fs: new cont helpers
[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 fastcall __lock_buffer(struct buffer_head *bh)
71 {
72         wait_on_bit_lock(&bh->b_state, BH_Lock, sync_buffer,
73                                                         TASK_UNINTERRUPTIBLE);
74 }
75 EXPORT_SYMBOL(__lock_buffer);
76
77 void fastcall unlock_buffer(struct buffer_head *bh)
78 {
79         smp_mb__before_clear_bit();
80         clear_buffer_locked(bh);
81         smp_mb__after_clear_bit();
82         wake_up_bit(&bh->b_state, BH_Lock);
83 }
84
85 /*
86  * Block until a buffer comes unlocked.  This doesn't stop it
87  * from becoming locked again - you have to lock it yourself
88  * if you want to preserve its state.
89  */
90 void __wait_on_buffer(struct buffer_head * bh)
91 {
92         wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
93 }
94
95 static void
96 __clear_page_buffers(struct page *page)
97 {
98         ClearPagePrivate(page);
99         set_page_private(page, 0);
100         page_cache_release(page);
101 }
102
103 static void buffer_io_error(struct buffer_head *bh)
104 {
105         char b[BDEVNAME_SIZE];
106
107         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
108                         bdevname(bh->b_bdev, b),
109                         (unsigned long long)bh->b_blocknr);
110 }
111
112 /*
113  * End-of-IO handler helper function which does not touch the bh after
114  * unlocking it.
115  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
116  * a race there is benign: unlock_buffer() only use the bh's address for
117  * hashing after unlocking the buffer, so it doesn't actually touch the bh
118  * itself.
119  */
120 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
121 {
122         if (uptodate) {
123                 set_buffer_uptodate(bh);
124         } else {
125                 /* This happens, due to failed READA attempts. */
126                 clear_buffer_uptodate(bh);
127         }
128         unlock_buffer(bh);
129 }
130
131 /*
132  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
133  * unlock the buffer. This is what ll_rw_block uses too.
134  */
135 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
136 {
137         __end_buffer_read_notouch(bh, uptodate);
138         put_bh(bh);
139 }
140
141 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
142 {
143         char b[BDEVNAME_SIZE];
144
145         if (uptodate) {
146                 set_buffer_uptodate(bh);
147         } else {
148                 if (!buffer_eopnotsupp(bh) && printk_ratelimit()) {
149                         buffer_io_error(bh);
150                         printk(KERN_WARNING "lost page write due to "
151                                         "I/O error on %s\n",
152                                        bdevname(bh->b_bdev, b));
153                 }
154                 set_buffer_write_io_error(bh);
155                 clear_buffer_uptodate(bh);
156         }
157         unlock_buffer(bh);
158         put_bh(bh);
159 }
160
161 /*
162  * Write out and wait upon all the dirty data associated with a block
163  * device via its mapping.  Does not take the superblock lock.
164  */
165 int sync_blockdev(struct block_device *bdev)
166 {
167         int ret = 0;
168
169         if (bdev)
170                 ret = filemap_write_and_wait(bdev->bd_inode->i_mapping);
171         return ret;
172 }
173 EXPORT_SYMBOL(sync_blockdev);
174
175 /*
176  * Write out and wait upon all dirty data associated with this
177  * device.   Filesystem data as well as the underlying block
178  * device.  Takes the superblock lock.
179  */
180 int fsync_bdev(struct block_device *bdev)
181 {
182         struct super_block *sb = get_super(bdev);
183         if (sb) {
184                 int res = fsync_super(sb);
185                 drop_super(sb);
186                 return res;
187         }
188         return sync_blockdev(bdev);
189 }
190
191 /**
192  * freeze_bdev  --  lock a filesystem and force it into a consistent state
193  * @bdev:       blockdevice to lock
194  *
195  * This takes the block device bd_mount_sem to make sure no new mounts
196  * happen on bdev until thaw_bdev() is called.
197  * If a superblock is found on this device, we take the s_umount semaphore
198  * on it to make sure nobody unmounts until the snapshot creation is done.
199  */
200 struct super_block *freeze_bdev(struct block_device *bdev)
201 {
202         struct super_block *sb;
203
204         down(&bdev->bd_mount_sem);
205         sb = get_super(bdev);
206         if (sb && !(sb->s_flags & MS_RDONLY)) {
207                 sb->s_frozen = SB_FREEZE_WRITE;
208                 smp_wmb();
209
210                 __fsync_super(sb);
211
212                 sb->s_frozen = SB_FREEZE_TRANS;
213                 smp_wmb();
214
215                 sync_blockdev(sb->s_bdev);
216
217                 if (sb->s_op->write_super_lockfs)
218                         sb->s_op->write_super_lockfs(sb);
219         }
220
221         sync_blockdev(bdev);
222         return sb;      /* thaw_bdev releases s->s_umount and bd_mount_sem */
223 }
224 EXPORT_SYMBOL(freeze_bdev);
225
226 /**
227  * thaw_bdev  -- unlock filesystem
228  * @bdev:       blockdevice to unlock
229  * @sb:         associated superblock
230  *
231  * Unlocks the filesystem and marks it writeable again after freeze_bdev().
232  */
233 void thaw_bdev(struct block_device *bdev, struct super_block *sb)
234 {
235         if (sb) {
236                 BUG_ON(sb->s_bdev != bdev);
237
238                 if (sb->s_op->unlockfs)
239                         sb->s_op->unlockfs(sb);
240                 sb->s_frozen = SB_UNFROZEN;
241                 smp_wmb();
242                 wake_up(&sb->s_wait_unfrozen);
243                 drop_super(sb);
244         }
245
246         up(&bdev->bd_mount_sem);
247 }
248 EXPORT_SYMBOL(thaw_bdev);
249
250 /*
251  * Various filesystems appear to want __find_get_block to be non-blocking.
252  * But it's the page lock which protects the buffers.  To get around this,
253  * we get exclusion from try_to_free_buffers with the blockdev mapping's
254  * private_lock.
255  *
256  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
257  * may be quite high.  This code could TryLock the page, and if that
258  * succeeds, there is no need to take private_lock. (But if
259  * private_lock is contended then so is mapping->tree_lock).
260  */
261 static struct buffer_head *
262 __find_get_block_slow(struct block_device *bdev, sector_t block)
263 {
264         struct inode *bd_inode = bdev->bd_inode;
265         struct address_space *bd_mapping = bd_inode->i_mapping;
266         struct buffer_head *ret = NULL;
267         pgoff_t index;
268         struct buffer_head *bh;
269         struct buffer_head *head;
270         struct page *page;
271         int all_mapped = 1;
272
273         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
274         page = find_get_page(bd_mapping, index);
275         if (!page)
276                 goto out;
277
278         spin_lock(&bd_mapping->private_lock);
279         if (!page_has_buffers(page))
280                 goto out_unlock;
281         head = page_buffers(page);
282         bh = head;
283         do {
284                 if (bh->b_blocknr == block) {
285                         ret = bh;
286                         get_bh(bh);
287                         goto out_unlock;
288                 }
289                 if (!buffer_mapped(bh))
290                         all_mapped = 0;
291                 bh = bh->b_this_page;
292         } while (bh != head);
293
294         /* we might be here because some of the buffers on this page are
295          * not mapped.  This is due to various races between
296          * file io on the block device and getblk.  It gets dealt with
297          * elsewhere, don't buffer_error if we had some unmapped buffers
298          */
299         if (all_mapped) {
300                 printk("__find_get_block_slow() failed. "
301                         "block=%llu, b_blocknr=%llu\n",
302                         (unsigned long long)block,
303                         (unsigned long long)bh->b_blocknr);
304                 printk("b_state=0x%08lx, b_size=%zu\n",
305                         bh->b_state, bh->b_size);
306                 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
307         }
308 out_unlock:
309         spin_unlock(&bd_mapping->private_lock);
310         page_cache_release(page);
311 out:
312         return ret;
313 }
314
315 /* If invalidate_buffers() will trash dirty buffers, it means some kind
316    of fs corruption is going on. Trashing dirty data always imply losing
317    information that was supposed to be just stored on the physical layer
318    by the user.
319
320    Thus invalidate_buffers in general usage is not allwowed to trash
321    dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
322    be preserved.  These buffers are simply skipped.
323   
324    We also skip buffers which are still in use.  For example this can
325    happen if a userspace program is reading the block device.
326
327    NOTE: In the case where the user removed a removable-media-disk even if
328    there's still dirty data not synced on disk (due a bug in the device driver
329    or due an error of the user), by not destroying the dirty buffers we could
330    generate corruption also on the next media inserted, thus a parameter is
331    necessary to handle this case in the most safe way possible (trying
332    to not corrupt also the new disk inserted with the data belonging to
333    the old now corrupted disk). Also for the ramdisk the natural thing
334    to do in order to release the ramdisk memory is to destroy dirty buffers.
335
336    These are two special cases. Normal usage imply the device driver
337    to issue a sync on the device (without waiting I/O completion) and
338    then an invalidate_buffers call that doesn't trash dirty buffers.
339
340    For handling cache coherency with the blkdev pagecache the 'update' case
341    is been introduced. It is needed to re-read from disk any pinned
342    buffer. NOTE: re-reading from disk is destructive so we can do it only
343    when we assume nobody is changing the buffercache under our I/O and when
344    we think the disk contains more recent information than the buffercache.
345    The update == 1 pass marks the buffers we need to update, the update == 2
346    pass does the actual I/O. */
347 void invalidate_bdev(struct block_device *bdev)
348 {
349         struct address_space *mapping = bdev->bd_inode->i_mapping;
350
351         if (mapping->nrpages == 0)
352                 return;
353
354         invalidate_bh_lrus();
355         invalidate_mapping_pages(mapping, 0, -1);
356 }
357
358 /*
359  * Kick pdflush then try to free up some ZONE_NORMAL memory.
360  */
361 static void free_more_memory(void)
362 {
363         struct zone **zones;
364         pg_data_t *pgdat;
365
366         wakeup_pdflush(1024);
367         yield();
368
369         for_each_online_pgdat(pgdat) {
370                 zones = pgdat->node_zonelists[gfp_zone(GFP_NOFS)].zones;
371                 if (*zones)
372                         try_to_free_pages(zones, 0, GFP_NOFS);
373         }
374 }
375
376 /*
377  * I/O completion handler for block_read_full_page() - pages
378  * which come unlocked at the end of I/O.
379  */
380 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
381 {
382         unsigned long flags;
383         struct buffer_head *first;
384         struct buffer_head *tmp;
385         struct page *page;
386         int page_uptodate = 1;
387
388         BUG_ON(!buffer_async_read(bh));
389
390         page = bh->b_page;
391         if (uptodate) {
392                 set_buffer_uptodate(bh);
393         } else {
394                 clear_buffer_uptodate(bh);
395                 if (printk_ratelimit())
396                         buffer_io_error(bh);
397                 SetPageError(page);
398         }
399
400         /*
401          * Be _very_ careful from here on. Bad things can happen if
402          * two buffer heads end IO at almost the same time and both
403          * decide that the page is now completely done.
404          */
405         first = page_buffers(page);
406         local_irq_save(flags);
407         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
408         clear_buffer_async_read(bh);
409         unlock_buffer(bh);
410         tmp = bh;
411         do {
412                 if (!buffer_uptodate(tmp))
413                         page_uptodate = 0;
414                 if (buffer_async_read(tmp)) {
415                         BUG_ON(!buffer_locked(tmp));
416                         goto still_busy;
417                 }
418                 tmp = tmp->b_this_page;
419         } while (tmp != bh);
420         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
421         local_irq_restore(flags);
422
423         /*
424          * If none of the buffers had errors and they are all
425          * uptodate then we can set the page uptodate.
426          */
427         if (page_uptodate && !PageError(page))
428                 SetPageUptodate(page);
429         unlock_page(page);
430         return;
431
432 still_busy:
433         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
434         local_irq_restore(flags);
435         return;
436 }
437
438 /*
439  * Completion handler for block_write_full_page() - pages which are unlocked
440  * during I/O, and which have PageWriteback cleared upon I/O completion.
441  */
442 static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
443 {
444         char b[BDEVNAME_SIZE];
445         unsigned long flags;
446         struct buffer_head *first;
447         struct buffer_head *tmp;
448         struct page *page;
449
450         BUG_ON(!buffer_async_write(bh));
451
452         page = bh->b_page;
453         if (uptodate) {
454                 set_buffer_uptodate(bh);
455         } else {
456                 if (printk_ratelimit()) {
457                         buffer_io_error(bh);
458                         printk(KERN_WARNING "lost page write due to "
459                                         "I/O error on %s\n",
460                                bdevname(bh->b_bdev, b));
461                 }
462                 set_bit(AS_EIO, &page->mapping->flags);
463                 set_buffer_write_io_error(bh);
464                 clear_buffer_uptodate(bh);
465                 SetPageError(page);
466         }
467
468         first = page_buffers(page);
469         local_irq_save(flags);
470         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
471
472         clear_buffer_async_write(bh);
473         unlock_buffer(bh);
474         tmp = bh->b_this_page;
475         while (tmp != bh) {
476                 if (buffer_async_write(tmp)) {
477                         BUG_ON(!buffer_locked(tmp));
478                         goto still_busy;
479                 }
480                 tmp = tmp->b_this_page;
481         }
482         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
483         local_irq_restore(flags);
484         end_page_writeback(page);
485         return;
486
487 still_busy:
488         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
489         local_irq_restore(flags);
490         return;
491 }
492
493 /*
494  * If a page's buffers are under async readin (end_buffer_async_read
495  * completion) then there is a possibility that another thread of
496  * control could lock one of the buffers after it has completed
497  * but while some of the other buffers have not completed.  This
498  * locked buffer would confuse end_buffer_async_read() into not unlocking
499  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
500  * that this buffer is not under async I/O.
