_changed)
2156                 mark_inode_dirty(inode);
2157
2158         return copied;
2159 }
2160 EXPORT_SYMBOL(generic_write_end);
2161
2162 /*
2163  * block_is_partially_uptodate checks whether buffers within a page are
2164  * uptodate or not.
2165  *
2166  * Returns true if all buffers which correspond to a file portion
2167  * we want to read are uptodate.
2168  */
2169 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2170                                         unsigned long from)
2171 {
2172         struct inode *inode = page->mapping->host;
2173         unsigned block_start, block_end, blocksize;
2174         unsigned to;
2175         struct buffer_head *bh, *head;
2176         int ret = 1;
2177
2178         if (!page_has_buffers(page))
2179                 return 0;
2180
2181         blocksize = 1 << inode->i_blkbits;
2182         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2183         to = from + to;
2184         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2185                 return 0;
2186
2187         head = page_buffers(page);
2188         bh = head;
2189         block_start = 0;
2190         do {
2191                 block_end = block_start + blocksize;
2192                 if (block_end > from && block_start < to) {
2193                         if (!buffer_uptodate(bh)) {
2194                                 ret = 0;
2195                                 break;
2196                         }
2197                         if (block_end >= to)
2198                                 break;
2199                 }
2200                 block_start = block_end;
2201                 bh = bh->b_this_page;
2202         } while (bh != head);
2203
2204         return ret;
2205 }
2206 EXPORT_SYMBOL(block_is_partially_uptodate);
2207
2208 /*
2209  * Generic "read page" function for block devices that have the normal
2210  * get_block functionality. This is most of the block device filesystems.
2211  * Reads the page asynchronously --- the unlock_buffer() and
2212  * set/clear_buffer_uptodate() functions propagate buffer state into the
2213  * page struct once IO has completed.
2214  */
2215 int block_read_full_page(struct page *page, get_block_t *get_block)
2216 {
2217         struct inode *inode = page->mapping->host;
2218         sector_t iblock, lblock;
2219         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2220         unsigned int blocksize;
2221         int nr, i;
2222         int fully_mapped = 1;
2223
2224         BUG_ON(!PageLocked(page));
2225         blocksize = 1 << inode->i_blkbits;
2226         if (!page_has_buffers(page))
2227                 create_empty_buffers(page, blocksize, 0);
2228         head = page_buffers(page);
2229
2230         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2231         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2232         bh = head;
2233         nr = 0;
2234         i = 0;
2235
2236         do {
2237                 if (buffer_uptodate(bh))
2238                         continue;
2239
2240                 if (!buffer_mapped(bh)) {
2241                         int err = 0;
2242
2243                         fully_mapped = 0;
2244                         if (iblock < lblock) {
2245                                 WARN_ON(bh->b_size != blocksize);
2246                                 err = get_block(inode, iblock, bh, 0);
2247                                 if (err)
2248                                         SetPageError(page);
2249                         }
2250                         if (!buffer_mapped(bh)) {
2251                                 zero_user(page, i * blocksize, blocksize);
2252                                 if (!err)
2253                                         set_buffer_uptodate(bh);
2254                                 continue;
2255                         }
2256                         /*
2257                          * get_block() might have updated the buffer
2258                          * synchronously
2259                          */
2260                         if (buffer_uptodate(bh))
2261                                 continue;
2262                 }
2263                 arr[nr++] = bh;
2264         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2265
2266         if (fully_mapped)
2267                 SetPageMappedToDisk(page);
2268
2269         if (!nr) {
2270                 /*
2271                  * All buffers are uptodate - we can set the page uptodate
2272                  * as well. But not if get_block() returned an error.
2273                  */
2274                 if (!PageError(page))
2275                         SetPageUptodate(page);
2276                 unlock_page(page);
2277                 return 0;
2278         }
2279
2280         /* Stage two: lock the buffers */
2281         for (i = 0; i < nr; i++) {
2282                 bh = arr[i];
2283                 lock_buffer(bh);
2284                 mark_buffer_async_read(bh);
2285         }
2286
2287         /*
2288          * Stage 3: start the IO.  Check for uptodateness
2289          * inside the buffer lock in case another process reading
2290          * the underlying blockdev brought it uptodate (the sct fix).
2291          */
2292         for (i = 0; i < nr; i++) {
2293                 bh = arr[i];
2294                 if (buffer_uptodate(bh))
2295                         end_buffer_async_read(bh, 1);
2296                 else
2297                         submit_bh(READ, bh);
2298         }
2299         return 0;
2300 }
2301
2302 /* utility function for filesystems that need to do work on expanding
2303  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2304  * deal with the hole.  
