mm: spinlock tree_lock
[linux-2.6.git] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/capability.h>
28 #include <linux/blkdev.h>
29 #include <linux/file.h>
30 #include <linux/quotaops.h>
31 #include <linux/highmem.h>
32 #include <linux/module.h>
33 #include <linux/writeback.h>
34 #include <linux/hash.h>
35 #include <linux/suspend.h>
36 #include <linux/buffer_head.h>
37 #include <linux/task_io_accounting_ops.h>
38 #include <linux/bio.h>
39 #include <linux/notifier.h>
40 #include <linux/cpu.h>
41 #include <linux/bitops.h>
42 #include <linux/mpage.h>
43 #include <linux/bit_spinlock.h>
44
45 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
46
47 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
48
49 inline void
50 init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
51 {
52         bh->b_end_io = handler;
53         bh->b_private = private;
54 }
55
56 static int sync_buffer(void *word)
57 {
58         struct block_device *bd;
59         struct buffer_head *bh
60                 = container_of(word, struct buffer_head, b_state);
61
62         smp_mb();
63         bd = bh->b_bdev;
64         if (bd)
65                 blk_run_address_space(bd->bd_inode->i_mapping);
66         io_schedule();
67         return 0;
68 }
69
70 void __lock_buffer(struct buffer_head *bh)
71 {
72         wait_on_bit_lock(&bh->b_state, BH_Lock, sync_buffer,
73                                                         TASK_UNINTERRUPTIBLE);
74 }
75 EXPORT_SYMBOL(__lock_buffer);
76
77 void unlock_buffer(struct buffer_head *bh)
78 {
79         smp_mb__before_clear_bit();
80         clear_buffer_locked(bh);
81         smp_mb__after_clear_bit();
82         wake_up_bit(&bh->b_state, BH_Lock);
83 }
84
85 /*
86  * Block until a buffer comes unlocked.  This doesn't stop it
87  * from becoming locked again - you have to lock it yourself
88  * if you want to preserve its state.
89  */
90 void __wait_on_buffer(struct buffer_head * bh)
91 {
92         wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
93 }
94
95 static void
96 __clear_page_buffers(struct page *page)
97 {
98         ClearPagePrivate(page);
99         set_page_private(page, 0);
100         page_cache_release(page);
101 }
102
103 static void buffer_io_error(struct buffer_head *bh)
104 {
105         char b[BDEVNAME_SIZE];
106
107         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
108                         bdevname(bh->b_bdev, b),
109                         (unsigned long long)bh->b_blocknr);
110 }
111
112 /*
113  * End-of-IO handler helper function which does not touch the bh after
114  * unlocking it.
115  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
116  * a race there is benign: unlock_buffer() only use the bh's address for
117  * hashing after unlocking the buffer, so it doesn't actually touch the bh
118  * itself.
119  */
120 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
121 {
122         if (uptodate) {
123                 set_buffer_uptodate(bh);
124         } else {
125                 /* This happens, due to failed READA attempts. */
126                 clear_buffer_uptodate(bh);
127         }
128         unlock_buffer(bh);
129 }
130
131 /*
132  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
133  * unlock the buffer. This is what ll_rw_block uses too.
134  */
135 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
136 {
137         __end_buffer_read_notouch(bh, uptodate);
138         put_bh(bh);
139 }
140
141 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
142 {
143         char b[BDEVNAME_SIZE];
144
145         if (uptodate) {
146                 set_buffer_uptodate(bh);
147         } else {
148                 if (!buffer_eopnotsupp(bh) && printk_ratelimit()) {
149                         buffer_io_error(bh);
150                         printk(KERN_WARNING "lost page write due to "
151                                         "I/O error on %s\n",
152                                        bdevname(bh->b_bdev, b));
153                 }
154                 set_buffer_write_io_error(bh);
155                 clear_buffer_uptodate(bh);
156         }
157         unlock_buffer(bh);
158         put_bh(bh);
159 }
160
161 /*
162  * Write out and wait upon all the dirty data associated with a block
163  * device via its mapping.  Does not take the superblock lock.
164  */
165 int sync_blockdev(struct block_device *bdev)
166 {
167         int ret = 0;
168
169         if (bdev)
170                 ret = filemap_write_and_wait(bdev->bd_inode->i_mapping);
171         return ret;
172 }
173 EXPORT_SYMBOL(sync_blockdev);
174
175 /*
176  * Write out and wait upon all dirty data associated with this
177  * device.   Filesystem data as well as the underlying block
178  * device.  Takes the superblock lock.
179  */
180 int fsync_bdev(struct block_device *bdev)
181 {
182         struct super_block *sb = get_super(bdev);
183         if (sb) {
184                 int res = fsync_super(sb);
185                 drop_super(sb);
186                 return res;
187         }
188         return sync_blockdev(bdev);
189 }
190
191 /**
192  * freeze_bdev  --  lock a filesystem and force it into a consistent state
193  * @bdev:       blockdevice to lock
194  *
195  * This takes the block device bd_mount_sem to make sure no new mounts
196  * happen on bdev until thaw_bdev() is called.
197  * If a superblock is found on this device, we take the s_umount semaphore
198  * on it to make sure nobody unmounts until the snapshot creation is done.
199  */
200 struct super_block *freeze_bdev(struct block_device *bdev)
201 {
202         struct super_block *sb;
203
204         down(&bdev->bd_mount_sem);
205         sb = get_super(bdev);
206         if (sb && !(sb->s_flags & MS_RDONLY)) {
207                 sb->s_frozen = SB_FREEZE_WRITE;
208                 smp_wmb();
209
210                 __fsync_super(sb);
211
212                 sb->s_frozen = SB_FREEZE_TRANS;
213                 smp_wmb();
214
215                 sync_blockdev(sb->s_bdev);
216
217                 if (sb->s_op->write_super_lockfs)
218                         sb->s_op->write_super_lockfs(sb);
219         }
220
221         sync_blockdev(bdev);
222         return sb;      /* thaw_bdev releases s->s_umount and bd_mount_sem */
223 }
224 EXPORT_SYMBOL(freeze_bdev);
225
226 /**
227  * thaw_bdev  -- unlock filesystem
228  * @bdev:       blockdevice to unlock
229  * @sb:         associated superblock
230  *
231  * Unlocks the filesystem and marks it writeable again after freeze_bdev().
232  */
233 void thaw_bdev(struct block_device *bdev, struct super_block *sb)
234 {
235         if (sb) {
236                 BUG_ON(sb->s_bdev != bdev);
237
238                 if (sb->s_op->unlockfs)
239                         sb->s_op->unlockfs(sb);
240                 sb->s_frozen = SB_UNFROZEN;
241                 smp_wmb();
242                 wake_up(&sb->s_wait_unfrozen);
243                 drop_super(sb);
244         }
245
246         up(&bdev->bd_mount_sem);
247 }
248 EXPORT_SYMBOL(thaw_bdev);
249
250 /*
251  * Various filesystems appear to want __find_get_block to be non-blocking.
252  * But it's the page lock which protects the buffers.  To get around this,
253  * we get exclusion from try_to_free_buffers with the blockdev mapping's
254  * private_lock.
255  *
256  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
257  * may be quite high.  This code could TryLock the page, and if that
258  * succeeds, there is no need to take private_lock. (But if
259  * private_lock is contended then so is mapping->tree_lock).
260  */
261 static struct buffer_head *
262 __find_get_block_slow(struct block_device *bdev, sector_t block)
263 {
264         struct inode *bd_inode = bdev->bd_inode;
265         struct address_space *bd_mapping = bd_inode->i_mapping;
266         struct buffer_head *ret = NULL;
267         pgoff_t index;
268         struct buffer_head *bh;
269         struct buffer_head *head;
270         struct page *page;
271         int all_mapped = 1;
272
273         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
274         page = find_get_page(bd_mapping, index);
275         if (!page)
276                 goto out;
277
278         spin_lock(&bd_mapping->private_lock);
279         if (!page_has_buffers(page))
280                 goto out_unlock;
281         head = page_buffers(page);
282         bh = head;
283         do {
284                 if (bh->b_blocknr == block) {
285                         ret = bh;
286                         get_bh(bh);
287                         goto out_unlock;
288                 }
289                 if (!buffer_mapped(bh))
290                         all_mapped = 0;
291                 bh = bh->b_this_page;
292         } while (bh != head);
293
294         /* we might be here because some of the buffers on this page are
295          * not mapped.  This is due to various races between
296          * file io on the block device and getblk.  It gets dealt with
297          * elsewhere, don't buffer_error if we had some unmapped buffers
298          */
299         if (all_mapped) {
300                 printk("__find_get_block_slow() failed. "
301                         "block=%llu, b_blocknr=%llu\n",
302                         (unsigned long long)block,
303                         (unsigned long long)bh->b_blocknr);
304                 printk("b_state=0x%08lx, b_size=%zu\n",
305                         bh->b_state, bh->b_size);
306                 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
307         }
308 out_unlock:
309         spin_unlock(&bd_mapping->private_lock);
310         page_cache_release(page);
311 out:
312         return ret;
313 }
314
315 /* If invalidate_buffers() will trash dirty buffers, it means some kind
316    of fs corruption is going on. Trashing dirty data always imply losing
317    information that was supposed to be just stored on the physical layer
318    by the user.
319
320    Thus invalidate_buffers in general usage is not allwowed to trash
321    dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
322    be preserved.  These buffers are simply skipped.
323   
324    We also skip buffers which are still in use.  For example this can
325    happen if a userspace program is reading the block device.
326
327    NOTE: In the case where the user removed a removable-media-disk even if
328    there's still dirty data not synced on disk (due a bug in the device driver
329    or due an error of the user), by not destroying the dirty buffers we could
330    generate corruption also on the next media inserted, thus a parameter is
331    necessary to handle this case in the most safe way possible (trying
332    to not corrupt also the new disk inserted with the data belonging to
333    the old now corrupted disk). Also for the ramdisk the natural thing
334    to do in order to release the ramdisk memory is to destroy dirty buffers.
335
336    These are two special cases. Normal usage imply the device driver
337    to issue a sync on the device (without waiting I/O completion) and
338    then an invalidate_buffers call that doesn't trash dirty buffers.
339
340    For handling cache coherency with the blkdev pagecache the 'update' case
341    is been introduced. It is needed to re-read from disk any pinned
342    buffer. NOTE: re-reading from disk is destructive so we can do it only
343    when we assume nobody is changing the buffercache under our I/O and when
344    we think the disk contains more recent information than the buffercache.
345    The update == 1 pass marks the buffers we need to update, the update == 2
346    pass does the actual I/O. */
347 void invalidate_bdev(struct block_device *bdev)
348 {
349         struct address_space *mapping = bdev->bd_inode->i_mapping;
350
351         if (mapping->nrpages == 0)
352                 return;
353
354         invalidate_bh_lrus();
355         invalidate_mapping_pages(mapping, 0, -1);
356 }
357
358 /*
359  * Kick pdflush then try to free up some ZONE_NORMAL memory.
360  */
361 static void free_more_memory(void)
362 {
363         struct zone *zone;
364         int nid;
365
366         wakeup_pdflush(1024);
367         yield();
368
369         for_each_online_node(nid) {
370                 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
371                                                 gfp_zone(GFP_NOFS), NULL,
372                                                 &zone);
373                 if (zone)
374                         try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
375                                                 GFP_NOFS);
376         }
377 }
378
379 /*
380  * I/O completion handler for block_read_full_page() - pages
381  * which come unlocked at the end of I/O.
382  */
383 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
384 {
385         unsigned long flags;
386         struct buffer_head *first;
387         struct buffer_head *tmp;
388         struct page *page;
389         int page_uptodate = 1;
390
391         BUG_ON(!buffer_async_read(bh));
392
393         page = bh->b_page;
394         if (uptodate) {
395                 set_buffer_uptodate(bh);
396         } else {
397                 clear_buffer_uptodate(bh);
398                 if (printk_ratelimit())
399                         buffer_io_error(bh);
400                 SetPageError(page);
401         }
402
403         /*
404          * Be _very_ careful from here on. Bad things can happen if
405          * two buffer heads end IO at almost the same time and both
406          * decide that the page is now completely done.
407          */
408         first = page_buffers(page);
409         local_irq_save(flags);
410         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
411         clear_buffer_async_read(bh);
412         unlock_buffer(bh);
413         tmp = bh;
414         do {
415                 if (!buffer_uptodate(tmp))
416                         page_uptodate = 0;
417                 if (buffer_async_read(tmp)) {
418                         BUG_ON(!buffer_locked(tmp));
419                         goto still_busy;
420                 }
421                 tmp = tmp->b_this_page;
422         } while (tmp != bh);
423         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
424         local_irq_restore(flags);
425
426         /*
427          * If none of the buffers had errors and they are all
428          * uptodate then we can set the page uptodate.
429          */
430         if (page_uptodate && !PageError(page))
431                 SetPageUptodate(page);
432         unlock_page(page);
433         return;
434
435 still_busy:
436         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
437         local_irq_restore(flags);
438         return;
439 }
440
441 /*
442  * Completion handler for block_write_full_page() - pages which are unlocked
443  * during I/O, and which have PageWriteback cleared upon I/O completion.