501  *
502  * The page comes unlocked when it has no locked buffer_async buffers
503  * left.
504  *
505  * PageLocked prevents anyone starting new async I/O reads any of
506  * the buffers.
507  *
508  * PageWriteback is used to prevent simultaneous writeout of the same
509  * page.
510  *
511  * PageLocked prevents anyone from starting writeback of a page which is
512  * under read I/O (PageWriteback is only ever set against a locked page).
513  */
514 static void mark_buffer_async_read(struct buffer_head *bh)
515 {
516         bh->b_end_io = end_buffer_async_read;
517         set_buffer_async_read(bh);
518 }
519
520 void mark_buffer_async_write(struct buffer_head *bh)
521 {
522         bh->b_end_io = end_buffer_async_write;
523         set_buffer_async_write(bh);
524 }
525 EXPORT_SYMBOL(mark_buffer_async_write);
526
527
528 /*
529  * fs/buffer.c contains helper functions for buffer-backed address space's
530  * fsync functions.  A common requirement for buffer-based filesystems is
531  * that certain data from the backing blockdev needs to be written out for
532  * a successful fsync().  For example, ext2 indirect blocks need to be
533  * written back and waited upon before fsync() returns.
534  *
535  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
536  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
537  * management of a list of dependent buffers at ->i_mapping->private_list.
538  *
539  * Locking is a little subtle: try_to_free_buffers() will remove buffers
540  * from their controlling inode's queue when they are being freed.  But
541  * try_to_free_buffers() will be operating against the *blockdev* mapping
542  * at the time, not against the S_ISREG file which depends on those buffers.
543  * So the locking for private_list is via the private_lock in the address_space
544  * which backs the buffers.  Which is different from the address_space 
545  * against which the buffers are listed.  So for a particular address_space,
546  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
547  * mapping->private_list will always be protected by the backing blockdev's
548  * ->private_lock.
549  *
550  * Which introduces a requirement: all buffers on an address_space's
551  * ->private_list must be from the same address_space: the blockdev's.
552  *
553  * address_spaces which do not place buffers at ->private_list via these
554  * utility functions are free to use private_lock and private_list for
555  * whatever they want.  The only requirement is that list_empty(private_list)
556  * be true at clear_inode() time.
557  *
558  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
559  * filesystems should do that.  invalidate_inode_buffers() should just go
560  * BUG_ON(!list_empty).
561  *
562  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
563  * take an address_space, not an inode.  And it should be called
564  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
565  * queued up.
566  *
567  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
568  * list if it is already on a list.  Because if the buffer is on a list,
569  * it *must* already be on the right one.  If not, the filesystem is being
570  * silly.  This will save a ton of locking.  But first we have to ensure
571  * that buffers are taken *off* the old inode's list when they are freed
572  * (presumably in truncate).  That requires careful auditing of all
573  * filesystems (do it inside bforget()).  It could also be done by bringing
574  * b_inode back.
575  */
576
577 /*
578  * The buffer's backing address_space's private_lock must be held
579  */
580 static inline void __remove_assoc_queue(struct buffer_head *bh)
581 {
582         list_del_init(&bh->b_assoc_buffers);
583         WARN_ON(!bh->b_assoc_map);
584         if (buffer_write_io_error(bh))
585                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
586         bh->b_assoc_map = NULL;
587 }
588
589 int inode_has_buffers(struct inode *inode)
590 {
591         return !list_empty(&inode->i_data.private_list);
592 }
593
594 /*
595  * osync is designed to support O_SYNC io.  It waits synchronously for
596  * all already-submitted IO to complete, but does not queue any new
597  * writes to the disk.
598  *
599  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
600  * you dirty the buffers, and then use osync_inode_buffers to wait for
601  * completion.  Any other dirty buffers which are not yet queued for
602  * write will not be flushed to disk by the osync.
603  */
604 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
605 {
606         struct buffer_head *bh;
607         struct list_head *p;
608         int err = 0;
609
610         spin_lock(lock);
611 repeat:
612         list_for_each_prev(p, list) {
613                 bh = BH_ENTRY(p);
614                 if (buffer_locked(bh)) {
615                         get_bh(bh);
616                         spin_unlock(lock);
617                         wait_on_buffer(bh);
618                         if (!buffer_uptodate(bh))
619                                 err = -EIO;
620                         brelse(bh);
621                         spin_lock(lock);
622                         goto repeat;
623                 }
624         }
625         spin_unlock(lock);
626         return err;
627 }
628
629 /**
630  * sync_mapping_buffers - write out and wait upon a mapping's "associated"
631  *                        buffers
632  * @mapping: the mapping which wants those buffers written
633  *
634  * Starts I/O against the buffers at mapping->private_list, and waits upon
635  * that I/O.
636  *
637  * Basically, this is a convenience function for fsync().
638  * @mapping is a file or directory which needs those buffers to be written for
639  * a successful fsync().
640  */
641 int sync_mapping_buffers(struct address_space *mapping)
642 {
643         struct address_space *buffer_mapping = mapping->assoc_mapping;
644
645         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
646                 return 0;
647
648         return fsync_buffers_list(&buffer_mapping->private_lock,
649                                         &mapping->private_list);
650 }
651 EXPORT_SYMBOL(sync_mapping_buffers);
652
653 /*
654  * Called when we've recently written block `bblock', and it is known that
655  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
656  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
657  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
658  */
659 void write_boundary_block(struct block_device *bdev,
660                         sector_t bblock, unsigned blocksize)
661 {
662         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
663         if (bh) {
664                 if (buffer_dirty(bh))
665                         ll_rw_block(WRITE, 1, &bh);
666                 put_bh(bh);
667         }
668 }
669
670 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
671 {
672         struct address_space *mapping = inode->i_mapping;
673         struct address_space *buffer_mapping = bh->b_page->mapping;
674
675         mark_buffer_dirty(bh);
676         if (!mapping->assoc_mapping) {
677                 mapping->assoc_mapping = buffer_mapping;
678         } else {
679                 BUG_ON(mapping->assoc_mapping != buffer_mapping);
680         }
681         if (list_empty(&bh->b_assoc_buffers)) {
682                 spin_lock(&buffer_mapping->private_lock);
683                 list_move_tail(&bh->b_assoc_buffers,
684                                 &mapping->private_list);
685                 bh->b_assoc_map = mapping;
686                 spin_unlock(&buffer_mapping->private_lock);
687         }
688 }
689 EXPORT_SYMBOL(mark_buffer_dirty_inode);
690
691 /*
692  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
693  * dirty.
694  *
695  * If warn is true, then emit a warning if the page is not uptodate and has
696  * not been truncated.
697  */
698 static int __set_page_dirty(struct page *page,
699                 struct address_space *mapping, int warn)
700 {
701         if (unlikely(!mapping))
702                 return !TestSetPageDirty(page);
703
704         if (TestSetPageDirty(page))
705                 return 0;
706
707         write_lock_irq(&mapping->tree_lock);
708         if (page->mapping) {    /* Race with truncate? */
709                 WARN_ON_ONCE(warn && !PageUptodate(page));
710
711                 if (mapping_cap_account_dirty(mapping)) {
712                         __inc_zone_page_state(page, NR_FILE_DIRTY);
713                         task_io_account_write(PAGE_CACHE_SIZE);
714                 }
715                 radix_tree_tag_set(&mapping->page_tree,
716                                 page_index(page), PAGECACHE_TAG_DIRTY);
717         }
718         write_unlock_irq(&mapping->tree_lock);
719         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
720
721         return 1;
722 }
723
724 /*
725  * Add a page to the dirty page list.
726  *
727  * It is a sad fact of life that this function is called from several places
728  * deeply under spinlocking.  It may not sleep.
729  *
730  * If the page has buffers, the uptodate buffers are set dirty, to preserve
731  * dirty-state coherency between the page and the buffers.  It the page does
732  * not have buffers then when they are later attached they will all be set
733  * dirty.
734  *
735  * The buffers are dirtied before the page is dirtied.  There's a small race
736  * window in which a writepage caller may see the page cleanness but not the
737  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
738  * before the buffers, a concurrent writepage caller could clear the page dirty
739  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
740  * page on the dirty page list.
741  *
742  * We use private_lock to lock against try_to_free_buffers while using the
743  * page's buffer list.  Also use this to protect against clean buffers being
744  * added to the page after it was set dirty.
745  *
746  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
747  * address_space though.
748  */
749 int __set_page_dirty_buffers(struct page *page)
750 {
751         struct address_space *mapping = page_mapping(page);
752
753         if (unlikely(!mapping))
754                 return !TestSetPageDirty(page);
755
756         spin_lock(&mapping->private_lock);
757         if (page_has_buffers(page)) {
758                 struct buffer_head *head = page_buffers(page);
759                 struct buffer_head *bh = head;
760
761                 do {
762                         set_buffer_dirty(bh);
763                         bh = bh->b_this_page;
764                 } while (bh != head);
765         }
766         spin_unlock(&mapping->private_lock);
767
768         return __set_page_dirty(page, mapping, 1);
769 }
770 EXPORT_SYMBOL(__set_page_dirty_buffers);
771
772 /*
773  * Write out and wait upon a list of buffers.
774  *
775  * We have conflicting pressures: we want to make sure that all
776  * initially dirty buffers get waited on, but that any subsequently
777  * dirtied buffers don't.  After all, we don't want fsync to last
778  * forever if somebody is actively writing to the file.
779  *
780  * Do this in two main stages: first we copy dirty buffers to a
781  * temporary inode list, queueing the writes as we go.  Then we clean
782  * up, waiting for those writes to complete.
783  * 
784  * During this second stage, any subsequent updates to the file may end
785  * up refiling the buffer on the original inode's dirty list again, so
786  * there is a chance we will end up with a buffer queued for write but
787  * not yet completed on that list.  So, as a final cleanup we go through
788  * the osync code to catch these locked, dirty buffers without requeuing
789  * any newly dirty buffers for write.
790  */
791 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
792 {
793         struct buffer_head *bh;
794         struct list_head tmp;
795         int err = 0, err2;
796
797         INIT_LIST_HEAD(&tmp);
798
799         spin_lock(lock);
800         while (!list_empty(list)) {
801                 bh = BH_ENTRY(list->next);
802                 __remove_assoc_queue(bh);
803                 if (buffer_dirty(bh) || buffer_locked(bh)) {
804                         list_add(&bh->b_assoc_buffers, &tmp);
805                         if (buffer_dirty(bh)) {
806                                 get_bh(bh);
807                                 spin_unlock(lock);
808                                 /*
809                                  * Ensure any pending I/O completes so that
810                                  * ll_rw_block() actually writes the current
811                                  * contents - it is a noop if I/O is still in
812                                  * flight on potentially older contents.
813                                  */
814                                 ll_rw_block(SWRITE, 1, &bh);
815                                 brelse(bh);
816                                 spin_lock(lock);
817                         }
818                 }
819         }
820
821         while (!list_empty(&tmp)) {
822                 bh = BH_ENTRY(tmp.prev);
823                 list_del_init(&bh->b_assoc_buffers);
824                 get_bh(bh);
825                 spin_unlock(lock);
826                 wait_on_buffer(bh);
827                 if (!buffer_uptodate(bh))
828                         err = -EIO;
829                 brelse(bh);
830                 spin_lock(lock);
831         }
832         
833         spin_unlock(lock);
834         err2 = osync_buffers_list(lock, list);
835         if (err)
836                 return err;
837         else
838                 return err2;
839 }
840
841 /*
842  * Invalidate any and all dirty buffers on a given inode.  We are
843  * probably unmounting the fs, but that doesn't mean we have already
844  * done a sync().  Just drop the buffers from the inode list.
845  *
846  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
847  * assumes that all the buffers are against the blockdev.  Not true
848  * for reiserfs.
849  */
850 void invalidate_inode_buffers(struct inode *inode)
851 {
852         if (inode_has_buffers(inode)) {
853                 struct address_space *mapping = &inode->i_data;
854                 struct list_head *list = &mapping->private_list;
855                 struct address_space *buffer_mapping = mapping->assoc_mapping;
856
857                 spin_lock(&buffer_mapping->private_lock);
858                 while (!list_empty(list))
859                         __remove_assoc_queue(BH_ENTRY(list->next));
860                 spin_unlock(&buffer_mapping->private_lock);
861         }
862 }
863
864 /*
865  * Remove any clean buffers from the inode's buffer list.  This is called
866  * when we're trying to free the inode itself.  Those buffers can pin it.
