ddfdd2c80bf9ba22c20901e5a74f9cd449d80ccd
[linux-2.6.git] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/capability.h>
28 #include <linux/blkdev.h>
29 #include <linux/file.h>
30 #include <linux/quotaops.h>
31 #include <linux/highmem.h>
32 #include <linux/module.h>
33 #include <linux/writeback.h>
34 #include <linux/hash.h>
35 #include <linux/suspend.h>
36 #include <linux/buffer_head.h>
37 #include <linux/task_io_accounting_ops.h>
38 #include <linux/bio.h>
39 #include <linux/notifier.h>
40 #include <linux/cpu.h>
41 #include <linux/bitops.h>
42 #include <linux/mpage.h>
43 #include <linux/bit_spinlock.h>
44
45 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
46
47 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
48
49 inline void
50 init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
51 {
52         bh->b_end_io = handler;
53         bh->b_private = private;
54 }
55
56 static int sync_buffer(void *word)
57 {
58         struct block_device *bd;
59         struct buffer_head *bh
60                 = container_of(word, struct buffer_head, b_state);
61
62         smp_mb();
63         bd = bh->b_bdev;
64         if (bd)
65                 blk_run_address_space(bd->bd_inode->i_mapping);
66         io_schedule();
67         return 0;
68 }
69
70 void __lock_buffer(struct buffer_head *bh)
71 {
72         wait_on_bit_lock(&bh->b_state, BH_Lock, sync_buffer,
73                                                         TASK_UNINTERRUPTIBLE);
74 }
75 EXPORT_SYMBOL(__lock_buffer);
76
77 void unlock_buffer(struct buffer_head *bh)
78 {
79         smp_mb__before_clear_bit();
80         clear_buffer_locked(bh);
81         smp_mb__after_clear_bit();
82         wake_up_bit(&bh->b_state, BH_Lock);
83 }
84
85 /*
86  * Block until a buffer comes unlocked.  This doesn't stop it
87  * from becoming locked again - you have to lock it yourself
88  * if you want to preserve its state.
89  */
90 void __wait_on_buffer(struct buffer_head * bh)
91 {
92         wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
93 }
94
95 static void
96 __clear_page_buffers(struct page *page)
97 {
98         ClearPagePrivate(page);
99         set_page_private(page, 0);
100         page_cache_release(page);
101 }
102
103 static void buffer_io_error(struct buffer_head *bh)
104 {
105         char b[BDEVNAME_SIZE];
106
107         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
108                         bdevname(bh->b_bdev, b),
109                         (unsigned long long)bh->b_blocknr);
110 }
111
112 /*
113  * End-of-IO handler helper function which does not touch the bh after
114  * unlocking it.
115  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
116  * a race there is benign: unlock_buffer() only use the bh's address for
117  * hashing after unlocking the buffer, so it doesn't actually touch the bh
118  * itself.
119  */
120 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
121 {
122         if (uptodate) {
123                 set_buffer_uptodate(bh);
124         } else {
125                 /* This happens, due to failed READA attempts. */
126                 clear_buffer_uptodate(bh);
127         }
128         unlock_buffer(bh);
129 }
130
131 /*
132  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
133  * unlock the buffer. This is what ll_rw_block uses too.
134  */
135 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
136 {
137         __end_buffer_read_notouch(bh, uptodate);
138         put_bh(bh);
139 }
140
141 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
142 {
143         char b[BDEVNAME_SIZE];
144
145         if (uptodate) {
146                 set_buffer_uptodate(bh);
147         } else {
148                 if (!buffer_eopnotsupp(bh) && printk_ratelimit()) {
149                         buffer_io_error(bh);
150                         printk(KERN_WARNING "lost page write due to "
151                                         "I/O error on %s\n",
152                                        bdevname(bh->b_bdev, b));
153                 }
154                 set_buffer_write_io_error(bh);
155                 clear_buffer_uptodate(bh);
156         }
157         unlock_buffer(bh);
158         put_bh(bh);
159 }
160
161 /*
162  * Write out and wait upon all the dirty data associated with a block
163  * device via its mapping.  Does not take the superblock lock.
164  */
165 int sync_blockdev(struct block_device *bdev)
166 {
167         int ret = 0;
168
169         if (bdev)
170                 ret = filemap_write_and_wait(bdev->bd_inode->i_mapping);
171         return ret;
172 }
173 EXPORT_SYMBOL(sync_blockdev);
174
175 /*
176  * Write out and wait upon all dirty data associated with this
177  * device.   Filesystem data as well as the underlying block
178  * device.  Takes the superblock lock.
179  */
180 int fsync_bdev(struct block_device *bdev)
181 {
182         struct super_block *sb = get_super(bdev);
183         if (sb) {
184                 int res = fsync_super(sb);
185                 drop_super(sb);
186                 return res;
187         }
188         return sync_blockdev(bdev);
189 }
190
191 /**
192  * freeze_bdev  --  lock a filesystem and force it into a consistent state
193  * @bdev:       blockdevice to lock
194  *
195  * This takes the block device bd_mount_sem to make sure no new mounts
196  * happen on bdev until thaw_bdev() is called.
197  * If a superblock is found on this device, we take the s_umount semaphore
198  * on it to make sure nobody unmounts until the snapshot creation is done.
199  */
200 struct super_block *freeze_bdev(struct block_device *bdev)
201 {
202         struct super_block *sb;
203
204         down(&bdev->bd_mount_sem);
205         sb = get_super(bdev);
206         if (sb && !(sb->s_flags & MS_RDONLY)) {
207                 sb->s_frozen = SB_FREEZE_WRITE;
208                 smp_wmb();
209
210                 __fsync_super(sb);
211
212                 sb->s_frozen = SB_FREEZE_TRANS;
213                 smp_wmb();
214
215                 sync_blockdev(sb->s_bdev);
216
217                 if (sb->s_op->write_super_lockfs)
218                         sb->s_op->write_super_lockfs(sb);
219         }
220
221         sync_blockdev(bdev);
222         return sb;      /* thaw_bdev releases s->s_umount and bd_mount_sem */
223 }
224 EXPORT_SYMBOL(freeze_bdev);
225
226 /**
227  * thaw_bdev  -- unlock filesystem
228  * @bdev:       blockdevice to unlock
229  * @sb:         associated superblock
230  *
231  * Unlocks the filesystem and marks it writeable again after freeze_bdev().
232  */
233 void thaw_bdev(struct block_device *bdev, struct super_block *sb)
234 {
235         if (sb) {
236                 BUG_ON(sb->s_bdev != bdev);
237
238                 if (sb->s_op->unlockfs)
239                         sb->s_op->unlockfs(sb);
240                 sb->s_frozen = SB_UNFROZEN;
241                 smp_wmb();
242                 wake_up(&sb->s_wait_unfrozen);
243                 drop_super(sb);
244         }
245
246         up(&bdev->bd_mount_sem);
247 }
248 EXPORT_SYMBOL(thaw_bdev);
249
250 /*
251  * Various filesystems appear to want __find_get_block to be non-blocking.
252  * But it's the page lock which protects the buffers.  To get around this,
253  * we get exclusion from try_to_free_buffers with the blockdev mapping's
254  * private_lock.
255  *
256  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
257  * may be quite high.  This code could TryLock the page, and if that
258  * succeeds, there is no need to take private_lock. (But if
259  * private_lock is contended then so is mapping->tree_lock).
260  */
261 static struct buffer_head *
262 __find_get_block_slow(struct block_device *bdev, sector_t block)
263 {
264         struct inode *bd_inode = bdev->bd_inode;
265         struct address_space *bd_mapping = bd_inode->i_mapping;
266         struct buffer_head *ret = NULL;
267         pgoff_t index;
268         struct buffer_head *bh;
269         struct buffer_head *head;
270         struct page *page;
271         int all_mapped = 1;
272
273         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
274         page = find_get_page(bd_mapping, index);
275         if (!page)
276                 goto out;
277
278         spin_lock(&bd_mapping->private_lock);
279         if (!page_has_buffers(page))
280                 goto out_unlock;
281         head = page_buffers(page);
282         bh = head;
283         do {
284                 if (bh->b_blocknr == block) {
285                         ret = bh;
286                         get_bh(bh);
287                         goto out_unlock;
288                 }
289                 if (!buffer_mapped(bh))
290                         all_mapped = 0;
291                 bh = bh->b_this_page;
292         } while (bh != head);
293
294         /* we might be here because some of the buffers on this page are
295          * not mapped.  This is due to various races between
296          * file io on the block device and getblk.  It gets dealt with
297          * elsewhere, don't buffer_error if we had some unmapped buffers
298          */
299         if (all_mapped) {
300                 printk("__find_get_block_slow() failed. "
301                         "block=%llu, b_blocknr=%llu\n",
302                         (unsigned long long)block,
303                         (unsigned long long)bh->b_blocknr);
304                 printk("b_state=0x%08lx, b_size=%zu\n",
305                         bh->b_state, bh->b_size);
306                 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
307         }
308 out_unlock:
309         spin_unlock(&bd_mapping->private_lock);
310         page_cache_release(page);
311 out:
312         return ret;
313 }
314
315 /* If invalidate_buffers() will trash dirty buffers, it means some kind
316    of fs corruption is going on. Trashing dirty data always imply losing
317    information that was supposed to be just stored on the physical layer
318    by the user.
319
320    Thus invalidate_buffers in general usage is not allwowed to trash
321    dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
322    be preserved.  These buffers are simply skipped.
323   
324    We also skip buffers which are still in use.  For example this can
325    happen if a userspace program is reading the block device.
326
327    NOTE: In the case where the user removed a removable-media-disk even if
328    there's still dirty data not synced on disk (due a bug in the device driver
329    or due an error of the user), by not destroying the dirty buffers we could
330    generate corruption also on the next media inserted, thus a parameter is
331    necessary to handle this case in the most safe way possible (trying
332    to not corrupt also the new disk inserted with the data belonging to
333    the old now corrupted disk). Also for the ramdisk the natural thing
334    to do in order to release the ramdisk memory is to destroy dirty buffers.
335
336    These are two special cases. Normal usage imply the device driver
337    to issue a sync on the device (without waiting I/O completion) and
338    then an invalidate_buffers call that doesn't trash dirty buffers.
339
340    For handling cache coherency with the blkdev pagecache the 'update' case
341    is been introduced. It is needed to re-read from disk any pinned
342    buffer. NOTE: re-reading from disk is destructive so we can do it only
343    when we assume nobody is changing the buffercache under our I/O and when
344    we think the disk contains more recent information than the buffercache.
345    The update == 1 pass marks the buffers we need to update, the update == 2
346    pass does the actual I/O. */
347 void invalidate_bdev(struct block_device *bdev)
348 {
349         struct address_space *mapping = bdev->bd_inode->i_mapping;
350
351         if (mapping->nrpages == 0)
352                 return;
353
354         invalidate_bh_lrus();
355         invalidate_mapping_pages(mapping, 0, -1);
356 }
357
358 /*
359  * Kick pdflush then try to free up some ZONE_NORMAL memory.
360  */
361 static void free_more_memory(void)
362 {
363         struct zone **zones;
364         pg_data_t *pgdat;
365
366         wakeup_pdflush(1024);
367         yield();
368
369         for_each_online_pgdat(pgdat) {
370                 zones = pgdat->node_zonelists[gfp_zone(GFP_NOFS)].zones;
371                 if (*zones)
372                         try_to_free_pages(zones, 0, GFP_NOFS);
373         }
374 }
375
376 /*
377  * I/O completion handler for block_read_full_page() - pages
378  * which come unlocked at the end of I/O.
379  */
380 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
381 {
382         unsigned long flags;
383         struct buffer_head *first;
384         struct buffer_head *tmp;
385         struct page *page;
386         int page_uptodate = 1;
387
388         BUG_ON(!buffer_async_read(bh));
389
390         page = bh->b_page;
391         if (uptodate) {
392                 set_buffer_uptodate(bh);
393         } else {
394                 clear_buffer_uptodate(bh);
395                 if (printk_ratelimit())
396                         buffer_io_error(bh);
397                 SetPageError(page);
398         }
399
400         /*
401          * Be _very_ careful from here on. Bad things can happen if
402          * two buffer heads end IO at almost the same time and both
403          * decide that the page is now completely done.
404          */
405         first = page_buffers(page);
406         local_irq_save(flags);
407         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
408         clear_buffer_async_read(bh);
409         unlock_buffer(bh);
410         tmp = bh;
411         do {
412                 if (!buffer_uptodate(tmp))
413                         page_uptodate = 0;
414                 if (buffer_async_read(tmp)) {
415                         BUG_ON(!buffer_locked(tmp));
416                         goto still_busy;
417                 }
418                 tmp = tmp->b_this_page;
419         } while (tmp != bh);
420         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
421         local_irq_restore(flags);
422
423         /*
424          * If none of the buffers had errors and they are all
425          * uptodate then we can set the page uptodate.
426          */
427         if (page_uptodate && !PageError(page))
428                 SetPageUptodate(page);
429         unlock_page(page);
430         return;
431
432 still_busy:
433         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
434         local_irq_restore(flags);
435         return;
436 }
437
438 /*
439  * Completion handler for block_write_full_page() - pages which are unlocked
440  * during I/O, and which have PageWriteback cleared upon I/O completion.