444  */
445 static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
446 {
447         char b[BDEVNAME_SIZE];
448         unsigned long flags;
449         struct buffer_head *first;
450         struct buffer_head *tmp;
451         struct page *page;
452
453         BUG_ON(!buffer_async_write(bh));
454
455         page = bh->b_page;
456         if (uptodate) {
457                 set_buffer_uptodate(bh);
458         } else {
459                 if (printk_ratelimit()) {
460                         buffer_io_error(bh);
461                         printk(KERN_WARNING "lost page write due to "
462                                         "I/O error on %s\n",
463                                bdevname(bh->b_bdev, b));
464                 }
465                 set_bit(AS_EIO, &page->mapping->flags);
466                 set_buffer_write_io_error(bh);
467                 clear_buffer_uptodate(bh);
468                 SetPageError(page);
469         }
470
471         first = page_buffers(page);
472         local_irq_save(flags);
473         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
474
475         clear_buffer_async_write(bh);
476         unlock_buffer(bh);
477         tmp = bh->b_this_page;
478         while (tmp != bh) {
479                 if (buffer_async_write(tmp)) {
480                         BUG_ON(!buffer_locked(tmp));
481                         goto still_busy;
482                 }
483                 tmp = tmp->b_this_page;
484         }
485         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
486         local_irq_restore(flags);
487         end_page_writeback(page);
488         return;
489
490 still_busy:
491         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
492         local_irq_restore(flags);
493         return;
494 }
495
496 /*
497  * If a page's buffers are under async readin (end_buffer_async_read
498  * completion) then there is a possibility that another thread of
499  * control could lock one of the buffers after it has completed
500  * but while some of the other buffers have not completed.  This
501  * locked buffer would confuse end_buffer_async_read() into not unlocking
502  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
503  * that this buffer is not under async I/O.
504  *
505  * The page comes unlocked when it has no locked buffer_async buffers
506  * left.
507  *
508  * PageLocked prevents anyone starting new async I/O reads any of
509  * the buffers.
510  *
511  * PageWriteback is used to prevent simultaneous writeout of the same
512  * page.
513  *
514  * PageLocked prevents anyone from starting writeback of a page which is
515  * under read I/O (PageWriteback is only ever set against a locked page).
516  */
517 static void mark_buffer_async_read(struct buffer_head *bh)
518 {
519         bh->b_end_io = end_buffer_async_read;
520         set_buffer_async_read(bh);
521 }
522
523 void mark_buffer_async_write(struct buffer_head *bh)
524 {
525         bh->b_end_io = end_buffer_async_write;
526         set_buffer_async_write(bh);
527 }
528 EXPORT_SYMBOL(mark_buffer_async_write);
529
530
531 /*
532  * fs/buffer.c contains helper functions for buffer-backed address space's
533  * fsync functions.  A common requirement for buffer-based filesystems is
534  * that certain data from the backing blockdev needs to be written out for
535  * a successful fsync().  For example, ext2 indirect blocks need to be
536  * written back and waited upon before fsync() returns.
537  *
538  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
539  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
540  * management of a list of dependent buffers at ->i_mapping->private_list.
541  *
542  * Locking is a little subtle: try_to_free_buffers() will remove buffers
543  * from their controlling inode's queue when they are being freed.  But
544  * try_to_free_buffers() will be operating against the *blockdev* mapping
545  * at the time, not against the S_ISREG file which depends on those buffers.
546  * So the locking for private_list is via the private_lock in the address_space
547  * which backs the buffers.  Which is different from the address_space 
548  * against which the buffers are listed.  So for a particular address_space,
549  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
550  * mapping->private_list will always be protected by the backing blockdev's
551  * ->private_lock.
552  *
553  * Which introduces a requirement: all buffers on an address_space's
554  * ->private_list must be from the same address_space: the blockdev's.
555  *
556  * address_spaces which do not place buffers at ->private_list via these
557  * utility functions are free to use private_lock and private_list for
558  * whatever they want.  The only requirement is that list_empty(private_list)
559  * be true at clear_inode() time.
560  *
561  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
562  * filesystems should do that.  invalidate_inode_buffers() should just go
563  * BUG_ON(!list_empty).
564  *
565  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
566  * take an address_space, not an inode.  And it should be called
567  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
568  * queued up.
569  *
570  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
571  * list if it is already on a list.  Because if the buffer is on a list,
572  * it *must* already be on the right one.  If not, the filesystem is being
573  * silly.  This will save a ton of locking.  But first we have to ensure
574  * that buffers are taken *off* the old inode's list when they are freed
575  * (presumably in truncate).  That requires careful auditing of all
576  * filesystems (do it inside bforget()).  It could also be done by bringing
577  * b_inode back.
578  */
579
580 /*
581  * The buffer's backing address_space's private_lock must be held
582  */
583 static inline void __remove_assoc_queue(struct buffer_head *bh)
584 {
585         list_del_init(&bh->b_assoc_buffers);
586         WARN_ON(!bh->b_assoc_map);
587         if (buffer_write_io_error(bh))
588                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
589         bh->b_assoc_map = NULL;
590 }
591
592 int inode_has_buffers(struct inode *inode)
593 {
594         return !list_empty(&inode->i_data.private_list);
595 }
596
597 /*
598  * osync is designed to support O_SYNC io.  It waits synchronously for
599  * all already-submitted IO to complete, but does not queue any new
600  * writes to the disk.
601  *
602  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
603  * you dirty the buffers, and then use osync_inode_buffers to wait for
604  * completion.  Any other dirty buffers which are not yet queued for
605  * write will not be flushed to disk by the osync.
606  */
607 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
608 {
609         struct buffer_head *bh;
610         struct list_head *p;
611         int err = 0;
612
613         spin_lock(lock);
614 repeat:
615         list_for_each_prev(p, list) {
616                 bh = BH_ENTRY(p);
617                 if (buffer_locked(bh)) {
618                         get_bh(bh);
619                         spin_unlock(lock);
620                         wait_on_buffer(bh);
621                         if (!buffer_uptodate(bh))
622                                 err = -EIO;
623                         brelse(bh);
624                         spin_lock(lock);
625                         goto repeat;
626                 }
627         }
628         spin_unlock(lock);
629         return err;
630 }
631
632 /**
633  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
634  * @mapping: the mapping which wants those buffers written
635  *
636  * Starts I/O against the buffers at mapping->private_list, and waits upon
637  * that I/O.
638  *
639  * Basically, this is a convenience function for fsync().
640  * @mapping is a file or directory which needs those buffers to be written for
641  * a successful fsync().
642  */
643 int sync_mapping_buffers(struct address_space *mapping)
644 {
645         struct address_space *buffer_mapping = mapping->assoc_mapping;
646
647         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
648                 return 0;
649
650         return fsync_buffers_list(&buffer_mapping->private_lock,
651                                         &mapping->private_list);
652 }
653 EXPORT_SYMBOL(sync_mapping_buffers);
654
655 /*
656  * Called when we've recently written block `bblock', and it is known that
657  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
658  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
659  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
660  */
661 void write_boundary_block(struct block_device *bdev,
662                         sector_t bblock, unsigned blocksize)
663 {
664         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
665         if (bh) {
666                 if (buffer_dirty(bh))
667                         ll_rw_block(WRITE, 1, &bh);
668                 put_bh(bh);
669         }
670 }
671
672 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
673 {
674         struct address_space *mapping = inode->i_mapping;
675         struct address_space *buffer_mapping = bh->b_page->mapping;
676
677         mark_buffer_dirty(bh);
678         if (!mapping->assoc_mapping) {
679                 mapping->assoc_mapping = buffer_mapping;
680         } else {
681                 BUG_ON(mapping->assoc_mapping != buffer_mapping);
682         }
683         if (!bh->b_assoc_map) {
684                 spin_lock(&buffer_mapping->private_lock);
685                 list_move_tail(&bh->b_assoc_buffers,
686                                 &mapping->private_list);
687                 bh->b_assoc_map = mapping;
688                 spin_unlock(&buffer_mapping->private_lock);
689         }
690 }
691 EXPORT_SYMBOL(mark_buffer_dirty_inode);
692
693 /*
694  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
695  * dirty.
696  *
697  * If warn is true, then emit a warning if the page is not uptodate and has
698  * not been truncated.
699  */
700 static int __set_page_dirty(struct page *page,
701                 struct address_space *mapping, int warn)
702 {
703         if (unlikely(!mapping))
704                 return !TestSetPageDirty(page);
705
706         if (TestSetPageDirty(page))
707                 return 0;
708
709         spin_lock_irq(&mapping->tree_lock);
710         if (page->mapping) {    /* Race with truncate? */
711                 WARN_ON_ONCE(warn && !PageUptodate(page));
712
713                 if (mapping_cap_account_dirty(mapping)) {
714                         __inc_zone_page_state(page, NR_FILE_DIRTY);
715                         __inc_bdi_stat(mapping->backing_dev_info,
716                                         BDI_RECLAIMABLE);
717                         task_io_account_write(PAGE_CACHE_SIZE);
718                 }
719                 radix_tree_tag_set(&mapping->page_tree,
720                                 page_index(page), PAGECACHE_TAG_DIRTY);
721         }
722         spin_unlock_irq(&mapping->tree_lock);
723         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
724
725         return 1;
726 }
727
728 /*
729  * Add a page to the dirty page list.
730  *
731  * It is a sad fact of life that this function is called from several places
732  * deeply under spinlocking.  It may not sleep.
733  *
734  * If the page has buffers, the uptodate buffers are set dirty, to preserve
735  * dirty-state coherency between the page and the buffers.  It the page does
736  * not have buffers then when they are later attached they will all be set
737  * dirty.
738  *
739  * The buffers are dirtied before the page is dirtied.  There's a small race
740  * window in which a writepage caller may see the page cleanness but not the
741  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
742  * before the buffers, a concurrent writepage caller could clear the page dirty
743  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
744  * page on the dirty page list.
745  *
746  * We use private_lock to lock against try_to_free_buffers while using the
747  * page's buffer list.  Also use this to protect against clean buffers being
748  * added to the page after it was set dirty.
749  *
750  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
751  * address_space though.
752  */
753 int __set_page_dirty_buffers(struct page *page)
754 {
755         struct address_space *mapping = page_mapping(page);
756
757         if (unlikely(!mapping))
758                 return !TestSetPageDirty(page);
759
760         spin_lock(&mapping->private_lock);
761         if (page_has_buffers(page)) {
762                 struct buffer_head *head = page_buffers(page);
763                 struct buffer_head *bh = head;
764
765                 do {
766                         set_buffer_dirty(bh);
767                         bh = bh->b_this_page;
768                 } while (bh != head);
769         }
770         spin_unlock(&mapping->private_lock);
771
772         return __set_page_dirty(page, mapping, 1);
773 }
774 EXPORT_SYMBOL(__set_page_dirty_buffers);
775
776 /*
777  * Write out and wait upon a list of buffers.
778  *
779  * We have conflicting pressures: we want to make sure that all
780  * initially dirty buffers get waited on, but that any subsequently
781  * dirtied buffers don't.  After all, we don't want fsync to last
782  * forever if somebody is actively writing to the file.
783  *
784  * Do this in two main stages: first we copy dirty buffers to a
785  * temporary inode list, queueing the writes as we go.  Then we clean
786  * up, waiting for those writes to complete.
787  * 
788  * During this second stage, any subsequent updates to the file may end
789  * up refiling the buffer on the original inode's dirty list again, so
790  * there is a chance we will end up with a buffer queued for write but
791  * not yet completed on that list.  So, as a final cleanup we go through
792  * the osync code to catch these locked, dirty buffers without requeuing
793  * any newly dirty buffers for write.
794  */
795 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
796 {
797         struct buffer_head *bh;
798         struct list_head tmp;
799         struct address_space *mapping;
800         int err = 0, err2;
801
802         INIT_LIST_HEAD(&tmp);
803
804         spin_lock(lock);
805         while (!list_empty(list)) {
806                 bh = BH_ENTRY(list->next);
807                 mapping = bh->b_assoc_map;
808                 __remove_assoc_queue(bh);
809                 /* Avoid race with mark_buffer_dirty_inode() which does
810                  * a lockless check and we rely on seeing the dirty bit */
811                 smp_mb();
812                 if (buffer_dirty(bh) || buffer_locked(bh)) {
813                         list_add(&bh->b_assoc_buffers, &tmp);
814                         bh->b_assoc_map = mapping;
815                         if (buffer_dirty(bh)) {
816                                 get_bh(bh);
817                                 spin_unlock(lock);
818                                 /*
819                                  * Ensure any pending I/O completes so that
820                                  * ll_rw_block() actually writes the current
821                                  * contents - it is a noop if I/O is still in
822                                  * flight on potentially older contents.
823                                  */
824                                 ll_rw_block(SWRITE_SYNC, 1, &bh);
825                                 brelse(bh);
826                                 spin_lock(lock);
827                         }
828                 }
829         }
830
831         while (!list_empty(&tmp)) {
832                 bh = BH_ENTRY(tmp.prev);
833                 get_bh(bh);
834                 mapping = bh->b_assoc_map;
835                 __remove_assoc_queue(bh);
836                 /* Avoid race with mark_buffer_dirty_inode() which does
837                  * a lockless check and we rely on seeing the dirty bit */
838                 smp_mb();
839                 if (buffer_dirty(bh)) {
840                         list_add(&bh->b_assoc_buffers,
841                                  &mapping->private_list);
842                         bh->b_assoc_map = mapping;
843                 }
844                 spin_unlock(lock);
845                 wait_on_buffer(bh);
846                 if (!buffer_uptodate(bh))
847                         err = -EIO;
848                 brelse(bh);
849                 spin_lock(lock);
850         }
851         
852         spin_unlock(lock);
853         err2 = osync_buffers_list(lock, list);
854         if (err)
855                 return err;
856         else
857                 return err2;
858 }
859
860 /*
861  * Invalidate any and all dirty buffers on a given inode.  We are
862  * probably unmounting the fs, but that doesn't mean we have already
863  * done a sync().  Just drop the buffers from the inode list.