867  *
868  * Returns true if all buffers were removed.
869  */
870 int remove_inode_buffers(struct inode *inode)
871 {
872         int ret = 1;
873
874         if (inode_has_buffers(inode)) {
875                 struct address_space *mapping = &inode->i_data;
876                 struct list_head *list = &mapping->private_list;
877                 struct address_space *buffer_mapping = mapping->assoc_mapping;
878
879                 spin_lock(&buffer_mapping->private_lock);
880                 while (!list_empty(list)) {
881                         struct buffer_head *bh = BH_ENTRY(list->next);
882                         if (buffer_dirty(bh)) {
883                                 ret = 0;
884                                 break;
885                         }
886                         __remove_assoc_queue(bh);
887                 }
888                 spin_unlock(&buffer_mapping->private_lock);
889         }
890         return ret;
891 }
892
893 /*
894  * Create the appropriate buffers when given a page for data area and
895  * the size of each buffer.. Use the bh->b_this_page linked list to
896  * follow the buffers created.  Return NULL if unable to create more
897  * buffers.
898  *
899  * The retry flag is used to differentiate async IO (paging, swapping)
900  * which may not fail from ordinary buffer allocations.
901  */
902 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
903                 int retry)
904 {
905         struct buffer_head *bh, *head;
906         long offset;
907
908 try_again:
909         head = NULL;
910         offset = PAGE_SIZE;
911         while ((offset -= size) >= 0) {
912                 bh = alloc_buffer_head(GFP_NOFS);
913                 if (!bh)
914                         goto no_grow;
915
916                 bh->b_bdev = NULL;
917                 bh->b_this_page = head;
918                 bh->b_blocknr = -1;
919                 head = bh;
920
921                 bh->b_state = 0;
922                 atomic_set(&bh->b_count, 0);
923                 bh->b_private = NULL;
924                 bh->b_size = size;
925
926                 /* Link the buffer to its page */
927                 set_bh_page(bh, page, offset);
928
929                 init_buffer(bh, NULL, NULL);
930         }
931         return head;
932 /*
933  * In case anything failed, we just free everything we got.
934  */
935 no_grow:
936         if (head) {
937                 do {
938                         bh = head;
939                         head = head->b_this_page;
940                         free_buffer_head(bh);
941                 } while (head);
942         }
943
944         /*
945          * Return failure for non-async IO requests.  Async IO requests
946          * are not allowed to fail, so we have to wait until buffer heads
947          * become available.  But we don't want tasks sleeping with 
948          * partially complete buffers, so all were released above.
949          */
950         if (!retry)
951                 return NULL;
952
953         /* We're _really_ low on memory. Now we just
954          * wait for old buffer heads to become free due to
955          * finishing IO.  Since this is an async request and
956          * the reserve list is empty, we're sure there are 
957          * async buffer heads in use.
958          */
959         free_more_memory();
960         goto try_again;
961 }
962 EXPORT_SYMBOL_GPL(alloc_page_buffers);
963
964 static inline void
965 link_dev_buffers(struct page *page, struct buffer_head *head)
966 {
967         struct buffer_head *bh, *tail;
968
969         bh = head;
970         do {
971                 tail = bh;
972                 bh = bh->b_this_page;
973         } while (bh);
974         tail->b_this_page = head;
975         attach_page_buffers(page, head);
976 }
977
978 /*
979  * Initialise the state of a blockdev page's buffers.
980  */ 
981 static void
982 init_page_buffers(struct page *page, struct block_device *bdev,
983                         sector_t block, int size)
984 {
985         struct buffer_head *head = page_buffers(page);
986         struct buffer_head *bh = head;
987         int uptodate = PageUptodate(page);
988
989         do {
990                 if (!buffer_mapped(bh)) {
991                         init_buffer(bh, NULL, NULL);
992                         bh->b_bdev = bdev;
993                         bh->b_blocknr = block;
994                         if (uptodate)
995                                 set_buffer_uptodate(bh);
996                         set_buffer_mapped(bh);
997                 }
998                 block++;
999                 bh = bh->b_this_page;
1000         } while (bh != head);
1001 }
1002
1003 /*
1004  * Create the page-cache page that contains the requested block.
1005  *
1006  * This is user purely for blockdev mappings.
1007  */
1008 static struct page *
1009 grow_dev_page(struct block_device *bdev, sector_t block,
1010                 pgoff_t index, int size)
1011 {
1012         struct inode *inode = bdev->bd_inode;
1013         struct page *page;
1014         struct buffer_head *bh;
1015
1016         page = find_or_create_page(inode->i_mapping, index,
1017                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
1018         if (!page)
1019                 return NULL;
1020
1021         BUG_ON(!PageLocked(page));
1022
1023         if (page_has_buffers(page)) {
1024                 bh = page_buffers(page);
1025                 if (bh->b_size == size) {
1026                         init_page_buffers(page, bdev, block, size);
1027                         return page;
1028                 }
1029                 if (!try_to_free_buffers(page))
1030                         goto failed;
1031         }
1032
1033         /*
1034          * Allocate some buffers for this page
1035          */
1036         bh = alloc_page_buffers(page, size, 0);
1037         if (!bh)
1038                 goto failed;
1039
1040         /*
1041          * Link the page to the buffers and initialise them.  Take the
1042          * lock to be atomic wrt __find_get_block(), which does not
1043          * run under the page lock.
1044          */
1045         spin_lock(&inode->i_mapping->private_lock);
1046         link_dev_buffers(page, bh);
1047         init_page_buffers(page, bdev, block, size);
1048         spin_unlock(&inode->i_mapping->private_lock);
1049         return page;
1050
1051 failed:
1052         BUG();
1053         unlock_page(page);
1054         page_cache_release(page);
1055         return NULL;
1056 }
1057
1058 /*
1059  * Create buffers for the specified block device block's page.  If
1060  * that page was dirty, the buffers are set dirty also.
1061  */
1062 static int
1063 grow_buffers(struct block_device *bdev, sector_t block, int size)
1064 {
1065         struct page *page;
1066         pgoff_t index;
1067         int sizebits;
1068
1069         sizebits = -1;
1070         do {
1071                 sizebits++;
1072         } while ((size << sizebits) < PAGE_SIZE);
1073
1074         index = block >> sizebits;
1075
1076         /*
1077          * Check for a block which wants to lie outside our maximum possible
1078          * pagecache index.  (this comparison is done using sector_t types).
1079          */
1080         if (unlikely(index != block >> sizebits)) {
1081                 char b[BDEVNAME_SIZE];
1082
1083                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1084                         "device %s\n",
1085                         __FUNCTION__, (unsigned long long)block,
1086                         bdevname(bdev, b));
1087                 return -EIO;
1088         }
1089         block = index << sizebits;
1090         /* Create a page with the proper size buffers.. */
1091         page = grow_dev_page(bdev, block, index, size);
1092         if (!page)
1093                 return 0;
1094         unlock_page(page);
1095         page_cache_release(page);
1096         return 1;
1097 }
1098
1099 static struct buffer_head *
1100 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1101 {
1102         /* Size must be multiple of hard sectorsize */
1103         if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1104                         (size < 512 || size > PAGE_SIZE))) {
1105                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1106                                         size);
1107                 printk(KERN_ERR "hardsect size: %d\n",
1108                                         bdev_hardsect_size(bdev));
1109
1110                 dump_stack();
1111                 return NULL;
1112         }
1113
1114         for (;;) {
1115                 struct buffer_head * bh;
1116                 int ret;
1117
1118                 bh = __find_get_block(bdev, block, size);
1119                 if (bh)
1120                         return bh;
1121
1122                 ret = grow_buffers(bdev, block, size);
1123                 if (ret < 0)
1124                         return NULL;
1125                 if (ret == 0)
1126                         free_more_memory();
1127         }
1128 }
1129
1130 /*
1131  * The relationship between dirty buffers and dirty pages:
1132  *
1133  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1134  * the page is tagged dirty in its radix tree.
1135  *
1136  * At all times, the dirtiness of the buffers represents the dirtiness of
1137  * subsections of the page.  If the page has buffers, the page dirty bit is
1138  * merely a hint about the true dirty state.
1139  *
1140  * When a page is set dirty in its entirety, all its buffers are marked dirty
1141  * (if the page has buffers).
1142  *
1143  * When a buffer is marked dirty, its page is dirtied, but the page's other
1144  * buffers are not.
1145  *
1146  * Also.  When blockdev buffers are explicitly read with bread(), they
1147  * individually become uptodate.  But their backing page remains not
1148  * uptodate - even if all of its buffers are uptodate.  A subsequent
1149  * block_read_full_page() against that page will discover all the uptodate
1150  * buffers, will set the page uptodate and will perform no I/O.
1151  */
1152
1153 /**
1154  * mark_buffer_dirty - mark a buffer_head as needing writeout
1155  * @bh: the buffer_head to mark dirty
1156  *
1157  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1158  * backing page dirty, then tag the page as dirty in its address_space's radix
1159  * tree and then attach the address_space's inode to its superblock's dirty
1160  * inode list.
1161  *
1162  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1163  * mapping->tree_lock and the global inode_lock.
1164  */
1165 void fastcall mark_buffer_dirty(struct buffer_head *bh)
1166 {
1167         WARN_ON_ONCE(!buffer_uptodate(bh));
1168         if (!buffer_dirty(bh) && !test_set_buffer_dirty(bh))
1169                 __set_page_dirty(bh->b_page, page_mapping(bh->b_page), 0);
1170 }
1171
1172 /*
1173  * Decrement a buffer_head's reference count.  If all buffers against a page
1174  * have zero reference count, are clean and unlocked, and if the page is clean
1175  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1176  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1177  * a page but it ends up not being freed, and buffers may later be reattached).
1178  */
1179 void __brelse(struct buffer_head * buf)
1180 {
1181         if (atomic_read(&buf->b_count)) {
1182                 put_bh(buf);
1183                 return;
1184         }
1185         printk(KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1186         WARN_ON(1);
1187 }
1188
1189 /*
1190  * bforget() is like brelse(), except it discards any
1191  * potentially dirty data.
1192  */
1193 void __bforget(struct buffer_head *bh)
1194 {
1195         clear_buffer_dirty(bh);
1196         if (!list_empty(&bh->b_assoc_buffers)) {
1197                 struct address_space *buffer_mapping = bh->b_page->mapping;
1198
1199                 spin_lock(&buffer_mapping->private_lock);
1200                 list_del_init(&bh->b_assoc_buffers);
1201                 bh->b_assoc_map = NULL;
1202                 spin_unlock(&buffer_mapping->private_lock);
1203         }
1204         __brelse(bh);
1205 }
1206
1207 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1208 {
1209         lock_buffer(bh);
1210         if (buffer_uptodate(bh)) {
1211                 unlock_buffer(bh);
1212                 return bh;
1213         } else {
1214                 get_bh(bh);
1215                 bh->b_end_io = end_buffer_read_sync;
1216                 submit_bh(READ, bh);
1217                 wait_on_buffer(bh);
1218                 if (buffer_uptodate(bh))
1219                         return bh;
1220         }
1221         brelse(bh);
1222         return NULL;
1223 }
1224
1225 /*
1226  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1227  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1228  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1229  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1230  * CPU's LRUs at the same time.
1231  *
1232  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1233  * sb_find_get_block().
1234  *
1235  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1236  * a local interrupt disable for that.
1237  */
1238
1239 #define BH_LRU_SIZE     8
1240
1241 struct bh_lru {
1242         struct buffer_head *bhs[BH_LRU_SIZE];
1243 };
1244
1245 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1246
1247 #ifdef CONFIG_SMP
1248 #define bh_lru_lock()   local_irq_disable()
1249 #define bh_lru_unlock() local_irq_enable()
1250 #else
1251 #define bh_lru_lock()   preempt_disable()
1252 #define bh_lru_unlock() preempt_enable()
1253 #endif
1254
1255 static inline void check_irqs_on(void)
1256 {
1257 #ifdef irqs_disabled
1258         BUG_ON(irqs_disabled());
1259 #endif
1260 }
1261
1262 /*
1263  * The LRU management algorithm is dopey-but-simple.  Sorry.