441  */
442 static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
443 {
444         char b[BDEVNAME_SIZE];
445         unsigned long flags;
446         struct buffer_head *first;
447         struct buffer_head *tmp;
448         struct page *page;
449
450         BUG_ON(!buffer_async_write(bh));
451
452         page = bh->b_page;
453         if (uptodate) {
454                 set_buffer_uptodate(bh);
455         } else {
456                 if (printk_ratelimit()) {
457                         buffer_io_error(bh);
458                         printk(KERN_WARNING "lost page write due to "
459                                         "I/O error on %s\n",
460                                bdevname(bh->b_bdev, b));
461                 }
462                 set_bit(AS_EIO, &page->mapping->flags);
463                 set_buffer_write_io_error(bh);
464                 clear_buffer_uptodate(bh);
465                 SetPageError(page);
466         }
467
468         first = page_buffers(page);
469         local_irq_save(flags);
470         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
471
472         clear_buffer_async_write(bh);
473         unlock_buffer(bh);
474         tmp = bh->b_this_page;
475         while (tmp != bh) {
476                 if (buffer_async_write(tmp)) {
477                         BUG_ON(!buffer_locked(tmp));
478                         goto still_busy;
479                 }
480                 tmp = tmp->b_this_page;
481         }
482         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
483         local_irq_restore(flags);
484         end_page_writeback(page);
485         return;
486
487 still_busy:
488         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
489         local_irq_restore(flags);
490         return;
491 }
492
493 /*
494  * If a page's buffers are under async readin (end_buffer_async_read
495  * completion) then there is a possibility that another thread of
496  * control could lock one of the buffers after it has completed
497  * but while some of the other buffers have not completed.  This
498  * locked buffer would confuse end_buffer_async_read() into not unlocking
499  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
500  * that this buffer is not under async I/O.
501  *
502  * The page comes unlocked when it has no locked buffer_async buffers
503  * left.
504  *
505  * PageLocked prevents anyone starting new async I/O reads any of
506  * the buffers.
507  *
508  * PageWriteback is used to prevent simultaneous writeout of the same
509  * page.
510  *
511  * PageLocked prevents anyone from starting writeback of a page which is
512  * under read I/O (PageWriteback is only ever set against a locked page).
513  */
514 static void mark_buffer_async_read(struct buffer_head *bh)
515 {
516         bh->b_end_io = end_buffer_async_read;
517         set_buffer_async_read(bh);
518 }
519
520 void mark_buffer_async_write(struct buffer_head *bh)
521 {
522         bh->b_end_io = end_buffer_async_write;
523         set_buffer_async_write(bh);
524 }
525 EXPORT_SYMBOL(mark_buffer_async_write);
526
527
528 /*
529  * fs/buffer.c contains helper functions for buffer-backed address space's
530  * fsync functions.  A common requirement for buffer-based filesystems is
531  * that certain data from the backing blockdev needs to be written out for
532  * a successful fsync().  For example, ext2 indirect blocks need to be
533  * written back and waited upon before fsync() returns.
534  *
535  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
536  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
537  * management of a list of dependent buffers at ->i_mapping->private_list.
538  *
539  * Locking is a little subtle: try_to_free_buffers() will remove buffers
540  * from their controlling inode's queue when they are being freed.  But
541  * try_to_free_buffers() will be operating against the *blockdev* mapping
542  * at the time, not against the S_ISREG file which depends on those buffers.
543  * So the locking for private_list is via the private_lock in the address_space
544  * which backs the buffers.  Which is different from the address_space 
545  * against which the buffers are listed.  So for a particular address_space,
546  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
547  * mapping->private_list will always be protected by the backing blockdev's
548  * ->private_lock.
549  *
550  * Which introduces a requirement: all buffers on an address_space's
551  * ->private_list must be from the same address_space: the blockdev's.
552  *
553  * address_spaces which do not place buffers at ->private_list via these
554  * utility functions are free to use private_lock and private_list for
555  * whatever they want.  The only requirement is that list_empty(private_list)
556  * be true at clear_inode() time.
557  *
558  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
559  * filesystems should do that.  invalidate_inode_buffers() should just go
560  * BUG_ON(!list_empty).
561  *
562  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
563  * take an address_space, not an inode.  And it should be called
564  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
565  * queued up.
566  *
567  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
568  * list if it is already on a list.  Because if the buffer is on a list,
569  * it *must* already be on the right one.  If not, the filesystem is being
570  * silly.  This will save a ton of locking.  But first we have to ensure
571  * that buffers are taken *off* the old inode's list when they are freed
572  * (presumably in truncate).  That requires careful auditing of all
573  * filesystems (do it inside bforget()).  It could also be done by bringing
574  * b_inode back.
575  */
576
577 /*
578  * The buffer's backing address_space's private_lock must be held
579  */
580 static inline void __remove_assoc_queue(struct buffer_head *bh)
581 {
582         list_del_init(&bh->b_assoc_buffers);
583         WARN_ON(!bh->b_assoc_map);
584         if (buffer_write_io_error(bh))
585                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
586         bh->b_assoc_map = NULL;
587 }
588
589 int inode_has_buffers(struct inode *inode)
590 {
591         return !list_empty(&inode->i_data.private_list);
592 }
593
594 /*
595  * osync is designed to support O_SYNC io.  It waits synchronously for
596  * all already-submitted IO to complete, but does not queue any new
597  * writes to the disk.
598  *
599  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
600  * you dirty the buffers, and then use osync_inode_buffers to wait for
601  * completion.  Any other dirty buffers which are not yet queued for
602  * write will not be flushed to disk by the osync.
603  */
604 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
605 {
606         struct buffer_head *bh;
607         struct list_head *p;
608         int err = 0;
609
610         spin_lock(lock);
611 repeat:
612         list_for_each_prev(p, list) {
613                 bh = BH_ENTRY(p);
614                 if (buffer_locked(bh)) {
615                         get_bh(bh);
616                         spin_unlock(lock);
617                         wait_on_buffer(bh);
618                         if (!buffer_uptodate(bh))
619                                 err = -EIO;
620                         brelse(bh);
621                         spin_lock(lock);
622                         goto repeat;
623                 }
624         }
625         spin_unlock(lock);
626         return err;
627 }
628
629 /**
630  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
631  * @mapping: the mapping which wants those buffers written
632  *
633  * Starts I/O against the buffers at mapping->private_list, and waits upon
634  * that I/O.
635  *
636  * Basically, this is a convenience function for fsync().
637  * @mapping is a file or directory which needs those buffers to be written for
638  * a successful fsync().
639  */
640 int sync_mapping_buffers(struct address_space *mapping)
641 {
642         struct address_space *buffer_mapping = mapping->assoc_mapping;
643
644         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
645                 return 0;
646
647         return fsync_buffers_list(&buffer_mapping->private_lock,
648                                         &mapping->private_list);
649 }
650 EXPORT_SYMBOL(sync_mapping_buffers);
651
652 /*
653  * Called when we've recently written block `bblock', and it is known that
654  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
655  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
656  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
657  */
658 void write_boundary_block(struct block_device *bdev,
659                         sector_t bblock, unsigned blocksize)
660 {
661         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
662         if (bh) {
663                 if (buffer_dirty(bh))
664                         ll_rw_block(WRITE, 1, &bh);
665                 put_bh(bh);
666         }
667 }
668
669 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
670 {
671         struct address_space *mapping = inode->i_mapping;
672         struct address_space *buffer_mapping = bh->b_page->mapping;
673
674         mark_buffer_dirty(bh);
675         if (!mapping->assoc_mapping) {
676                 mapping->assoc_mapping = buffer_mapping;
677         } else {
678                 BUG_ON(mapping->assoc_mapping != buffer_mapping);
679         }
680         if (!bh->b_assoc_map) {
681                 spin_lock(&buffer_mapping->private_lock);
682                 list_move_tail(&bh->b_assoc_buffers,
683                                 &mapping->private_list);
684                 bh->b_assoc_map = mapping;
685                 spin_unlock(&buffer_mapping->private_lock);
686         }
687 }
688 EXPORT_SYMBOL(mark_buffer_dirty_inode);
689
690 /*
691  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
692  * dirty.
693  *
694  * If warn is true, then emit a warning if the page is not uptodate and has
695  * not been truncated.
696  */
697 static int __set_page_dirty(struct page *page,
698                 struct address_space *mapping, int warn)
699 {
700         if (unlikely(!mapping))
701                 return !TestSetPageDirty(page);
702
703         if (TestSetPageDirty(page))
704                 return 0;
705
706         write_lock_irq(&mapping->tree_lock);
707         if (page->mapping) {    /* Race with truncate? */
708                 WARN_ON_ONCE(warn && !PageUptodate(page));
709
710                 if (mapping_cap_account_dirty(mapping)) {
711                         __inc_zone_page_state(page, NR_FILE_DIRTY);
712                         __inc_bdi_stat(mapping->backing_dev_info,
713                                         BDI_RECLAIMABLE);
714                         task_io_account_write(PAGE_CACHE_SIZE);
715                 }
716                 radix_tree_tag_set(&mapping->page_tree,
717                                 page_index(page), PAGECACHE_TAG_DIRTY);
718         }
719         write_unlock_irq(&mapping->tree_lock);
720         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
721
722         return 1;
723 }
724
725 /*
726  * Add a page to the dirty page list.
727  *
728  * It is a sad fact of life that this function is called from several places
729  * deeply under spinlocking.  It may not sleep.
730  *
731  * If the page has buffers, the uptodate buffers are set dirty, to preserve
732  * dirty-state coherency between the page and the buffers.  It the page does
733  * not have buffers then when they are later attached they will all be set
734  * dirty.
735  *
736  * The buffers are dirtied before the page is dirtied.  There's a small race
737  * window in which a writepage caller may see the page cleanness but not the
738  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
739  * before the buffers, a concurrent writepage caller could clear the page dirty
740  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
741  * page on the dirty page list.
742  *
743  * We use private_lock to lock against try_to_free_buffers while using the
744  * page's buffer list.  Also use this to protect against clean buffers being
745  * added to the page after it was set dirty.
746  *
747  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
748  * address_space though.
749  */
750 int __set_page_dirty_buffers(struct page *page)
751 {
752         struct address_space *mapping = page_mapping(page);
753
754         if (unlikely(!mapping))
755                 return !TestSetPageDirty(page);
756
757         spin_lock(&mapping->private_lock);
758         if (page_has_buffers(page)) {
759                 struct buffer_head *head = page_buffers(page);
760                 struct buffer_head *bh = head;
761
762                 do {
763                         set_buffer_dirty(bh);
764                         bh = bh->b_this_page;
765                 } while (bh != head);
766         }
767         spin_unlock(&mapping->private_lock);
768
769         return __set_page_dirty(page, mapping, 1);
770 }
771 EXPORT_SYMBOL(__set_page_dirty_buffers);
772
773 /*
774  * Write out and wait upon a list of buffers.
775  *
776  * We have conflicting pressures: we want to make sure that all
777  * initially dirty buffers get waited on, but that any subsequently
778  * dirtied buffers don't.  After all, we don't want fsync to last
779  * forever if somebody is actively writing to the file.
780  *
781  * Do this in two main stages: first we copy dirty buffers to a
782  * temporary inode list, queueing the writes as we go.  Then we clean
783  * up, waiting for those writes to complete.
784  * 
785  * During this second stage, any subsequent updates to the file may end
786  * up refiling the buffer on the original inode's dirty list again, so
787  * there is a chance we will end up with a buffer queued for write but
788  * not yet completed on that list.  So, as a final cleanup we go through
789  * the osync code to catch these locked, dirty buffers without requeuing
790  * any newly dirty buffers for write.
791  */
792 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
793 {
794         struct buffer_head *bh;
795         struct list_head tmp;
796         struct address_space *mapping;
797         int err = 0, err2;
798
799         INIT_LIST_HEAD(&tmp);
800
801         spin_lock(lock);
802         while (!list_empty(list)) {
803                 bh = BH_ENTRY(list->next);
804                 mapping = bh->b_assoc_map;
805                 __remove_assoc_queue(bh);
806                 /* Avoid race with mark_buffer_dirty_inode() which does
807                  * a lockless check and we rely on seeing the dirty bit */
808                 smp_mb();
809                 if (buffer_dirty(bh) || buffer_locked(bh)) {
810                         list_add(&bh->b_assoc_buffers, &tmp);
811                         bh->b_assoc_map = mapping;
812                         if (buffer_dirty(bh)) {
813                                 get_bh(bh);
814                                 spin_unlock(lock);
815                                 /*
816                                  * Ensure any pending I/O completes so that
817                                  * ll_rw_block() actually writes the current
818                                  * contents - it is a noop if I/O is still in
819                                  * flight on potentially older contents.
820                                  */
821                                 ll_rw_block(SWRITE, 1, &bh);
822                                 brelse(bh);
823                                 spin_lock(lock);
824                         }
825                 }
826         }
827
828         while (!list_empty(&tmp)) {
829                 bh = BH_ENTRY(tmp.prev);
830                 get_bh(bh);
831                 mapping = bh->b_assoc_map;
832                 __remove_assoc_queue(bh);
833                 /* Avoid race with mark_buffer_dirty_inode() which does
834                  * a lockless check and we rely on seeing the dirty bit */
835                 smp_mb();
836                 if (buffer_dirty(bh)) {
837                         list_add(&bh->b_assoc_buffers,
838                                  &mapping->private_list);
839                         bh->b_assoc_map = mapping;
840                 }
841                 spin_unlock(lock);
842                 wait_on_buffer(bh);
843                 if (!buffer_uptodate(bh))
844                         err = -EIO;
845                 brelse(bh);
846                 spin_lock(lock);
847         }
848         
849         spin_unlock(lock);
850         err2 = osync_buffers_list(lock, list);
851         if (err)
852                 return err;
853         else
854                 return err2;
855 }
856
857 /*
858  * Invalidate any and all dirty buffers on a given inode.  We are
859  * probably unmounting the fs, but that doesn't mean we have already
860  * done a sync().  Just drop the buffers from the inode list.