864  *
865  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
866  * assumes that all the buffers are against the blockdev.  Not true
867  * for reiserfs.
868  */
869 void invalidate_inode_buffers(struct inode *inode)
870 {
871         if (inode_has_buffers(inode)) {
872                 struct address_space *mapping = &inode->i_data;
873                 struct list_head *list = &mapping->private_list;
874                 struct address_space *buffer_mapping = mapping->assoc_mapping;
875
876                 spin_lock(&buffer_mapping->private_lock);
877                 while (!list_empty(list))
878                         __remove_assoc_queue(BH_ENTRY(list->next));
879                 spin_unlock(&buffer_mapping->private_lock);
880         }
881 }
882
883 /*
884  * Remove any clean buffers from the inode's buffer list.  This is called
885  * when we're trying to free the inode itself.  Those buffers can pin it.
886  *
887  * Returns true if all buffers were removed.
888  */
889 int remove_inode_buffers(struct inode *inode)
890 {
891         int ret = 1;
892
893         if (inode_has_buffers(inode)) {
894                 struct address_space *mapping = &inode->i_data;
895                 struct list_head *list = &mapping->private_list;
896                 struct address_space *buffer_mapping = mapping->assoc_mapping;
897
898                 spin_lock(&buffer_mapping->private_lock);
899                 while (!list_empty(list)) {
900                         struct buffer_head *bh = BH_ENTRY(list->next);
901                         if (buffer_dirty(bh)) {
902                                 ret = 0;
903                                 break;
904                         }
905                         __remove_assoc_queue(bh);
906                 }
907                 spin_unlock(&buffer_mapping->private_lock);
908         }
909         return ret;
910 }
911
912 /*
913  * Create the appropriate buffers when given a page for data area and
914  * the size of each buffer.. Use the bh->b_this_page linked list to
915  * follow the buffers created.  Return NULL if unable to create more
916  * buffers.
917  *
918  * The retry flag is used to differentiate async IO (paging, swapping)
919  * which may not fail from ordinary buffer allocations.
920  */
921 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
922                 int retry)
923 {
924         struct buffer_head *bh, *head;
925         long offset;
926
927 try_again:
928         head = NULL;
929         offset = PAGE_SIZE;
930         while ((offset -= size) >= 0) {
931                 bh = alloc_buffer_head(GFP_NOFS);
932                 if (!bh)
933                         goto no_grow;
934
935                 bh->b_bdev = NULL;
936                 bh->b_this_page = head;
937                 bh->b_blocknr = -1;
938                 head = bh;
939
940                 bh->b_state = 0;
941                 atomic_set(&bh->b_count, 0);
942                 bh->b_private = NULL;
943                 bh->b_size = size;
944
945                 /* Link the buffer to its page */
946                 set_bh_page(bh, page, offset);
947
948                 init_buffer(bh, NULL, NULL);
949         }
950         return head;
951 /*
952  * In case anything failed, we just free everything we got.
953  */
954 no_grow:
955         if (head) {
956                 do {
957                         bh = head;
958                         head = head->b_this_page;
959                         free_buffer_head(bh);
960                 } while (head);
961         }
962
963         /*
964          * Return failure for non-async IO requests.  Async IO requests
965          * are not allowed to fail, so we have to wait until buffer heads
966          * become available.  But we don't want tasks sleeping with 
967          * partially complete buffers, so all were released above.
968          */
969         if (!retry)
970                 return NULL;
971
972         /* We're _really_ low on memory. Now we just
973          * wait for old buffer heads to become free due to
974          * finishing IO.  Since this is an async request and
975          * the reserve list is empty, we're sure there are 
976          * async buffer heads in use.
977          */
978         free_more_memory();
979         goto try_again;
980 }
981 EXPORT_SYMBOL_GPL(alloc_page_buffers);
982
983 static inline void
984 link_dev_buffers(struct page *page, struct buffer_head *head)
985 {
986         struct buffer_head *bh, *tail;
987
988         bh = head;
989         do {
990                 tail = bh;
991                 bh = bh->b_this_page;
992         } while (bh);
993         tail->b_this_page = head;
994         attach_page_buffers(page, head);
995 }
996
997 /*
998  * Initialise the state of a blockdev page's buffers.
999  */ 
1000 static void
1001 init_page_buffers(struct page *page, struct block_device *bdev,
1002                         sector_t block, int size)
1003 {
1004         struct buffer_head *head = page_buffers(page);
1005         struct buffer_head *bh = head;
1006         int uptodate = PageUptodate(page);
1007
1008         do {
1009                 if (!buffer_mapped(bh)) {
1010                         init_buffer(bh, NULL, NULL);
1011                         bh->b_bdev = bdev;
1012                         bh->b_blocknr = block;
1013                         if (uptodate)
1014                                 set_buffer_uptodate(bh);
1015                         set_buffer_mapped(bh);
1016                 }
1017                 block++;
1018                 bh = bh->b_this_page;
1019         } while (bh != head);
1020 }
1021
1022 /*
1023  * Create the page-cache page that contains the requested block.
1024  *
1025  * This is user purely for blockdev mappings.
1026  */
1027 static struct page *
1028 grow_dev_page(struct block_device *bdev, sector_t block,
1029                 pgoff_t index, int size)
1030 {
1031         struct inode *inode = bdev->bd_inode;
1032         struct page *page;
1033         struct buffer_head *bh;
1034
1035         page = find_or_create_page(inode->i_mapping, index,
1036                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
1037         if (!page)
1038                 return NULL;
1039
1040         BUG_ON(!PageLocked(page));
1041
1042         if (page_has_buffers(page)) {
1043                 bh = page_buffers(page);
1044                 if (bh->b_size == size) {
1045                         init_page_buffers(page, bdev, block, size);
1046                         return page;
1047                 }
1048                 if (!try_to_free_buffers(page))
1049                         goto failed;
1050         }
1051
1052         /*
1053          * Allocate some buffers for this page
1054          */
1055         bh = alloc_page_buffers(page, size, 0);
1056         if (!bh)
1057                 goto failed;
1058
1059         /*
1060          * Link the page to the buffers and initialise them.  Take the
1061          * lock to be atomic wrt __find_get_block(), which does not
1062          * run under the page lock.
1063          */
1064         spin_lock(&inode->i_mapping->private_lock);
1065         link_dev_buffers(page, bh);
1066         init_page_buffers(page, bdev, block, size);
1067         spin_unlock(&inode->i_mapping->private_lock);
1068         return page;
1069
1070 failed:
1071         BUG();
1072         unlock_page(page);
1073         page_cache_release(page);
1074         return NULL;
1075 }
1076
1077 /*
1078  * Create buffers for the specified block device block's page.  If
1079  * that page was dirty, the buffers are set dirty also.
1080  */
1081 static int
1082 grow_buffers(struct block_device *bdev, sector_t block, int size)
1083 {
1084         struct page *page;
1085         pgoff_t index;
1086         int sizebits;
1087
1088         sizebits = -1;
1089         do {
1090                 sizebits++;
1091         } while ((size << sizebits) < PAGE_SIZE);
1092
1093         index = block >> sizebits;
1094
1095         /*
1096          * Check for a block which wants to lie outside our maximum possible
1097          * pagecache index.  (this comparison is done using sector_t types).
1098          */
1099         if (unlikely(index != block >> sizebits)) {
1100                 char b[BDEVNAME_SIZE];
1101
1102                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1103                         "device %s\n",
1104                         __func__, (unsigned long long)block,
1105                         bdevname(bdev, b));
1106                 return -EIO;
1107         }
1108         block = index << sizebits;
1109         /* Create a page with the proper size buffers.. */
1110         page = grow_dev_page(bdev, block, index, size);
1111         if (!page)
1112                 return 0;
1113         unlock_page(page);
1114         page_cache_release(page);
1115         return 1;
1116 }
1117
1118 static struct buffer_head *
1119 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1120 {
1121         /* Size must be multiple of hard sectorsize */
1122         if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1123                         (size < 512 || size > PAGE_SIZE))) {
1124                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1125                                         size);
1126                 printk(KERN_ERR "hardsect size: %d\n",
1127                                         bdev_hardsect_size(bdev));
1128
1129                 dump_stack();
1130                 return NULL;
1131         }
1132
1133         for (;;) {
1134                 struct buffer_head * bh;
1135                 int ret;
1136
1137                 bh = __find_get_block(bdev, block, size);
1138                 if (bh)
1139                         return bh;
1140
1141                 ret = grow_buffers(bdev, block, size);
1142                 if (ret < 0)
1143                         return NULL;
1144                 if (ret == 0)
1145                         free_more_memory();
1146         }
1147 }
1148
1149 /*
1150  * The relationship between dirty buffers and dirty pages:
1151  *
1152  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1153  * the page is tagged dirty in its radix tree.
1154  *
1155  * At all times, the dirtiness of the buffers represents the dirtiness of
1156  * subsections of the page.  If the page has buffers, the page dirty bit is
1157  * merely a hint about the true dirty state.
1158  *
1159  * When a page is set dirty in its entirety, all its buffers are marked dirty
1160  * (if the page has buffers).
1161  *
1162  * When a buffer is marked dirty, its page is dirtied, but the page's other
1163  * buffers are not.
1164  *
1165  * Also.  When blockdev buffers are explicitly read with bread(), they
1166  * individually become uptodate.  But their backing page remains not
1167  * uptodate - even if all of its buffers are uptodate.  A subsequent
1168  * block_read_full_page() against that page will discover all the uptodate
1169  * buffers, will set the page uptodate and will perform no I/O.
1170  */
1171
1172 /**
1173  * mark_buffer_dirty - mark a buffer_head as needing writeout
1174  * @bh: the buffer_head to mark dirty
1175  *
1176  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1177  * backing page dirty, then tag the page as dirty in its address_space's radix
1178  * tree and then attach the address_space's inode to its superblock's dirty
1179  * inode list.
1180  *
1181  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1182  * mapping->tree_lock and the global inode_lock.
1183  */
1184 void mark_buffer_dirty(struct buffer_head *bh)
1185 {
1186         WARN_ON_ONCE(!buffer_uptodate(bh));
1187
1188         /*
1189          * Very *carefully* optimize the it-is-already-dirty case.
1190          *
1191          * Don't let the final "is it dirty" escape to before we
1192          * perhaps modified the buffer.
1193          */
1194         if (buffer_dirty(bh)) {
1195                 smp_mb();
1196                 if (buffer_dirty(bh))
1197                         return;
1198         }
1199
1200         if (!test_set_buffer_dirty(bh))
1201                 __set_page_dirty(bh->b_page, page_mapping(bh->b_page), 0);
1202 }
1203
1204 /*
1205  * Decrement a buffer_head's reference count.  If all buffers against a page
1206  * have zero reference count, are clean and unlocked, and if the page is clean
1207  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1208  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1209  * a page but it ends up not being freed, and buffers may later be reattached).
1210  */
1211 void __brelse(struct buffer_head * buf)
1212 {
1213         if (atomic_read(&buf->b_count)) {
1214                 put_bh(buf);
1215                 return;
1216         }
1217         printk(KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1218         WARN_ON(1);
1219 }
1220
1221 /*
1222  * bforget() is like brelse(), except it discards any
1223  * potentially dirty data.
1224  */
1225 void __bforget(struct buffer_head *bh)
1226 {
1227         clear_buffer_dirty(bh);
1228         if (bh->b_assoc_map) {
1229                 struct address_space *buffer_mapping = bh->b_page->mapping;
1230
1231                 spin_lock(&buffer_mapping->private_lock);
1232                 list_del_init(&bh->b_assoc_buffers);
1233                 bh->b_assoc_map = NULL;
1234                 spin_unlock(&buffer_mapping->private_lock);
1235         }
1236         __brelse(bh);
1237 }
1238
1239 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1240 {
1241         lock_buffer(bh);
1242         if (buffer_uptodate(bh)) {
1243                 unlock_buffer(bh);
1244                 return bh;
1245         } else {
1246                 get_bh(bh);
1247                 bh->b_end_io = end_buffer_read_sync;
1248                 submit_bh(READ, bh);
1249                 wait_on_buffer(bh);
1250                 if (buffer_uptodate(bh))
1251                         return bh;
1252         }
1253         brelse(bh);
1254         return NULL;
1255 }
1256
1257 /*
1258  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1259  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1260  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1261  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1262  * CPU's LRUs at the same time.
1263  *
1264  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1265  * sb_find_get_block().
1266  *
1267  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1268  * a local interrupt disable for that.
1269  */
1270
1271 #define BH_LRU_SIZE     8
1272
1273 struct bh_lru {
1274         struct buffer_head *bhs[BH_LRU_SIZE];
1275 };
1276
1277 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1278
1279 #ifdef CONFIG_SMP
1280 #define bh_lru_lock()   local_irq_disable()
1281 #define bh_lru_unlock() local_irq_enable()
1282 #else
1283 #define bh_lru_lock()   preempt_disable()
1284 #define bh_lru_unlock() preempt_enable()
1285 #endif
1286
1287 static inline void check_irqs_on(void)
1288 {
1289 #ifdef irqs_disabled
1290         BUG_ON(irqs_disabled());
1291 #endif
1292 }
1293
1294 /*
1295  * The LRU management algorithm is dopey-but-simple.  Sorry.