1264  */
1265 static void bh_lru_install(struct buffer_head *bh)
1266 {
1267         struct buffer_head *evictee = NULL;
1268         struct bh_lru *lru;
1269
1270         check_irqs_on();
1271         bh_lru_lock();
1272         lru = &__get_cpu_var(bh_lrus);
1273         if (lru->bhs[0] != bh) {
1274                 struct buffer_head *bhs[BH_LRU_SIZE];
1275                 int in;
1276                 int out = 0;
1277
1278                 get_bh(bh);
1279                 bhs[out++] = bh;
1280                 for (in = 0; in < BH_LRU_SIZE; in++) {
1281                         struct buffer_head *bh2 = lru->bhs[in];
1282
1283                         if (bh2 == bh) {
1284                                 __brelse(bh2);
1285                         } else {
1286                                 if (out >= BH_LRU_SIZE) {
1287                                         BUG_ON(evictee != NULL);
1288                                         evictee = bh2;
1289                                 } else {
1290                                         bhs[out++] = bh2;
1291                                 }
1292                         }
1293                 }
1294                 while (out < BH_LRU_SIZE)
1295                         bhs[out++] = NULL;
1296                 memcpy(lru->bhs, bhs, sizeof(bhs));
1297         }
1298         bh_lru_unlock();
1299
1300         if (evictee)
1301                 __brelse(evictee);
1302 }
1303
1304 /*
1305  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1306  */
1307 static struct buffer_head *
1308 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1309 {
1310         struct buffer_head *ret = NULL;
1311         struct bh_lru *lru;
1312         unsigned int i;
1313
1314         check_irqs_on();
1315         bh_lru_lock();
1316         lru = &__get_cpu_var(bh_lrus);
1317         for (i = 0; i < BH_LRU_SIZE; i++) {
1318                 struct buffer_head *bh = lru->bhs[i];
1319
1320                 if (bh && bh->b_bdev == bdev &&
1321                                 bh->b_blocknr == block && bh->b_size == size) {
1322                         if (i) {
1323                                 while (i) {
1324                                         lru->bhs[i] = lru->bhs[i - 1];
1325                                         i--;
1326                                 }
1327                                 lru->bhs[0] = bh;
1328                         }
1329                         get_bh(bh);
1330                         ret = bh;
1331                         break;
1332                 }
1333         }
1334         bh_lru_unlock();
1335         return ret;
1336 }
1337
1338 /*
1339  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1340  * it in the LRU and mark it as accessed.  If it is not present then return
1341  * NULL
1342  */
1343 struct buffer_head *
1344 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1345 {
1346         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1347
1348         if (bh == NULL) {
1349                 bh = __find_get_block_slow(bdev, block);
1350                 if (bh)
1351                         bh_lru_install(bh);
1352         }
1353         if (bh)
1354                 touch_buffer(bh);
1355         return bh;
1356 }
1357 EXPORT_SYMBOL(__find_get_block);
1358
1359 /*
1360  * __getblk will locate (and, if necessary, create) the buffer_head
1361  * which corresponds to the passed block_device, block and size. The
1362  * returned buffer has its reference count incremented.
1363  *
1364  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1365  * illegal block number, __getblk() will happily return a buffer_head
1366  * which represents the non-existent block.  Very weird.
1367  *
1368  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1369  * attempt is failing.  FIXME, perhaps?
1370  */
1371 struct buffer_head *
1372 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1373 {
1374         struct buffer_head *bh = __find_get_block(bdev, block, size);
1375
1376         might_sleep();
1377         if (bh == NULL)
1378                 bh = __getblk_slow(bdev, block, size);
1379         return bh;
1380 }
1381 EXPORT_SYMBOL(__getblk);
1382
1383 /*
1384  * Do async read-ahead on a buffer..
1385  */
1386 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1387 {
1388         struct buffer_head *bh = __getblk(bdev, block, size);
1389         if (likely(bh)) {
1390                 ll_rw_block(READA, 1, &bh);
1391                 brelse(bh);
1392         }
1393 }
1394 EXPORT_SYMBOL(__breadahead);
1395
1396 /**
1397  *  __bread() - reads a specified block and returns the bh
1398  *  @bdev: the block_device to read from
1399  *  @block: number of block
1400  *  @size: size (in bytes) to read
1401  * 
1402  *  Reads a specified block, and returns buffer head that contains it.
1403  *  It returns NULL if the block was unreadable.
1404  */
1405 struct buffer_head *
1406 __bread(struct block_device *bdev, sector_t block, unsigned size)
1407 {
1408         struct buffer_head *bh = __getblk(bdev, block, size);
1409
1410         if (likely(bh) && !buffer_uptodate(bh))
1411                 bh = __bread_slow(bh);
1412         return bh;
1413 }
1414 EXPORT_SYMBOL(__bread);
1415
1416 /*
1417  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1418  * This doesn't race because it runs in each cpu either in irq
1419  * or with preempt disabled.
1420  */
1421 static void invalidate_bh_lru(void *arg)
1422 {
1423         struct bh_lru *b = &get_cpu_var(bh_lrus);
1424         int i;
1425
1426         for (i = 0; i < BH_LRU_SIZE; i++) {
1427                 brelse(b->bhs[i]);
1428                 b->bhs[i] = NULL;
1429         }
1430         put_cpu_var(bh_lrus);
1431 }
1432         
1433 void invalidate_bh_lrus(void)
1434 {
1435         on_each_cpu(invalidate_bh_lru, NULL, 1, 1);
1436 }
1437
1438 void set_bh_page(struct buffer_head *bh,
1439                 struct page *page, unsigned long offset)
1440 {
1441         bh->b_page = page;
1442         BUG_ON(offset >= PAGE_SIZE);
1443         if (PageHighMem(page))
1444                 /*
1445                  * This catches illegal uses and preserves the offset:
1446                  */
1447                 bh->b_data = (char *)(0 + offset);
1448         else
1449                 bh->b_data = page_address(page) + offset;
1450 }
1451 EXPORT_SYMBOL(set_bh_page);
1452
1453 /*
1454  * Called when truncating a buffer on a page completely.
1455  */
1456 static void discard_buffer(struct buffer_head * bh)
1457 {
1458         lock_buffer(bh);
1459         clear_buffer_dirty(bh);
1460         bh->b_bdev = NULL;
1461         clear_buffer_mapped(bh);
1462         clear_buffer_req(bh);
1463         clear_buffer_new(bh);
1464         clear_buffer_delay(bh);
1465         clear_buffer_unwritten(bh);
1466         unlock_buffer(bh);
1467 }
1468
1469 /**
1470  * block_invalidatepage - invalidate part of all of a buffer-backed page
1471  *
1472  * @page: the page which is affected
1473  * @offset: the index of the truncation point
1474  *
1475  * block_invalidatepage() is called when all or part of the page has become
1476  * invalidatedby a truncate operation.
1477  *
1478  * block_invalidatepage() does not have to release all buffers, but it must
1479  * ensure that no dirty buffer is left outside @offset and that no I/O
1480  * is underway against any of the blocks which are outside the truncation
1481  * point.  Because the caller is about to free (and possibly reuse) those
1482  * blocks on-disk.
1483  */
1484 void block_invalidatepage(struct page *page, unsigned long offset)
1485 {
1486         struct buffer_head *head, *bh, *next;
1487         unsigned int curr_off = 0;
1488
1489         BUG_ON(!PageLocked(page));
1490         if (!page_has_buffers(page))
1491                 goto out;
1492
1493         head = page_buffers(page);
1494         bh = head;
1495         do {
1496                 unsigned int next_off = curr_off + bh->b_size;
1497                 next = bh->b_this_page;
1498
1499                 /*
1500                  * is this block fully invalidated?
1501                  */
1502                 if (offset <= curr_off)
1503                         discard_buffer(bh);
1504                 curr_off = next_off;
1505                 bh = next;
1506         } while (bh != head);
1507
1508         /*
1509          * We release buffers only if the entire page is being invalidated.
1510          * The get_block cached value has been unconditionally invalidated,
1511          * so real IO is not possible anymore.
1512          */
1513         if (offset == 0)
1514                 try_to_release_page(page, 0);
1515 out:
1516         return;
1517 }
1518 EXPORT_SYMBOL(block_invalidatepage);
1519
1520 /*
1521  * We attach and possibly dirty the buffers atomically wrt
1522  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1523  * is already excluded via the page lock.
1524  */
1525 void create_empty_buffers(struct page *page,
1526                         unsigned long blocksize, unsigned long b_state)
1527 {
1528         struct buffer_head *bh, *head, *tail;
1529
1530         head = alloc_page_buffers(page, blocksize, 1);
1531         bh = head;
1532         do {
1533                 bh->b_state |= b_state;
1534                 tail = bh;
1535                 bh = bh->b_this_page;
1536         } while (bh);
1537         tail->b_this_page = head;
1538
1539         spin_lock(&page->mapping->private_lock);
1540         if (PageUptodate(page) || PageDirty(page)) {
1541                 bh = head;
1542                 do {
1543                         if (PageDirty(page))
1544                                 set_buffer_dirty(bh);
1545                         if (PageUptodate(page))
1546                                 set_buffer_uptodate(bh);
1547                         bh = bh->b_this_page;
1548                 } while (bh != head);
1549         }
1550         attach_page_buffers(page, head);
1551         spin_unlock(&page->mapping->private_lock);
1552 }
1553 EXPORT_SYMBOL(create_empty_buffers);
1554
1555 /*
1556  * We are taking a block for data and we don't want any output from any
1557  * buffer-cache aliases starting from return from that function and
1558  * until the moment when something will explicitly mark the buffer
1559  * dirty (hopefully that will not happen until we will free that block ;-)
1560  * We don't even need to mark it not-uptodate - nobody can expect
1561  * anything from a newly allocated buffer anyway. We used to used
1562  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1563  * don't want to mark the alias unmapped, for example - it would confuse
1564  * anyone who might pick it with bread() afterwards...
1565  *
1566  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1567  * be writeout I/O going on against recently-freed buffers.  We don't
1568  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1569  * only if we really need to.  That happens here.
1570  */
1571 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1572 {
1573         struct buffer_head *old_bh;
1574
1575         might_sleep();
1576
1577         old_bh = __find_get_block_slow(bdev, block);
1578         if (old_bh) {
1579                 clear_buffer_dirty(old_bh);
1580                 wait_on_buffer(old_bh);
1581                 clear_buffer_req(old_bh);
1582                 __brelse(old_bh);
1583         }
1584 }
1585 EXPORT_SYMBOL(unmap_underlying_metadata);
1586
1587 /*
1588  * NOTE! All mapped/uptodate combinations are valid:
1589  *
1590  *      Mapped  Uptodate        Meaning
1591  *
1592  *      No      No              "unknown" - must do get_block()
1593  *      No      Yes             "hole" - zero-filled
1594  *      Yes     No              "allocated" - allocated on disk, not read in
1595  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1596  *
1597  * "Dirty" is valid only with the last case (mapped+uptodate).
1598  */
1599
1600 /*
1601  * While block_write_full_page is writing back the dirty buffers under
1602  * the page lock, whoever dirtied the buffers may decide to clean them
1603  * again at any time.  We handle that by only looking at the buffer
1604  * state inside lock_buffer().
1605  *
1606  * If block_write_full_page() is called for regular writeback
1607  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1608  * locked buffer.   This only can happen if someone has written the buffer
1609  * directly, with submit_bh().  At the address_space level PageWriteback
1610  * prevents this contention from occurring.
1611  */
1612 static int __block_write_full_page(struct inode *inode, struct page *page,
1613                         get_block_t *get_block, struct writeback_control *wbc)
1614 {
1615         int err;
1616         sector_t block;
1617         sector_t last_block;
1618         struct buffer_head *bh, *head;
1619         const unsigned blocksize = 1 << inode->i_blkbits;
1620         int nr_underway = 0;
1621
1622         BUG_ON(!PageLocked(page));
1623
1624         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1625
1626         if (!page_has_buffers(page)) {
1627                 create_empty_buffers(page, blocksize,
1628                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1629         }
1630
1631         /*
1632          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1633          * here, and the (potentially unmapped) buffers may become dirty at
1634          * any time.  If a buffer becomes dirty here after we've inspected it
1635          * then we just miss that fact, and the page stays dirty.
1636          *
1637          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1638          * handle that here by just cleaning them.
1639          */
1640
1641         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1642         head = page_buffers(page);
1643         bh = head;
1644
1645         /*
1646          * Get all the dirty buffers mapped to disk addresses and
1647          * handle any aliases from the underlying blockdev's mapping.
1648          */
1649         do {
1650                 if (block > last_block) {
1651                         /*
1652                          * mapped buffers outside i_size will occur, because
1653                          * this page can be outside i_size when there is a
1654                          * truncate in progress.
1655                          */
1656                         /*
1657                          * The buffer was zeroed by block_write_full_page()
1658                          */
1659                         clear_buffer_dirty(bh);
1660                         set_buffer_uptodate(bh);
1661                 } else if (!buffer_mapped(bh) && buffer_dirty(bh)) {
1662                         WARN_ON(bh->b_size != blocksize);
1663                         err = get_block(inode, block, bh, 1);
1664                         if (err)
1665                                 goto recover;
1666                         if (buffer_new(bh)) {
1667                                 /* blockdev mappings never come here */
1668                                 clear_buffer_new(bh);
1669                                 unmap_underlying_metadata(bh->b_bdev,
1670                                                         bh->b_blocknr);
1671                         }
1672                 }
1673                 bh = bh->b_this_page;
1674                 block++;
1675         } while (bh != head);
1676
1677         do {
1678                 if (!buffer_mapped(bh))
1679                         continue;
1680                 /*
1681                  * If it's a fully non-blocking write attempt and we cannot
1682                  * lock the buffer then redirty the page.  Note that this can
1683                  * potentially cause a busy-wait loop from pdflush and kswapd
1684                  * activity, but those code paths have their own higher-level
1685                  * throttling.