861  *
862  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
863  * assumes that all the buffers are against the blockdev.  Not true
864  * for reiserfs.
865  */
866 void invalidate_inode_buffers(struct inode *inode)
867 {
868         if (inode_has_buffers(inode)) {
869                 struct address_space *mapping = &inode->i_data;
870                 struct list_head *list = &mapping->private_list;
871                 struct address_space *buffer_mapping = mapping->assoc_mapping;
872
873                 spin_lock(&buffer_mapping->private_lock);
874                 while (!list_empty(list))
875                         __remove_assoc_queue(BH_ENTRY(list->next));
876                 spin_unlock(&buffer_mapping->private_lock);
877         }
878 }
879
880 /*
881  * Remove any clean buffers from the inode's buffer list.  This is called
882  * when we're trying to free the inode itself.  Those buffers can pin it.
883  *
884  * Returns true if all buffers were removed.
885  */
886 int remove_inode_buffers(struct inode *inode)
887 {
888         int ret = 1;
889
890         if (inode_has_buffers(inode)) {
891                 struct address_space *mapping = &inode->i_data;
892                 struct list_head *list = &mapping->private_list;
893                 struct address_space *buffer_mapping = mapping->assoc_mapping;
894
895                 spin_lock(&buffer_mapping->private_lock);
896                 while (!list_empty(list)) {
897                         struct buffer_head *bh = BH_ENTRY(list->next);
898                         if (buffer_dirty(bh)) {
899                                 ret = 0;
900                                 break;
901                         }
902                         __remove_assoc_queue(bh);
903                 }
904                 spin_unlock(&buffer_mapping->private_lock);
905         }
906         return ret;
907 }
908
909 /*
910  * Create the appropriate buffers when given a page for data area and
911  * the size of each buffer.. Use the bh->b_this_page linked list to
912  * follow the buffers created.  Return NULL if unable to create more
913  * buffers.
914  *
915  * The retry flag is used to differentiate async IO (paging, swapping)
916  * which may not fail from ordinary buffer allocations.
917  */
918 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
919                 int retry)
920 {
921         struct buffer_head *bh, *head;
922         long offset;
923
924 try_again:
925         head = NULL;
926         offset = PAGE_SIZE;
927         while ((offset -= size) >= 0) {
928                 bh = alloc_buffer_head(GFP_NOFS);
929                 if (!bh)
930                         goto no_grow;
931
932                 bh->b_bdev = NULL;
933                 bh->b_this_page = head;
934                 bh->b_blocknr = -1;
935                 head = bh;
936
937                 bh->b_state = 0;
938                 atomic_set(&bh->b_count, 0);
939                 bh->b_private = NULL;
940                 bh->b_size = size;
941
942                 /* Link the buffer to its page */
943                 set_bh_page(bh, page, offset);
944
945                 init_buffer(bh, NULL, NULL);
946         }
947         return head;
948 /*
949  * In case anything failed, we just free everything we got.
950  */
951 no_grow:
952         if (head) {
953                 do {
954                         bh = head;
955                         head = head->b_this_page;
956                         free_buffer_head(bh);
957                 } while (head);
958         }
959
960         /*
961          * Return failure for non-async IO requests.  Async IO requests
962          * are not allowed to fail, so we have to wait until buffer heads
963          * become available.  But we don't want tasks sleeping with 
964          * partially complete buffers, so all were released above.
965          */
966         if (!retry)
967                 return NULL;
968
969         /* We're _really_ low on memory. Now we just
970          * wait for old buffer heads to become free due to
971          * finishing IO.  Since this is an async request and
972          * the reserve list is empty, we're sure there are 
973          * async buffer heads in use.
974          */
975         free_more_memory();
976         goto try_again;
977 }
978 EXPORT_SYMBOL_GPL(alloc_page_buffers);
979
980 static inline void
981 link_dev_buffers(struct page *page, struct buffer_head *head)
982 {
983         struct buffer_head *bh, *tail;
984
985         bh = head;
986         do {
987                 tail = bh;
988                 bh = bh->b_this_page;
989         } while (bh);
990         tail->b_this_page = head;
991         attach_page_buffers(page, head);
992 }
993
994 /*
995  * Initialise the state of a blockdev page's buffers.
996  */ 
997 static void
998 init_page_buffers(struct page *page, struct block_device *bdev,
999                         sector_t block, int size)
1000 {
1001         struct buffer_head *head = page_buffers(page);
1002         struct buffer_head *bh = head;
1003         int uptodate = PageUptodate(page);
1004
1005         do {
1006                 if (!buffer_mapped(bh)) {
1007                         init_buffer(bh, NULL, NULL);
1008                         bh->b_bdev = bdev;
1009                         bh->b_blocknr = block;
1010                         if (uptodate)
1011                                 set_buffer_uptodate(bh);
1012                         set_buffer_mapped(bh);
1013                 }
1014                 block++;
1015                 bh = bh->b_this_page;
1016         } while (bh != head);
1017 }
1018
1019 /*
1020  * Create the page-cache page that contains the requested block.
1021  *
1022  * This is user purely for blockdev mappings.
1023  */
1024 static struct page *
1025 grow_dev_page(struct block_device *bdev, sector_t block,
1026                 pgoff_t index, int size)
1027 {
1028         struct inode *inode = bdev->bd_inode;
1029         struct page *page;
1030         struct buffer_head *bh;
1031
1032         page = find_or_create_page(inode->i_mapping, index,
1033                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
1034         if (!page)
1035                 return NULL;
1036
1037         BUG_ON(!PageLocked(page));
1038
1039         if (page_has_buffers(page)) {
1040                 bh = page_buffers(page);
1041                 if (bh->b_size == size) {
1042                         init_page_buffers(page, bdev, block, size);
1043                         return page;
1044                 }
1045                 if (!try_to_free_buffers(page))
1046                         goto failed;
1047         }
1048
1049         /*
1050          * Allocate some buffers for this page
1051          */
1052         bh = alloc_page_buffers(page, size, 0);
1053         if (!bh)
1054                 goto failed;
1055
1056         /*
1057          * Link the page to the buffers and initialise them.  Take the
1058          * lock to be atomic wrt __find_get_block(), which does not
1059          * run under the page lock.
1060          */
1061         spin_lock(&inode->i_mapping->private_lock);
1062         link_dev_buffers(page, bh);
1063         init_page_buffers(page, bdev, block, size);
1064         spin_unlock(&inode->i_mapping->private_lock);
1065         return page;
1066
1067 failed:
1068         BUG();
1069         unlock_page(page);
1070         page_cache_release(page);
1071         return NULL;
1072 }
1073
1074 /*
1075  * Create buffers for the specified block device block's page.  If
1076  * that page was dirty, the buffers are set dirty also.
1077  */
1078 static int
1079 grow_buffers(struct block_device *bdev, sector_t block, int size)
1080 {
1081         struct page *page;
1082         pgoff_t index;
1083         int sizebits;
1084
1085         sizebits = -1;
1086         do {
1087                 sizebits++;
1088         } while ((size << sizebits) < PAGE_SIZE);
1089
1090         index = block >> sizebits;
1091
1092         /*
1093          * Check for a block which wants to lie outside our maximum possible
1094          * pagecache index.  (this comparison is done using sector_t types).
1095          */
1096         if (unlikely(index != block >> sizebits)) {
1097                 char b[BDEVNAME_SIZE];
1098
1099                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1100                         "device %s\n",
1101                         __FUNCTION__, (unsigned long long)block,
1102                         bdevname(bdev, b));
1103                 return -EIO;
1104         }
1105         block = index << sizebits;
1106         /* Create a page with the proper size buffers.. */
1107         page = grow_dev_page(bdev, block, index, size);
1108         if (!page)
1109                 return 0;
1110         unlock_page(page);
1111         page_cache_release(page);
1112         return 1;
1113 }
1114
1115 static struct buffer_head *
1116 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1117 {
1118         /* Size must be multiple of hard sectorsize */
1119         if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1120                         (size < 512 || size > PAGE_SIZE))) {
1121                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1122                                         size);
1123                 printk(KERN_ERR "hardsect size: %d\n",
1124                                         bdev_hardsect_size(bdev));
1125
1126                 dump_stack();
1127                 return NULL;
1128         }
1129
1130         for (;;) {
1131                 struct buffer_head * bh;
1132                 int ret;
1133
1134                 bh = __find_get_block(bdev, block, size);
1135                 if (bh)
1136                         return bh;
1137
1138                 ret = grow_buffers(bdev, block, size);
1139                 if (ret < 0)
1140                         return NULL;
1141                 if (ret == 0)
1142                         free_more_memory();
1143         }
1144 }
1145
1146 /*
1147  * The relationship between dirty buffers and dirty pages:
1148  *
1149  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1150  * the page is tagged dirty in its radix tree.
1151  *
1152  * At all times, the dirtiness of the buffers represents the dirtiness of
1153  * subsections of the page.  If the page has buffers, the page dirty bit is
1154  * merely a hint about the true dirty state.
1155  *
1156  * When a page is set dirty in its entirety, all its buffers are marked dirty
1157  * (if the page has buffers).
1158  *
1159  * When a buffer is marked dirty, its page is dirtied, but the page's other
1160  * buffers are not.
1161  *
1162  * Also.  When blockdev buffers are explicitly read with bread(), they
1163  * individually become uptodate.  But their backing page remains not
1164  * uptodate - even if all of its buffers are uptodate.  A subsequent
1165  * block_read_full_page() against that page will discover all the uptodate
1166  * buffers, will set the page uptodate and will perform no I/O.
1167  */
1168
1169 /**
1170  * mark_buffer_dirty - mark a buffer_head as needing writeout
1171  * @bh: the buffer_head to mark dirty
1172  *
1173  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1174  * backing page dirty, then tag the page as dirty in its address_space's radix
1175  * tree and then attach the address_space's inode to its superblock's dirty
1176  * inode list.
1177  *
1178  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1179  * mapping->tree_lock and the global inode_lock.
1180  */
1181 void mark_buffer_dirty(struct buffer_head *bh)
1182 {
1183         WARN_ON_ONCE(!buffer_uptodate(bh));
1184         if (!buffer_dirty(bh) && !test_set_buffer_dirty(bh))
1185                 __set_page_dirty(bh->b_page, page_mapping(bh->b_page), 0);
1186 }
1187
1188 /*
1189  * Decrement a buffer_head's reference count.  If all buffers against a page
1190  * have zero reference count, are clean and unlocked, and if the page is clean
1191  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1192  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1193  * a page but it ends up not being freed, and buffers may later be reattached).
1194  */
1195 void __brelse(struct buffer_head * buf)
1196 {
1197         if (atomic_read(&buf->b_count)) {
1198                 put_bh(buf);
1199                 return;
1200         }
1201         printk(KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1202         WARN_ON(1);
1203 }
1204
1205 /*
1206  * bforget() is like brelse(), except it discards any
1207  * potentially dirty data.
1208  */
1209 void __bforget(struct buffer_head *bh)
1210 {
1211         clear_buffer_dirty(bh);
1212         if (bh->b_assoc_map) {
1213                 struct address_space *buffer_mapping = bh->b_page->mapping;
1214
1215                 spin_lock(&buffer_mapping->private_lock);
1216                 list_del_init(&bh->b_assoc_buffers);
1217                 bh->b_assoc_map = NULL;
1218                 spin_unlock(&buffer_mapping->private_lock);
1219         }
1220         __brelse(bh);
1221 }
1222
1223 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1224 {
1225         lock_buffer(bh);
1226         if (buffer_uptodate(bh)) {
1227                 unlock_buffer(bh);
1228                 return bh;
1229         } else {
1230                 get_bh(bh);
1231                 bh->b_end_io = end_buffer_read_sync;
1232                 submit_bh(READ, bh);
1233                 wait_on_buffer(bh);
1234                 if (buffer_uptodate(bh))
1235                         return bh;
1236         }
1237         brelse(bh);
1238         return NULL;
1239 }
1240
1241 /*
1242  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1243  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1244  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1245  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1246  * CPU's LRUs at the same time.
1247  *
1248  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1249  * sb_find_get_block().
1250  *
1251  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1252  * a local interrupt disable for that.
1253  */
1254
1255 #define BH_LRU_SIZE     8
1256
1257 struct bh_lru {
1258         struct buffer_head *bhs[BH_LRU_SIZE];
1259 };
1260
1261 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1262
1263 #ifdef CONFIG_SMP
1264 #define bh_lru_lock()   local_irq_disable()
1265 #define bh_lru_unlock() local_irq_enable()
1266 #else
1267 #define bh_lru_lock()   preempt_disable()
1268 #define bh_lru_unlock() preempt_enable()
1269 #endif
1270
1271 static inline void check_irqs_on(void)
1272 {
1273 #ifdef irqs_disabled
1274         BUG_ON(irqs_disabled());
1275 #endif
1276 }
1277
1278 /*
1279  * The LRU management algorithm is dopey-but-simple.  Sorry.