2305  */
2306 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2307 {
2308         struct address_space *mapping = inode->i_mapping;
2309         struct page *page;
2310         void *fsdata;
2311         unsigned long limit;
2312         int err;
2313
2314         err = -EFBIG;
2315         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2316         if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2317                 send_sig(SIGXFSZ, current, 0);
2318                 goto out;
2319         }
2320         if (size > inode->i_sb->s_maxbytes)
2321                 goto out;
2322
2323         err = pagecache_write_begin(NULL, mapping, size, 0,
2324                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2325                                 &page, &fsdata);
2326         if (err)
2327                 goto out;
2328
2329         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2330         BUG_ON(err > 0);
2331
2332 out:
2333         return err;
2334 }
2335
2336 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2337                             loff_t pos, loff_t *bytes)
2338 {
2339         struct inode *inode = mapping->host;
2340         unsigned blocksize = 1 << inode->i_blkbits;
2341         struct page *page;
2342         void *fsdata;
2343         pgoff_t index, curidx;
2344         loff_t curpos;
2345         unsigned zerofrom, offset, len;
2346         int err = 0;
2347
2348         index = pos >> PAGE_CACHE_SHIFT;
2349         offset = pos & ~PAGE_CACHE_MASK;
2350
2351         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2352                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2353                 if (zerofrom & (blocksize-1)) {
2354                         *bytes |= (blocksize-1);
2355                         (*bytes)++;
2356                 }
2357                 len = PAGE_CACHE_SIZE - zerofrom;
2358
2359                 err = pagecache_write_begin(file, mapping, curpos, len,
2360                                                 AOP_FLAG_UNINTERRUPTIBLE,
2361                                                 &page, &fsdata);
2362                 if (err)
2363                         goto out;
2364                 zero_user(page, zerofrom, len);
2365                 err = pagecache_write_end(file, mapping, curpos, len, len,
2366                                                 page, fsdata);
2367                 if (err < 0)
2368                         goto out;
2369                 BUG_ON(err != len);
2370                 err = 0;
2371
2372                 balance_dirty_pages_ratelimited(mapping);
2373         }
2374
2375         /* page covers the boundary, find the boundary offset */
2376         if (index == curidx) {
2377                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2378                 /* if we will expand the thing last block will be filled */
2379                 if (offset <= zerofrom) {
2380                         goto out;
2381                 }
2382                 if (zerofrom & (blocksize-1)) {
2383                         *bytes |= (blocksize-1);
2384                         (*bytes)++;
2385                 }
2386                 len = offset - zerofrom;
2387
2388                 err = pagecache_write_begin(file, mapping, curpos, len,
2389                                                 AOP_FLAG_UNINTERRUPTIBLE,
2390                                                 &page, &fsdata);
2391                 if (err)
2392                         goto out;
2393                 zero_user(page, zerofrom, len);
2394                 err = pagecache_write_end(file, mapping, curpos, len, len,
2395                                                 page, fsdata);
2396                 if (err < 0)
2397                         goto out;
2398                 BUG_ON(err != len);
2399                 err = 0;
2400         }
2401 out:
2402         return err;
2403 }
2404
2405 /*
2406  * For moronic filesystems that do not allow holes in file.
2407  * We may have to extend the file.
2408  */
2409 int cont_write_begin(struct file *file, struct address_space *mapping,
2410                         loff_t pos, unsigned len, unsigned flags,
2411                         struct page **pagep, void **fsdata,
2412                         get_block_t *get_block, loff_t *bytes)
2413 {
2414         struct inode *inode = mapping->host;
2415         unsigned blocksize = 1 << inode->i_blkbits;
2416         unsigned zerofrom;
2417         int err;
2418
2419         err = cont_expand_zero(file, mapping, pos, bytes);
2420         if (err)
2421                 goto out;
2422
2423         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2424         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2425                 *bytes |= (blocksize-1);
2426                 (*bytes)++;
2427         }
2428
2429         *pagep = NULL;
2430         err = block_write_begin(file, mapping, pos, len,
2431                                 flags, pagep, fsdata, get_block);
2432 out:
2433         return err;
2434 }
2435
2436 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2437                         get_block_t *get_block)
2438 {
2439         struct inode *inode = page->mapping->host;
2440         int err = __block_prepare_write(inode, page, from, to, get_block);
2441         if (err)
2442                 ClearPageUptodate(page);
2443         return err;
2444 }
2445
2446 int block_commit_write(struct page *page, unsigned from, unsigned to)
2447 {
2448         struct inode *inode = page->mapping->host;
2449         __block_commit_write(inode,page,from,to);
2450         return 0;
2451 }
2452
2453 /*
2454  * block_page_mkwrite() is not allowed to change the file size as it gets
2455  * called from a page fault handler when a page is first dirtied. Hence we must
2456  * be careful to check for EOF conditions here. We set the page up correctly
2457  * for a written page which means we get ENOSPC checking when writing into
2458  * holes and correct delalloc and unwritten extent mapping on filesystems that
2459  * support these features.
2460  *
2461  * We are not allowed to take the i_mutex here so we have to play games to
2462  * protect against truncate races as the page could now be beyond EOF.  Because
2463  * vmtruncate() writes the inode size before removing pages, once we have the
2464  * page lock we can determine safely if the page is beyond EOF. If it is not
2465  * beyond EOF, then the page is guaranteed safe against truncation until we
2466  * unlock the page.
2467  */
2468 int
2469 block_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2470                    get_block_t get_block)
2471 {
2472         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2473         unsigned long end;
2474         loff_t size;
2475         int ret = -EINVAL;
2476
2477         lock_page(page);
2478         size = i_size_read(inode);
2479         if ((page->mapping != inode->i_mapping) ||
2480             (page_offset(page) > size)) {
2481                 /* page got truncated out from underneath us */
2482                 goto out_unlock;
2483         }
2484
2485         /* page is wholly or partially inside EOF */
2486         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2487                 end = size & ~PAGE_CACHE_MASK;
2488         else
2489                 end = PAGE_CACHE_SIZE;
2490
2491         ret = block_prepare_write(page, 0, end, get_block);
2492         if (!ret)
2493                 ret = block_commit_write(page, 0, end);
2494
2495 out_unlock:
2496         unlock_page(page);
2497         return ret;
2498 }
2499
2500 /*
2501  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2502  * immediately, while under the page lock.  So it needs a special end_io
2503  * handler which does not touch the bh after unlocking it.