1296  */
1297 static void bh_lru_install(struct buffer_head *bh)
1298 {
1299         struct buffer_head *evictee = NULL;
1300         struct bh_lru *lru;
1301
1302         check_irqs_on();
1303         bh_lru_lock();
1304         lru = &__get_cpu_var(bh_lrus);
1305         if (lru->bhs[0] != bh) {
1306                 struct buffer_head *bhs[BH_LRU_SIZE];
1307                 int in;
1308                 int out = 0;
1309
1310                 get_bh(bh);
1311                 bhs[out++] = bh;
1312                 for (in = 0; in < BH_LRU_SIZE; in++) {
1313                         struct buffer_head *bh2 = lru->bhs[in];
1314
1315                         if (bh2 == bh) {
1316                                 __brelse(bh2);
1317                         } else {
1318                                 if (out >= BH_LRU_SIZE) {
1319                                         BUG_ON(evictee != NULL);
1320                                         evictee = bh2;
1321                                 } else {
1322                                         bhs[out++] = bh2;
1323                                 }
1324                         }
1325                 }
1326                 while (out < BH_LRU_SIZE)
1327                         bhs[out++] = NULL;
1328                 memcpy(lru->bhs, bhs, sizeof(bhs));
1329         }
1330         bh_lru_unlock();
1331
1332         if (evictee)
1333                 __brelse(evictee);
1334 }
1335
1336 /*
1337  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1338  */
1339 static struct buffer_head *
1340 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1341 {
1342         struct buffer_head *ret = NULL;
1343         struct bh_lru *lru;
1344         unsigned int i;
1345
1346         check_irqs_on();
1347         bh_lru_lock();
1348         lru = &__get_cpu_var(bh_lrus);
1349         for (i = 0; i < BH_LRU_SIZE; i++) {
1350                 struct buffer_head *bh = lru->bhs[i];
1351
1352                 if (bh && bh->b_bdev == bdev &&
1353                                 bh->b_blocknr == block && bh->b_size == size) {
1354                         if (i) {
1355                                 while (i) {
1356                                         lru->bhs[i] = lru->bhs[i - 1];
1357                                         i--;
1358                                 }
1359                                 lru->bhs[0] = bh;
1360                         }
1361                         get_bh(bh);
1362                         ret = bh;
1363                         break;
1364                 }
1365         }
1366         bh_lru_unlock();
1367         return ret;
1368 }
1369
1370 /*
1371  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1372  * it in the LRU and mark it as accessed.  If it is not present then return
1373  * NULL
1374  */
1375 struct buffer_head *
1376 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1377 {
1378         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1379
1380         if (bh == NULL) {
1381                 bh = __find_get_block_slow(bdev, block);
1382                 if (bh)
1383                         bh_lru_install(bh);
1384         }
1385         if (bh)
1386                 touch_buffer(bh);
1387         return bh;
1388 }
1389 EXPORT_SYMBOL(__find_get_block);
1390
1391 /*
1392  * __getblk will locate (and, if necessary, create) the buffer_head
1393  * which corresponds to the passed block_device, block and size. The
1394  * returned buffer has its reference count incremented.
1395  *
1396  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1397  * illegal block number, __getblk() will happily return a buffer_head
1398  * which represents the non-existent block.  Very weird.
1399  *
1400  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1401  * attempt is failing.  FIXME, perhaps?
1402  */
1403 struct buffer_head *
1404 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1405 {
1406         struct buffer_head *bh = __find_get_block(bdev, block, size);
1407
1408         might_sleep();
1409         if (bh == NULL)
1410                 bh = __getblk_slow(bdev, block, size);
1411         return bh;
1412 }
1413 EXPORT_SYMBOL(__getblk);
1414
1415 /*
1416  * Do async read-ahead on a buffer..
1417  */
1418 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1419 {
1420         struct buffer_head *bh = __getblk(bdev, block, size);
1421         if (likely(bh)) {
1422                 ll_rw_block(READA, 1, &bh);
1423                 brelse(bh);
1424         }
1425 }
1426 EXPORT_SYMBOL(__breadahead);
1427
1428 /**
1429  *  __bread() - reads a specified block and returns the bh
1430  *  @bdev: the block_device to read from
1431  *  @block: number of block
1432  *  @size: size (in bytes) to read
1433  * 
1434  *  Reads a specified block, and returns buffer head that contains it.
1435  *  It returns NULL if the block was unreadable.
1436  */
1437 struct buffer_head *
1438 __bread(struct block_device *bdev, sector_t block, unsigned size)
1439 {
1440         struct buffer_head *bh = __getblk(bdev, block, size);
1441
1442         if (likely(bh) && !buffer_uptodate(bh))
1443                 bh = __bread_slow(bh);
1444         return bh;
1445 }
1446 EXPORT_SYMBOL(__bread);
1447
1448 /*
1449  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1450  * This doesn't race because it runs in each cpu either in irq
1451  * or with preempt disabled.
1452  */
1453 static void invalidate_bh_lru(void *arg)
1454 {
1455         struct bh_lru *b = &get_cpu_var(bh_lrus);
1456         int i;
1457
1458         for (i = 0; i < BH_LRU_SIZE; i++) {
1459                 brelse(b->bhs[i]);
1460                 b->bhs[i] = NULL;
1461         }
1462         put_cpu_var(bh_lrus);
1463 }
1464         
1465 void invalidate_bh_lrus(void)
1466 {
1467         on_each_cpu(invalidate_bh_lru, NULL, 1);
1468 }
1469 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1470
1471 void set_bh_page(struct buffer_head *bh,
1472                 struct page *page, unsigned long offset)
1473 {
1474         bh->b_page = page;
1475         BUG_ON(offset >= PAGE_SIZE);
1476         if (PageHighMem(page))
1477                 /*
1478                  * This catches illegal uses and preserves the offset:
1479                  */
1480                 bh->b_data = (char *)(0 + offset);
1481         else
1482                 bh->b_data = page_address(page) + offset;
1483 }
1484 EXPORT_SYMBOL(set_bh_page);
1485
1486 /*
1487  * Called when truncating a buffer on a page completely.
1488  */
1489 static void discard_buffer(struct buffer_head * bh)
1490 {
1491         lock_buffer(bh);
1492         clear_buffer_dirty(bh);
1493         bh->b_bdev = NULL;
1494         clear_buffer_mapped(bh);
1495         clear_buffer_req(bh);
1496         clear_buffer_new(bh);
1497         clear_buffer_delay(bh);
1498         clear_buffer_unwritten(bh);
1499         unlock_buffer(bh);
1500 }
1501
1502 /**
1503  * block_invalidatepage - invalidate part of all of a buffer-backed page
1504  *
1505  * @page: the page which is affected
1506  * @offset: the index of the truncation point
1507  *
1508  * block_invalidatepage() is called when all or part of the page has become
1509  * invalidatedby a truncate operation.
1510  *
1511  * block_invalidatepage() does not have to release all buffers, but it must
1512  * ensure that no dirty buffer is left outside @offset and that no I/O
1513  * is underway against any of the blocks which are outside the truncation
1514  * point.  Because the caller is about to free (and possibly reuse) those
1515  * blocks on-disk.
1516  */
1517 void block_invalidatepage(struct page *page, unsigned long offset)
1518 {
1519         struct buffer_head *head, *bh, *next;
1520         unsigned int curr_off = 0;
1521
1522         BUG_ON(!PageLocked(page));
1523         if (!page_has_buffers(page))
1524                 goto out;
1525
1526         head = page_buffers(page);
1527         bh = head;
1528         do {
1529                 unsigned int next_off = curr_off + bh->b_size;
1530                 next = bh->b_this_page;
1531
1532                 /*
1533                  * is this block fully invalidated?
1534                  */
1535                 if (offset <= curr_off)
1536                         discard_buffer(bh);
1537                 curr_off = next_off;
1538                 bh = next;
1539         } while (bh != head);
1540
1541         /*
1542          * We release buffers only if the entire page is being invalidated.
1543          * The get_block cached value has been unconditionally invalidated,
1544          * so real IO is not possible anymore.
1545          */
1546         if (offset == 0)
1547                 try_to_release_page(page, 0);
1548 out:
1549         return;
1550 }
1551 EXPORT_SYMBOL(block_invalidatepage);
1552
1553 /*
1554  * We attach and possibly dirty the buffers atomically wrt
1555  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1556  * is already excluded via the page lock.
1557  */
1558 void create_empty_buffers(struct page *page,
1559                         unsigned long blocksize, unsigned long b_state)
1560 {
1561         struct buffer_head *bh, *head, *tail;
1562
1563         head = alloc_page_buffers(page, blocksize, 1);
1564         bh = head;
1565         do {
1566                 bh->b_state |= b_state;
1567                 tail = bh;
1568                 bh = bh->b_this_page;
1569         } while (bh);
1570         tail->b_this_page = head;
1571
1572         spin_lock(&page->mapping->private_lock);
1573         if (PageUptodate(page) || PageDirty(page)) {
1574                 bh = head;
1575                 do {
1576                         if (PageDirty(page))
1577                                 set_buffer_dirty(bh);
1578                         if (PageUptodate(page))
1579                                 set_buffer_uptodate(bh);
1580                         bh = bh->b_this_page;
1581                 } while (bh != head);
1582         }
1583         attach_page_buffers(page, head);
1584         spin_unlock(&page->mapping->private_lock);
1585 }
1586 EXPORT_SYMBOL(create_empty_buffers);
1587
1588 /*
1589  * We are taking a block for data and we don't want any output from any
1590  * buffer-cache aliases starting from return from that function and
1591  * until the moment when something will explicitly mark the buffer
1592  * dirty (hopefully that will not happen until we will free that block ;-)
1593  * We don't even need to mark it not-uptodate - nobody can expect
1594  * anything from a newly allocated buffer anyway. We used to used
1595  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1596  * don't want to mark the alias unmapped, for example - it would confuse
1597  * anyone who might pick it with bread() afterwards...
1598  *
1599  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1600  * be writeout I/O going on against recently-freed buffers.  We don't
1601  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1602  * only if we really need to.  That happens here.
1603  */
1604 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1605 {
1606         struct buffer_head *old_bh;
1607
1608         might_sleep();
1609
1610         old_bh = __find_get_block_slow(bdev, block);
1611         if (old_bh) {
1612                 clear_buffer_dirty(old_bh);
1613                 wait_on_buffer(old_bh);
1614                 clear_buffer_req(old_bh);
1615                 __brelse(old_bh);
1616         }
1617 }
1618 EXPORT_SYMBOL(unmap_underlying_metadata);
1619
1620 /*
1621  * NOTE! All mapped/uptodate combinations are valid:
1622  *
1623  *      Mapped  Uptodate        Meaning
1624  *
1625  *      No      No              "unknown" - must do get_block()
1626  *      No      Yes             "hole" - zero-filled
1627  *      Yes     No              "allocated" - allocated on disk, not read in
1628  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1629  *
1630  * "Dirty" is valid only with the last case (mapped+uptodate).
1631  */
1632
1633 /*
1634  * While block_write_full_page is writing back the dirty buffers under
1635  * the page lock, whoever dirtied the buffers may decide to clean them
1636  * again at any time.  We handle that by only looking at the buffer
1637  * state inside lock_buffer().
1638  *
1639  * If block_write_full_page() is called for regular writeback
1640  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1641  * locked buffer.   This only can happen if someone has written the buffer
1642  * directly, with submit_bh().  At the address_space level PageWriteback
1643  * prevents this contention from occurring.
1644  */
1645 static int __block_write_full_page(struct inode *inode, struct page *page,
1646                         get_block_t *get_block, struct writeback_control *wbc)
1647 {
1648         int err;
1649         sector_t block;
1650         sector_t last_block;
1651         struct buffer_head *bh, *head;
1652         const unsigned blocksize = 1 << inode->i_blkbits;
1653         int nr_underway = 0;
1654
1655         BUG_ON(!PageLocked(page));
1656
1657         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1658
1659         if (!page_has_buffers(page)) {
1660                 create_empty_buffers(page, blocksize,
1661                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1662         }
1663
1664         /*
1665          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1666          * here, and the (potentially unmapped) buffers may become dirty at
1667          * any time.  If a buffer becomes dirty here after we've inspected it
1668          * then we just miss that fact, and the page stays dirty.
1669          *
1670          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1671          * handle that here by just cleaning them.
1672          */
1673
1674         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1675         head = page_buffers(page);
1676         bh = head;
1677
1678         /*
1679          * Get all the dirty buffers mapped to disk addresses and
1680          * handle any aliases from the underlying blockdev's mapping.
1681          */
1682         do {
1683                 if (block > last_block) {
1684                         /*
1685                          * mapped buffers outside i_size will occur, because
1686                          * this page can be outside i_size when there is a
1687                          * truncate in progress.