1280  */
1281 static void bh_lru_install(struct buffer_head *bh)
1282 {
1283         struct buffer_head *evictee = NULL;
1284         struct bh_lru *lru;
1285
1286         check_irqs_on();
1287         bh_lru_lock();
1288         lru = &__get_cpu_var(bh_lrus);
1289         if (lru->bhs[0] != bh) {
1290                 struct buffer_head *bhs[BH_LRU_SIZE];
1291                 int in;
1292                 int out = 0;
1293
1294                 get_bh(bh);
1295                 bhs[out++] = bh;
1296                 for (in = 0; in < BH_LRU_SIZE; in++) {
1297                         struct buffer_head *bh2 = lru->bhs[in];
1298
1299                         if (bh2 == bh) {
1300                                 __brelse(bh2);
1301                         } else {
1302                                 if (out >= BH_LRU_SIZE) {
1303                                         BUG_ON(evictee != NULL);
1304                                         evictee = bh2;
1305                                 } else {
1306                                         bhs[out++] = bh2;
1307                                 }
1308                         }
1309                 }
1310                 while (out < BH_LRU_SIZE)
1311                         bhs[out++] = NULL;
1312                 memcpy(lru->bhs, bhs, sizeof(bhs));
1313         }
1314         bh_lru_unlock();
1315
1316         if (evictee)
1317                 __brelse(evictee);
1318 }
1319
1320 /*
1321  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1322  */
1323 static struct buffer_head *
1324 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1325 {
1326         struct buffer_head *ret = NULL;
1327         struct bh_lru *lru;
1328         unsigned int i;
1329
1330         check_irqs_on();
1331         bh_lru_lock();
1332         lru = &__get_cpu_var(bh_lrus);
1333         for (i = 0; i < BH_LRU_SIZE; i++) {
1334                 struct buffer_head *bh = lru->bhs[i];
1335
1336                 if (bh && bh->b_bdev == bdev &&
1337                                 bh->b_blocknr == block && bh->b_size == size) {
1338                         if (i) {
1339                                 while (i) {
1340                                         lru->bhs[i] = lru->bhs[i - 1];
1341                                         i--;
1342                                 }
1343                                 lru->bhs[0] = bh;
1344                         }
1345                         get_bh(bh);
1346                         ret = bh;
1347                         break;
1348                 }
1349         }
1350         bh_lru_unlock();
1351         return ret;
1352 }
1353
1354 /*
1355  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1356  * it in the LRU and mark it as accessed.  If it is not present then return
1357  * NULL
1358  */
1359 struct buffer_head *
1360 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1361 {
1362         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1363
1364         if (bh == NULL) {
1365                 bh = __find_get_block_slow(bdev, block);
1366                 if (bh)
1367                         bh_lru_install(bh);
1368         }
1369         if (bh)
1370                 touch_buffer(bh);
1371         return bh;
1372 }
1373 EXPORT_SYMBOL(__find_get_block);
1374
1375 /*
1376  * __getblk will locate (and, if necessary, create) the buffer_head
1377  * which corresponds to the passed block_device, block and size. The
1378  * returned buffer has its reference count incremented.
1379  *
1380  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1381  * illegal block number, __getblk() will happily return a buffer_head
1382  * which represents the non-existent block.  Very weird.
1383  *
1384  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1385  * attempt is failing.  FIXME, perhaps?
1386  */
1387 struct buffer_head *
1388 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1389 {
1390         struct buffer_head *bh = __find_get_block(bdev, block, size);
1391
1392         might_sleep();
1393         if (bh == NULL)
1394                 bh = __getblk_slow(bdev, block, size);
1395         return bh;
1396 }
1397 EXPORT_SYMBOL(__getblk);
1398
1399 /*
1400  * Do async read-ahead on a buffer..
1401  */
1402 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1403 {
1404         struct buffer_head *bh = __getblk(bdev, block, size);
1405         if (likely(bh)) {
1406                 ll_rw_block(READA, 1, &bh);
1407                 brelse(bh);
1408         }
1409 }
1410 EXPORT_SYMBOL(__breadahead);
1411
1412 /**
1413  *  __bread() - reads a specified block and returns the bh
1414  *  @bdev: the block_device to read from
1415  *  @block: number of block
1416  *  @size: size (in bytes) to read
1417  * 
1418  *  Reads a specified block, and returns buffer head that contains it.
1419  *  It returns NULL if the block was unreadable.
1420  */
1421 struct buffer_head *
1422 __bread(struct block_device *bdev, sector_t block, unsigned size)
1423 {
1424         struct buffer_head *bh = __getblk(bdev, block, size);
1425
1426         if (likely(bh) && !buffer_uptodate(bh))
1427                 bh = __bread_slow(bh);
1428         return bh;
1429 }
1430 EXPORT_SYMBOL(__bread);
1431
1432 /*
1433  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1434  * This doesn't race because it runs in each cpu either in irq
1435  * or with preempt disabled.
1436  */
1437 static void invalidate_bh_lru(void *arg)
1438 {
1439         struct bh_lru *b = &get_cpu_var(bh_lrus);
1440         int i;
1441
1442         for (i = 0; i < BH_LRU_SIZE; i++) {
1443                 brelse(b->bhs[i]);
1444                 b->bhs[i] = NULL;
1445         }
1446         put_cpu_var(bh_lrus);
1447 }
1448         
1449 void invalidate_bh_lrus(void)
1450 {
1451         on_each_cpu(invalidate_bh_lru, NULL, 1, 1);
1452 }
1453 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1454
1455 void set_bh_page(struct buffer_head *bh,
1456                 struct page *page, unsigned long offset)
1457 {
1458         bh->b_page = page;
1459         BUG_ON(offset >= PAGE_SIZE);
1460         if (PageHighMem(page))
1461                 /*
1462                  * This catches illegal uses and preserves the offset:
1463                  */
1464                 bh->b_data = (char *)(0 + offset);
1465         else
1466                 bh->b_data = page_address(page) + offset;
1467 }
1468 EXPORT_SYMBOL(set_bh_page);
1469
1470 /*
1471  * Called when truncating a buffer on a page completely.
1472  */
1473 static void discard_buffer(struct buffer_head * bh)
1474 {
1475         lock_buffer(bh);
1476         clear_buffer_dirty(bh);
1477         bh->b_bdev = NULL;
1478         clear_buffer_mapped(bh);
1479         clear_buffer_req(bh);
1480         clear_buffer_new(bh);
1481         clear_buffer_delay(bh);
1482         clear_buffer_unwritten(bh);
1483         unlock_buffer(bh);
1484 }
1485
1486 /**
1487  * block_invalidatepage - invalidate part of all of a buffer-backed page
1488  *
1489  * @page: the page which is affected
1490  * @offset: the index of the truncation point
1491  *
1492  * block_invalidatepage() is called when all or part of the page has become
1493  * invalidatedby a truncate operation.
1494  *
1495  * block_invalidatepage() does not have to release all buffers, but it must
1496  * ensure that no dirty buffer is left outside @offset and that no I/O
1497  * is underway against any of the blocks which are outside the truncation
1498  * point.  Because the caller is about to free (and possibly reuse) those
1499  * blocks on-disk.
1500  */
1501 void block_invalidatepage(struct page *page, unsigned long offset)
1502 {
1503         struct buffer_head *head, *bh, *next;
1504         unsigned int curr_off = 0;
1505
1506         BUG_ON(!PageLocked(page));
1507         if (!page_has_buffers(page))
1508                 goto out;
1509
1510         head = page_buffers(page);
1511         bh = head;
1512         do {
1513                 unsigned int next_off = curr_off + bh->b_size;
1514                 next = bh->b_this_page;
1515
1516                 /*
1517                  * is this block fully invalidated?
1518                  */
1519                 if (offset <= curr_off)
1520                         discard_buffer(bh);
1521                 curr_off = next_off;
1522                 bh = next;
1523         } while (bh != head);
1524
1525         /*
1526          * We release buffers only if the entire page is being invalidated.
1527          * The get_block cached value has been unconditionally invalidated,
1528          * so real IO is not possible anymore.
1529          */
1530         if (offset == 0)
1531                 try_to_release_page(page, 0);
1532 out:
1533         return;
1534 }
1535 EXPORT_SYMBOL(block_invalidatepage);
1536
1537 /*
1538  * We attach and possibly dirty the buffers atomically wrt
1539  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1540  * is already excluded via the page lock.
1541  */
1542 void create_empty_buffers(struct page *page,
1543                         unsigned long blocksize, unsigned long b_state)
1544 {
1545         struct buffer_head *bh, *head, *tail;
1546
1547         head = alloc_page_buffers(page, blocksize, 1);
1548         bh = head;
1549         do {
1550                 bh->b_state |= b_state;
1551                 tail = bh;
1552                 bh = bh->b_this_page;
1553         } while (bh);
1554         tail->b_this_page = head;
1555
1556         spin_lock(&page->mapping->private_lock);
1557         if (PageUptodate(page) || PageDirty(page)) {
1558                 bh = head;
1559                 do {
1560                         if (PageDirty(page))
1561                                 set_buffer_dirty(bh);
1562                         if (PageUptodate(page))
1563                                 set_buffer_uptodate(bh);
1564                         bh = bh->b_this_page;
1565                 } while (bh != head);
1566         }
1567         attach_page_buffers(page, head);
1568         spin_unlock(&page->mapping->private_lock);
1569 }
1570 EXPORT_SYMBOL(create_empty_buffers);
1571
1572 /*
1573  * We are taking a block for data and we don't want any output from any
1574  * buffer-cache aliases starting from return from that function and
1575  * until the moment when something will explicitly mark the buffer
1576  * dirty (hopefully that will not happen until we will free that block ;-)
1577  * We don't even need to mark it not-uptodate - nobody can expect
1578  * anything from a newly allocated buffer anyway. We used to used
1579  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1580  * don't want to mark the alias unmapped, for example - it would confuse
1581  * anyone who might pick it with bread() afterwards...
1582  *
1583  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1584  * be writeout I/O going on against recently-freed buffers.  We don't
1585  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1586  * only if we really need to.  That happens here.
1587  */
1588 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1589 {
1590         struct buffer_head *old_bh;
1591
1592         might_sleep();
1593
1594         old_bh = __find_get_block_slow(bdev, block);
1595         if (old_bh) {
1596                 clear_buffer_dirty(old_bh);
1597                 wait_on_buffer(old_bh);
1598                 clear_buffer_req(old_bh);
1599                 __brelse(old_bh);
1600         }
1601 }
1602 EXPORT_SYMBOL(unmap_underlying_metadata);
1603
1604 /*
1605  * NOTE! All mapped/uptodate combinations are valid:
1606  *
1607  *      Mapped  Uptodate        Meaning
1608  *
1609  *      No      No              "unknown" - must do get_block()
1610  *      No      Yes             "hole" - zero-filled
1611  *      Yes     No              "allocated" - allocated on disk, not read in
1612  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1613  *
1614  * "Dirty" is valid only with the last case (mapped+uptodate).
1615  */
1616
1617 /*
1618  * While block_write_full_page is writing back the dirty buffers under
1619  * the page lock, whoever dirtied the buffers may decide to clean them
1620  * again at any time.  We handle that by only looking at the buffer
1621  * state inside lock_buffer().
1622  *
1623  * If block_write_full_page() is called for regular writeback
1624  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1625  * locked buffer.   This only can happen if someone has written the buffer
1626  * directly, with submit_bh().  At the address_space level PageWriteback
1627  * prevents this contention from occurring.
1628  */
1629 static int __block_write_full_page(struct inode *inode, struct page *page,
1630                         get_block_t *get_block, struct writeback_control *wbc)
1631 {
1632         int err;
1633         sector_t block;
1634         sector_t last_block;
1635         struct buffer_head *bh, *head;
1636         const unsigned blocksize = 1 << inode->i_blkbits;
1637         int nr_underway = 0;
1638
1639         BUG_ON(!PageLocked(page));
1640
1641         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1642
1643         if (!page_has_buffers(page)) {
1644                 create_empty_buffers(page, blocksize,
1645                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1646         }
1647
1648         /*
1649          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1650          * here, and the (potentially unmapped) buffers may become dirty at
1651          * any time.  If a buffer becomes dirty here after we've inspected it
1652          * then we just miss that fact, and the page stays dirty.
1653          *
1654          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1655          * handle that here by just cleaning them.
1656          */
1657
1658         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1659         head = page_buffers(page);
1660         bh = head;
1661
1662         /*
1663          * Get all the dirty buffers mapped to disk addresses and
1664          * handle any aliases from the underlying blockdev's mapping.
1665          */
1666         do {
1667                 if (block > last_block) {
1668                         /*
1669                          * mapped buffers outside i_size will occur, because
1670                          * this page can be outside i_size when there is a
1671                          * truncate in progress.
1672                          */
1673                         /*
1674                          * The buffer was zeroed by block_write_full_page()
1675                          */
1676                         clear_buffer_dirty(bh);
1677                         set_buffer_uptodate(bh);
1678                 } else if (!buffer_mapped(bh) && buffer_dirty(bh)) {
1679                         WARN_ON(bh->b_size != blocksize);
1680                         err = get_block(inode, block, bh, 1);
1681                         if (err)
1682                                 goto recover;
1683                         if (buffer_new(bh)) {
1684                                 /* blockdev mappings never come here */
1685                                 clear_buffer_new(bh);
1686                                 unmap_underlying_metadata(bh->b_bdev,
1687                                                         bh->b_blocknr);
1688                         }
1689                 }
1690                 bh = bh->b_this_page;
1691                 block++;
1692         } while (bh != head);
1693
1694         do {
1695                 if (!buffer_mapped(bh))
1696                         continue;
1697                 /*
1698                  * If it's a fully non-blocking write attempt and we cannot
1699                  * lock the buffer then redirty the page.  Note that this can
1700                  * potentially cause a busy-wait loop from pdflush and kswapd
1701                  * activity, but those code paths have their own higher-level
1702                  * throttling.