2504  */
2505 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2506 {
2507         __end_buffer_read_notouch(bh, uptodate);
2508 }
2509
2510 /*
2511  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2512  * the page (converting it to circular linked list and taking care of page
2513  * dirty races).
2514  */
2515 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2516 {
2517         struct buffer_head *bh;
2518
2519         BUG_ON(!PageLocked(page));
2520
2521         spin_lock(&page->mapping->private_lock);
2522         bh = head;
2523         do {
2524                 if (PageDirty(page))
2525                         set_buffer_dirty(bh);
2526                 if (!bh->b_this_page)
2527                         bh->b_this_page = head;
2528                 bh = bh->b_this_page;
2529         } while (bh != head);
2530         attach_page_buffers(page, head);
2531         spin_unlock(&page->mapping->private_lock);
2532 }
2533
2534 /*
2535  * On entry, the page is fully not uptodate.
2536  * On exit the page is fully uptodate in the areas outside (from,to)
2537  */
2538 int nobh_write_begin(struct file *file, struct address_space *mapping,
2539                         loff_t pos, unsigned len, unsigned flags,
2540                         struct page **pagep, void **fsdata,
2541                         get_block_t *get_block)
2542 {
2543         struct inode *inode = mapping->host;
2544         const unsigned blkbits = inode->i_blkbits;
2545         const unsigned blocksize = 1 << blkbits;
2546         struct buffer_head *head, *bh;
2547         struct page *page;
2548         pgoff_t index;
2549         unsigned from, to;
2550         unsigned block_in_page;
2551         unsigned block_start, block_end;
2552         sector_t block_in_file;
2553         int nr_reads = 0;
2554         int ret = 0;
2555         int is_mapped_to_disk = 1;
2556
2557         index = pos >> PAGE_CACHE_SHIFT;
2558         from = pos & (PAGE_CACHE_SIZE - 1);
2559         to = from + len;
2560
2561         page = grab_cache_page_write_begin(mapping, index, flags);
2562         if (!page)
2563                 return -ENOMEM;
2564         *pagep = page;
2565         *fsdata = NULL;
2566
2567         if (page_has_buffers(page)) {
2568                 unlock_page(page);
2569                 page_cache_release(page);
2570                 *pagep = NULL;
2571                 return block_write_begin(file, mapping, pos, len, flags, pagep,
2572                                         fsdata, get_block);
2573         }
2574
2575         if (PageMappedToDisk(page))
2576                 return 0;
2577
2578         /*
2579          * Allocate buffers so that we can keep track of state, and potentially
2580          * attach them to the page if an error occurs. In the common case of
2581          * no error, they will just be freed again without ever being attached
2582          * to the page (which is all OK, because we're under the page lock).
2583          *
2584          * Be careful: the buffer linked list is a NULL terminated one, rather
2585          * than the circular one we're used to.
2586          */
2587         head = alloc_page_buffers(page, blocksize, 0);
2588         if (!head) {
2589                 ret = -ENOMEM;
2590                 goto out_release;
2591         }
2592
2593         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2594
2595         /*
2596          * We loop across all blocks in the page, whether or not they are
2597          * part of the affected region.  This is so we can discover if the
2598          * page is fully mapped-to-disk.
2599          */
2600         for (block_start = 0, block_in_page = 0, bh = head;
2601                   block_start < PAGE_CACHE_SIZE;
2602                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2603                 int create;
2604
2605                 block_end = block_start + blocksize;
2606                 bh->b_state = 0;
2607                 create = 1;
2608                 if (block_start >= to)
2609                         create = 0;
2610                 ret = get_block(inode, block_in_file + block_in_page,
2611                                         bh, create);
2612                 if (ret)
2613                         goto failed;
2614                 if (!buffer_mapped(bh))
2615                         is_mapped_to_disk = 0;
2616                 if (buffer_new(bh))
2617                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2618                 if (PageUptodate(page)) {
2619                         set_buffer_uptodate(bh);
2620                         continue;
2621                 }
2622                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2623                         zero_user_segments(page, block_start, from,
2624                                                         to, block_end);
2625                         continue;
2626                 }
2627                 if (buffer_uptodate(bh))
2628                         continue;       /* reiserfs does this */
2629                 if (block_start < from || block_end > to) {
2630                         lock_buffer(bh);
2631                         bh->b_end_io = end_buffer_read_nobh;
2632                         submit_bh(READ, bh);
2633                         nr_reads++;
2634                 }
2635         }
2636
2637         if (nr_reads) {
2638                 /*
2639                  * The page is locked, so these buffers are protected from
2640                  * any VM or truncate activity.  Hence we don't need to care
2641                  * for the buffer_head refcounts.
2642                  */
2643                 for (bh = head; bh; bh = bh->b_this_page) {
2644                         wait_on_buffer(bh);
2645                         if (!buffer_uptodate(bh))
2646                                 ret = -EIO;
2647                 }
2648                 if (ret)
2649                         goto failed;
2650         }
2651
2652         if (is_mapped_to_disk)
2653                 SetPageMappedToDisk(page);
2654
2655         *fsdata = head; /* to be released by nobh_write_end */
2656
2657         return 0;
2658
2659 failed:
2660         BUG_ON(!ret);
2661         /*
2662          * Error recovery is a bit difficult. We need to zero out blocks that
2663          * were newly allocated, and dirty them to ensure they get written out.