1686                  */
1687                 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1688                         lock_buffer(bh);
1689                 } else if (test_set_buffer_locked(bh)) {
1690                         redirty_page_for_writepage(wbc, page);
1691                         continue;
1692                 }
1693                 if (test_clear_buffer_dirty(bh)) {
1694                         mark_buffer_async_write(bh);
1695                 } else {
1696                         unlock_buffer(bh);
1697                 }
1698         } while ((bh = bh->b_this_page) != head);
1699
1700         /*
1701          * The page and its buffers are protected by PageWriteback(), so we can
1702          * drop the bh refcounts early.
1703          */
1704         BUG_ON(PageWriteback(page));
1705         set_page_writeback(page);
1706
1707         do {
1708                 struct buffer_head *next = bh->b_this_page;
1709                 if (buffer_async_write(bh)) {
1710                         submit_bh(WRITE, bh);
1711                         nr_underway++;
1712                 }
1713                 bh = next;
1714         } while (bh != head);
1715         unlock_page(page);
1716
1717         err = 0;
1718 done:
1719         if (nr_underway == 0) {
1720                 /*
1721                  * The page was marked dirty, but the buffers were
1722                  * clean.  Someone wrote them back by hand with
1723                  * ll_rw_block/submit_bh.  A rare case.
1724                  */
1725                 end_page_writeback(page);
1726
1727                 /*
1728                  * The page and buffer_heads can be released at any time from
1729                  * here on.
1730                  */
1731                 wbc->pages_skipped++;   /* We didn't write this page */
1732         }
1733         return err;
1734
1735 recover:
1736         /*
1737          * ENOSPC, or some other error.  We may already have added some
1738          * blocks to the file, so we need to write these out to avoid
1739          * exposing stale data.
1740          * The page is currently locked and not marked for writeback
1741          */
1742         bh = head;
1743         /* Recovery: lock and submit the mapped buffers */
1744         do {
1745                 if (buffer_mapped(bh) && buffer_dirty(bh)) {
1746                         lock_buffer(bh);
1747                         mark_buffer_async_write(bh);
1748                 } else {
1749                         /*
1750                          * The buffer may have been set dirty during
1751                          * attachment to a dirty page.
1752                          */
1753                         clear_buffer_dirty(bh);
1754                 }
1755         } while ((bh = bh->b_this_page) != head);
1756         SetPageError(page);
1757         BUG_ON(PageWriteback(page));
1758         mapping_set_error(page->mapping, err);
1759         set_page_writeback(page);
1760         do {
1761                 struct buffer_head *next = bh->b_this_page;
1762                 if (buffer_async_write(bh)) {
1763                         clear_buffer_dirty(bh);
1764                         submit_bh(WRITE, bh);
1765                         nr_underway++;
1766                 }
1767                 bh = next;
1768         } while (bh != head);
1769         unlock_page(page);
1770         goto done;
1771 }
1772
1773 /*
1774  * If a page has any new buffers, zero them out here, and mark them uptodate
1775  * and dirty so they'll be written out (in order to prevent uninitialised
1776  * block data from leaking). And clear the new bit.
1777  */
1778 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1779 {
1780         unsigned int block_start, block_end;
1781         struct buffer_head *head, *bh;
1782
1783         BUG_ON(!PageLocked(page));
1784         if (!page_has_buffers(page))
1785                 return;
1786
1787         bh = head = page_buffers(page);
1788         block_start = 0;
1789         do {
1790                 block_end = block_start + bh->b_size;
1791
1792                 if (buffer_new(bh)) {
1793                         if (block_end > from && block_start < to) {
1794                                 if (!PageUptodate(page)) {
1795                                         unsigned start, size;
1796
1797                                         start = max(from, block_start);
1798                                         size = min(to, block_end) - start;
1799
1800                                         zero_user_page(page, start, size, KM_USER0);
1801                                         set_buffer_uptodate(bh);
1802                                 }
1803
1804                                 clear_buffer_new(bh);
1805                                 mark_buffer_dirty(bh);
1806                         }
1807                 }
1808
1809                 block_start = block_end;
1810                 bh = bh->b_this_page;
1811         } while (bh != head);
1812 }
1813 EXPORT_SYMBOL(page_zero_new_buffers);
1814
1815 static int __block_prepare_write(struct inode *inode, struct page *page,
1816                 unsigned from, unsigned to, get_block_t *get_block)
1817 {
1818         unsigned block_start, block_end;
1819         sector_t block;
1820         int err = 0;
1821         unsigned blocksize, bbits;
1822         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1823
1824         BUG_ON(!PageLocked(page));
1825         BUG_ON(from > PAGE_CACHE_SIZE);
1826         BUG_ON(to > PAGE_CACHE_SIZE);
1827         BUG_ON(from > to);
1828
1829         blocksize = 1 << inode->i_blkbits;
1830         if (!page_has_buffers(page))
1831                 create_empty_buffers(page, blocksize, 0);
1832         head = page_buffers(page);
1833
1834         bbits = inode->i_blkbits;
1835         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1836
1837         for(bh = head, block_start = 0; bh != head || !block_start;
1838             block++, block_start=block_end, bh = bh->b_this_page) {
1839                 block_end = block_start + blocksize;
1840                 if (block_end <= from || block_start >= to) {
1841                         if (PageUptodate(page)) {
1842                                 if (!buffer_uptodate(bh))
1843                                         set_buffer_uptodate(bh);
1844                         }
1845                         continue;
1846                 }
1847                 if (buffer_new(bh))
1848                         clear_buffer_new(bh);
1849                 if (!buffer_mapped(bh)) {
1850                         WARN_ON(bh->b_size != blocksize);
1851                         err = get_block(inode, block, bh, 1);
1852                         if (err)
1853                                 break;
1854                         if (buffer_new(bh)) {
1855                                 unmap_underlying_metadata(bh->b_bdev,
1856                                                         bh->b_blocknr);
1857                                 if (PageUptodate(page)) {
1858                                         clear_buffer_new(bh);
1859                                         set_buffer_uptodate(bh);
1860                                         mark_buffer_dirty(bh);
1861                                         continue;
1862                                 }
1863                                 if (block_end > to || block_start < from) {
1864                                         void *kaddr;
1865
1866                                         kaddr = kmap_atomic(page, KM_USER0);
1867                                         if (block_end > to)
1868                                                 memset(kaddr+to, 0,
1869                                                         block_end-to);
1870                                         if (block_start < from)
1871                                                 memset(kaddr+block_start,
1872                                                         0, from-block_start);
1873                                         flush_dcache_page(page);
1874                                         kunmap_atomic(kaddr, KM_USER0);
1875                                 }
1876                                 continue;
1877                         }
1878                 }
1879                 if (PageUptodate(page)) {
1880                         if (!buffer_uptodate(bh))
1881                                 set_buffer_uptodate(bh);
1882                         continue; 
1883                 }
1884                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1885                     !buffer_unwritten(bh) &&
1886                      (block_start < from || block_end > to)) {
1887                         ll_rw_block(READ, 1, &bh);
1888                         *wait_bh++=bh;
1889                 }
1890         }
1891         /*
1892          * If we issued read requests - let them complete.
1893          */
1894         while(wait_bh > wait) {
1895                 wait_on_buffer(*--wait_bh);
1896                 if (!buffer_uptodate(*wait_bh))
1897                         err = -EIO;
1898         }
1899         if (unlikely(err))
1900                 page_zero_new_buffers(page, from, to);
1901         return err;
1902 }
1903
1904 static int __block_commit_write(struct inode *inode, struct page *page,
1905                 unsigned from, unsigned to)
1906 {
1907         unsigned block_start, block_end;
1908         int partial = 0;
1909         unsigned blocksize;
1910         struct buffer_head *bh, *head;
1911
1912         blocksize = 1 << inode->i_blkbits;
1913
1914         for(bh = head = page_buffers(page), block_start = 0;
1915             bh != head || !block_start;
1916             block_start=block_end, bh = bh->b_this_page) {
1917                 block_end = block_start + blocksize;
1918                 if (block_end <= from || block_start >= to) {
1919                         if (!buffer_uptodate(bh))
1920                                 partial = 1;
1921                 } else {
1922                         set_buffer_uptodate(bh);
1923                         mark_buffer_dirty(bh);
1924                 }
1925                 clear_buffer_new(bh);
1926         }
1927
1928         /*
1929          * If this is a partial write which happened to make all buffers
1930          * uptodate then we can optimize away a bogus readpage() for
1931          * the next read(). Here we 'discover' whether the page went
1932          * uptodate as a result of this (potentially partial) write.
1933          */
1934         if (!partial)
1935                 SetPageUptodate(page);
1936         return 0;
1937 }
1938
1939 /*
1940  * block_write_begin takes care of the basic task of block allocation and
1941  * bringing partial write blocks uptodate first.
1942  *
1943  * If *pagep is not NULL, then block_write_begin uses the locked page
1944  * at *pagep rather than allocating its own. In this case, the page will
1945  * not be unlocked or deallocated on failure.
1946  */
1947 int block_write_begin(struct file *file, struct address_space *mapping,
1948                         loff_t pos, unsigned len, unsigned flags,
1949                         struct page **pagep, void **fsdata,
1950                         get_block_t *get_block)
1951 {
1952         struct inode *inode = mapping->host;
1953         int status = 0;
1954         struct page *page;
1955         pgoff_t index;
1956         unsigned start, end;
1957         int ownpage = 0;
1958
1959         index = pos >> PAGE_CACHE_SHIFT;
1960         start = pos & (PAGE_CACHE_SIZE - 1);
1961         end = start + len;
1962
1963         page = *pagep;
1964         if (page == NULL) {
1965                 ownpage = 1;
1966                 page = __grab_cache_page(mapping, index);
1967                 if (!page) {
1968                         status = -ENOMEM;
1969                         goto out;
1970                 }
1971                 *pagep = page;
1972         } else
1973                 BUG_ON(!PageLocked(page));
1974
1975         status = __block_prepare_write(inode, page, start, end, get_block);
1976         if (unlikely(status)) {
1977                 ClearPageUptodate(page);
1978
1979                 if (ownpage) {
1980                         unlock_page(page);
1981                         page_cache_release(page);
1982                         *pagep = NULL;
1983
1984                         /*
1985                          * prepare_write() may have instantiated a few blocks
1986                          * outside i_size.  Trim these off again. Don't need
1987                          * i_size_read because we hold i_mutex.
1988                          */
1989                         if (pos + len > inode->i_size)
1990                                 vmtruncate(inode, inode->i_size);
1991                 }
1992                 goto out;
1993         }
1994
1995 out:
1996         return status;
1997 }
1998 EXPORT_SYMBOL(block_write_begin);
1999
2000 int block_write_end(struct file *file, struct address_space *mapping,
2001                         loff_t pos, unsigned len, unsigned copied,
2002                         struct page *page, void *fsdata)
2003 {
2004         struct inode *inode = mapping->host;
2005         unsigned start;
2006
2007         start = pos & (PAGE_CACHE_SIZE - 1);
2008
2009         if (unlikely(copied < len)) {
2010                 /*
2011                  * The buffers that were written will now be uptodate, so we
2012                  * don't have to worry about a readpage reading them and
2013                  * overwriting a partial write. However if we have encountered
2014                  * a short write and only partially written into a buffer, it
2015                  * will not be marked uptodate, so a readpage might come in and
2016                  * destroy our partial write.
2017                  *
2018                  * Do the simplest thing, and just treat any short write to a
2019                  * non uptodate page as a zero-length write, and force the
2020                  * caller to redo the whole thing.
2021                  */
2022                 if (!PageUptodate(page))
2023                         copied = 0;
2024
2025                 page_zero_new_buffers(page, start+copied, start+len);
2026         }
2027         flush_dcache_page(page);
2028
2029         /* This could be a short (even 0-length) commit */
2030         __block_commit_write(inode, page, start, start+copied);
2031
2032         return copied;
2033 }
2034 EXPORT_SYMBOL(block_write_end);
2035
2036 int generic_write_end(struct file *file, struct address_space *mapping,
2037                         loff_t pos, unsigned len, unsigned copied,
2038                         struct page *page, void *fsdata)
2039 {
2040         struct inode *inode = mapping->host;
2041
2042         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2043
2044         /*
2045          * No need to use i_size_read() here, the i_size
2046          * cannot change under us because we hold i_mutex.
2047          *
2048          * But it's important to update i_size while still holding page lock:
2049          * page writeout could otherwise come in and zero beyond i_size.