1703                  */
1704                 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1705                         lock_buffer(bh);
1706                 } else if (test_set_buffer_locked(bh)) {
1707                         redirty_page_for_writepage(wbc, page);
1708                         continue;
1709                 }
1710                 if (test_clear_buffer_dirty(bh)) {
1711                         mark_buffer_async_write(bh);
1712                 } else {
1713                         unlock_buffer(bh);
1714                 }
1715         } while ((bh = bh->b_this_page) != head);
1716
1717         /*
1718          * The page and its buffers are protected by PageWriteback(), so we can
1719          * drop the bh refcounts early.
1720          */
1721         BUG_ON(PageWriteback(page));
1722         set_page_writeback(page);
1723
1724         do {
1725                 struct buffer_head *next = bh->b_this_page;
1726                 if (buffer_async_write(bh)) {
1727                         submit_bh(WRITE, bh);
1728                         nr_underway++;
1729                 }
1730                 bh = next;
1731         } while (bh != head);
1732         unlock_page(page);
1733
1734         err = 0;
1735 done:
1736         if (nr_underway == 0) {
1737                 /*
1738                  * The page was marked dirty, but the buffers were
1739                  * clean.  Someone wrote them back by hand with
1740                  * ll_rw_block/submit_bh.  A rare case.
1741                  */
1742                 end_page_writeback(page);
1743
1744                 /*
1745                  * The page and buffer_heads can be released at any time from
1746                  * here on.
1747                  */
1748         }
1749         return err;
1750
1751 recover:
1752         /*
1753          * ENOSPC, or some other error.  We may already have added some
1754          * blocks to the file, so we need to write these out to avoid
1755          * exposing stale data.
1756          * The page is currently locked and not marked for writeback
1757          */
1758         bh = head;
1759         /* Recovery: lock and submit the mapped buffers */
1760         do {
1761                 if (buffer_mapped(bh) && buffer_dirty(bh)) {
1762                         lock_buffer(bh);
1763                         mark_buffer_async_write(bh);
1764                 } else {
1765                         /*
1766                          * The buffer may have been set dirty during
1767                          * attachment to a dirty page.
1768                          */
1769                         clear_buffer_dirty(bh);
1770                 }
1771         } while ((bh = bh->b_this_page) != head);
1772         SetPageError(page);
1773         BUG_ON(PageWriteback(page));
1774         mapping_set_error(page->mapping, err);
1775         set_page_writeback(page);
1776         do {
1777                 struct buffer_head *next = bh->b_this_page;
1778                 if (buffer_async_write(bh)) {
1779                         clear_buffer_dirty(bh);
1780                         submit_bh(WRITE, bh);
1781                         nr_underway++;
1782                 }
1783                 bh = next;
1784         } while (bh != head);
1785         unlock_page(page);
1786         goto done;
1787 }
1788
1789 /*
1790  * If a page has any new buffers, zero them out here, and mark them uptodate
1791  * and dirty so they'll be written out (in order to prevent uninitialised
1792  * block data from leaking). And clear the new bit.
1793  */
1794 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1795 {
1796         unsigned int block_start, block_end;
1797         struct buffer_head *head, *bh;
1798
1799         BUG_ON(!PageLocked(page));
1800         if (!page_has_buffers(page))
1801                 return;
1802
1803         bh = head = page_buffers(page);
1804         block_start = 0;
1805         do {
1806                 block_end = block_start + bh->b_size;
1807
1808                 if (buffer_new(bh)) {
1809                         if (block_end > from && block_start < to) {
1810                                 if (!PageUptodate(page)) {
1811                                         unsigned start, size;
1812
1813                                         start = max(from, block_start);
1814                                         size = min(to, block_end) - start;
1815
1816                                         zero_user(page, start, size);
1817                                         set_buffer_uptodate(bh);
1818                                 }
1819
1820                                 clear_buffer_new(bh);
1821                                 mark_buffer_dirty(bh);
1822                         }
1823                 }
1824
1825                 block_start = block_end;
1826                 bh = bh->b_this_page;
1827         } while (bh != head);
1828 }
1829 EXPORT_SYMBOL(page_zero_new_buffers);
1830
1831 static int __block_prepare_write(struct inode *inode, struct page *page,
1832                 unsigned from, unsigned to, get_block_t *get_block)
1833 {
1834         unsigned block_start, block_end;
1835         sector_t block;
1836         int err = 0;
1837         unsigned blocksize, bbits;
1838         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1839
1840         BUG_ON(!PageLocked(page));
1841         BUG_ON(from > PAGE_CACHE_SIZE);
1842         BUG_ON(to > PAGE_CACHE_SIZE);
1843         BUG_ON(from > to);
1844
1845         blocksize = 1 << inode->i_blkbits;
1846         if (!page_has_buffers(page))
1847                 create_empty_buffers(page, blocksize, 0);
1848         head = page_buffers(page);
1849
1850         bbits = inode->i_blkbits;
1851         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1852
1853         for(bh = head, block_start = 0; bh != head || !block_start;
1854             block++, block_start=block_end, bh = bh->b_this_page) {
1855                 block_end = block_start + blocksize;
1856                 if (block_end <= from || block_start >= to) {
1857                         if (PageUptodate(page)) {
1858                                 if (!buffer_uptodate(bh))
1859                                         set_buffer_uptodate(bh);
1860                         }
1861                         continue;
1862                 }
1863                 if (buffer_new(bh))
1864                         clear_buffer_new(bh);
1865                 if (!buffer_mapped(bh)) {
1866                         WARN_ON(bh->b_size != blocksize);
1867                         err = get_block(inode, block, bh, 1);
1868                         if (err)
1869                                 break;
1870                         if (buffer_new(bh)) {
1871                                 unmap_underlying_metadata(bh->b_bdev,
1872                                                         bh->b_blocknr);
1873                                 if (PageUptodate(page)) {
1874                                         clear_buffer_new(bh);
1875                                         set_buffer_uptodate(bh);
1876                                         mark_buffer_dirty(bh);
1877                                         continue;
1878                                 }
1879                                 if (block_end > to || block_start < from)
1880                                         zero_user_segments(page,
1881                                                 to, block_end,
1882                                                 block_start, from);
1883                                 continue;
1884                         }
1885                 }
1886                 if (PageUptodate(page)) {
1887                         if (!buffer_uptodate(bh))
1888                                 set_buffer_uptodate(bh);
1889                         continue; 
1890                 }
1891                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1892                     !buffer_unwritten(bh) &&
1893                      (block_start < from || block_end > to)) {
1894                         ll_rw_block(READ, 1, &bh);
1895                         *wait_bh++=bh;
1896                 }
1897         }
1898         /*
1899          * If we issued read requests - let them complete.
1900          */
1901         while(wait_bh > wait) {
1902                 wait_on_buffer(*--wait_bh);
1903                 if (!buffer_uptodate(*wait_bh))
1904                         err = -EIO;
1905         }
1906         if (unlikely(err))
1907                 page_zero_new_buffers(page, from, to);
1908         return err;
1909 }
1910
1911 static int __block_commit_write(struct inode *inode, struct page *page,
1912                 unsigned from, unsigned to)
1913 {
1914         unsigned block_start, block_end;
1915         int partial = 0;
1916         unsigned blocksize;
1917         struct buffer_head *bh, *head;
1918
1919         blocksize = 1 << inode->i_blkbits;
1920
1921         for(bh = head = page_buffers(page), block_start = 0;
1922             bh != head || !block_start;
1923             block_start=block_end, bh = bh->b_this_page) {
1924                 block_end = block_start + blocksize;
1925                 if (block_end <= from || block_start >= to) {
1926                         if (!buffer_uptodate(bh))
1927                                 partial = 1;
1928                 } else {
1929                         set_buffer_uptodate(bh);
1930                         mark_buffer_dirty(bh);
1931                 }
1932                 clear_buffer_new(bh);
1933         }
1934
1935         /*
1936          * If this is a partial write which happened to make all buffers
1937          * uptodate then we can optimize away a bogus readpage() for
1938          * the next read(). Here we 'discover' whether the page went
1939          * uptodate as a result of this (potentially partial) write.
1940          */
1941         if (!partial)
1942                 SetPageUptodate(page);
1943         return 0;
1944 }
1945
1946 /*
1947  * block_write_begin takes care of the basic task of block allocation and
1948  * bringing partial write blocks uptodate first.
1949  *
1950  * If *pagep is not NULL, then block_write_begin uses the locked page
1951  * at *pagep rather than allocating its own. In this case, the page will
1952  * not be unlocked or deallocated on failure.
1953  */
1954 int block_write_begin(struct file *file, struct address_space *mapping,
1955                         loff_t pos, unsigned len, unsigned flags,
1956                         struct page **pagep, void **fsdata,
1957                         get_block_t *get_block)
1958 {
1959         struct inode *inode = mapping->host;
1960         int status = 0;
1961         struct page *page;
1962         pgoff_t index;
1963         unsigned start, end;
1964         int ownpage = 0;
1965
1966         index = pos >> PAGE_CACHE_SHIFT;
1967         start = pos & (PAGE_CACHE_SIZE - 1);
1968         end = start + len;
1969
1970         page = *pagep;
1971         if (page == NULL) {
1972                 ownpage = 1;
1973                 page = __grab_cache_page(mapping, index);
1974                 if (!page) {
1975                         status = -ENOMEM;
1976                         goto out;
1977                 }
1978                 *pagep = page;
1979         } else
1980                 BUG_ON(!PageLocked(page));
1981
1982         status = __block_prepare_write(inode, page, start, end, get_block);
1983         if (unlikely(status)) {
1984                 ClearPageUptodate(page);
1985
1986                 if (ownpage) {
1987                         unlock_page(page);
1988                         page_cache_release(page);
1989                         *pagep = NULL;
1990
1991                         /*
1992                          * prepare_write() may have instantiated a few blocks
1993                          * outside i_size.  Trim these off again. Don't need
1994                          * i_size_read because we hold i_mutex.
1995                          */
1996                         if (pos + len > inode->i_size)
1997                                 vmtruncate(inode, inode->i_size);
1998                 }
1999                 goto out;
2000         }
2001
2002 out:
2003         return status;
2004 }
2005 EXPORT_SYMBOL(block_write_begin);
2006
2007 int block_write_end(struct file *file, struct address_space *mapping,
2008                         loff_t pos, unsigned len, unsigned copied,
2009                         struct page *page, void *fsdata)
2010 {
2011         struct inode *inode = mapping->host;
2012         unsigned start;
2013
2014         start = pos & (PAGE_CACHE_SIZE - 1);
2015
2016         if (unlikely(copied < len)) {
2017                 /*
2018                  * The buffers that were written will now be uptodate, so we
2019                  * don't have to worry about a readpage reading them and
2020                  * overwriting a partial write. However if we have encountered
2021                  * a short write and only partially written into a buffer, it
2022                  * will not be marked uptodate, so a readpage might come in and
2023                  * destroy our partial write.
2024                  *
2025                  * Do the simplest thing, and just treat any short write to a
2026                  * non uptodate page as a zero-length write, and force the
2027                  * caller to redo the whole thing.
2028                  */
2029                 if (!PageUptodate(page))
2030                         copied = 0;
2031
2032                 page_zero_new_buffers(page, start+copied, start+len);
2033         }
2034         flush_dcache_page(page);
2035
2036         /* This could be a short (even 0-length) commit */
2037         __block_commit_write(inode, page, start, start+copied);
2038
2039         return copied;
2040 }
2041 EXPORT_SYMBOL(block_write_end);
2042
2043 int generic_write_end(struct file *file, struct address_space *mapping,
2044                         loff_t pos, unsigned len, unsigned copied,
2045                         struct page *page, void *fsdata)
2046 {
2047         struct inode *inode = mapping->host;
2048
2049         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2050
2051         /*
2052          * No need to use i_size_read() here, the i_size
2053          * cannot change under us because we hold i_mutex.
2054          *
2055          * But it's important to update i_size while still holding page lock:
2056          * page writeout could otherwise come in and zero beyond i_size.
2057          */
2058         if (pos+copied > inode->i_size) {
2059                 i_size_write(inode, pos+copied);
2060                 mark_inode_dirty(inode);
2061         }
2062
2063         unlock_page(page);
2064         page_cache_release(page);
2065
2066         return copied;
2067 }
2068 EXPORT_SYMBOL(generic_write_end);
2069
2070 /*
2071  * Generic "read page" function for block devices that have the normal
2072  * get_block functionality. This is most of the block device filesystems.
2073  * Reads the page asynchronously --- the unlock_buffer() and
2074  * set/clear_buffer_uptodate() functions propagate buffer state into the
2075  * page struct once IO has completed.