2664          * Buffers need to be attached to the page at this point, otherwise
2665          * the handling of potential IO errors during writeout would be hard
2666          * (could try doing synchronous writeout, but what if that fails too?)
2667          */
2668         attach_nobh_buffers(page, head);
2669         page_zero_new_buffers(page, from, to);
2670
2671 out_release:
2672         unlock_page(page);
2673         page_cache_release(page);
2674         *pagep = NULL;
2675
2676         if (pos + len > inode->i_size)
2677                 vmtruncate(inode, inode->i_size);
2678
2679         return ret;
2680 }
2681 EXPORT_SYMBOL(nobh_write_begin);
2682
2683 int nobh_write_end(struct file *file, struct address_space *mapping,
2684                         loff_t pos, unsigned len, unsigned copied,
2685                         struct page *page, void *fsdata)
2686 {
2687         struct inode *inode = page->mapping->host;
2688         struct buffer_head *head = fsdata;
2689         struct buffer_head *bh;
2690         BUG_ON(fsdata != NULL && page_has_buffers(page));
2691
2692         if (unlikely(copied < len) && head)
2693                 attach_nobh_buffers(page, head);
2694         if (page_has_buffers(page))
2695                 return generic_write_end(file, mapping, pos, len,
2696                                         copied, page, fsdata);
2697
2698         SetPageUptodate(page);
2699         set_page_dirty(page);
2700         if (pos+copied > inode->i_size) {
2701                 i_size_write(inode, pos+copied);
2702                 mark_inode_dirty(inode);
2703         }
2704
2705         unlock_page(page);
2706         page_cache_release(page);
2707
2708         while (head) {
2709                 bh = head;
2710                 head = head->b_this_page;
2711                 free_buffer_head(bh);
2712         }
2713
2714         return copied;
2715 }
2716 EXPORT_SYMBOL(nobh_write_end);
2717
2718 /*
2719  * nobh_writepage() - based on block_full_write_page() except
2720  * that it tries to operate without attaching bufferheads to
2721  * the page.
2722  */
2723 int nobh_writepage(struct page *page, get_block_t *get_block,
2724                         struct writeback_control *wbc)
2725 {
2726         struct inode * const inode = page->mapping->host;
2727         loff_t i_size = i_size_read(inode);
2728         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2729         unsigned offset;
2730         int ret;
2731
2732         /* Is the page fully inside i_size? */
2733         if (page->index < end_index)
2734                 goto out;
2735
2736         /* Is the page fully outside i_size? (truncate in progress) */
2737         offset = i_size & (PAGE_CACHE_SIZE-1);
2738         if (page->index >= end_index+1 || !offset) {
2739                 /*
2740                  * The page may have dirty, unmapped buffers.  For example,
2741                  * they may have been added in ext3_writepage().  Make them
2742                  * freeable here, so the page does not leak.
2743                  */
2744 #if 0
2745                 /* Not really sure about this  - do we need this ? */
2746                 if (page->mapping->a_ops->invalidatepage)
2747                         page->mapping->a_ops->invalidatepage(page, offset);
2748 #endif
2749                 unlock_page(page);
2750                 return 0; /* don't care */
2751         }
2752
2753         /*
2754          * The page straddles i_size.  It must be zeroed out on each and every
2755          * writepage invocation because it may be mmapped.  "A file is mapped
2756          * in multiples of the page size.  For a file that is not a multiple of
2757          * the  page size, the remaining memory is zeroed when mapped, and
2758          * writes to that region are not written out to the file."