2050          */
2051         if (pos+copied > inode->i_size) {
2052                 i_size_write(inode, pos+copied);
2053                 mark_inode_dirty(inode);
2054         }
2055
2056         unlock_page(page);
2057         page_cache_release(page);
2058
2059         return copied;
2060 }
2061 EXPORT_SYMBOL(generic_write_end);
2062
2063 /*
2064  * Generic "read page" function for block devices that have the normal
2065  * get_block functionality. This is most of the block device filesystems.
2066  * Reads the page asynchronously --- the unlock_buffer() and
2067  * set/clear_buffer_uptodate() functions propagate buffer state into the
2068  * page struct once IO has completed.
2069  */
2070 int block_read_full_page(struct page *page, get_block_t *get_block)
2071 {
2072         struct inode *inode = page->mapping->host;
2073         sector_t iblock, lblock;
2074         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2075         unsigned int blocksize;
2076         int nr, i;
2077         int fully_mapped = 1;
2078
2079         BUG_ON(!PageLocked(page));
2080         blocksize = 1 << inode->i_blkbits;
2081         if (!page_has_buffers(page))
2082                 create_empty_buffers(page, blocksize, 0);
2083         head = page_buffers(page);
2084
2085         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2086         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2087         bh = head;
2088         nr = 0;
2089         i = 0;
2090
2091         do {
2092                 if (buffer_uptodate(bh))
2093                         continue;
2094
2095                 if (!buffer_mapped(bh)) {
2096                         int err = 0;
2097
2098                         fully_mapped = 0;
2099                         if (iblock < lblock) {
2100                                 WARN_ON(bh->b_size != blocksize);
2101                                 err = get_block(inode, iblock, bh, 0);
2102                                 if (err)
2103                                         SetPageError(page);
2104                         }
2105                         if (!buffer_mapped(bh)) {
2106                                 zero_user_page(page, i * blocksize, blocksize,
2107                                                 KM_USER0);
2108                                 if (!err)
2109                                         set_buffer_uptodate(bh);
2110                                 continue;
2111                         }
2112                         /*
2113                          * get_block() might have updated the buffer
2114                          * synchronously
2115                          */
2116                         if (buffer_uptodate(bh))
2117                                 continue;
2118                 }
2119                 arr[nr++] = bh;
2120         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2121
2122         if (fully_mapped)
2123                 SetPageMappedToDisk(page);
2124
2125         if (!nr) {
2126                 /*
2127                  * All buffers are uptodate - we can set the page uptodate
2128                  * as well. But not if get_block() returned an error.
2129                  */
2130                 if (!PageError(page))
2131                         SetPageUptodate(page);
2132                 unlock_page(page);
2133                 return 0;
2134         }
2135
2136         /* Stage two: lock the buffers */
2137         for (i = 0; i < nr; i++) {
2138                 bh = arr[i];
2139                 lock_buffer(bh);
2140                 mark_buffer_async_read(bh);
2141         }
2142
2143         /*
2144          * Stage 3: start the IO.  Check for uptodateness
2145          * inside the buffer lock in case another process reading
2146          * the underlying blockdev brought it uptodate (the sct fix).
2147          */
2148         for (i = 0; i < nr; i++) {
2149                 bh = arr[i];
2150                 if (buffer_uptodate(bh))
2151                         end_buffer_async_read(bh, 1);
2152                 else
2153                         submit_bh(READ, bh);
2154         }
2155         return 0;
2156 }
2157
2158 /* utility function for filesystems that need to do work on expanding
2159  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2160  * deal with the hole.  
2161  */
2162 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2163 {
2164         struct address_space *mapping = inode->i_mapping;
2165         struct page *page;
2166         void *fsdata;
2167         unsigned long limit;
2168         int err;
2169
2170         err = -EFBIG;
2171         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2172         if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2173                 send_sig(SIGXFSZ, current, 0);
2174                 goto out;
2175         }
2176         if (size > inode->i_sb->s_maxbytes)
2177                 goto out;
2178
2179         err = pagecache_write_begin(NULL, mapping, size, 0,
2180                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2181                                 &page, &fsdata);
2182         if (err)
2183                 goto out;
2184
2185         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2186         BUG_ON(err > 0);
2187
2188 out:
2189         return err;
2190 }
2191
2192 int generic_cont_expand(struct inode *inode, loff_t size)
2193 {
2194         unsigned int offset;
2195
2196         offset = (size & (PAGE_CACHE_SIZE - 1)); /* Within page */
2197
2198         /* ugh.  in prepare/commit_write, if from==to==start of block, we
2199          * skip the prepare.  make sure we never send an offset for the start
2200          * of a block.
2201          * XXX: actually, this should be handled in those filesystems by
2202          * checking for the AOP_FLAG_CONT_EXPAND flag.
2203          */
2204         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
2205                 /* caller must handle this extra byte. */
2206                 size++;
2207         }
2208         return generic_cont_expand_simple(inode, size);
2209 }
2210
2211 int cont_expand_zero(struct file *file, struct address_space *mapping,
2212                         loff_t pos, loff_t *bytes)
2213 {
2214         struct inode *inode = mapping->host;
2215         unsigned blocksize = 1 << inode->i_blkbits;
2216         struct page *page;
2217         void *fsdata;
2218         pgoff_t index, curidx;
2219         loff_t curpos;
2220         unsigned zerofrom, offset, len;
2221         int err = 0;
2222
2223         index = pos >> PAGE_CACHE_SHIFT;
2224         offset = pos & ~PAGE_CACHE_MASK;
2225
2226         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2227                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2228                 if (zerofrom & (blocksize-1)) {
2229                         *bytes |= (blocksize-1);
2230                         (*bytes)++;
2231                 }
2232                 len = PAGE_CACHE_SIZE - zerofrom;
2233
2234                 err = pagecache_write_begin(file, mapping, curpos, len,
2235                                                 AOP_FLAG_UNINTERRUPTIBLE,
2236                                                 &page, &fsdata);
2237                 if (err)
2238                         goto out;
2239                 zero_user_page(page, zerofrom, len, KM_USER0);
2240                 err = pagecache_write_end(file, mapping, curpos, len, len,
2241                                                 page, fsdata);
2242                 if (err < 0)
2243                         goto out;
2244                 BUG_ON(err != len);
2245                 err = 0;
2246         }
2247
2248         /* page covers the boundary, find the boundary offset */
2249         if (index == curidx) {
2250                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2251                 /* if we will expand the thing last block will be filled */
2252                 if (offset <= zerofrom) {
2253                         goto out;
2254                 }
2255                 if (zerofrom & (blocksize-1)) {
2256                         *bytes |= (blocksize-1);
2257                         (*bytes)++;
2258                 }
2259                 len = offset - zerofrom;
2260
2261                 err = pagecache_write_begin(file, mapping, curpos, len,
2262                                                 AOP_FLAG_UNINTERRUPTIBLE,
2263                                                 &page, &fsdata);
2264                 if (err)
2265                         goto out;
2266                 zero_user_page(page, zerofrom, len, KM_USER0);
2267                 err = pagecache_write_end(file, mapping, curpos, len, len,
2268                                                 page, fsdata);
2269                 if (err < 0)
2270                         goto out;
2271                 BUG_ON(err != len);
2272                 err = 0;
2273         }
2274 out:
2275         return err;
2276 }
2277
2278 /*
2279  * For moronic filesystems that do not allow holes in file.
2280  * We may have to extend the file.
2281  */
2282 int cont_write_begin(struct file *file, struct address_space *mapping,
2283                         loff_t pos, unsigned len, unsigned flags,
2284                         struct page **pagep, void **fsdata,
2285                         get_block_t *get_block, loff_t *bytes)
2286 {
2287         struct inode *inode = mapping->host;
2288         unsigned blocksize = 1 << inode->i_blkbits;
2289         unsigned zerofrom;
2290         int err;
2291
2292         err = cont_expand_zero(file, mapping, pos, bytes);
2293         if (err)
2294                 goto out;
2295
2296         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2297         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2298                 *bytes |= (blocksize-1);
2299                 (*bytes)++;
2300         }
2301
2302         *pagep = NULL;
2303         err = block_write_begin(file, mapping, pos, len,
2304                                 flags, pagep, fsdata, get_block);
2305 out:
2306         return err;
2307 }
2308
2309 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2310                         get_block_t *get_block)
2311 {
2312         struct inode *inode = page->mapping->host;
2313         int err = __block_prepare_write(inode, page, from, to, get_block);
2314         if (err)
2315                 ClearPageUptodate(page);
2316         return err;
2317 }
2318
2319 int block_commit_write(struct page *page, unsigned from, unsigned to)
2320 {
2321         struct inode *inode = page->mapping->host;
2322         __block_commit_write(inode,page,from,to);
2323         return 0;
2324 }
2325
2326 int generic_commit_write(struct file *file, struct page *page,
2327                 unsigned from, unsigned to)
2328 {
2329         struct inode *inode = page->mapping->host;
2330         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2331         __block_commit_write(inode,page,from,to);
2332         /*
2333          * No need to use i_size_read() here, the i_size
2334          * cannot change under us because we hold i_mutex.
2335          */
2336         if (pos > inode->i_size) {
2337                 i_size_write(inode, pos);
2338                 mark_inode_dirty(inode);
2339         }
2340         return 0;
2341 }
2342
2343 /*
2344  * block_page_mkwrite() is not allowed to change the file size as it gets
2345  * called from a page fault handler when a page is first dirtied. Hence we must
2346  * be careful to check for EOF conditions here. We set the page up correctly
2347  * for a written page which means we get ENOSPC checking when writing into
2348  * holes and correct delalloc and unwritten extent mapping on filesystems that
2349  * support these features.
2350  *
2351  * We are not allowed to take the i_mutex here so we have to play games to
2352  * protect against truncate races as the page could now be beyond EOF.  Because
2353  * vmtruncate() writes the inode size before removing pages, once we have the
2354  * page lock we can determine safely if the page is beyond EOF. If it is not
2355  * beyond EOF, then the page is guaranteed safe against truncation until we
2356  * unlock the page.
2357  */
2358 int
2359 block_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2360                    get_block_t get_block)
2361 {
2362         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2363         unsigned long end;
2364         loff_t size;
2365         int ret = -EINVAL;
2366
2367         lock_page(page);
2368         size = i_size_read(inode);
2369         if ((page->mapping != inode->i_mapping) ||
2370             (page_offset(page) > size)) {
2371                 /* page got truncated out from underneath us */
2372                 goto out_unlock;
2373         }
2374
2375         /* page is wholly or partially inside EOF */
2376         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2377                 end = size & ~PAGE_CACHE_MASK;
2378         else
2379                 end = PAGE_CACHE_SIZE;
2380
2381         ret = block_prepare_write(page, 0, end, get_block);
2382         if (!ret)
2383                 ret = block_commit_write(page, 0, end);
2384
2385 out_unlock:
2386         unlock_page(page);
2387         return ret;
2388 }
2389
2390 /*
2391  * nobh_prepare_write()'s prereads are special: the buffer_heads are freed
2392  * immediately, while under the page lock.  So it needs a special end_io
2393  * handler which does not touch the bh after unlocking it.
2394  */
2395 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2396 {
2397         __end_buffer_read_notouch(bh, uptodate);
2398 }
2399
2400 /*
2401  * On entry, the page is fully not uptodate.
2402  * On exit the page is fully uptodate in the areas outside (from,to)
2403  */
2404 int nobh_prepare_write(struct page *page, unsigned from, unsigned to,
2405                         get_block_t *get_block)
2406 {
2407         struct inode *inode = page->mapping->host;
2408         const unsigned blkbits = inode->i_blkbits;
2409         const unsigned blocksize = 1 << blkbits;
2410         struct buffer_head *head, *bh;
2411         unsigned block_in_page;
2412         unsigned block_start, block_end;
2413         sector_t block_in_file;
2414         char *kaddr;
2415         int nr_reads = 0;
2416         int ret = 0;
2417         int is_mapped_to_disk = 1;
2418
2419         if (page_has_buffers(page))
2420                 return block_prepare_write(page, from, to, get_block);
2421
2422         if (PageMappedToDisk(page))
2423                 return 0;
2424
2425         /*
2426          * Allocate buffers so that we can keep track of state, and potentially
2427          * attach them to the page if an error occurs. In the common case of
2428          * no error, they will just be freed again without ever being attached
2429          * to the page (which is all OK, because we're under the page lock).
2430          *
2431          * Be careful: the buffer linked list is a NULL terminated one, rather
2432          * than the circular one we're used to.
2433          */
2434         head = alloc_page_buffers(page, blocksize, 0);
2435         if (!head)
2436                 return -ENOMEM;
2437
2438         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2439
2440         /*
2441          * We loop across all blocks in the page, whether or not they are
2442          * part of the affected region.  This is so we can discover if the
2443          * page is fully mapped-to-disk.