2076  */
2077 int block_read_full_page(struct page *page, get_block_t *get_block)
2078 {
2079         struct inode *inode = page->mapping->host;
2080         sector_t iblock, lblock;
2081         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2082         unsigned int blocksize;
2083         int nr, i;
2084         int fully_mapped = 1;
2085
2086         BUG_ON(!PageLocked(page));
2087         blocksize = 1 << inode->i_blkbits;
2088         if (!page_has_buffers(page))
2089                 create_empty_buffers(page, blocksize, 0);
2090         head = page_buffers(page);
2091
2092         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2093         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2094         bh = head;
2095         nr = 0;
2096         i = 0;
2097
2098         do {
2099                 if (buffer_uptodate(bh))
2100                         continue;
2101
2102                 if (!buffer_mapped(bh)) {
2103                         int err = 0;
2104
2105                         fully_mapped = 0;
2106                         if (iblock < lblock) {
2107                                 WARN_ON(bh->b_size != blocksize);
2108                                 err = get_block(inode, iblock, bh, 0);
2109                                 if (err)
2110                                         SetPageError(page);
2111                         }
2112                         if (!buffer_mapped(bh)) {
2113                                 zero_user(page, i * blocksize, blocksize);
2114                                 if (!err)
2115                                         set_buffer_uptodate(bh);
2116                                 continue;
2117                         }
2118                         /*
2119                          * get_block() might have updated the buffer
2120                          * synchronously
2121                          */
2122                         if (buffer_uptodate(bh))
2123                                 continue;
2124                 }
2125                 arr[nr++] = bh;
2126         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2127
2128         if (fully_mapped)
2129                 SetPageMappedToDisk(page);
2130
2131         if (!nr) {
2132                 /*
2133                  * All buffers are uptodate - we can set the page uptodate
2134                  * as well. But not if get_block() returned an error.
2135                  */
2136                 if (!PageError(page))
2137                         SetPageUptodate(page);
2138                 unlock_page(page);
2139                 return 0;
2140         }
2141
2142         /* Stage two: lock the buffers */
2143         for (i = 0; i < nr; i++) {
2144                 bh = arr[i];
2145                 lock_buffer(bh);
2146                 mark_buffer_async_read(bh);
2147         }
2148
2149         /*
2150          * Stage 3: start the IO.  Check for uptodateness
2151          * inside the buffer lock in case another process reading
2152          * the underlying blockdev brought it uptodate (the sct fix).
2153          */
2154         for (i = 0; i < nr; i++) {
2155                 bh = arr[i];
2156                 if (buffer_uptodate(bh))
2157                         end_buffer_async_read(bh, 1);
2158                 else
2159                         submit_bh(READ, bh);
2160         }
2161         return 0;
2162 }
2163
2164 /* utility function for filesystems that need to do work on expanding
2165  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2166  * deal with the hole.  
2167  */
2168 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2169 {
2170         struct address_space *mapping = inode->i_mapping;
2171         struct page *page;
2172         void *fsdata;
2173         unsigned long limit;
2174         int err;
2175
2176         err = -EFBIG;
2177         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2178         if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2179                 send_sig(SIGXFSZ, current, 0);
2180                 goto out;
2181         }
2182         if (size > inode->i_sb->s_maxbytes)
2183                 goto out;
2184
2185         err = pagecache_write_begin(NULL, mapping, size, 0,
2186                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2187                                 &page, &fsdata);
2188         if (err)
2189                 goto out;
2190
2191         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2192         BUG_ON(err > 0);
2193
2194 out:
2195         return err;
2196 }
2197
2198 int cont_expand_zero(struct file *file, struct address_space *mapping,
2199                         loff_t pos, loff_t *bytes)
2200 {
2201         struct inode *inode = mapping->host;
2202         unsigned blocksize = 1 << inode->i_blkbits;
2203         struct page *page;
2204         void *fsdata;
2205         pgoff_t index, curidx;
2206         loff_t curpos;
2207         unsigned zerofrom, offset, len;
2208         int err = 0;
2209
2210         index = pos >> PAGE_CACHE_SHIFT;
2211         offset = pos & ~PAGE_CACHE_MASK;
2212
2213         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2214                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2215                 if (zerofrom & (blocksize-1)) {
2216                         *bytes |= (blocksize-1);
2217                         (*bytes)++;
2218                 }
2219                 len = PAGE_CACHE_SIZE - zerofrom;
2220
2221                 err = pagecache_write_begin(file, mapping, curpos, len,
2222                                                 AOP_FLAG_UNINTERRUPTIBLE,
2223                                                 &page, &fsdata);
2224                 if (err)
2225                         goto out;
2226                 zero_user(page, zerofrom, len);
2227                 err = pagecache_write_end(file, mapping, curpos, len, len,
2228                                                 page, fsdata);
2229                 if (err < 0)
2230                         goto out;
2231                 BUG_ON(err != len);
2232                 err = 0;
2233         }
2234
2235         /* page covers the boundary, find the boundary offset */
2236         if (index == curidx) {
2237                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2238                 /* if we will expand the thing last block will be filled */
2239                 if (offset <= zerofrom) {
2240                         goto out;
2241                 }
2242                 if (zerofrom & (blocksize-1)) {
2243                         *bytes |= (blocksize-1);
2244                         (*bytes)++;
2245                 }
2246                 len = offset - zerofrom;
2247
2248                 err = pagecache_write_begin(file, mapping, curpos, len,
2249                                                 AOP_FLAG_UNINTERRUPTIBLE,
2250                                                 &page, &fsdata);
2251                 if (err)
2252                         goto out;
2253                 zero_user(page, zerofrom, len);
2254                 err = pagecache_write_end(file, mapping, curpos, len, len,
2255                                                 page, fsdata);
2256                 if (err < 0)
2257                         goto out;
2258                 BUG_ON(err != len);
2259                 err = 0;
2260         }
2261 out:
2262         return err;
2263 }
2264
2265 /*
2266  * For moronic filesystems that do not allow holes in file.
2267  * We may have to extend the file.
2268  */
2269 int cont_write_begin(struct file *file, struct address_space *mapping,
2270                         loff_t pos, unsigned len, unsigned flags,
2271                         struct page **pagep, void **fsdata,
2272                         get_block_t *get_block, loff_t *bytes)
2273 {
2274         struct inode *inode = mapping->host;
2275         unsigned blocksize = 1 << inode->i_blkbits;
2276         unsigned zerofrom;
2277         int err;
2278
2279         err = cont_expand_zero(file, mapping, pos, bytes);
2280         if (err)
2281                 goto out;
2282
2283         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2284         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2285                 *bytes |= (blocksize-1);
2286                 (*bytes)++;
2287         }
2288
2289         *pagep = NULL;
2290         err = block_write_begin(file, mapping, pos, len,
2291                                 flags, pagep, fsdata, get_block);
2292 out:
2293         return err;
2294 }
2295
2296 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2297                         get_block_t *get_block)
2298 {
2299         struct inode *inode = page->mapping->host;
2300         int err = __block_prepare_write(inode, page, from, to, get_block);
2301         if (err)
2302                 ClearPageUptodate(page);
2303         return err;
2304 }
2305
2306 int block_commit_write(struct page *page, unsigned from, unsigned to)
2307 {
2308         struct inode *inode = page->mapping->host;
2309         __block_commit_write(inode,page,from,to);
2310         return 0;
2311 }
2312
2313 int generic_commit_write(struct file *file, struct page *page,
2314                 unsigned from, unsigned to)
2315 {
2316         struct inode *inode = page->mapping->host;
2317         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2318         __block_commit_write(inode,page,from,to);
2319         /*
2320          * No need to use i_size_read() here, the i_size
2321          * cannot change under us because we hold i_mutex.
2322          */
2323         if (pos > inode->i_size) {
2324                 i_size_write(inode, pos);
2325                 mark_inode_dirty(inode);
2326         }
2327         return 0;
2328 }
2329
2330 /*
2331  * block_page_mkwrite() is not allowed to change the file size as it gets
2332  * called from a page fault handler when a page is first dirtied. Hence we must
2333  * be careful to check for EOF conditions here. We set the page up correctly
2334  * for a written page which means we get ENOSPC checking when writing into
2335  * holes and correct delalloc and unwritten extent mapping on filesystems that
2336  * support these features.
2337  *
2338  * We are not allowed to take the i_mutex here so we have to play games to
2339  * protect against truncate races as the page could now be beyond EOF.  Because
2340  * vmtruncate() writes the inode size before removing pages, once we have the
2341  * page lock we can determine safely if the page is beyond EOF. If it is not
2342  * beyond EOF, then the page is guaranteed safe against truncation until we
2343  * unlock the page.
2344  */
2345 int
2346 block_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2347                    get_block_t get_block)
2348 {
2349         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2350         unsigned long end;
2351         loff_t size;
2352         int ret = -EINVAL;
2353
2354         lock_page(page);
2355         size = i_size_read(inode);
2356         if ((page->mapping != inode->i_mapping) ||
2357             (page_offset(page) > size)) {
2358                 /* page got truncated out from underneath us */
2359                 goto out_unlock;
2360         }
2361
2362         /* page is wholly or partially inside EOF */
2363         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2364                 end = size & ~PAGE_CACHE_MASK;
2365         else
2366                 end = PAGE_CACHE_SIZE;
2367
2368         ret = block_prepare_write(page, 0, end, get_block);
2369         if (!ret)
2370                 ret = block_commit_write(page, 0, end);
2371
2372 out_unlock:
2373         unlock_page(page);
2374         return ret;
2375 }
2376
2377 /*
2378  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2379  * immediately, while under the page lock.  So it needs a special end_io
2380  * handler which does not touch the bh after unlocking it.
2381  */
2382 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2383 {
2384         __end_buffer_read_notouch(bh, uptodate);
2385 }
2386
2387 /*
2388  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2389  * the page (converting it to circular linked list and taking care of page
2390  * dirty races).
2391  */
2392 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2393 {
2394         struct buffer_head *bh;
2395
2396         BUG_ON(!PageLocked(page));
2397
2398         spin_lock(&page->mapping->private_lock);
2399         bh = head;
2400         do {
2401                 if (PageDirty(page))
2402                         set_buffer_dirty(bh);
2403                 if (!bh->b_this_page)
2404                         bh->b_this_page = head;
2405                 bh = bh->b_this_page;
2406         } while (bh != head);
2407         attach_page_buffers(page, head);
2408         spin_unlock(&page->mapping->private_lock);
2409 }
2410
2411 /*
2412  * On entry, the page is fully not uptodate.
2413  * On exit the page is fully uptodate in the areas outside (from,to)
2414  */
2415 int nobh_write_begin(struct file *file, struct address_space *mapping,
2416                         loff_t pos, unsigned len, unsigned flags,
2417                         struct page **pagep, void **fsdata,
2418                         get_block_t *get_block)
2419 {
2420         struct inode *inode = mapping->host;
2421         const unsigned blkbits = inode->i_blkbits;
2422         const unsigned blocksize = 1 << blkbits;
2423         struct buffer_head *head, *bh;
2424         struct page *page;
2425         pgoff_t index;
2426         unsigned from, to;
2427         unsigned block_in_page;
2428         unsigned block_start, block_end;
2429         sector_t block_in_file;
2430         int nr_reads = 0;
2431         int ret = 0;
2432         int is_mapped_to_disk = 1;
2433
2434         index = pos >> PAGE_CACHE_SHIFT;
2435         from = pos & (PAGE_CACHE_SIZE - 1);
2436         to = from + len;
2437
2438         page = __grab_cache_page(mapping, index);
2439         if (!page)
2440                 return -ENOMEM;
2441         *pagep = page;
2442         *fsdata = NULL;
2443
2444         if (page_has_buffers(page)) {
2445                 unlock_page(page);
2446                 page_cache_release(page);
2447                 *pagep = NULL;
2448                 return block_write_begin(file, mapping, pos, len, flags, pagep,
2449                                         fsdata, get_block);
2450         }
2451
2452         if (PageMappedToDisk(page))
2453                 return 0;
2454
2455         /*
2456          * Allocate buffers so that we can keep track of state, and potentially
2457          * attach them to the page if an error occurs. In the common case of
2458          * no error, they will just be freed again without ever being attached
2459          * to the page (which is all OK, because we're under the page lock).
2460          *
2461          * Be careful: the buffer linked list is a NULL terminated one, rather
2462          * than the circular one we're used to.
2463          */
2464         head = alloc_page_buffers(page, blocksize, 0);
2465         if (!head) {
2466                 ret = -ENOMEM;
2467                 goto out_release;
2468         }
2469
2470         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2471
2472         /*
2473          * We loop across all blocks in the page, whether or not they are
2474          * part of the affected region.  This is so we can discover if the
2475          * page is fully mapped-to-disk.
2476          */
2477         for (block_start = 0, block_in_page = 0, bh = head;
2478                   block_start < PAGE_CACHE_SIZE;
2479                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2480                 int create;
2481
2482                 block_end = block_start + blocksize;
2483                 bh->b_state = 0;
2484                 create = 1;
2485                 if (block_start >= to)
2486                         create = 0;
2487                 ret = get_block(inode, block_in_file + block_in_page,
2488                                         bh, create);
2489                 if (ret)
2490                         goto failed;
2491                 if (!buffer_mapped(bh))
2492                         is_mapped_to_disk = 0;
2493                 if (buffer_new(bh))
2494                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2495                 if (PageUptodate(page)) {
2496                         set_buffer_uptodate(bh);
2497                         continue;
2498                 }
2499                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2500                         zero_user_segments(page, block_start, from,
2501                                                         to, block_end);
2502                         continue;
2503                 }
2504                 if (buffer_uptodate(bh))
2505                         continue;       /* reiserfs does this */
2506                 if (block_start < from || block_end > to) {
2507                         lock_buffer(bh);
2508                         bh->b_end_io = end_buffer_read_nobh;
2509                         submit_bh(READ, bh);
2510                         nr_reads++;
2511                 }
2512         }
2513
2514         if (nr_reads) {
2515                 /*
2516                  * The page is locked, so these buffers are protected from
2517                  * any VM or truncate activity.  Hence we don't need to care
2518                  * for the buffer_head refcounts.