2759          */
2760         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2761 out:
2762         ret = mpage_writepage(page, get_block, wbc);
2763         if (ret == -EAGAIN)
2764                 ret = __block_write_full_page(inode, page, get_block, wbc);
2765         return ret;
2766 }
2767 EXPORT_SYMBOL(nobh_writepage);
2768
2769 int nobh_truncate_page(struct address_space *mapping,
2770                         loff_t from, get_block_t *get_block)
2771 {
2772         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2773         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2774         unsigned blocksize;
2775         sector_t iblock;
2776         unsigned length, pos;
2777         struct inode *inode = mapping->host;
2778         struct page *page;
2779         struct buffer_head map_bh;
2780         int err;
2781
2782         blocksize = 1 << inode->i_blkbits;
2783         length = offset & (blocksize - 1);
2784
2785         /* Block boundary? Nothing to do */
2786         if (!length)
2787                 return 0;
2788
2789         length = blocksize - length;
2790         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2791
2792         page = grab_cache_page(mapping, index);
2793         err = -ENOMEM;
2794         if (!page)
2795                 goto out;
2796
2797         if (page_has_buffers(page)) {
2798 has_buffers:
2799                 unlock_page(page);
2800                 page_cache_release(page);
2801                 return block_truncate_page(mapping, from, get_block);
2802         }
2803
2804         /* Find the buffer that contains "offset" */
2805         pos = blocksize;
2806         while (offset >= pos) {
2807                 iblock++;
2808                 pos += blocksize;
2809         }
2810
2811         err = get_block(inode, iblock, &map_bh, 0);
2812         if (err)
2813                 goto unlock;
2814         /* unmapped? It's a hole - nothing to do */
2815         if (!buffer_mapped(&map_bh))
2816                 goto unlock;
2817
2818         /* Ok, it's mapped. Make sure it's up-to-date */
2819         if (!PageUptodate(page)) {
2820                 err = mapping->a_ops->readpage(NULL, page);
2821                 if (err) {
2822                         page_cache_release(page);
2823                         goto out;
2824                 }
2825                 lock_page(page);
2826                 if (!PageUptodate(page)) {
2827                         err = -EIO;
2828                         goto unlock;
2829                 }
2830                 if (page_has_buffers(page))
2831                         goto has_buffers;
2832         }
2833         zero_user(page, offset, length);
2834         set_page_dirty(page);
2835         err = 0;
2836
2837 unlock:
2838         unlock_page(page);
2839         page_cache_release(page);
2840 out:
2841         return err;
2842 }
2843 EXPORT_SYMBOL(nobh_truncate_page);
2844
2845 int block_truncate_page(struct address_space *mapping,
2846                         loff_t from, get_block_t *get_block)
2847 {
2848         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2849         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2850         unsigned blocksize;
2851         sector_t iblock;
2852         unsigned length, pos;
2853         struct inode *inode = mapping->host;
2854         struct page *page;
2855         struct buffer_head *bh;
2856         int err;
2857
2858         blocksize = 1 << inode->i_blkbits;
2859         length = offset & (blocksize - 1);
2860
2861         /* Block boundary? Nothing to do */
2862         if (!length)
2863                 return 0;
2864
2865         length = blocksize - length;
2866         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2867         
2868         page = grab_cache_page(mapping, index);
2869         err = -ENOMEM;
2870         if (!page)
2871                 goto out;
2872
2873         if (!page_has_buffers(page))
2874                 create_empty_buffers(page, blocksize, 0);
2875
2876         /* Find the buffer that contains "offset" */
2877         bh = page_buffers(page);
2878         pos = blocksize;
2879         while (offset >= pos) {
2880                 bh = bh->b_this_page;
2881                 iblock++;
2882                 pos += blocksize;
2883         }
2884
2885         err = 0;
2886         if (!buffer_mapped(bh)) {
2887                 WARN_ON(bh->b_size != blocksize);
2888                 err = get_block(inode, iblock, bh, 0);
2889                 if (err)
2890                         goto unlock;
2891                 /* unmapped? It's a hole - nothing to do */
2892                 if (!buffer_mapped(bh))
2893                         goto unlock;
2894         }
2895
2896         /* Ok, it's mapped. Make sure it's up-to-date */
2897         if (PageUptodate(page))
2898                 set_buffer_uptodate(bh);
2899
2900         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2901                 err = -EIO;
2902                 ll_rw_block(READ, 1, &bh);
2903                 wait_on_buffer(bh);
2904                 /* Uhhuh. Read error. Complain and punt. */
2905                 if (!buffer_uptodate(bh))
2906                         goto unlock;
2907         }
2908
2909         zero_user(page, offset, length);
2910         mark_buffer_dirty(bh);
2911         err = 0;
2912
2913 unlock:
2914         unlock_page(page);
2915         page_cache_release(page);
2916 out:
2917         return err;
2918 }
2919
2920 /*
2921  * The generic ->writepage function for buffer-backed address_spaces
2922  */
2923 int block_write_full_page(struct page *page, get_block_t *get_block,
2924                         struct writeback_control *wbc)
2925 {
2926         struct inode * const inode = page->mapping->host;
2927         loff_t i_size = i_size_read(inode);
2928         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2929         unsigned offset;
2930
2931         /* Is the page fully inside i_size? */
2932         if (page->index < end_index)
2933                 return __block_write_full_page(inode, page, get_block, wbc);
2934
2935         /* Is the page fully outside i_size? (truncate in progress) */
2936         offset = i_size & (PAGE_CACHE_SIZE-1);
2937         if (page->index >= end_index+1 || !offset) {
2938                 /*
2939                  * The page may have dirty, unmapped buffers.  For example,
2940                  * they may have been added in ext3_writepage().  Make them
2941                  * freeable here, so the page does not leak.
2942                  */
2943                 do_invalidatepage(page, 0);
2944                 unlock_page(page);
2945                 return 0; /* don't care */
2946         }
2947
2948         /*
2949          * The page straddles i_size.  It must be zeroed out on each and every
2950          * writepage invokation because it may be mmapped.  "A file is mapped
2951          * in multiples of the page size.  For a file that is not a multiple of
2952          * the  page size, the remaining memory is zeroed when mapped, and
2953          * writes to that region are not written out to the file."