2444          */
2445         for (block_start = 0, block_in_page = 0, bh = head;
2446                   block_start < PAGE_CACHE_SIZE;
2447                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2448                 int create;
2449
2450                 block_end = block_start + blocksize;
2451                 bh->b_state = 0;
2452                 create = 1;
2453                 if (block_start >= to)
2454                         create = 0;
2455                 ret = get_block(inode, block_in_file + block_in_page,
2456                                         bh, create);
2457                 if (ret)
2458                         goto failed;
2459                 if (!buffer_mapped(bh))
2460                         is_mapped_to_disk = 0;
2461                 if (buffer_new(bh))
2462                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2463                 if (PageUptodate(page)) {
2464                         set_buffer_uptodate(bh);
2465                         continue;
2466                 }
2467                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2468                         kaddr = kmap_atomic(page, KM_USER0);
2469                         if (block_start < from)
2470                                 memset(kaddr+block_start, 0, from-block_start);
2471                         if (block_end > to)
2472                                 memset(kaddr + to, 0, block_end - to);
2473                         flush_dcache_page(page);
2474                         kunmap_atomic(kaddr, KM_USER0);
2475                         continue;
2476                 }
2477                 if (buffer_uptodate(bh))
2478                         continue;       /* reiserfs does this */
2479                 if (block_start < from || block_end > to) {
2480                         lock_buffer(bh);
2481                         bh->b_end_io = end_buffer_read_nobh;
2482                         submit_bh(READ, bh);
2483                         nr_reads++;
2484                 }
2485         }
2486
2487         if (nr_reads) {
2488                 /*
2489                  * The page is locked, so these buffers are protected from
2490                  * any VM or truncate activity.  Hence we don't need to care
2491                  * for the buffer_head refcounts.
2492                  */
2493                 for (bh = head; bh; bh = bh->b_this_page) {
2494                         wait_on_buffer(bh);
2495                         if (!buffer_uptodate(bh))
2496                                 ret = -EIO;
2497                 }
2498                 if (ret)
2499                         goto failed;
2500         }
2501
2502         if (is_mapped_to_disk)
2503                 SetPageMappedToDisk(page);
2504
2505         do {
2506                 bh = head;
2507                 head = head->b_this_page;
2508                 free_buffer_head(bh);
2509         } while (head);
2510
2511         return 0;
2512
2513 failed:
2514         /*
2515          * Error recovery is a bit difficult. We need to zero out blocks that
2516          * were newly allocated, and dirty them to ensure they get written out.
2517          * Buffers need to be attached to the page at this point, otherwise
2518          * the handling of potential IO errors during writeout would be hard
2519          * (could try doing synchronous writeout, but what if that fails too?)
2520          */
2521         spin_lock(&page->mapping->private_lock);
2522         bh = head;
2523         block_start = 0;
2524         do {
2525                 if (PageUptodate(page))
2526                         set_buffer_uptodate(bh);
2527                 if (PageDirty(page))
2528                         set_buffer_dirty(bh);
2529
2530                 block_end = block_start+blocksize;
2531                 if (block_end <= from)
2532                         goto next;
2533                 if (block_start >= to)
2534                         goto next;
2535
2536                 if (buffer_new(bh)) {
2537                         clear_buffer_new(bh);
2538                         if (!buffer_uptodate(bh)) {
2539                                 zero_user_page(page, block_start, bh->b_size, KM_USER0);
2540                                 set_buffer_uptodate(bh);
2541                         }
2542                         mark_buffer_dirty(bh);
2543                 }
2544 next:
2545                 block_start = block_end;
2546                 if (!bh->b_this_page)
2547                         bh->b_this_page = head;
2548                 bh = bh->b_this_page;
2549         } while (bh != head);
2550         attach_page_buffers(page, head);
2551         spin_unlock(&page->mapping->private_lock);
2552
2553         return ret;
2554 }
2555 EXPORT_SYMBOL(nobh_prepare_write);
2556
2557 /*
2558  * Make sure any changes to nobh_commit_write() are reflected in
2559  * nobh_truncate_page(), since it doesn't call commit_write().
2560  */
2561 int nobh_commit_write(struct file *file, struct page *page,
2562                 unsigned from, unsigned to)
2563 {
2564         struct inode *inode = page->mapping->host;
2565         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2566
2567         if (page_has_buffers(page))
2568                 return generic_commit_write(file, page, from, to);
2569
2570         SetPageUptodate(page);
2571         set_page_dirty(page);
2572         if (pos > inode->i_size) {
2573                 i_size_write(inode, pos);
2574                 mark_inode_dirty(inode);
2575         }
2576         return 0;
2577 }
2578 EXPORT_SYMBOL(nobh_commit_write);
2579
2580 /*
2581  * nobh_writepage() - based on block_full_write_page() except
2582  * that it tries to operate without attaching bufferheads to
2583  * the page.
2584  */
2585 int nobh_writepage(struct page *page, get_block_t *get_block,
2586                         struct writeback_control *wbc)
2587 {
2588         struct inode * const inode = page->mapping->host;
2589         loff_t i_size = i_size_read(inode);
2590         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2591         unsigned offset;
2592         int ret;
2593
2594         /* Is the page fully inside i_size? */
2595         if (page->index < end_index)
2596                 goto out;
2597
2598         /* Is the page fully outside i_size? (truncate in progress) */
2599         offset = i_size & (PAGE_CACHE_SIZE-1);
2600         if (page->index >= end_index+1 || !offset) {
2601                 /*
2602                  * The page may have dirty, unmapped buffers.  For example,
2603                  * they may have been added in ext3_writepage().  Make them
2604                  * freeable here, so the page does not leak.
2605                  */
2606 #if 0
2607                 /* Not really sure about this  - do we need this ? */
2608                 if (page->mapping->a_ops->invalidatepage)
2609                         page->mapping->a_ops->invalidatepage(page, offset);
2610 #endif
2611                 unlock_page(page);
2612                 return 0; /* don't care */
2613         }
2614
2615         /*
2616          * The page straddles i_size.  It must be zeroed out on each and every
2617          * writepage invocation because it may be mmapped.  "A file is mapped
2618          * in multiples of the page size.  For a file that is not a multiple of
2619          * the  page size, the remaining memory is zeroed when mapped, and
2620          * writes to that region are not written out to the file."
2621          */
2622         zero_user_page(page, offset, PAGE_CACHE_SIZE - offset, KM_USER0);
2623 out:
2624         ret = mpage_writepage(page, get_block, wbc);
2625         if (ret == -EAGAIN)
2626                 ret = __block_write_full_page(inode, page, get_block, wbc);
2627         return ret;
2628 }
2629 EXPORT_SYMBOL(nobh_writepage);
2630
2631 /*
2632  * This function assumes that ->prepare_write() uses nobh_prepare_write().
2633  */
2634 int nobh_truncate_page(struct address_space *mapping, loff_t from)
2635 {
2636         struct inode *inode = mapping->host;
2637         unsigned blocksize = 1 << inode->i_blkbits;
2638         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2639         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2640         unsigned to;
2641         struct page *page;
2642         const struct address_space_operations *a_ops = mapping->a_ops;
2643         int ret = 0;
2644
2645         if ((offset & (blocksize - 1)) == 0)
2646                 goto out;
2647
2648         ret = -ENOMEM;
2649         page = grab_cache_page(mapping, index);
2650         if (!page)
2651                 goto out;
2652
2653         to = (offset + blocksize) & ~(blocksize - 1);
2654         ret = a_ops->prepare_write(NULL, page, offset, to);
2655         if (ret == 0) {
2656                 zero_user_page(page, offset, PAGE_CACHE_SIZE - offset,
2657                                 KM_USER0);
2658                 /*
2659                  * It would be more correct to call aops->commit_write()
2660                  * here, but this is more efficient.
2661                  */
2662                 SetPageUptodate(page);
2663                 set_page_dirty(page);
2664         }
2665         unlock_page(page);
2666         page_cache_release(page);
2667 out:
2668         return ret;
2669 }
2670 EXPORT_SYMBOL(nobh_truncate_page);
2671
2672 int block_truncate_page(struct address_space *mapping,
2673                         loff_t from, get_block_t *get_block)
2674 {
2675         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2676         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2677         unsigned blocksize;
2678         sector_t iblock;
2679         unsigned length, pos;
2680         struct inode *inode = mapping->host;
2681         struct page *page;
2682         struct buffer_head *bh;
2683         int err;
2684
2685         blocksize = 1 << inode->i_blkbits;
2686         length = offset & (blocksize - 1);
2687
2688         /* Block boundary? Nothing to do */
2689         if (!length)
2690                 return 0;
2691
2692         length = blocksize - length;
2693         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2694         
2695         page = grab_cache_page(mapping, index);
2696         err = -ENOMEM;
2697         if (!page)
2698                 goto out;
2699
2700         if (!page_has_buffers(page))
2701                 create_empty_buffers(page, blocksize, 0);
2702
2703         /* Find the buffer that contains "offset" */
2704         bh = page_buffers(page);
2705         pos = blocksize;
2706         while (offset >= pos) {
2707                 bh = bh->b_this_page;
2708                 iblock++;
2709                 pos += blocksize;
2710         }
2711
2712         err = 0;
2713         if (!buffer_mapped(bh)) {
2714                 WARN_ON(bh->b_size != blocksize);
2715                 err = get_block(inode, iblock, bh, 0);
2716                 if (err)
2717                         goto unlock;
2718                 /* unmapped? It's a hole - nothing to do */
2719                 if (!buffer_mapped(bh))
2720                         goto unlock;
2721         }
2722
2723         /* Ok, it's mapped. Make sure it's up-to-date */
2724         if (PageUptodate(page))
2725                 set_buffer_uptodate(bh);
2726
2727         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2728                 err = -EIO;
2729                 ll_rw_block(READ, 1, &bh);
2730                 wait_on_buffer(bh);
2731                 /* Uhhuh. Read error. Complain and punt. */
2732                 if (!buffer_uptodate(bh))
2733                         goto unlock;
2734         }
2735
2736         zero_user_page(page, offset, length, KM_USER0);
2737         mark_buffer_dirty(bh);
2738         err = 0;
2739
2740 unlock:
2741         unlock_page(page);
2742         page_cache_release(page);
2743 out:
2744         return err;
2745 }
2746
2747 /*
2748  * The generic ->writepage function for buffer-backed address_spaces
2749  */
2750 int block_write_full_page(struct page *page, get_block_t *get_block,
2751                         struct writeback_control *wbc)
2752 {
2753         struct inode * const inode = page->mapping->host;
2754         loff_t i_size = i_size_read(inode);
2755         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2756         unsigned offset;
2757
2758         /* Is the page fully inside i_size? */
2759         if (page->index < end_index)
2760                 return __block_write_full_page(inode, page, get_block, wbc);
2761
2762         /* Is the page fully outside i_size? (truncate in progress) */
2763         offset = i_size & (PAGE_CACHE_SIZE-1);
2764         if (page->index >= end_index+1 || !offset) {
2765                 /*
2766                  * The page may have dirty, unmapped buffers.  For example,
2767                  * they may have been added in ext3_writepage().  Make them
2768                  * freeable here, so the page does not leak.
2769                  */
2770                 do_invalidatepage(page, 0);
2771                 unlock_page(page);
2772                 return 0; /* don't care */
2773         }
2774
2775         /*
2776          * The page straddles i_size.  It must be zeroed out on each and every
2777          * writepage invokation because it may be mmapped.  "A file is mapped
2778          * in multiples of the page size.  For a file that is not a multiple of
2779          * the  page size, the remaining memory is zeroed when mapped, and
2780          * writes to that region are not written out to the file."
2781          */
2782         zero_user_page(page, offset, PAGE_CACHE_SIZE - offset, KM_USER0);
2783         return __block_write_full_page(inode, page, get_block, wbc);
2784 }
2785
2786 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2787                             get_block_t *get_block)
2788 {
2789         struct buffer_head tmp;
2790         struct inode *inode = mapping->host;
2791         tmp.b_state = 0;
2792         tmp.b_blocknr = 0;
2793         tmp.b_size = 1 << inode->i_blkbits;
2794         get_block(inode, block, &tmp, 0);
2795         return tmp.b_blocknr;
2796 }
2797
2798 static void end_bio_bh_io_sync(struct bio *bio, int err)
2799 {
2800         struct buffer_head *bh = bio->bi_private;
2801
2802         if (err == -EOPNOTSUPP) {
2803                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2804                 set_bit(BH_Eopnotsupp, &bh->b_state);
2805         }
2806
2807         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2808         bio_put(bio);
2809 }
2810
2811 int submit_bh(int rw, struct buffer_head * bh)
2812 {
2813         struct bio *bio;
2814         int ret = 0;
2815
2816         BUG_ON(!buffer_locked(bh));
2817         BUG_ON(!buffer_mapped(bh));
2818         BUG_ON(!bh->b_end_io);
2819
2820         if (buffer_ordered(bh) && (rw == WRITE))
2821                 rw = WRITE_BARRIER;
2822
2823         /*
2824          * Only clear out a write error when rewriting, should this
2825          * include WRITE_SYNC as well?