2519                  */
2520                 for (bh = head; bh; bh = bh->b_this_page) {
2521                         wait_on_buffer(bh);
2522                         if (!buffer_uptodate(bh))
2523                                 ret = -EIO;
2524                 }
2525                 if (ret)
2526                         goto failed;
2527         }
2528
2529         if (is_mapped_to_disk)
2530                 SetPageMappedToDisk(page);
2531
2532         *fsdata = head; /* to be released by nobh_write_end */
2533
2534         return 0;
2535
2536 failed:
2537         BUG_ON(!ret);
2538         /*
2539          * Error recovery is a bit difficult. We need to zero out blocks that
2540          * were newly allocated, and dirty them to ensure they get written out.
2541          * Buffers need to be attached to the page at this point, otherwise
2542          * the handling of potential IO errors during writeout would be hard
2543          * (could try doing synchronous writeout, but what if that fails too?)
2544          */
2545         attach_nobh_buffers(page, head);
2546         page_zero_new_buffers(page, from, to);
2547
2548 out_release:
2549         unlock_page(page);
2550         page_cache_release(page);
2551         *pagep = NULL;
2552
2553         if (pos + len > inode->i_size)
2554                 vmtruncate(inode, inode->i_size);
2555
2556         return ret;
2557 }
2558 EXPORT_SYMBOL(nobh_write_begin);
2559
2560 int nobh_write_end(struct file *file, struct address_space *mapping,
2561                         loff_t pos, unsigned len, unsigned copied,
2562                         struct page *page, void *fsdata)
2563 {
2564         struct inode *inode = page->mapping->host;
2565         struct buffer_head *head = fsdata;
2566         struct buffer_head *bh;
2567
2568         if (!PageMappedToDisk(page)) {
2569                 if (unlikely(copied < len) && !page_has_buffers(page))
2570                         attach_nobh_buffers(page, head);
2571                 if (page_has_buffers(page))
2572                         return generic_write_end(file, mapping, pos, len,
2573                                                 copied, page, fsdata);
2574         }
2575
2576         SetPageUptodate(page);
2577         set_page_dirty(page);
2578         if (pos+copied > inode->i_size) {
2579                 i_size_write(inode, pos+copied);
2580                 mark_inode_dirty(inode);
2581         }
2582
2583         unlock_page(page);
2584         page_cache_release(page);
2585
2586         while (head) {
2587                 bh = head;
2588                 head = head->b_this_page;
2589                 free_buffer_head(bh);
2590         }
2591
2592         return copied;
2593 }
2594 EXPORT_SYMBOL(nobh_write_end);
2595
2596 /*
2597  * nobh_writepage() - based on block_full_write_page() except
2598  * that it tries to operate without attaching bufferheads to
2599  * the page.
2600  */
2601 int nobh_writepage(struct page *page, get_block_t *get_block,
2602                         struct writeback_control *wbc)
2603 {
2604         struct inode * const inode = page->mapping->host;
2605         loff_t i_size = i_size_read(inode);
2606         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2607         unsigned offset;
2608         int ret;
2609
2610         /* Is the page fully inside i_size? */
2611         if (page->index < end_index)
2612                 goto out;
2613
2614         /* Is the page fully outside i_size? (truncate in progress) */
2615         offset = i_size & (PAGE_CACHE_SIZE-1);
2616         if (page->index >= end_index+1 || !offset) {
2617                 /*
2618                  * The page may have dirty, unmapped buffers.  For example,
2619                  * they may have been added in ext3_writepage().  Make them
2620                  * freeable here, so the page does not leak.
2621                  */
2622 #if 0
2623                 /* Not really sure about this  - do we need this ? */
2624                 if (page->mapping->a_ops->invalidatepage)
2625                         page->mapping->a_ops->invalidatepage(page, offset);
2626 #endif
2627                 unlock_page(page);
2628                 return 0; /* don't care */
2629         }
2630
2631         /*
2632          * The page straddles i_size.  It must be zeroed out on each and every
2633          * writepage invocation because it may be mmapped.  "A file is mapped
2634          * in multiples of the page size.  For a file that is not a multiple of
2635          * the  page size, the remaining memory is zeroed when mapped, and
2636          * writes to that region are not written out to the file."
2637          */
2638         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2639 out:
2640         ret = mpage_writepage(page, get_block, wbc);
2641         if (ret == -EAGAIN)
2642                 ret = __block_write_full_page(inode, page, get_block, wbc);
2643         return ret;
2644 }
2645 EXPORT_SYMBOL(nobh_writepage);
2646
2647 int nobh_truncate_page(struct address_space *mapping,
2648                         loff_t from, get_block_t *get_block)
2649 {
2650         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2651         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2652         unsigned blocksize;
2653         sector_t iblock;
2654         unsigned length, pos;
2655         struct inode *inode = mapping->host;
2656         struct page *page;
2657         struct buffer_head map_bh;
2658         int err;
2659
2660         blocksize = 1 << inode->i_blkbits;
2661         length = offset & (blocksize - 1);
2662
2663         /* Block boundary? Nothing to do */
2664         if (!length)
2665                 return 0;
2666
2667         length = blocksize - length;
2668         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2669
2670         page = grab_cache_page(mapping, index);
2671         err = -ENOMEM;
2672         if (!page)
2673                 goto out;
2674
2675         if (page_has_buffers(page)) {
2676 has_buffers:
2677                 unlock_page(page);
2678                 page_cache_release(page);
2679                 return block_truncate_page(mapping, from, get_block);
2680         }
2681
2682         /* Find the buffer that contains "offset" */
2683         pos = blocksize;
2684         while (offset >= pos) {
2685                 iblock++;
2686                 pos += blocksize;
2687         }
2688
2689         err = get_block(inode, iblock, &map_bh, 0);
2690         if (err)
2691                 goto unlock;
2692         /* unmapped? It's a hole - nothing to do */
2693         if (!buffer_mapped(&map_bh))
2694                 goto unlock;
2695
2696         /* Ok, it's mapped. Make sure it's up-to-date */
2697         if (!PageUptodate(page)) {
2698                 err = mapping->a_ops->readpage(NULL, page);
2699                 if (err) {
2700                         page_cache_release(page);
2701                         goto out;
2702                 }
2703                 lock_page(page);
2704                 if (!PageUptodate(page)) {
2705                         err = -EIO;
2706                         goto unlock;
2707                 }
2708                 if (page_has_buffers(page))
2709                         goto has_buffers;
2710         }
2711         zero_user(page, offset, length);
2712         set_page_dirty(page);
2713         err = 0;
2714
2715 unlock:
2716         unlock_page(page);
2717         page_cache_release(page);
2718 out:
2719         return err;
2720 }
2721 EXPORT_SYMBOL(nobh_truncate_page);
2722
2723 int block_truncate_page(struct address_space *mapping,
2724                         loff_t from, get_block_t *get_block)
2725 {
2726         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2727         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2728         unsigned blocksize;
2729         sector_t iblock;
2730         unsigned length, pos;
2731         struct inode *inode = mapping->host;
2732         struct page *page;
2733         struct buffer_head *bh;
2734         int err;
2735
2736         blocksize = 1 << inode->i_blkbits;
2737         length = offset & (blocksize - 1);
2738
2739         /* Block boundary? Nothing to do */
2740         if (!length)
2741                 return 0;
2742
2743         length = blocksize - length;
2744         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2745         
2746         page = grab_cache_page(mapping, index);
2747         err = -ENOMEM;
2748         if (!page)
2749                 goto out;
2750
2751         if (!page_has_buffers(page))
2752                 create_empty_buffers(page, blocksize, 0);
2753
2754         /* Find the buffer that contains "offset" */
2755         bh = page_buffers(page);
2756         pos = blocksize;
2757         while (offset >= pos) {
2758                 bh = bh->b_this_page;
2759                 iblock++;
2760                 pos += blocksize;
2761         }
2762
2763         err = 0;
2764         if (!buffer_mapped(bh)) {
2765                 WARN_ON(bh->b_size != blocksize);
2766                 err = get_block(inode, iblock, bh, 0);
2767                 if (err)
2768                         goto unlock;
2769                 /* unmapped? It's a hole - nothing to do */
2770                 if (!buffer_mapped(bh))
2771                         goto unlock;
2772         }
2773
2774         /* Ok, it's mapped. Make sure it's up-to-date */
2775         if (PageUptodate(page))
2776                 set_buffer_uptodate(bh);
2777
2778         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2779                 err = -EIO;
2780                 ll_rw_block(READ, 1, &bh);
2781                 wait_on_buffer(bh);
2782                 /* Uhhuh. Read error. Complain and punt. */
2783                 if (!buffer_uptodate(bh))
2784                         goto unlock;
2785         }
2786
2787         zero_user(page, offset, length);
2788         mark_buffer_dirty(bh);
2789         err = 0;
2790
2791 unlock:
2792         unlock_page(page);
2793         page_cache_release(page);
2794 out:
2795         return err;
2796 }
2797
2798 /*
2799  * The generic ->writepage function for buffer-backed address_spaces
2800  */
2801 int block_write_full_page(struct page *page, get_block_t *get_block,
2802                         struct writeback_control *wbc)
2803 {
2804         struct inode * const inode = page->mapping->host;
2805         loff_t i_size = i_size_read(inode);
2806         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2807         unsigned offset;
2808
2809         /* Is the page fully inside i_size? */
2810         if (page->index < end_index)
2811                 return __block_write_full_page(inode, page, get_block, wbc);
2812
2813         /* Is the page fully outside i_size? (truncate in progress) */
2814         offset = i_size & (PAGE_CACHE_SIZE-1);
2815         if (page->index >= end_index+1 || !offset) {
2816                 /*
2817                  * The page may have dirty, unmapped buffers.  For example,
2818                  * they may have been added in ext3_writepage().  Make them
2819                  * freeable here, so the page does not leak.
2820                  */
2821                 do_invalidatepage(page, 0);
2822                 unlock_page(page);
2823                 return 0; /* don't care */
2824         }
2825
2826         /*
2827          * The page straddles i_size.  It must be zeroed out on each and every
2828          * writepage invokation because it may be mmapped.  "A file is mapped
2829          * in multiples of the page size.  For a file that is not a multiple of
2830          * the  page size, the remaining memory is zeroed when mapped, and
2831          * writes to that region are not written out to the file."
2832          */
2833         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2834         return __block_write_full_page(inode, page, get_block, wbc);
2835 }
2836
2837 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2838                             get_block_t *get_block)
2839 {
2840         struct buffer_head tmp;
2841         struct inode *inode = mapping->host;
2842         tmp.b_state = 0;
2843         tmp.b_blocknr = 0;
2844         tmp.b_size = 1 << inode->i_blkbits;
2845         get_block(inode, block, &tmp, 0);
2846         return tmp.b_blocknr;
2847 }
2848
2849 static void end_bio_bh_io_sync(struct bio *bio, int err)
2850 {
2851         struct buffer_head *bh = bio->bi_private;
2852
2853         if (err == -EOPNOTSUPP) {
2854                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2855                 set_bit(BH_Eopnotsupp, &bh->b_state);
2856         }
2857
2858         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2859         bio_put(bio);
2860 }
2861
2862 int submit_bh(int rw, struct buffer_head * bh)
2863 {
2864         struct bio *bio;
2865         int ret = 0;
2866
2867         BUG_ON(!buffer_locked(bh));
2868         BUG_ON(!buffer_mapped(bh));
2869         BUG_ON(!bh->b_end_io);
2870
2871         if (buffer_ordered(bh) && (rw == WRITE))
2872                 rw = WRITE_BARRIER;
2873
2874         /*
2875          * Only clear out a write error when rewriting, should this
2876          * include WRITE_SYNC as well?
2877          */
2878         if (test_set_buffer_req(bh) && (rw == WRITE || rw == WRITE_BARRIER))
2879                 clear_buffer_write_io_error(bh);
2880
2881         /*
2882          * from here on down, it's all bio -- do the initial mapping,
2883          * submit_bio -> generic_make_request may further map this bio around
2884          */
2885         bio = bio_alloc(GFP_NOIO, 1);
2886
2887         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2888         bio->bi_bdev = bh->b_bdev;
2889         bio->bi_io_vec[0].bv_page = bh->b_page;
2890         bio->bi_io_vec[0].bv_len = bh->b_size;
2891         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2892
2893         bio->bi_vcnt = 1;
2894         bio->bi_idx = 0;
2895         bio->bi_size = bh->b_size;
2896
2897         bio->bi_end_io = end_bio_bh_io_sync;
2898         bio->bi_private = bh;
2899
2900         bio_get(bio);
2901         submit_bio(rw, bio);
2902
2903         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2904                 ret = -EOPNOTSUPP;
2905
2906         bio_put(bio);
2907         return ret;
2908 }
2909
2910 /**
2911  * ll_rw_block: low-level access to block devices (DEPRECATED)
2912  * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
2913  * @nr: number of &struct buffer_heads in the array
2914  * @bhs: array of pointers to &struct buffer_head
2915  *
2916  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2917  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2918  * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2919  * are sent to disk. The fourth %READA option is described in the documentation
2920  * for generic_make_request() which ll_rw_block() calls.