2954          */
2955         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2956         return __block_write_full_page(inode, page, get_block, wbc);
2957 }
2958
2959 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2960                             get_block_t *get_block)
2961 {
2962         struct buffer_head tmp;
2963         struct inode *inode = mapping->host;
2964         tmp.b_state = 0;
2965         tmp.b_blocknr = 0;
2966         tmp.b_size = 1 << inode->i_blkbits;
2967         get_block(inode, block, &tmp, 0);
2968         return tmp.b_blocknr;
2969 }
2970
2971 static void end_bio_bh_io_sync(struct bio *bio, int err)
2972 {
2973         struct buffer_head *bh = bio->bi_private;
2974
2975         if (err == -EOPNOTSUPP) {
2976                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2977                 set_bit(BH_Eopnotsupp, &bh->b_state);
2978         }
2979
2980         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2981                 set_bit(BH_Quiet, &bh->b_state);
2982
2983         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2984         bio_put(bio);
2985 }
2986
2987 int submit_bh(int rw, struct buffer_head * bh)
2988 {
2989         struct bio *bio;
2990         int ret = 0;
2991
2992         BUG_ON(!buffer_locked(bh));
2993         BUG_ON(!buffer_mapped(bh));
2994         BUG_ON(!bh->b_end_io);
2995
2996         /*
2997          * Mask in barrier bit for a write (could be either a WRITE or a
2998          * WRITE_SYNC
2999          */
3000         if (buffer_ordered(bh) && (rw & WRITE))
3001                 rw |= WRITE_BARRIER;
3002
3003         /*
3004          * Only clear out a write error when rewriting
3005          */
3006         if (test_set_buffer_req(bh) && (rw & WRITE))
3007                 clear_buffer_write_io_error(bh);
3008
3009         /*
3010          * from here on down, it's all bio -- do the initial mapping,
3011          * submit_bio -> generic_make_request may further map this bio around
3012          */
3013         bio = bio_alloc(GFP_NOIO, 1);
3014
3015         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3016         bio->bi_bdev = bh->b_bdev;
3017         bio->bi_io_vec[0].bv_page = bh->b_page;
3018         bio->bi_io_vec[0].bv_len = bh->b_size;
3019         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
3020
3021         bio->bi_vcnt = 1;
3022         bio->bi_idx = 0;
3023         bio->bi_size = bh->b_size;
3024
3025         bio->bi_end_io = end_bio_bh_io_sync;
3026         bio->bi_private = bh;
3027
3028         bio_get(bio);
3029         submit_bio(rw, bio);
3030
3031         if (bio_flagged(bio, BIO_EOPNOTSUPP))
3032                 ret = -EOPNOTSUPP;
3033
3034         bio_put(bio);
3035         return ret;
3036 }
3037
3038 /**
3039  * ll_rw_block: low-level access to block devices (DEPRECATED)
3040  * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
3041  * @nr: number of &struct buffer_heads in the array
3042  * @bhs: array of pointers to &struct buffer_head
3043  *
3044  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3045  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
3046  * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
3047  * are sent to disk. The fourth %READA option is described in the documentation
3048  * for generic_make_request() which ll_rw_block() calls.
3049  *
3050  * This function drops any buffer that it cannot get a lock on (with the
3051  * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
3052  * clean when doing a write request, and any buffer that appears to be
3053  * up-to-date when doing read request.  Further it marks as clean buffers that
3054  * are processed for writing (the buffer cache won't assume that they are
3055  * actually clean until the buffer gets unlocked).
3056  *
3057  * ll_rw_block sets b_end_io to simple completion handler that marks
3058  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
3059  * any waiters. 
3060  *
3061  * All of the buffers must be for the same device, and must also be a
3062  * multiple of the current approved size for the device.
3063  */
3064 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
3065 {
3066         int i;
3067
3068         for (i = 0; i < nr; i++) {
3069                 struct buffer_head *bh = bhs[i];
3070
3071                 if (rw == SWRITE || rw == SWRITE_SYNC)
3072                         lock_buffer(bh);
3073                 else if (!trylock_buffer(bh))
3074                         continue;
3075
3076                 if (rw == WRITE || rw == SWRITE || rw == SWRITE_SYNC) {
3077                         if (test_clear_buffer_dirty(bh)) {
3078                                 bh->b_end_io = end_buffer_write_sync;
3079                                 get_bh(bh);
3080                                 if (rw == SWRITE_SYNC)
3081                                         submit_bh(WRITE_SYNC, bh);
3082                                 else
3083                                         submit_bh(WRITE, bh);
3084                                 continue;
3085                         }
3086                 } else {
3087                         if (!buffer_uptodate(bh)) {
3088                                 bh->b_end_io = end_buffer_read_sync;
3089                                 get_bh(bh);
3090                                 submit_bh(rw, bh);
3091                                 continue;
3092                         }
3093                 }
3094                 unlock_buffer(bh);
3095         }
3096 }
3097
3098 /*
3099  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3100  * and then start new I/O and then wait upon it.  The caller must have a ref on
3101  * the buffer_head.
3102  */
3103 int sync_dirty_buffer(struct buffer_head *bh)
3104 {
3105         int ret = 0;
3106
3107         WARN_ON(atomic_read(&bh->b_count) < 1);
3108         lock_buffer(bh);
3109         if (test_clear_buffer_dirty(bh)) {
3110                 get_bh(bh);
3111                 bh->b_end_io = end_buffer_write_sync;
3112                 ret = submit_bh(WRITE, bh);
3113                 wait_on_buffer(bh);
3114                 if (buffer_eopnotsupp(bh)) {
3115                         clear_buffer_eopnotsupp(bh);
3116                         ret = -EOPNOTSUPP;
3117                 }
3118                 if (!ret && !buffer_uptodate(bh))
3119                         ret = -EIO;
3120         } else {
3121                 unlock_buffer(bh);
3122         }
3123         return ret;
3124 }
3125
3126 /*
3127  * try_to_free_buffers() checks if all the buffers on this particular page
3128  * are unused, and releases them if so.