1688                          */
1689                         /*
1690                          * The buffer was zeroed by block_write_full_page()
1691                          */
1692                         clear_buffer_dirty(bh);
1693                         set_buffer_uptodate(bh);
1694                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1695                            buffer_dirty(bh)) {
1696                         WARN_ON(bh->b_size != blocksize);
1697                         err = get_block(inode, block, bh, 1);
1698                         if (err)
1699                                 goto recover;
1700                         clear_buffer_delay(bh);
1701                         if (buffer_new(bh)) {
1702                                 /* blockdev mappings never come here */
1703                                 clear_buffer_new(bh);
1704                                 unmap_underlying_metadata(bh->b_bdev,
1705                                                         bh->b_blocknr);
1706                         }
1707                 }
1708                 bh = bh->b_this_page;
1709                 block++;
1710         } while (bh != head);
1711
1712         do {
1713                 if (!buffer_mapped(bh))
1714                         continue;
1715                 /*
1716                  * If it's a fully non-blocking write attempt and we cannot
1717                  * lock the buffer then redirty the page.  Note that this can
1718                  * potentially cause a busy-wait loop from pdflush and kswapd
1719                  * activity, but those code paths have their own higher-level
1720                  * throttling.
1721                  */
1722                 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1723                         lock_buffer(bh);
1724                 } else if (test_set_buffer_locked(bh)) {
1725                         redirty_page_for_writepage(wbc, page);
1726                         continue;
1727                 }
1728                 if (test_clear_buffer_dirty(bh)) {
1729                         mark_buffer_async_write(bh);
1730                 } else {
1731                         unlock_buffer(bh);
1732                 }
1733         } while ((bh = bh->b_this_page) != head);
1734
1735         /*
1736          * The page and its buffers are protected by PageWriteback(), so we can
1737          * drop the bh refcounts early.
1738          */
1739         BUG_ON(PageWriteback(page));
1740         set_page_writeback(page);
1741
1742         do {
1743                 struct buffer_head *next = bh->b_this_page;
1744                 if (buffer_async_write(bh)) {
1745                         submit_bh(WRITE, bh);
1746                         nr_underway++;
1747                 }
1748                 bh = next;
1749         } while (bh != head);
1750         unlock_page(page);
1751
1752         err = 0;
1753 done:
1754         if (nr_underway == 0) {
1755                 /*
1756                  * The page was marked dirty, but the buffers were
1757                  * clean.  Someone wrote them back by hand with
1758                  * ll_rw_block/submit_bh.  A rare case.
1759                  */
1760                 end_page_writeback(page);
1761
1762                 /*
1763                  * The page and buffer_heads can be released at any time from
1764                  * here on.
1765                  */
1766         }
1767         return err;
1768
1769 recover:
1770         /*
1771          * ENOSPC, or some other error.  We may already have added some
1772          * blocks to the file, so we need to write these out to avoid
1773          * exposing stale data.
1774          * The page is currently locked and not marked for writeback
1775          */
1776         bh = head;
1777         /* Recovery: lock and submit the mapped buffers */
1778         do {
1779                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1780                     !buffer_delay(bh)) {
1781                         lock_buffer(bh);
1782                         mark_buffer_async_write(bh);
1783                 } else {
1784                         /*
1785                          * The buffer may have been set dirty during
1786                          * attachment to a dirty page.
1787                          */
1788                         clear_buffer_dirty(bh);
1789                 }
1790         } while ((bh = bh->b_this_page) != head);
1791         SetPageError(page);
1792         BUG_ON(PageWriteback(page));
1793         mapping_set_error(page->mapping, err);
1794         set_page_writeback(page);
1795         do {
1796                 struct buffer_head *next = bh->b_this_page;
1797                 if (buffer_async_write(bh)) {
1798                         clear_buffer_dirty(bh);
1799                         submit_bh(WRITE, bh);
1800                         nr_underway++;
1801                 }
1802                 bh = next;
1803         } while (bh != head);
1804         unlock_page(page);
1805         goto done;
1806 }
1807
1808 /*
1809  * If a page has any new buffers, zero them out here, and mark them uptodate
1810  * and dirty so they'll be written out (in order to prevent uninitialised
1811  * block data from leaking). And clear the new bit.
1812  */
1813 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1814 {
1815         unsigned int block_start, block_end;
1816         struct buffer_head *head, *bh;
1817
1818         BUG_ON(!PageLocked(page));
1819         if (!page_has_buffers(page))
1820                 return;
1821
1822         bh = head = page_buffers(page);
1823         block_start = 0;
1824         do {
1825                 block_end = block_start + bh->b_size;
1826
1827                 if (buffer_new(bh)) {
1828                         if (block_end > from && block_start < to) {
1829                                 if (!PageUptodate(page)) {
1830                                         unsigned start, size;
1831
1832                                         start = max(from, block_start);
1833                                         size = min(to, block_end) - start;
1834
1835                                         zero_user(page, start, size);
1836                                         set_buffer_uptodate(bh);
1837                                 }
1838
1839                                 clear_buffer_new(bh);
1840                                 mark_buffer_dirty(bh);
1841                         }
1842                 }
1843
1844                 block_start = block_end;
1845                 bh = bh->b_this_page;
1846         } while (bh != head);
1847 }
1848 EXPORT_SYMBOL(page_zero_new_buffers);
1849
1850 static int __block_prepare_write(struct inode *inode, struct page *page,
1851                 unsigned from, unsigned to, get_block_t *get_block)
1852 {
1853         unsigned block_start, block_end;
1854         sector_t block;
1855         int err = 0;
1856         unsigned blocksize, bbits;
1857         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1858
1859         BUG_ON(!PageLocked(page));
1860         BUG_ON(from > PAGE_CACHE_SIZE);
1861         BUG_ON(to > PAGE_CACHE_SIZE);
1862         BUG_ON(from > to);
1863
1864         blocksize = 1 << inode->i_blkbits;
1865         if (!page_has_buffers(page))
1866                 create_empty_buffers(page, blocksize, 0);
1867         head = page_buffers(page);
1868
1869         bbits = inode->i_blkbits;
1870         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1871
1872         for(bh = head, block_start = 0; bh != head || !block_start;
1873             block++, block_start=block_end, bh = bh->b_this_page) {
1874                 block_end = block_start + blocksize;
1875                 if (block_end <= from || block_start >= to) {
1876                         if (PageUptodate(page)) {
1877                                 if (!buffer_uptodate(bh))
1878                                         set_buffer_uptodate(bh);
1879                         }
1880                         continue;
1881                 }
1882                 if (buffer_new(bh))
1883                         clear_buffer_new(bh);
1884                 if (!buffer_mapped(bh)) {
1885                         WARN_ON(bh->b_size != blocksize);
1886                         err = get_block(inode, block, bh, 1);
1887                         if (err)
1888                                 break;
1889                         if (buffer_new(bh)) {
1890                                 unmap_underlying_metadata(bh->b_bdev,
1891                                                         bh->b_blocknr);
1892                                 if (PageUptodate(page)) {
1893                                         clear_buffer_new(bh);
1894                                         set_buffer_uptodate(bh);
1895                                         mark_buffer_dirty(bh);
1896                                         continue;
1897                                 }
1898                                 if (block_end > to || block_start < from)
1899                                         zero_user_segments(page,
1900                                                 to, block_end,
1901                                                 block_start, from);
1902                                 continue;
1903                         }
1904                 }
1905                 if (PageUptodate(page)) {
1906                         if (!buffer_uptodate(bh))
1907                                 set_buffer_uptodate(bh);
1908                         continue; 
1909                 }
1910                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1911                     !buffer_unwritten(bh) &&
1912                      (block_start < from || block_end > to)) {
1913                         ll_rw_block(READ, 1, &bh);
1914                         *wait_bh++=bh;
1915                 }
1916         }
1917         /*
1918          * If we issued read requests - let them complete.
1919          */
1920         while(wait_bh > wait) {
1921                 wait_on_buffer(*--wait_bh);
1922                 if (!buffer_uptodate(*wait_bh))
1923                         err = -EIO;
1924         }
1925         if (unlikely(err))
1926                 page_zero_new_buffers(page, from, to);
1927         return err;
1928 }
1929
1930 static int __block_commit_write(struct inode *inode, struct page *page,
1931                 unsigned from, unsigned to)
1932 {
1933         unsigned block_start, block_end;
1934         int partial = 0;
1935         unsigned blocksize;
1936         struct buffer_head *bh, *head;
1937
1938         blocksize = 1 << inode->i_blkbits;
1939
1940         for(bh = head = page_buffers(page), block_start = 0;
1941             bh != head || !block_start;
1942             block_start=block_end, bh = bh->b_this_page) {
1943                 block_end = block_start + blocksize;
1944                 if (block_end <= from || block_start >= to) {
1945                         if (!buffer_uptodate(bh))
1946                                 partial = 1;
1947                 } else {
1948                         set_buffer_uptodate(bh);
1949                         mark_buffer_dirty(bh);
1950                 }
1951                 clear_buffer_new(bh);
1952         }
1953
1954         /*
1955          * If this is a partial write which happened to make all buffers
1956          * uptodate then we can optimize away a bogus readpage() for
1957          * the next read(). Here we 'discover' whether the page went
1958          * uptodate as a result of this (potentially partial) write.
1959          */
1960         if (!partial)
1961                 SetPageUptodate(page);
1962         return 0;
1963 }
1964
1965 /*
1966  * block_write_begin takes care of the basic task of block allocation and
1967  * bringing partial write blocks uptodate first.
1968  *
1969  * If *pagep is not NULL, then block_write_begin uses the locked page
1970  * at *pagep rather than allocating its own. In this case, the page will
1971  * not be unlocked or deallocated on failure.
1972  */
1973 int block_write_begin(struct file *file, struct address_space *mapping,
1974                         loff_t pos, unsigned len, unsigned flags,
1975                         struct page **pagep, void **fsdata,
1976                         get_block_t *get_block)
1977 {
1978         struct inode *inode = mapping->host;
1979         int status = 0;
1980         struct page *page;
1981         pgoff_t index;
1982         unsigned start, end;
1983         int ownpage = 0;
1984
1985         index = pos >> PAGE_CACHE_SHIFT;
1986         start = pos & (PAGE_CACHE_SIZE - 1);
1987         end = start + len;
1988
1989         page = *pagep;
1990         if (page == NULL) {
1991                 ownpage = 1;
1992                 page = __grab_cache_page(mapping, index);
1993                 if (!page) {
1994                         status = -ENOMEM;
1995                         goto out;
1996                 }
1997                 *pagep = page;
1998         } else
1999                 BUG_ON(!PageLocked(page));
2000
2001         status = __block_prepare_write(inode, page, start, end, get_block);
2002         if (unlikely(status)) {
2003                 ClearPageUptodate(page);
2004
2005                 if (ownpage) {
2006                         unlock_page(page);
2007                         page_cache_release(page);
2008                         *pagep = NULL;
2009
2010                         /*
2011                          * prepare_write() may have instantiated a few blocks
2012                          * outside i_size.  Trim these off again. Don't need
2013                          * i_size_read because we hold i_mutex.
2014                          */
2015                         if (pos + len > inode->i_size)
2016                                 vmtruncate(inode, inode->i_size);
2017                 }
2018                 goto out;
2019         }
2020
2021 out:
2022         return status;
2023 }
2024 EXPORT_SYMBOL(block_write_begin);
2025
2026 int block_write_end(struct file *file, struct address_space *mapping,
2027                         loff_t pos, unsigned len, unsigned copied,
2028                         struct page *page, void *fsdata)
2029 {
2030         struct inode *inode = mapping->host;
2031         unsigned start;
2032
2033         start = pos & (PAGE_CACHE_SIZE - 1);
2034
2035         if (unlikely(copied < len)) {
2036                 /*
2037                  * The buffers that were written will now be uptodate, so we
2038                  * don't have to worry about a readpage reading them and
2039                  * overwriting a partial write. However if we have encountered
2040                  * a short write and only partially written into a buffer, it
2041                  * will not be marked uptodate, so a readpage might come in and
2042                  * destroy our partial write.
2043                  *
2044                  * Do the simplest thing, and just treat any short write to a
2045                  * non uptodate page as a zero-length write, and force the
2046                  * caller to redo the whole thing.
2047                  */
2048                 if (!PageUptodate(page))
2049                         copied = 0;
2050
2051                 page_zero_new_buffers(page, start+copied, start+len);
2052         }
2053         flush_dcache_page(page);
2054
2055         /* This could be a short (even 0-length) commit */
2056         __block_commit_write(inode, page, start, start+copied);
2057
2058         return copied;
2059 }
2060 EXPORT_SYMBOL(block_write_end);
2061
2062 int generic_write_end(struct file *file, struct address_space *mapping,
2063                         loff_t pos, unsigned len, unsigned copied,
2064                         struct page *page, void *fsdata)
2065 {
2066         struct inode *inode = mapping->host;
2067         int i_size_changed = 0;
2068
2069         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2070
2071         /*
2072          * No need to use i_size_read() here, the i_size
2073          * cannot change under us because we hold i_mutex.
2074          *
2075          * But it's important to update i_size while still holding page lock:
2076          * page writeout could otherwise come in and zero beyond i_size.
2077          */
2078         if (pos+copied > inode->i_size) {
2079                 i_size_write(inode, pos+copied);
2080                 i_size_changed = 1;
2081         }
2082
2083         unlock_page(page);
2084         page_cache_release(page);
2085
2086         /*
2087          * Don't mark the inode dirty under page lock. First, it unnecessarily
2088          * makes the holding time of page lock longer. Second, it forces lock
2089          * ordering of page lock and transaction start for journaling
2090          * filesystems.
2091          */
2092         if (i_size_changed)
2093                 mark_inode_dirty(inode);
2094
2095         return copied;
2096 }
2097 EXPORT_SYMBOL(generic_write_end);
2098
2099 /*
2100  * Generic "read page" function for block devices that have the normal
2101  * get_block functionality. This is most of the block device filesystems.