2826          */
2827         if (test_set_buffer_req(bh) && (rw == WRITE || rw == WRITE_BARRIER))
2828                 clear_buffer_write_io_error(bh);
2829
2830         /*
2831          * from here on down, it's all bio -- do the initial mapping,
2832          * submit_bio -> generic_make_request may further map this bio around
2833          */
2834         bio = bio_alloc(GFP_NOIO, 1);
2835
2836         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2837         bio->bi_bdev = bh->b_bdev;
2838         bio->bi_io_vec[0].bv_page = bh->b_page;
2839         bio->bi_io_vec[0].bv_len = bh->b_size;
2840         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2841
2842         bio->bi_vcnt = 1;
2843         bio->bi_idx = 0;
2844         bio->bi_size = bh->b_size;
2845
2846         bio->bi_end_io = end_bio_bh_io_sync;
2847         bio->bi_private = bh;
2848
2849         bio_get(bio);
2850         submit_bio(rw, bio);
2851
2852         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2853                 ret = -EOPNOTSUPP;
2854
2855         bio_put(bio);
2856         return ret;
2857 }
2858
2859 /**
2860  * ll_rw_block: low-level access to block devices (DEPRECATED)
2861  * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
2862  * @nr: number of &struct buffer_heads in the array
2863  * @bhs: array of pointers to &struct buffer_head
2864  *
2865  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2866  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2867  * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2868  * are sent to disk. The fourth %READA option is described in the documentation
2869  * for generic_make_request() which ll_rw_block() calls.
2870  *
2871  * This function drops any buffer that it cannot get a lock on (with the
2872  * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
2873  * clean when doing a write request, and any buffer that appears to be
2874  * up-to-date when doing read request.  Further it marks as clean buffers that
2875  * are processed for writing (the buffer cache won't assume that they are
2876  * actually clean until the buffer gets unlocked).
2877  *
2878  * ll_rw_block sets b_end_io to simple completion handler that marks
2879  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2880  * any waiters. 
2881  *
2882  * All of the buffers must be for the same device, and must also be a
2883  * multiple of the current approved size for the device.
2884  */
2885 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2886 {
2887         int i;
2888
2889         for (i = 0; i < nr; i++) {
2890                 struct buffer_head *bh = bhs[i];
2891
2892                 if (rw == SWRITE)
2893                         lock_buffer(bh);
2894                 else if (test_set_buffer_locked(bh))
2895                         continue;
2896
2897                 if (rw == WRITE || rw == SWRITE) {
2898                         if (test_clear_buffer_dirty(bh)) {
2899                                 bh->b_end_io = end_buffer_write_sync;
2900                                 get_bh(bh);
2901                                 submit_bh(WRITE, bh);
2902                                 continue;
2903                         }
2904                 } else {
2905                         if (!buffer_uptodate(bh)) {
2906                                 bh->b_end_io = end_buffer_read_sync;
2907                                 get_bh(bh);
2908                                 submit_bh(rw, bh);
2909                                 continue;
2910                         }
2911                 }
2912                 unlock_buffer(bh);
2913         }
2914 }
2915
2916 /*
2917  * For a data-integrity writeout, we need to wait upon any in-progress I/O
2918  * and then start new I/O and then wait upon it.  The caller must have a ref on
2919  * the buffer_head.
2920  */
2921 int sync_dirty_buffer(struct buffer_head *bh)
2922 {
2923         int ret = 0;
2924
2925         WARN_ON(atomic_read(&bh->b_count) < 1);
2926         lock_buffer(bh);
2927         if (test_clear_buffer_dirty(bh)) {
2928                 get_bh(bh);
2929                 bh->b_end_io = end_buffer_write_sync;
2930                 ret = submit_bh(WRITE, bh);
2931                 wait_on_buffer(bh);
2932                 if (buffer_eopnotsupp(bh)) {
2933                         clear_buffer_eopnotsupp(bh);
2934                         ret = -EOPNOTSUPP;
2935                 }
2936                 if (!ret && !buffer_uptodate(bh))
2937                         ret = -EIO;
2938         } else {
2939                 unlock_buffer(bh);
2940         }
2941         return ret;
2942 }
2943
2944 /*
2945  * try_to_free_buffers() checks if all the buffers on this particular page
2946  * are unused, and releases them if so.
2947  *
2948  * Exclusion against try_to_free_buffers may be obtained by either
2949  * locking the page or by holding its mapping's private_lock.
2950  *
2951  * If the page is dirty but all the buffers are clean then we need to
2952  * be sure to mark the page clean as well.  This is because the page
2953  * may be against a block device, and a later reattachment of buffers
2954  * to a dirty page will set *all* buffers dirty.  Which would corrupt
2955  * filesystem data on the same device.
2956  *
2957  * The same applies to regular filesystem pages: if all the buffers are
2958  * clean then we set the page clean and proceed.  To do that, we require
2959  * total exclusion from __set_page_dirty_buffers().  That is obtained with
2960  * private_lock.
2961  *
2962  * try_to_free_buffers() is non-blocking.
2963  */
2964 static inline int buffer_busy(struct buffer_head *bh)
2965 {
2966         return atomic_read(&bh->b_count) |
2967                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
2968 }
2969
2970 static int
2971 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
2972 {
2973         struct buffer_head *head = page_buffers(page);
2974         struct buffer_head *bh;
2975
2976         bh = head;
2977         do {
2978                 if (buffer_write_io_error(bh) && page->mapping)
2979                         set_bit(AS_EIO, &page->mapping->flags);
2980                 if (buffer_busy(bh))
2981                         goto failed;
2982                 bh = bh->b_this_page;
2983         } while (bh != head);
2984
2985         do {
2986                 struct buffer_head *next = bh->b_this_page;
2987
2988                 if (!list_empty(&bh->b_assoc_buffers))
2989                         __remove_assoc_queue(bh);
2990                 bh = next;
2991         } while (bh != head);
2992         *buffers_to_free = head;
2993         __clear_page_buffers(page);
2994         return 1;
2995 failed:
2996         return 0;
2997 }
2998
2999 int try_to_free_buffers(struct page *page)
3000 {
3001         struct address_space * const mapping = page->mapping;
3002         struct buffer_head *buffers_to_free = NULL;
3003         int ret = 0;
3004
3005         BUG_ON(!PageLocked(page));
3006         if (PageWriteback(page))
3007                 return 0;
3008
3009         if (mapping == NULL) {          /* can this still happen? */
3010                 ret = drop_buffers(page, &buffers_to_free);
3011                 goto out;
3012         }
3013
3014         spin_lock(&mapping->private_lock);
3015         ret = drop_buffers(page, &buffers_to_free);
3016
3017         /*
3018          * If the filesystem writes its buffers by hand (eg ext3)
3019          * then we can have clean buffers against a dirty page.  We
3020          * clean the page here; otherwise the VM will never notice
3021          * that the filesystem did any IO at all.
3022          *
3023          * Also, during truncate, discard_buffer will have marked all
3024          * the page's buffers clean.  We discover that here and clean
3025          * the page also.
3026          *
3027          * private_lock must be held over this entire operation in order
3028          * to synchronise against __set_page_dirty_buffers and prevent the
3029          * dirty bit from being lost.
3030          */
3031         if (ret)
3032                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3033         spin_unlock(&mapping->private_lock);
3034 out:
3035         if (buffers_to_free) {
3036                 struct buffer_head *bh = buffers_to_free;
3037
3038                 do {
3039                         struct buffer_head *next = bh->b_this_page;
3040                         free_buffer_head(bh);
3041                         bh = next;
3042                 } while (bh != buffers_to_free);
3043         }
3044         return ret;
3045 }
3046 EXPORT_SYMBOL(try_to_free_buffers);
3047
3048 void block_sync_page(struct page *page)
3049 {
3050         struct address_space *mapping;
3051
3052         smp_mb();
3053         mapping = page_mapping(page);
3054         if (mapping)
3055                 blk_run_backing_dev(mapping->backing_dev_info, page);
3056 }
3057
3058 /*
3059  * There are no bdflush tunables left.  But distributions are
3060  * still running obsolete flush daemons, so we terminate them here.
3061  *
3062  * Use of bdflush() is deprecated and will be removed in a future kernel.
3063  * The `pdflush' kernel threads fully replace bdflush daemons and this call.
3064  */
3065 asmlinkage long sys_bdflush(int func, long data)
3066 {
3067         static int msg_count;
3068
3069         if (!capable(CAP_SYS_ADMIN))
3070                 return -EPERM;
3071
3072         if (msg_count < 5) {
3073                 msg_count++;
3074                 printk(KERN_INFO
3075                         "warning: process `%s' used the obsolete bdflush"
3076                         " system call\n", current->comm);
3077                 printk(KERN_INFO "Fix your initscripts?\n");
3078         }
3079
3080         if (func == 1)
3081                 do_exit(0);
3082         return 0;
3083 }
3084
3085 /*
3086  * Buffer-head allocation
3087  */
3088 static struct kmem_cache *bh_cachep;
3089
3090 /*
3091  * Once the number of bh's in the machine exceeds this level, we start
3092  * stripping them in writeback.
3093  */
3094 static int max_buffer_heads;
3095
3096 int buffer_heads_over_limit;
3097
3098 struct bh_accounting {
3099         int nr;                 /* Number of live bh's */
3100         int ratelimit;          /* Limit cacheline bouncing */
3101 };
3102
3103 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3104
3105 static void recalc_bh_state(void)
3106 {
3107         int i;
3108         int tot = 0;
3109
3110         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3111                 return;
3112         __get_cpu_var(bh_accounting).ratelimit = 0;
3113         for_each_online_cpu(i)
3114                 tot += per_cpu(bh_accounting, i).nr;
3115         buffer_heads_over_limit = (tot > max_buffer_heads);
3116 }
3117         
3118 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3119 {
3120         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3121         if (ret) {
3122                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3123                 get_cpu_var(bh_accounting).nr++;
3124                 recalc_bh_state();
3125                 put_cpu_var(bh_accounting);
3126         }
3127         return ret;
3128 }
3129 EXPORT_SYMBOL(alloc_buffer_head);
3130
3131 void free_buffer_head(struct buffer_head *bh)
3132 {
3133         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3134         kmem_cache_free(bh_cachep, bh);
3135         get_cpu_var(bh_accounting).nr--;
3136         recalc_bh_state();
3137         put_cpu_var(bh_accounting);
3138 }
3139 EXPORT_SYMBOL(free_buffer_head);
3140
3141 static void buffer_exit_cpu(int cpu)
3142 {
3143         int i;
3144         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3145
3146         for (i = 0; i < BH_LRU_SIZE; i++) {
3147                 brelse(b->bhs[i]);
3148                 b->bhs[i] = NULL;
3149         }
3150         get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3151         per_cpu(bh_accounting, cpu).nr = 0;
3152         put_cpu_var(bh_accounting);
3153 }
3154
3155 static int buffer_cpu_notify(struct notifier_block *self,
3156                               unsigned long action, void *hcpu)
3157 {
3158         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3159                 buffer_exit_cpu((unsigned long)hcpu);
3160         return NOTIFY_OK;
3161 }
3162
3163 void __init buffer_init(void)
3164 {
3165         int nrpages;
3166
3167         bh_cachep = KMEM_CACHE(buffer_head,
3168                         SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD);
3169
3170         /*
3171          * Limit the bh occupancy to 10% of ZONE_NORMAL
3172          */
3173         nrpages = (nr_free_buffer_pages() * 10) / 100;
3174         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3175         hotcpu_notifier(buffer_cpu_notify, 0);
3176 }
3177
3178 EXPORT_SYMBOL(__bforget);
3179 EXPORT_SYMBOL(__brelse);
3180 EXPORT_SYMBOL(__wait_on_buffer);
3181 EXPORT_SYMBOL(block_commit_write);
3182 EXPORT_SYMBOL(block_prepare_write);
3183 EXPORT_SYMBOL(block_page_mkwrite);
3184 EXPORT_SYMBOL(block_read_full_page);
3185 EXPORT_SYMBOL(block_sync_page);
3186 EXPORT_SYMBOL(block_truncate_page);
3187 EXPORT_SYMBOL(block_write_full_page);
3188 EXPORT_SYMBOL(cont_write_begin);
3189 EXPORT_SYMBOL(end_buffer_read_sync);
3190 EXPORT_SYMBOL(end_buffer_write_sync);
3191 EXPORT_SYMBOL(file_fsync);
3192 EXPORT_SYMBOL(fsync_bdev);
3193 EXPORT_SYMBOL(generic_block_bmap);
3194 EXPORT_SYMBOL(generic_commit_write);
3195 EXPORT_SYMBOL(generic_cont_expand);
3196 EXPORT_SYMBOL(generic_cont_expand_simple);
3197 EXPORT_SYMBOL(init_buffer);
3198 EXPORT_SYMBOL(invalidate_bdev);
3199 EXPORT_SYMBOL(ll_rw_block);
3200 EXPORT_SYMBOL(mark_buffer_dirty);
3201 EXPORT_SYMBOL(submit_bh);
3202 EXPORT_SYMBOL(sync_dirty_buffer);
3203 EXPORT_SYMBOL(unlock_buffer);