3129  *
3130  * Exclusion against try_to_free_buffers may be obtained by either
3131  * locking the page or by holding its mapping's private_lock.
3132  *
3133  * If the page is dirty but all the buffers are clean then we need to
3134  * be sure to mark the page clean as well.  This is because the page
3135  * may be against a block device, and a later reattachment of buffers
3136  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3137  * filesystem data on the same device.
3138  *
3139  * The same applies to regular filesystem pages: if all the buffers are
3140  * clean then we set the page clean and proceed.  To do that, we require
3141  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3142  * private_lock.
3143  *
3144  * try_to_free_buffers() is non-blocking.
3145  */
3146 static inline int buffer_busy(struct buffer_head *bh)
3147 {
3148         return atomic_read(&bh->b_count) |
3149                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3150 }
3151
3152 static int
3153 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3154 {
3155         struct buffer_head *head = page_buffers(page);
3156         struct buffer_head *bh;
3157
3158         bh = head;
3159         do {
3160                 if (buffer_write_io_error(bh) && page->mapping)
3161                         set_bit(AS_EIO, &page->mapping->flags);
3162                 if (buffer_busy(bh))
3163                         goto failed;
3164                 bh = bh->b_this_page;
3165         } while (bh != head);
3166
3167         do {
3168                 struct buffer_head *next = bh->b_this_page;
3169
3170                 if (bh->b_assoc_map)
3171                         __remove_assoc_queue(bh);
3172                 bh = next;
3173         } while (bh != head);
3174         *buffers_to_free = head;
3175         __clear_page_buffers(page);
3176         return 1;
3177 failed:
3178         return 0;
3179 }
3180
3181 int try_to_free_buffers(struct page *page)
3182 {
3183         struct address_space * const mapping = page->mapping;
3184         struct buffer_head *buffers_to_free = NULL;
3185         int ret = 0;
3186
3187         BUG_ON(!PageLocked(page));
3188         if (PageWriteback(page))
3189                 return 0;
3190
3191         if (mapping == NULL) {          /* can this still happen? */
3192                 ret = drop_buffers(page, &buffers_to_free);
3193                 goto out;
3194         }
3195
3196         spin_lock(&mapping->private_lock);
3197         ret = drop_buffers(page, &buffers_to_free);
3198
3199         /*
3200          * If the filesystem writes its buffers by hand (eg ext3)
3201          * then we can have clean buffers against a dirty page.  We
3202          * clean the page here; otherwise the VM will never notice
3203          * that the filesystem did any IO at all.
3204          *
3205          * Also, during truncate, discard_buffer will have marked all
3206          * the page's buffers clean.  We discover that here and clean
3207          * the page also.
3208          *
3209          * private_lock must be held over this entire operation in order
3210          * to synchronise against __set_page_dirty_buffers and prevent the
3211          * dirty bit from being lost.
3212          */
3213         if (ret)
3214                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3215         spin_unlock(&mapping->private_lock);
3216 out:
3217         if (buffers_to_free) {
3218                 struct buffer_head *bh = buffers_to_free;
3219
3220                 do {
3221                         struct buffer_head *next = bh->b_this_page;
3222                         free_buffer_head(bh);
3223                         bh = next;
3224                 } while (bh != buffers_to_free);
3225         }
3226         return ret;
3227 }
3228 EXPORT_SYMBOL(try_to_free_buffers);
3229
3230 void block_sync_page(struct page *page)
3231 {
3232         struct address_space *mapping;
3233
3234         smp_mb();
3235         mapping = page_mapping(page);
3236         if (mapping)
3237                 blk_run_backing_dev(mapping->backing_dev_info, page);
3238 }
3239
3240 /*
3241  * There are no bdflush tunables left.  But distributions are
3242  * still running obsolete flush daemons, so we terminate them here.
3243  *
3244  * Use of bdflush() is deprecated and will be removed in a future kernel.
3245  * The `pdflush' kernel threads fully replace bdflush daemons and this call.
3246  */
3247 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3248 {
3249         static int msg_count;
3250
3251         if (!capable(CAP_SYS_ADMIN))
3252                 return -EPERM;
3253
3254         if (msg_count < 5) {
3255                 msg_count++;
3256                 printk(KERN_INFO
3257                         "warning: process `%s' used the obsolete bdflush"
3258                         " system call\n", current->comm);
3259                 printk(KERN_INFO "Fix your initscripts?\n");
3260         }
3261
3262         if (func == 1)
3263                 do_exit(0);
3264         return 0;
3265 }
3266
3267 /*
3268  * Buffer-head allocation
3269  */
3270 static struct kmem_cache *bh_cachep;
3271
3272 /*
3273  * Once the number of bh's in the machine exceeds this level, we start
3274  * stripping them in writeback.