2102  * Reads the page asynchronously --- the unlock_buffer() and
2103  * set/clear_buffer_uptodate() functions propagate buffer state into the
2104  * page struct once IO has completed.
2105  */
2106 int block_read_full_page(struct page *page, get_block_t *get_block)
2107 {
2108         struct inode *inode = page->mapping->host;
2109         sector_t iblock, lblock;
2110         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2111         unsigned int blocksize;
2112         int nr, i;
2113         int fully_mapped = 1;
2114
2115         BUG_ON(!PageLocked(page));
2116         blocksize = 1 << inode->i_blkbits;
2117         if (!page_has_buffers(page))
2118                 create_empty_buffers(page, blocksize, 0);
2119         head = page_buffers(page);
2120
2121         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2122         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2123         bh = head;
2124         nr = 0;
2125         i = 0;
2126
2127         do {
2128                 if (buffer_uptodate(bh))
2129                         continue;
2130
2131                 if (!buffer_mapped(bh)) {
2132                         int err = 0;
2133
2134                         fully_mapped = 0;
2135                         if (iblock < lblock) {
2136                                 WARN_ON(bh->b_size != blocksize);
2137                                 err = get_block(inode, iblock, bh, 0);
2138                                 if (err)
2139                                         SetPageError(page);
2140                         }
2141                         if (!buffer_mapped(bh)) {
2142                                 zero_user(page, i * blocksize, blocksize);
2143                                 if (!err)
2144                                         set_buffer_uptodate(bh);
2145                                 continue;
2146                         }
2147                         /*
2148                          * get_block() might have updated the buffer
2149                          * synchronously
2150                          */
2151                         if (buffer_uptodate(bh))
2152                                 continue;
2153                 }
2154                 arr[nr++] = bh;
2155         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2156
2157         if (fully_mapped)
2158                 SetPageMappedToDisk(page);
2159
2160         if (!nr) {
2161                 /*
2162                  * All buffers are uptodate - we can set the page uptodate
2163                  * as well. But not if get_block() returned an error.
2164                  */
2165                 if (!PageError(page))
2166                         SetPageUptodate(page);
2167                 unlock_page(page);
2168                 return 0;
2169         }
2170
2171         /* Stage two: lock the buffers */
2172         for (i = 0; i < nr; i++) {
2173                 bh = arr[i];
2174                 lock_buffer(bh);
2175                 mark_buffer_async_read(bh);
2176         }
2177
2178         /*
2179          * Stage 3: start the IO.  Check for uptodateness
2180          * inside the buffer lock in case another process reading
2181          * the underlying blockdev brought it uptodate (the sct fix).
2182          */
2183         for (i = 0; i < nr; i++) {
2184                 bh = arr[i];
2185                 if (buffer_uptodate(bh))
2186                         end_buffer_async_read(bh, 1);
2187                 else
2188                         submit_bh(READ, bh);
2189         }
2190         return 0;
2191 }
2192
2193 /* utility function for filesystems that need to do work on expanding
2194  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2195  * deal with the hole.  
2196  */
2197 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2198 {
2199         struct address_space *mapping = inode->i_mapping;
2200         struct page *page;
2201         void *fsdata;
2202         unsigned long limit;
2203         int err;
2204
2205         err = -EFBIG;
2206         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2207         if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2208                 send_sig(SIGXFSZ, current, 0);
2209                 goto out;
2210         }
2211         if (size > inode->i_sb->s_maxbytes)
2212                 goto out;
2213
2214         err = pagecache_write_begin(NULL, mapping, size, 0,
2215                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2216                                 &page, &fsdata);
2217         if (err)
2218                 goto out;
2219
2220         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2221         BUG_ON(err > 0);
2222
2223 out:
2224         return err;
2225 }
2226
2227 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2228                             loff_t pos, loff_t *bytes)
2229 {
2230         struct inode *inode = mapping->host;
2231         unsigned blocksize = 1 << inode->i_blkbits;
2232         struct page *page;
2233         void *fsdata;
2234         pgoff_t index, curidx;
2235         loff_t curpos;
2236         unsigned zerofrom, offset, len;
2237         int err = 0;
2238
2239         index = pos >> PAGE_CACHE_SHIFT;
2240         offset = pos & ~PAGE_CACHE_MASK;
2241
2242         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2243                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2244                 if (zerofrom & (blocksize-1)) {
2245                         *bytes |= (blocksize-1);
2246                         (*bytes)++;
2247                 }
2248                 len = PAGE_CACHE_SIZE - zerofrom;
2249
2250                 err = pagecache_write_begin(file, mapping, curpos, len,
2251                                                 AOP_FLAG_UNINTERRUPTIBLE,
2252                                                 &page, &fsdata);
2253                 if (err)
2254                         goto out;
2255                 zero_user(page, zerofrom, len);
2256                 err = pagecache_write_end(file, mapping, curpos, len, len,
2257                                                 page, fsdata);
2258                 if (err < 0)
2259                         goto out;
2260                 BUG_ON(err != len);
2261                 err = 0;
2262
2263                 balance_dirty_pages_ratelimited(mapping);
2264         }
2265
2266         /* page covers the boundary, find the boundary offset */
2267         if (index == curidx) {
2268                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2269                 /* if we will expand the thing last block will be filled */
2270                 if (offset <= zerofrom) {
2271                         goto out;
2272                 }
2273                 if (zerofrom & (blocksize-1)) {
2274                         *bytes |= (blocksize-1);
2275                         (*bytes)++;
2276                 }
2277                 len = offset - zerofrom;
2278
2279                 err = pagecache_write_begin(file, mapping, curpos, len,
2280                                                 AOP_FLAG_UNINTERRUPTIBLE,
2281                                                 &page, &fsdata);
2282                 if (err)
2283                         goto out;
2284                 zero_user(page, zerofrom, len);
2285                 err = pagecache_write_end(file, mapping, curpos, len, len,
2286                                                 page, fsdata);
2287                 if (err < 0)
2288                         goto out;
2289                 BUG_ON(err != len);
2290                 err = 0;
2291         }
2292 out:
2293         return err;
2294 }
2295
2296 /*
2297  * For moronic filesystems that do not allow holes in file.
2298  * We may have to extend the file.
2299  */
2300 int cont_write_begin(struct file *file, struct address_space *mapping,
2301                         loff_t pos, unsigned len, unsigned flags,
2302                         struct page **pagep, void **fsdata,
2303                         get_block_t *get_block, loff_t *bytes)
2304 {
2305         struct inode *inode = mapping->host;
2306         unsigned blocksize = 1 << inode->i_blkbits;
2307         unsigned zerofrom;
2308         int err;
2309
2310         err = cont_expand_zero(file, mapping, pos, bytes);
2311         if (err)
2312                 goto out;
2313
2314         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2315         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2316                 *bytes |= (blocksize-1);
2317                 (*bytes)++;
2318         }
2319
2320         *pagep = NULL;
2321         err = block_write_begin(file, mapping, pos, len,
2322                                 flags, pagep, fsdata, get_block);
2323 out:
2324         return err;
2325 }
2326
2327 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2328                         get_block_t *get_block)
2329 {
2330         struct inode *inode = page->mapping->host;
2331         int err = __block_prepare_write(inode, page, from, to, get_block);
2332         if (err)
2333                 ClearPageUptodate(page);
2334         return err;
2335 }
2336
2337 int block_commit_write(struct page *page, unsigned from, unsigned to)
2338 {
2339         struct inode *inode = page->mapping->host;
2340         __block_commit_write(inode,page,from,to);
2341         return 0;
2342 }
2343
2344 /*
2345  * block_page_mkwrite() is not allowed to change the file size as it gets
2346  * called from a page fault handler when a page is first dirtied. Hence we must
2347  * be careful to check for EOF conditions here. We set the page up correctly
2348  * for a written page which means we get ENOSPC checking when writing into
2349  * holes and correct delalloc and unwritten extent mapping on filesystems that
2350  * support these features.
2351  *
2352  * We are not allowed to take the i_mutex here so we have to play games to
2353  * protect against truncate races as the page could now be beyond EOF.  Because
2354  * vmtruncate() writes the inode size before removing pages, once we have the
2355  * page lock we can determine safely if the page is beyond EOF. If it is not
2356  * beyond EOF, then the page is guaranteed safe against truncation until we
2357  * unlock the page.
2358  */
2359 int
2360 block_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2361                    get_block_t get_block)
2362 {
2363         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2364         unsigned long end;
2365         loff_t size;
2366         int ret = -EINVAL;
2367
2368         lock_page(page);
2369         size = i_size_read(inode);
2370         if ((page->mapping != inode->i_mapping) ||
2371             (page_offset(page) > size)) {
2372                 /* page got truncated out from underneath us */
2373                 goto out_unlock;
2374         }
2375
2376         /* page is wholly or partially inside EOF */
2377         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2378                 end = size & ~PAGE_CACHE_MASK;
2379         else
2380                 end = PAGE_CACHE_SIZE;
2381
2382         ret = block_prepare_write(page, 0, end, get_block);
2383         if (!ret)
2384                 ret = block_commit_write(page, 0, end);
2385
2386 out_unlock:
2387         unlock_page(page);
2388         return ret;
2389 }
2390
2391 /*
2392  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2393  * immediately, while under the page lock.  So it needs a special end_io
2394  * handler which does not touch the bh after unlocking it.
2395  */
2396 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2397 {
2398         __end_buffer_read_notouch(bh, uptodate);
2399 }
2400
2401 /*
2402  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2403  * the page (converting it to circular linked list and taking care of page
2404  * dirty races).
2405  */
2406 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2407 {
2408         struct buffer_head *bh;
2409
2410         BUG_ON(!PageLocked(page));
2411
2412         spin_lock(&page->mapping->private_lock);
2413         bh = head;
2414         do {
2415                 if (PageDirty(page))
2416                         set_buffer_dirty(bh);
2417                 if (!bh->b_this_page)
2418                         bh->b_this_page = head;
2419                 bh = bh->b_this_page;
2420         } while (bh != head);
2421         attach_page_buffers(page, head);
2422         spin_unlock(&page->mapping->private_lock);
2423 }
2424
2425 /*
2426  * On entry, the page is fully not uptodate.
2427  * On exit the page is fully uptodate in the areas outside (from,to)
2428  */
2429 int nobh_write_begin(struct file *file, struct address_space *mapping,
2430                         loff_t pos, unsigned len, unsigned flags,
2431                         struct page **pagep, void **fsdata,
2432                         get_block_t *get_block)
2433 {
2434         struct inode *inode = mapping->host;
2435         const unsigned blkbits = inode->i_blkbits;
2436         const unsigned blocksize = 1 << blkbits;
2437         struct buffer_head *head, *bh;
2438         struct page *page;
2439         pgoff_t index;
2440         unsigned from, to;
2441         unsigned block_in_page;
2442         unsigned block_start, block_end;
2443         sector_t block_in_file;
2444         int nr_reads = 0;
2445         int ret = 0;
2446         int is_mapped_to_disk = 1;
2447
2448         index = pos >> PAGE_CACHE_SHIFT;
2449         from = pos & (PAGE_CACHE_SIZE - 1);
2450         to = from + len;
2451
2452         page = __grab_cache_page(mapping, index);
2453         if (!page)
2454                 return -ENOMEM;
2455         *pagep = page;
2456         *fsdata = NULL;
2457
2458         if (page_has_buffers(page)) {
2459                 unlock_page(page);
2460                 page_cache_release(page);
2461                 *pagep = NULL;
2462                 return block_write_begin(file, mapping, pos, len, flags, pagep,
2463                                         fsdata, get_block);
2464         }
2465
2466         if (PageMappedToDisk(page))
2467                 return 0;
2468
2469         /*
2470          * Allocate buffers so that we can keep track of state, and potentially
2471          * attach them to the page if an error occurs. In the common case of
2472          * no error, they will just be freed again without ever being attached
2473          * to the page (which is all OK, because we're under the page lock).
2474          *
2475          * Be careful: the buffer linked list is a NULL terminated one, rather
2476          * than the circular one we're used to.
2477          */
2478         head = alloc_page_buffers(page, blocksize, 0);
2479         if (!head) {
2480                 ret = -ENOMEM;
2481                 goto out_release;
2482         }
2483
2484         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2485
2486         /*
2487          * We loop across all blocks in the page, whether or not they are
2488          * part of the affected region.  This is so we can discover if the
2489          * page is fully mapped-to-disk.
2490          */
2491         for (block_start = 0, block_in_page = 0, bh = head;
2492                   block_start < PAGE_CACHE_SIZE;
2493                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2494                 int create;
2495
2496                 block_end = block_start + blocksize;
2497                 bh->b_state = 0;
2498                 create = 1;
2499                 if (block_start >= to)
2500                         create = 0;
2501                 ret = get_block(inode, block_in_file + block_in_page,
2502                                         bh, create);
2503                 if (ret)
2504                         goto failed;
2505                 if (!buffer_mapped(bh))
2506                         is_mapped_to_disk = 0;
2507                 if (buffer_new(bh))
2508                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2509                 if (PageUptodate(page)) {
2510                         set_buffer_uptodate(bh);
2511                         continue;
2512                 }
2513                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2514                         zero_user_segments(page, block_start, from,
2515                                                         to, block_end);
2516                         continue;
2517                 }
2518                 if (buffer_uptodate(bh))
2519                         continue;       /* reiserfs does this */
2520                 if (block_start < from || block_end > to) {
2521                         lock_buffer(bh);
2522                         bh->b_end_io = end_buffer_read_nobh;
2523                         submit_bh(READ, bh);
2524                         nr_reads++;
2525                 }
2526         }
2527
2528         if (nr_reads) {
2529                 /*
2530                  * The page is locked, so these buffers are protected from
2531                  * any VM or truncate activity.  Hence we don't need to care
2532                  * for the buffer_head refcounts.