2921  *
2922  * This function drops any buffer that it cannot get a lock on (with the
2923  * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
2924  * clean when doing a write request, and any buffer that appears to be
2925  * up-to-date when doing read request.  Further it marks as clean buffers that
2926  * are processed for writing (the buffer cache won't assume that they are
2927  * actually clean until the buffer gets unlocked).
2928  *
2929  * ll_rw_block sets b_end_io to simple completion handler that marks
2930  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2931  * any waiters. 
2932  *
2933  * All of the buffers must be for the same device, and must also be a
2934  * multiple of the current approved size for the device.
2935  */
2936 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2937 {
2938         int i;
2939
2940         for (i = 0; i < nr; i++) {
2941                 struct buffer_head *bh = bhs[i];
2942
2943                 if (rw == SWRITE)
2944                         lock_buffer(bh);
2945                 else if (test_set_buffer_locked(bh))
2946                         continue;
2947
2948                 if (rw == WRITE || rw == SWRITE) {
2949                         if (test_clear_buffer_dirty(bh)) {
2950                                 bh->b_end_io = end_buffer_write_sync;
2951                                 get_bh(bh);
2952                                 submit_bh(WRITE, bh);
2953                                 continue;
2954                         }
2955                 } else {
2956                         if (!buffer_uptodate(bh)) {
2957                                 bh->b_end_io = end_buffer_read_sync;
2958                                 get_bh(bh);
2959                                 submit_bh(rw, bh);
2960                                 continue;
2961                         }
2962                 }
2963                 unlock_buffer(bh);
2964         }
2965 }
2966
2967 /*
2968  * For a data-integrity writeout, we need to wait upon any in-progress I/O
2969  * and then start new I/O and then wait upon it.  The caller must have a ref on
2970  * the buffer_head.
2971  */
2972 int sync_dirty_buffer(struct buffer_head *bh)
2973 {
2974         int ret = 0;
2975
2976         WARN_ON(atomic_read(&bh->b_count) < 1);
2977         lock_buffer(bh);
2978         if (test_clear_buffer_dirty(bh)) {
2979                 get_bh(bh);
2980                 bh->b_end_io = end_buffer_write_sync;
2981                 ret = submit_bh(WRITE, bh);
2982                 wait_on_buffer(bh);
2983                 if (buffer_eopnotsupp(bh)) {
2984                         clear_buffer_eopnotsupp(bh);
2985                         ret = -EOPNOTSUPP;
2986                 }
2987                 if (!ret && !buffer_uptodate(bh))
2988                         ret = -EIO;
2989         } else {
2990                 unlock_buffer(bh);
2991         }
2992         return ret;
2993 }
2994
2995 /*
2996  * try_to_free_buffers() checks if all the buffers on this particular page
2997  * are unused, and releases them if so.
2998  *
2999  * Exclusion against try_to_free_buffers may be obtained by either
3000  * locking the page or by holding its mapping's private_lock.
3001  *
3002  * If the page is dirty but all the buffers are clean then we need to
3003  * be sure to mark the page clean as well.  This is because the page
3004  * may be against a block device, and a later reattachment of buffers
3005  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3006  * filesystem data on the same device.
3007  *
3008  * The same applies to regular filesystem pages: if all the buffers are
3009  * clean then we set the page clean and proceed.  To do that, we require
3010  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3011  * private_lock.
3012  *
3013  * try_to_free_buffers() is non-blocking.
3014  */
3015 static inline int buffer_busy(struct buffer_head *bh)
3016 {
3017         return atomic_read(&bh->b_count) |
3018                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3019 }
3020
3021 static int
3022 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3023 {
3024         struct buffer_head *head = page_buffers(page);
3025         struct buffer_head *bh;
3026
3027         bh = head;
3028         do {
3029                 if (buffer_write_io_error(bh) && page->mapping)
3030                         set_bit(AS_EIO, &page->mapping->flags);
3031                 if (buffer_busy(bh))
3032                         goto failed;
3033                 bh = bh->b_this_page;
3034         } while (bh != head);
3035
3036         do {
3037                 struct buffer_head *next = bh->b_this_page;
3038
3039                 if (bh->b_assoc_map)
3040                         __remove_assoc_queue(bh);
3041                 bh = next;
3042         } while (bh != head);
3043         *buffers_to_free = head;
3044         __clear_page_buffers(page);
3045         return 1;
3046 failed:
3047         return 0;
3048 }
3049
3050 int try_to_free_buffers(struct page *page)
3051 {
3052         struct address_space * const mapping = page->mapping;
3053         struct buffer_head *buffers_to_free = NULL;
3054         int ret = 0;
3055
3056         BUG_ON(!PageLocked(page));
3057         if (PageWriteback(page))
3058                 return 0;
3059
3060         if (mapping == NULL) {          /* can this still happen? */
3061                 ret = drop_buffers(page, &buffers_to_free);
3062                 goto out;
3063         }
3064
3065         spin_lock(&mapping->private_lock);
3066         ret = drop_buffers(page, &buffers_to_free);
3067
3068         /*
3069          * If the filesystem writes its buffers by hand (eg ext3)
3070          * then we can have clean buffers against a dirty page.  We
3071          * clean the page here; otherwise the VM will never notice
3072          * that the filesystem did any IO at all.
3073          *
3074          * Also, during truncate, discard_buffer will have marked all
3075          * the page's buffers clean.  We discover that here and clean
3076          * the page also.
3077          *
3078          * private_lock must be held over this entire operation in order
3079          * to synchronise against __set_page_dirty_buffers and prevent the
3080          * dirty bit from being lost.
3081          */
3082         if (ret)
3083                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3084         spin_unlock(&mapping->private_lock);
3085 out:
3086         if (buffers_to_free) {
3087                 struct buffer_head *bh = buffers_to_free;
3088
3089                 do {
3090                         struct buffer_head *next = bh->b_this_page;
3091                         free_buffer_head(bh);
3092                         bh = next;
3093                 } while (bh != buffers_to_free);
3094         }
3095         return ret;
3096 }
3097 EXPORT_SYMBOL(try_to_free_buffers);
3098
3099 void block_sync_page(struct page *page)
3100 {
3101         struct address_space *mapping;
3102
3103         smp_mb();
3104         mapping = page_mapping(page);
3105         if (mapping)
3106                 blk_run_backing_dev(mapping->backing_dev_info, page);
3107 }
3108
3109 /*
3110  * There are no bdflush tunables left.  But distributions are
3111  * still running obsolete flush daemons, so we terminate them here.
3112  *
3113  * Use of bdflush() is deprecated and will be removed in a future kernel.
3114  * The `pdflush' kernel threads fully replace bdflush daemons and this call.
3115  */
3116 asmlinkage long sys_bdflush(int func, long data)
3117 {
3118         static int msg_count;
3119
3120         if (!capable(CAP_SYS_ADMIN))
3121                 return -EPERM;
3122
3123         if (msg_count < 5) {
3124                 msg_count++;
3125                 printk(KERN_INFO
3126                         "warning: process `%s' used the obsolete bdflush"
3127                         " system call\n", current->comm);
3128                 printk(KERN_INFO "Fix your initscripts?\n");
3129         }
3130
3131         if (func == 1)
3132                 do_exit(0);
3133         return 0;
3134 }
3135
3136 /*
3137  * Buffer-head allocation
3138  */
3139 static struct kmem_cache *bh_cachep;
3140
3141 /*
3142  * Once the number of bh's in the machine exceeds this level, we start
3143  * stripping them in writeback.
3144  */
3145 static int max_buffer_heads;
3146
3147 int buffer_heads_over_limit;
3148
3149 struct bh_accounting {
3150         int nr;                 /* Number of live bh's */
3151         int ratelimit;          /* Limit cacheline bouncing */
3152 };
3153
3154 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3155
3156 static void recalc_bh_state(void)
3157 {
3158         int i;
3159         int tot = 0;
3160
3161         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3162                 return;
3163         __get_cpu_var(bh_accounting).ratelimit = 0;
3164         for_each_online_cpu(i)
3165                 tot += per_cpu(bh_accounting, i).nr;
3166         buffer_heads_over_limit = (tot > max_buffer_heads);
3167 }
3168         
3169 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3170 {
3171         struct buffer_head *ret = kmem_cache_alloc(bh_cachep,
3172                                 set_migrateflags(gfp_flags, __GFP_RECLAIMABLE));
3173         if (ret) {
3174                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3175                 get_cpu_var(bh_accounting).nr++;
3176                 recalc_bh_state();
3177                 put_cpu_var(bh_accounting);
3178         }
3179         return ret;
3180 }
3181 EXPORT_SYMBOL(alloc_buffer_head);
3182
3183 void free_buffer_head(struct buffer_head *bh)
3184 {
3185         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3186         kmem_cache_free(bh_cachep, bh);
3187         get_cpu_var(bh_accounting).nr--;
3188         recalc_bh_state();
3189         put_cpu_var(bh_accounting);
3190 }
3191 EXPORT_SYMBOL(free_buffer_head);
3192
3193 static void buffer_exit_cpu(int cpu)
3194 {
3195         int i;
3196         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3197
3198         for (i = 0; i < BH_LRU_SIZE; i++) {
3199                 brelse(b->bhs[i]);
3200                 b->bhs[i] = NULL;
3201         }
3202         get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3203         per_cpu(bh_accounting, cpu).nr = 0;
3204         put_cpu_var(bh_accounting);
3205 }
3206
3207 static int buffer_cpu_notify(struct notifier_block *self,
3208                               unsigned long action, void *hcpu)
3209 {
3210         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3211                 buffer_exit_cpu((unsigned long)hcpu);
3212         return NOTIFY_OK;
3213 }
3214
3215 /**
3216  * bh_uptodate_or_lock: Test whether the buffer is uptodate
3217  * @bh: struct buffer_head
3218  *
3219  * Return true if the buffer is up-to-date and false,
3220  * with the buffer locked, if not.
3221  */
3222 int bh_uptodate_or_lock(struct buffer_head *bh)
3223 {
3224         if (!buffer_uptodate(bh)) {
3225                 lock_buffer(bh);
3226                 if (!buffer_uptodate(bh))
3227                         return 0;
3228                 unlock_buffer(bh);
3229         }
3230         return 1;
3231 }
3232 EXPORT_SYMBOL(bh_uptodate_or_lock);
3233
3234 /**
3235  * bh_submit_read: Submit a locked buffer for reading
3236  * @bh: struct buffer_head
3237  *
3238  * Returns zero on success and -EIO on error.
3239  */
3240 int bh_submit_read(struct buffer_head *bh)
3241 {
3242         BUG_ON(!buffer_locked(bh));
3243
3244         if (buffer_uptodate(bh)) {
3245                 unlock_buffer(bh);
3246                 return 0;
3247         }
3248
3249         get_bh(bh);
3250         bh->b_end_io = end_buffer_read_sync;
3251         submit_bh(READ, bh);
3252         wait_on_buffer(bh);
3253         if (buffer_uptodate(bh))
3254                 return 0;
3255         return -EIO;
3256 }
3257 EXPORT_SYMBOL(bh_submit_read);
3258
3259 static void
3260 init_buffer_head(struct kmem_cache *cachep, void *data)
3261 {
3262         struct buffer_head *bh = data;
3263
3264         memset(bh, 0, sizeof(*bh));
3265         INIT_LIST_HEAD(&bh->b_assoc_buffers);
3266 }
3267
3268 void __init buffer_init(void)
3269 {
3270         int nrpages;
3271
3272         bh_cachep = kmem_cache_create("buffer_head",
3273                         sizeof(struct buffer_head), 0,
3274                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3275                                 SLAB_MEM_SPREAD),
3276                                 init_buffer_head);
3277
3278         /*
3279          * Limit the bh occupancy to 10% of ZONE_NORMAL
3280          */
3281         nrpages = (nr_free_buffer_pages() * 10) / 100;
3282         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3283         hotcpu_notifier(buffer_cpu_notify, 0);
3284 }
3285
3286 EXPORT_SYMBOL(__bforget);
3287 EXPORT_SYMBOL(__brelse);
3288 EXPORT_SYMBOL(__wait_on_buffer);
3289 EXPORT_SYMBOL(block_commit_write);
3290 EXPORT_SYMBOL(block_prepare_write);
3291 EXPORT_SYMBOL(block_page_mkwrite);
3292 EXPORT_SYMBOL(block_read_full_page);
3293 EXPORT_SYMBOL(block_sync_page);
3294 EXPORT_SYMBOL(block_truncate_page);
3295 EXPORT_SYMBOL(block_write_full_page);
3296 EXPORT_SYMBOL(cont_write_begin);
3297 EXPORT_SYMBOL(end_buffer_read_sync);
3298 EXPORT_SYMBOL(end_buffer_write_sync);
3299 EXPORT_SYMBOL(file_fsync);
3300 EXPORT_SYMBOL(fsync_bdev);
3301 EXPORT_SYMBOL(generic_block_bmap);
3302 EXPORT_SYMBOL(generic_commit_write);
3303 EXPORT_SYMBOL(generic_cont_expand_simple);
3304 EXPORT_SYMBOL(init_buffer);
3305 EXPORT_SYMBOL(invalidate_bdev);
3306 EXPORT_SYMBOL(ll_rw_block);
3307 EXPORT_SYMBOL(mark_buffer_dirty);
3308 EXPORT_SYMBOL(submit_bh);
3309 EXPORT_SYMBOL(sync_dirty_buffer);
3310 EXPORT_SYMBOL(unlock_buffer);