3275  */
3276 static int max_buffer_heads;
3277
3278 int buffer_heads_over_limit;
3279
3280 struct bh_accounting {
3281         int nr;                 /* Number of live bh's */
3282         int ratelimit;          /* Limit cacheline bouncing */
3283 };
3284
3285 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3286
3287 static void recalc_bh_state(void)
3288 {
3289         int i;
3290         int tot = 0;
3291
3292         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3293                 return;
3294         __get_cpu_var(bh_accounting).ratelimit = 0;
3295         for_each_online_cpu(i)
3296                 tot += per_cpu(bh_accounting, i).nr;
3297         buffer_heads_over_limit = (tot > max_buffer_heads);
3298 }
3299         
3300 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3301 {
3302         struct buffer_head *ret = kmem_cache_alloc(bh_cachep, gfp_flags);
3303         if (ret) {
3304                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3305                 get_cpu_var(bh_accounting).nr++;
3306                 recalc_bh_state();
3307                 put_cpu_var(bh_accounting);
3308         }
3309         return ret;
3310 }
3311 EXPORT_SYMBOL(alloc_buffer_head);
3312
3313 void free_buffer_head(struct buffer_head *bh)
3314 {
3315         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3316         kmem_cache_free(bh_cachep, bh);
3317         get_cpu_var(bh_accounting).nr--;
3318         recalc_bh_state();
3319         put_cpu_var(bh_accounting);
3320 }
3321 EXPORT_SYMBOL(free_buffer_head);
3322
3323 static void buffer_exit_cpu(int cpu)
3324 {
3325         int i;
3326         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3327
3328         for (i = 0; i < BH_LRU_SIZE; i++) {
3329                 brelse(b->bhs[i]);
3330                 b->bhs[i] = NULL;
3331         }
3332         get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3333         per_cpu(bh_accounting, cpu).nr = 0;
3334         put_cpu_var(bh_accounting);
3335 }
3336
3337 static int buffer_cpu_notify(struct notifier_block *self,
3338                               unsigned long action, void *hcpu)
3339 {
3340         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3341                 buffer_exit_cpu((unsigned long)hcpu);
3342         return NOTIFY_OK;
3343 }
3344
3345 /**
3346  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3347  * @bh: struct buffer_head
3348  *
3349  * Return true if the buffer is up-to-date and false,
3350  * with the buffer locked, if not.
3351  */
3352 int bh_uptodate_or_lock(struct buffer_head *bh)
3353 {
3354         if (!buffer_uptodate(bh)) {
3355                 lock_buffer(bh);
3356                 if (!buffer_uptodate(bh))
3357                         return 0;
3358                 unlock_buffer(bh);
3359         }
3360         return 1;
3361 }
3362 EXPORT_SYMBOL(bh_uptodate_or_lock);
3363
3364 /**
3365  * bh_submit_read - Submit a locked buffer for reading
3366  * @bh: struct buffer_head
3367  *
3368  * Returns zero on success and -EIO on error.
3369  */
3370 int bh_submit_read(struct buffer_head *bh)
3371 {
3372         BUG_ON(!buffer_locked(bh));
3373
3374         if (buffer_uptodate(bh)) {
3375                 unlock_buffer(bh);
3376                 return 0;
3377         }
3378
3379         get_bh(bh);
3380         bh->b_end_io = end_buffer_read_sync;
3381         submit_bh(READ, bh);
3382         wait_on_buffer(bh);
3383         if (buffer_uptodate(bh))
3384                 return 0;
3385         return -EIO;
3386 }
3387 EXPORT_SYMBOL(bh_submit_read);
3388
3389 static void
3390 init_buffer_head(void *data)
3391 {
3392         struct buffer_head *bh = data;
3393
3394         memset(bh, 0, sizeof(*bh));
3395         INIT_LIST_HEAD(&bh->b_assoc_buffers);
3396 }
3397
3398 void __init buffer_init(void)
3399 {
3400         int nrpages;
3401
3402         bh_cachep = kmem_cache_create("buffer_head",
3403                         sizeof(struct buffer_head), 0,
3404                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3405                                 SLAB_MEM_SPREAD),
3406                                 init_buffer_head);
3407
3408         /*
3409          * Limit the bh occupancy to 10% of ZONE_NORMAL
3410          */
3411         nrpages = (nr_free_buffer_pages() * 10) / 100;
3412         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3413         hotcpu_notifier(buffer_cpu_notify, 0);
3414 }
3415
3416 EXPORT_SYMBOL(__bforget);
3417 EXPORT_SYMBOL(__brelse);
3418 EXPORT_SYMBOL(__wait_on_buffer);
3419 EXPORT_SYMBOL(block_commit_write);
3420 EXPORT_SYMBOL(block_prepare_write);
3421 EXPORT_SYMBOL(block_page_mkwrite);
3422 EXPORT_SYMBOL(block_read_full_page);
3423 EXPORT_SYMBOL(block_sync_page);
3424 EXPORT_SYMBOL(block_truncate_page);
3425 EXPORT_SYMBOL(block_write_full_page);
3426 EXPORT_SYMBOL(cont_write_begin);
3427 EXPORT_SYMBOL(end_buffer_read_sync);
3428 EXPORT_SYMBOL(end_buffer_write_sync);
3429 EXPORT_SYMBOL(file_fsync);
3430 EXPORT_SYMBOL(fsync_bdev);
3431 EXPORT_SYMBOL(generic_block_bmap);
3432 EXPORT_SYMBOL(generic_cont_expand_simple);
3433 EXPORT_SYMBOL(init_buffer);
3434 EXPORT_SYMBOL(invalidate_bdev);
3435 EXPORT_SYMBOL(ll_rw_block);
3436 EXPORT_SYMBOL(mark_buffer_dirty);
3437 EXPORT_SYMBOL(submit_bh);
3438 EXPORT_SYMBOL(sync_dirty_buffer);
3439 EXPORT_SYMBOL(unlock_buffer);