2533                  */
2534                 for (bh = head; bh; bh = bh->b_this_page) {
2535                         wait_on_buffer(bh);
2536                         if (!buffer_uptodate(bh))
2537                                 ret = -EIO;
2538                 }
2539                 if (ret)
2540                         goto failed;
2541         }
2542
2543         if (is_mapped_to_disk)
2544                 SetPageMappedToDisk(page);
2545
2546         *fsdata = head; /* to be released by nobh_write_end */
2547
2548         return 0;
2549
2550 failed:
2551         BUG_ON(!ret);
2552         /*
2553          * Error recovery is a bit difficult. We need to zero out blocks that
2554          * were newly allocated, and dirty them to ensure they get written out.
2555          * Buffers need to be attached to the page at this point, otherwise
2556          * the handling of potential IO errors during writeout would be hard
2557          * (could try doing synchronous writeout, but what if that fails too?)
2558          */
2559         attach_nobh_buffers(page, head);
2560         page_zero_new_buffers(page, from, to);
2561
2562 out_release:
2563         unlock_page(page);
2564         page_cache_release(page);
2565         *pagep = NULL;
2566
2567         if (pos + len > inode->i_size)
2568                 vmtruncate(inode, inode->i_size);
2569
2570         return ret;
2571 }
2572 EXPORT_SYMBOL(nobh_write_begin);
2573
2574 int nobh_write_end(struct file *file, struct address_space *mapping,
2575                         loff_t pos, unsigned len, unsigned copied,
2576                         struct page *page, void *fsdata)
2577 {
2578         struct inode *inode = page->mapping->host;
2579         struct buffer_head *head = fsdata;
2580         struct buffer_head *bh;
2581         BUG_ON(fsdata != NULL && page_has_buffers(page));
2582
2583         if (unlikely(copied < len) && !page_has_buffers(page))
2584                 attach_nobh_buffers(page, head);
2585         if (page_has_buffers(page))
2586                 return generic_write_end(file, mapping, pos, len,
2587                                         copied, page, fsdata);
2588
2589         SetPageUptodate(page);
2590         set_page_dirty(page);
2591         if (pos+copied > inode->i_size) {
2592                 i_size_write(inode, pos+copied);
2593                 mark_inode_dirty(inode);
2594         }
2595
2596         unlock_page(page);
2597         page_cache_release(page);
2598
2599         while (head) {
2600                 bh = head;
2601                 head = head->b_this_page;
2602                 free_buffer_head(bh);
2603         }
2604
2605         return copied;
2606 }
2607 EXPORT_SYMBOL(nobh_write_end);
2608
2609 /*
2610  * nobh_writepage() - based on block_full_write_page() except
2611  * that it tries to operate without attaching bufferheads to
2612  * the page.
2613  */
2614 int nobh_writepage(struct page *page, get_block_t *get_block,
2615                         struct writeback_control *wbc)
2616 {
2617         struct inode * const inode = page->mapping->host;
2618         loff_t i_size = i_size_read(inode);
2619         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2620         unsigned offset;
2621         int ret;
2622
2623         /* Is the page fully inside i_size? */
2624         if (page->index < end_index)
2625                 goto out;
2626
2627         /* Is the page fully outside i_size? (truncate in progress) */
2628         offset = i_size & (PAGE_CACHE_SIZE-1);
2629         if (page->index >= end_index+1 || !offset) {
2630                 /*
2631                  * The page may have dirty, unmapped buffers.  For example,
2632                  * they may have been added in ext3_writepage().  Make them
2633                  * freeable here, so the page does not leak.
2634                  */
2635 #if 0
2636                 /* Not really sure about this  - do we need this ? */
2637                 if (page->mapping->a_ops->invalidatepage)
2638                         page->mapping->a_ops->invalidatepage(page, offset);
2639 #endif
2640                 unlock_page(page);
2641                 return 0; /* don't care */
2642         }
2643
2644         /*
2645          * The page straddles i_size.  It must be zeroed out on each and every
2646          * writepage invocation because it may be mmapped.  "A file is mapped
2647          * in multiples of the page size.  For a file that is not a multiple of
2648          * the  page size, the remaining memory is zeroed when mapped, and
2649          * writes to that region are not written out to the file."
2650          */
2651         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2652 out:
2653         ret = mpage_writepage(page, get_block, wbc);
2654         if (ret == -EAGAIN)
2655                 ret = __block_write_full_page(inode, page, get_block, wbc);
2656         return ret;
2657 }
2658 EXPORT_SYMBOL(nobh_writepage);
2659
2660 int nobh_truncate_page(struct address_space *mapping,
2661                         loff_t from, get_block_t *get_block)
2662 {
2663         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2664         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2665         unsigned blocksize;
2666         sector_t iblock;
2667         unsigned length, pos;
2668         struct inode *inode = mapping->host;
2669         struct page *page;
2670         struct buffer_head map_bh;
2671         int err;
2672
2673         blocksize = 1 << inode->i_blkbits;
2674         length = offset & (blocksize - 1);
2675
2676         /* Block boundary? Nothing to do */
2677         if (!length)
2678                 return 0;
2679
2680         length = blocksize - length;
2681         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2682
2683         page = grab_cache_page(mapping, index);
2684         err = -ENOMEM;
2685         if (!page)
2686                 goto out;
2687
2688         if (page_has_buffers(page)) {
2689 has_buffers:
2690                 unlock_page(page);
2691                 page_cache_release(page);
2692                 return block_truncate_page(mapping, from, get_block);
2693         }
2694
2695         /* Find the buffer that contains "offset" */
2696         pos = blocksize;
2697         while (offset >= pos) {
2698                 iblock++;
2699                 pos += blocksize;
2700         }
2701
2702         err = get_block(inode, iblock, &map_bh, 0);
2703         if (err)
2704                 goto unlock;
2705         /* unmapped? It's a hole - nothing to do */
2706         if (!buffer_mapped(&map_bh))
2707                 goto unlock;
2708
2709         /* Ok, it's mapped. Make sure it's up-to-date */
2710         if (!PageUptodate(page)) {
2711                 err = mapping->a_ops->readpage(NULL, page);
2712                 if (err) {
2713                         page_cache_release(page);
2714                         goto out;
2715                 }
2716                 lock_page(page);
2717                 if (!PageUptodate(page)) {
2718                         err = -EIO;
2719                         goto unlock;
2720                 }
2721                 if (page_has_buffers(page))
2722                         goto has_buffers;
2723         }
2724         zero_user(page, offset, length);
2725         set_page_dirty(page);
2726         err = 0;
2727
2728 unlock:
2729         unlock_page(page);
2730         page_cache_release(page);
2731 out:
2732         return err;
2733 }
2734 EXPORT_SYMBOL(nobh_truncate_page);
2735
2736 int block_truncate_page(struct address_space *mapping,
2737                         loff_t from, get_block_t *get_block)
2738 {
2739         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2740         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2741         unsigned blocksize;
2742         sector_t iblock;
2743         unsigned length, pos;
2744         struct inode *inode = mapping->host;
2745         struct page *page;
2746         struct buffer_head *bh;
2747         int err;
2748
2749         blocksize = 1 << inode->i_blkbits;
2750         length = offset & (blocksize - 1);
2751
2752         /* Block boundary? Nothing to do */
2753         if (!length)
2754                 return 0;
2755
2756         length = blocksize - length;
2757         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2758         
2759         page = grab_cache_page(mapping, index);
2760         err = -ENOMEM;
2761         if (!page)
2762                 goto out;
2763
2764         if (!page_has_buffers(page))
2765                 create_empty_buffers(page, blocksize, 0);
2766
2767         /* Find the buffer that contains "offset" */
2768         bh = page_buffers(page);
2769         pos = blocksize;
2770         while (offset >= pos) {
2771                 bh = bh->b_this_page;
2772                 iblock++;
2773                 pos += blocksize;
2774         }
2775
2776         err = 0;
2777         if (!buffer_mapped(bh)) {
2778                 WARN_ON(bh->b_size != blocksize);
2779                 err = get_block(inode, iblock, bh, 0);
2780                 if (err)
2781                         goto unlock;
2782                 /* unmapped? It's a hole - nothing to do */
2783                 if (!buffer_mapped(bh))
2784                         goto unlock;
2785         }
2786
2787         /* Ok, it's mapped. Make sure it's up-to-date */
2788         if (PageUptodate(page))
2789                 set_buffer_uptodate(bh);
2790
2791         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2792                 err = -EIO;
2793                 ll_rw_block(READ, 1, &bh);
2794                 wait_on_buffer(bh);
2795                 /* Uhhuh. Read error. Complain and punt. */
2796                 if (!buffer_uptodate(bh))
2797                         goto unlock;
2798         }
2799
2800         zero_user(page, offset, length);
2801         mark_buffer_dirty(bh);
2802         err = 0;
2803
2804 unlock:
2805         unlock_page(page);
2806         page_cache_release(page);
2807 out:
2808         return err;
2809 }
2810
2811 /*
2812  * The generic ->writepage function for buffer-backed address_spaces
2813  */
2814 int block_write_full_page(struct page *page, get_block_t *get_block,
2815                         struct writeback_control *wbc)
2816 {
2817         struct inode * const inode = page->mapping->host;
2818         loff_t i_size = i_size_read(inode);
2819         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2820         unsigned offset;
2821
2822         /* Is the page fully inside i_size? */
2823         if (page->index < end_index)
2824                 return __block_write_full_page(inode, page, get_block, wbc);
2825
2826         /* Is the page fully outside i_size? (truncate in progress) */
2827         offset = i_size & (PAGE_CACHE_SIZE-1);
2828         if (page->index >= end_index+1 || !offset) {
2829                 /*
2830                  * The page may have dirty, unmapped buffers.  For example,
2831                  * they may have been added in ext3_writepage().  Make them
2832                  * freeable here, so the page does not leak.
2833                  */
2834                 do_invalidatepage(page, 0);
2835                 unlock_page(page);
2836                 return 0; /* don't care */
2837         }
2838
2839         /*
2840          * The page straddles i_size.  It must be zeroed out on each and every
2841          * writepage invokation because it may be mmapped.  "A file is mapped
2842          * in multiples of the page size.  For a file that is not a multiple of
2843          * the  page size, the remaining memory is zeroed when mapped, and
2844          * writes to that region are not written out to the file."
2845          */
2846         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2847         return __block_write_full_page(inode, page, get_block, wbc);
2848 }
2849
2850 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2851                             get_block_t *get_block)
2852 {
2853         struct buffer_head tmp;
2854         struct inode *inode = mapping->host;
2855         tmp.b_state = 0;
2856         tmp.b_blocknr = 0;
2857         tmp.b_size = 1 << inode->i_blkbits;
2858         get_block(inode, block, &tmp, 0);
2859         return tmp.b_blocknr;
2860 }
2861
2862 static void end_bio_bh_io_sync(struct bio *bio, int err)
2863 {
2864         struct buffer_head *bh = bio->bi_private;
2865
2866         if (err == -EOPNOTSUPP) {
2867                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2868                 set_bit(BH_Eopnotsupp, &bh->b_state);
2869         }
2870
2871         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2872         bio_put(bio);
2873 }
2874
2875 int submit_bh(int rw, struct buffer_head * bh)
2876 {
2877         struct bio *bio;
2878         int ret = 0;
2879
2880         BUG_ON(!buffer_locked(bh));
2881         BUG_ON(!buffer_mapped(bh));
2882         BUG_ON(!bh->b_end_io);
2883
2884         if (buffer_ordered(bh) && (rw == WRITE))
2885                 rw = WRITE_BARRIER;
2886
2887         /*
2888          * Only clear out a write error when rewriting, should this
2889          * include WRITE_SYNC as well?
2890          */
2891         if (test_set_buffer_req(bh) && (rw == WRITE || rw == WRITE_BARRIER))
2892                 clear_buffer_write_io_error(bh);
2893
2894         /*
2895          * from here on down, it's all bio -- do the initial mapping,
2896          * submit_bio -> generic_make_request may further map this bio around
2897          */
2898         bio = bio_alloc(GFP_NOIO, 1);
2899
2900         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2901         bio->bi_bdev = bh->b_bdev;
2902         bio->bi_io_vec[0].bv_page = bh->b_page;
2903         bio->bi_io_vec[0].bv_len = bh->b_size;
2904         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2905
2906         bio->bi_vcnt = 1;
2907         bio->bi_idx = 0;
2908         bio->bi_size = bh->b_size;
2909
2910         bio->bi_end_io = end_bio_bh_io_sync;
2911         bio->bi_private = bh;
2912
2913         bio_get(bio);
2914         submit_bio(rw, bio);
2915
2916         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2917                 ret = -EOPNOTSUPP;
2918
2919         bio_put(bio);
2920         return ret;
2921 }
2922
2923 /**
2924  * ll_rw_block: low-level access to block devices (DEPRECATED)
2925  * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
2926  * @nr: number of &struct buffer_heads in the array
2927  * @bhs: array of pointers to &struct buffer_head
2928  *
2929  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2930  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2931  * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2932  * are sent to disk. The fourth %READA option is described in the documentation
2933  * for generic_make_request() which ll_rw_block() calls.
2934  *
2935  * This function drops any buffer that it cannot get a lock on (with the
2936  * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
2937  * clean when doing a write request, and any buffer that appears to be
2938  * up-to-date when doing read request.  Further it marks as clean buffers that
2939  * are processed for writing (the buffer cache won't assume that they are
2940  * actually clean until the buffer gets unlocked).
2941  *
2942  * ll_rw_block sets b_end_io to simple completion handler that marks
2943  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2944  * any waiters. 
2945  *
2946  * All of the buffers must be for the same device, and must also be a
2947  * multiple of the current approved size for the device.
2948  */
2949 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2950 {
2951         int i;
2952
2953         for (i = 0; i < nr; i++) {
2954                 struct buffer_head *bh = bhs[i];
2955
2956                 if (rw == SWRITE || rw == SWRITE_SYNC)
2957                         lock_buffer(bh);
2958                 else if (test_set_buffer_locked(bh))
2959                         continue;
2960
2961                 if (rw == WRITE || rw == SWRITE || rw == SWRITE_SYNC) {
2962                         if (test_clear_buffer_dirty(bh)) {
2963                                 bh->b_end_io = end_buffer_write_sync;
2964                                 get_bh(bh);
2965                                 if (rw == SWRITE_SYNC)
2966                                         submit_bh(WRITE_SYNC, bh);
2967                                 else
2968                                         submit_bh(WRITE, bh);
2969                                 continue;
2970                         }
2971                 } else {
2972                         if (!buffer_uptodate(bh)) {
2973                                 bh->b_end_io = end_buffer_read_sync;
2974                                 get_bh(bh);
2975                                 submit_bh(rw, bh);
2976                                 continue;
2977                         }
2978                 }
2979                 unlock_buffer(bh);
2980         }
2981 }
2982
2983 /*
2984  * For a data-integrity writeout, we need to wait upon any in-progress I/O
2985  * and then start new I/O and then wait upon it.  The caller must have a ref on
2986  * the buffer_head.
2987  */
2988 int sync_dirty_buffer(struct buffer_head *bh)
2989 {
2990         int ret = 0;
2991
2992         WARN_ON(atomic_read(&bh->b_count) < 1);
2993         lock_buffer(bh);
2994         if (test_clear_buffer_dirty(bh)) {
2995                 get_bh(bh);
2996                 bh->b_end_io = end_buffer_write_sync;
2997                 ret = submit_bh(WRITE_SYNC, bh);
2998                 wait_on_buffer(bh);
2999                 if (buffer_eopnotsupp(bh)) {
3000                         clear_buffer_eopnotsupp(bh);
3001                         ret = -EOPNOTSUPP;
3002                 }
3003                 if (!ret && !buffer_uptodate(bh))
3004                         ret = -EIO;
3005         } else {
3006                 unlock_buffer(bh);
3007         }
3008         return ret;
3009 }
3010
3011 /*
3012  * try_to_free_buffers() checks if all the buffers on this particular page
3013  * are unused, and releases them if so.
3014  *
3015  * Exclusion against try_to_free_buffers may be obtained by either
3016  * locking the page or by holding its mapping's private_lock.
3017  *
3018  * If the page is dirty but all the buffers are clean then we need to
3019  * be sure to mark the page clean as well.  This is because the page
3020  * may be against a block device, and a later reattachment of buffers
3021  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3022  * filesystem data on the same device.
3023  *
3024  * The same applies to regular filesystem pages: if all the buffers are
3025  * clean then we set the page clean and proceed.  To do that, we require
3026  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3027  * private_lock.
3028  *
3029  * try_to_free_buffers() is non-blocking.
3030  */
3031 static inline int buffer_busy(struct buffer_head *bh)
3032 {
3033         return atomic_read(&bh->b_count) |
3034                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3035 }
3036
3037 static int
3038 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3039 {
3040         struct buffer_head *head = page_buffers(page);
3041         struct buffer_head *bh;
3042
3043         bh = head;
3044         do {
3045                 if (buffer_write_io_error(bh) && page->mapping)
3046                         set_bit(AS_EIO, &page->mapping->flags);
3047                 if (buffer_busy(bh))
3048                         goto failed;
3049                 bh = bh->b_this_page;
3050         } while (bh != head);
3051
3052         do {
3053                 struct buffer_head *next = bh->b_this_page;
3054
3055                 if (bh->b_assoc_map)
3056                         __remove_assoc_queue(bh);
3057                 bh = next;
3058         } while (bh != head);
3059         *buffers_to_free = head;
3060         __clear_page_buffers(page);
3061         return 1;
3062 failed:
3063         return 0;
3064 }
3065
3066 int try_to_free_buffers(struct page *page)
3067 {
3068         struct address_space * const mapping = page->mapping;
3069         struct buffer_head *buffers_to_free = NULL;
3070         int ret = 0;
3071
3072         BUG_ON(!PageLocked(page));
3073         if (PageWriteback(page))
3074                 return 0;
3075
3076         if (mapping == NULL) {          /* can this still happen? */
3077                 ret = drop_buffers(page, &buffers_to_free);
3078                 goto out;
3079         }
3080
3081         spin_lock(&mapping->private_lock);
3082         ret = drop_buffers(page, &buffers_to_free);
3083
3084         /*
3085          * If the filesystem writes its buffers by hand (eg ext3)
3086          * then we can have clean buffers against a dirty page.  We
3087          * clean the page here; otherwise the VM will never notice
3088          * that the filesystem did any IO at all.
3089          *
3090          * Also, during truncate, discard_buffer will have marked all
3091          * the page's buffers clean.  We discover that here and clean
3092          * the page also.
3093          *
3094          * private_lock must be held over this entire operation in order
3095          * to synchronise against __set_page_dirty_buffers and prevent the
3096          * dirty bit from being lost.
3097          */
3098         if (ret)
3099                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3100         spin_unlock(&mapping->private_lock);
3101 out:
3102         if (buffers_to_free) {
3103                 struct buffer_head *bh = buffers_to_free;
3104
3105                 do {
3106                         struct buffer_head *next = bh->b_this_page;
3107                         free_buffer_head(bh);
3108                         bh = next;
3109                 } while (bh != buffers_to_free);
3110         }
3111         return ret;
3112 }
3113 EXPORT_SYMBOL(try_to_free_buffers);
3114
3115 void block_sync_page(struct page *page)
3116 {
3117         struct address_space *mapping;
3118
3119         smp_mb();
3120         mapping = page_mapping(page);
3121         if (mapping)
3122                 blk_run_backing_dev(mapping->backing_dev_info, page);
3123 }
3124
3125 /*
3126  * There are no bdflush tunables left.  But distributions are
3127  * still running obsolete flush daemons, so we terminate them here.
3128  *
3129  * Use of bdflush() is deprecated and will be removed in a future kernel.
3130  * The `pdflush' kernel threads fully replace bdflush daemons and this call.
3131  */
3132 asmlinkage long sys_bdflush(int func, long data)
3133 {
3134         static int msg_count;
3135
3136         if (!capable(CAP_SYS_ADMIN))
3137                 return -EPERM;
3138
3139         if (msg_count < 5) {
3140                 msg_count++;
3141                 printk(KERN_INFO
3142                         "warning: process `%s' used the obsolete bdflush"
3143                         " system call\n", current->comm);
3144                 printk(KERN_INFO "Fix your initscripts?\n");
3145         }
3146
3147         if (func == 1)
3148                 do_exit(0);
3149         return 0;
3150 }
3151
3152 /*
3153  * Buffer-head allocation
3154  */
3155 static struct kmem_cache *bh_cachep;
3156
3157 /*
3158  * Once the number of bh's in the machine exceeds this level, we start
3159  * stripping them in writeback.
3160  */
3161 static int max_buffer_heads;
3162
3163 int buffer_heads_over_limit;
3164
3165 struct bh_accounting {
3166         int nr;                 /* Number of live bh's */
3167         int ratelimit;          /* Limit cacheline bouncing */
3168 };
3169
3170 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3171
3172 static void recalc_bh_state(void)
3173 {
3174         int i;
3175         int tot = 0;
3176
3177         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3178                 return;
3179         __get_cpu_var(bh_accounting).ratelimit = 0;
3180         for_each_online_cpu(i)
3181                 tot += per_cpu(bh_accounting, i).nr;
3182         buffer_heads_over_limit = (tot > max_buffer_heads);
3183 }
3184         
3185 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3186 {
3187         struct buffer_head *ret = kmem_cache_alloc(bh_cachep, gfp_flags);
3188         if (ret) {
3189                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3190                 get_cpu_var(bh_accounting).nr++;
3191                 recalc_bh_state();
3192                 put_cpu_var(bh_accounting);
3193         }
3194         return ret;
3195 }
3196 EXPORT_SYMBOL(alloc_buffer_head);
3197
3198 void free_buffer_head(struct buffer_head *bh)
3199 {
3200         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3201         kmem_cache_free(bh_cachep, bh);
3202         get_cpu_var(bh_accounting).nr--;
3203         recalc_bh_state();
3204         put_cpu_var(bh_accounting);
3205 }
3206 EXPORT_SYMBOL(free_buffer_head);
3207
3208 static void buffer_exit_cpu(int cpu)
3209 {
3210         int i;
3211         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3212
3213         for (i = 0; i < BH_LRU_SIZE; i++) {
3214                 brelse(b->bhs[i]);
3215                 b->bhs[i] = NULL;
3216         }
3217         get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3218         per_cpu(bh_accounting, cpu).nr = 0;
3219         put_cpu_var(bh_accounting);
3220 }
3221
3222 static int buffer_cpu_notify(struct notifier_block *self,
3223                               unsigned long action, void *hcpu)
3224 {
3225         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3226                 buffer_exit_cpu((unsigned long)hcpu);
3227         return NOTIFY_OK;
3228 }
3229
3230 /**
3231  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3232  * @bh: struct buffer_head
3233  *
3234  * Return true if the buffer is up-to-date and false,
3235  * with the buffer locked, if not.
3236  */
3237 int bh_uptodate_or_lock(struct buffer_head *bh)
3238 {
3239         if (!buffer_uptodate(bh)) {
3240                 lock_buffer(bh);
3241                 if (!buffer_uptodate(bh))
3242                         return 0;
3243                 unlock_buffer(bh);
3244         }
3245         return 1;
3246 }
3247 EXPORT_SYMBOL(bh_uptodate_or_lock);
3248
3249 /**
3250  * bh_submit_read - Submit a locked buffer for reading
3251  * @bh: struct buffer_head
3252  *
3253  * Returns zero on success and -EIO on error.
3254  */
3255 int bh_submit_read(struct buffer_head *bh)
3256 {
3257         BUG_ON(!buffer_locked(bh));
3258
3259         if (buffer_uptodate(bh)) {
3260                 unlock_buffer(bh);
3261                 return 0;
3262         }
3263
3264         get_bh(bh);
3265         bh->b_end_io = end_buffer_read_sync;
3266         submit_bh(READ, bh);
3267         wait_on_buffer(bh);
3268         if (buffer_uptodate(bh))
3269                 return 0;
3270         return -EIO;
3271 }
3272 EXPORT_SYMBOL(bh_submit_read);
3273
3274 static void
3275 init_buffer_head(struct kmem_cache *cachep, void *data)
3276 {
3277         struct buffer_head *bh = data;
3278
3279         memset(bh, 0, sizeof(*bh));
3280         INIT_LIST_HEAD(&bh->b_assoc_buffers);
3281 }
3282
3283 void __init buffer_init(void)
3284 {
3285         int nrpages;
3286
3287         bh_cachep = kmem_cache_create("buffer_head",
3288                         sizeof(struct buffer_head), 0,
3289                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3290                                 SLAB_MEM_SPREAD),
3291                                 init_buffer_head);
3292
3293         /*
3294          * Limit the bh occupancy to 10% of ZONE_NORMAL
3295          */
3296         nrpages = (nr_free_buffer_pages() * 10) / 100;
3297         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3298         hotcpu_notifier(buffer_cpu_notify, 0);
3299 }
3300
3301 EXPORT_SYMBOL(__bforget);
3302 EXPORT_SYMBOL(__brelse);
3303 EXPORT_SYMBOL(__wait_on_buffer);
3304 EXPORT_SYMBOL(block_commit_write);
3305 EXPORT_SYMBOL(block_prepare_write);
3306 EXPORT_SYMBOL(block_page_mkwrite);
3307 EXPORT_SYMBOL(block_read_full_page);
3308 EXPORT_SYMBOL(block_sync_page);
3309 EXPORT_SYMBOL(block_truncate_page);
3310 EXPORT_SYMBOL(block_write_full_page);
3311 EXPORT_SYMBOL(cont_write_begin);
3312 EXPORT_SYMBOL(end_buffer_read_sync);
3313 EXPORT_SYMBOL(end_buffer_write_sync);
3314 EXPORT_SYMBOL(file_fsync);
3315 EXPORT_SYMBOL(fsync_bdev);
3316 EXPORT_SYMBOL(generic_block_bmap);
3317 EXPORT_SYMBOL(generic_cont_expand_simple);
3318 EXPORT_SYMBOL(init_buffer);
3319 EXPORT_SYMBOL(invalidate_bdev);
3320 EXPORT_SYMBOL(ll_rw_block);
3321 EXPORT_SYMBOL(mark_buffer_dirty);
3322 EXPORT_SYMBOL(submit_bh);
3323 EXPORT_SYMBOL(sync_dirty_buffer);
3324 EXPORT_SYMBOL(unlock_buffer);