Pull percpu-dtc into release branch
[linux-2.6] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/smp_lock.h>
28 #include <linux/capability.h>
29 #include <linux/blkdev.h>
30 #include <linux/file.h>
31 #include <linux/quotaops.h>
32 #include <linux/highmem.h>
33 #include <linux/module.h>
34 #include <linux/writeback.h>
35 #include <linux/hash.h>
36 #include <linux/suspend.h>
37 #include <linux/buffer_head.h>
38 #include <linux/task_io_accounting_ops.h>
39 #include <linux/bio.h>
40 #include <linux/notifier.h>
41 #include <linux/cpu.h>
42 #include <linux/bitops.h>
43 #include <linux/mpage.h>
44 #include <linux/bit_spinlock.h>
45
46 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
47 static void invalidate_bh_lrus(void);
48
49 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
50
51 inline void
52 init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
53 {
54         bh->b_end_io = handler;
55         bh->b_private = private;
56 }
57
58 static int sync_buffer(void *word)
59 {
60         struct block_device *bd;
61         struct buffer_head *bh
62                 = container_of(word, struct buffer_head, b_state);
63
64         smp_mb();
65         bd = bh->b_bdev;
66         if (bd)
67                 blk_run_address_space(bd->bd_inode->i_mapping);
68         io_schedule();
69         return 0;
70 }
71
72 void fastcall __lock_buffer(struct buffer_head *bh)
73 {
74         wait_on_bit_lock(&bh->b_state, BH_Lock, sync_buffer,
75                                                         TASK_UNINTERRUPTIBLE);
76 }
77 EXPORT_SYMBOL(__lock_buffer);
78
79 void fastcall unlock_buffer(struct buffer_head *bh)
80 {
81         smp_mb__before_clear_bit();
82         clear_buffer_locked(bh);
83         smp_mb__after_clear_bit();
84         wake_up_bit(&bh->b_state, BH_Lock);
85 }
86
87 /*
88  * Block until a buffer comes unlocked.  This doesn't stop it
89  * from becoming locked again - you have to lock it yourself
90  * if you want to preserve its state.
91  */
92 void __wait_on_buffer(struct buffer_head * bh)
93 {
94         wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
95 }
96
97 static void
98 __clear_page_buffers(struct page *page)
99 {
100         ClearPagePrivate(page);
101         set_page_private(page, 0);
102         page_cache_release(page);
103 }
104
105 static void buffer_io_error(struct buffer_head *bh)
106 {
107         char b[BDEVNAME_SIZE];
108
109         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
110                         bdevname(bh->b_bdev, b),
111                         (unsigned long long)bh->b_blocknr);
112 }
113
114 /*
115  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
116  * unlock the buffer. This is what ll_rw_block uses too.
117  */
118 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
119 {
120         if (uptodate) {
121                 set_buffer_uptodate(bh);
122         } else {
123                 /* This happens, due to failed READA attempts. */
124                 clear_buffer_uptodate(bh);
125         }
126         unlock_buffer(bh);
127         put_bh(bh);
128 }
129
130 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
131 {
132         char b[BDEVNAME_SIZE];
133
134         if (uptodate) {
135                 set_buffer_uptodate(bh);
136         } else {
137                 if (!buffer_eopnotsupp(bh) && printk_ratelimit()) {
138                         buffer_io_error(bh);
139                         printk(KERN_WARNING "lost page write due to "
140                                         "I/O error on %s\n",
141                                        bdevname(bh->b_bdev, b));
142                 }
143                 set_buffer_write_io_error(bh);
144                 clear_buffer_uptodate(bh);
145         }
146         unlock_buffer(bh);
147         put_bh(bh);
148 }
149
150 /*
151  * Write out and wait upon all the dirty data associated with a block
152  * device via its mapping.  Does not take the superblock lock.
153  */
154 int sync_blockdev(struct block_device *bdev)
155 {
156         int ret = 0;
157
158         if (bdev)
159                 ret = filemap_write_and_wait(bdev->bd_inode->i_mapping);
160         return ret;
161 }
162 EXPORT_SYMBOL(sync_blockdev);
163
164 /*
165  * Write out and wait upon all dirty data associated with this
166  * device.   Filesystem data as well as the underlying block
167  * device.  Takes the superblock lock.
168  */
169 int fsync_bdev(struct block_device *bdev)
170 {
171         struct super_block *sb = get_super(bdev);
172         if (sb) {
173                 int res = fsync_super(sb);
174                 drop_super(sb);
175                 return res;
176         }
177         return sync_blockdev(bdev);
178 }
179
180 /**
181  * freeze_bdev  --  lock a filesystem and force it into a consistent state
182  * @bdev:       blockdevice to lock
183  *
184  * This takes the block device bd_mount_sem to make sure no new mounts
185  * happen on bdev until thaw_bdev() is called.
186  * If a superblock is found on this device, we take the s_umount semaphore
187  * on it to make sure nobody unmounts until the snapshot creation is done.
188  */
189 struct super_block *freeze_bdev(struct block_device *bdev)
190 {
191         struct super_block *sb;
192
193         down(&bdev->bd_mount_sem);
194         sb = get_super(bdev);
195         if (sb && !(sb->s_flags & MS_RDONLY)) {
196                 sb->s_frozen = SB_FREEZE_WRITE;
197                 smp_wmb();
198
199                 __fsync_super(sb);
200
201                 sb->s_frozen = SB_FREEZE_TRANS;
202                 smp_wmb();
203
204                 sync_blockdev(sb->s_bdev);
205
206                 if (sb->s_op->write_super_lockfs)
207                         sb->s_op->write_super_lockfs(sb);
208         }
209
210         sync_blockdev(bdev);
211         return sb;      /* thaw_bdev releases s->s_umount and bd_mount_sem */
212 }
213 EXPORT_SYMBOL(freeze_bdev);
214
215 /**
216  * thaw_bdev  -- unlock filesystem
217  * @bdev:       blockdevice to unlock
218  * @sb:         associated superblock
219  *
220  * Unlocks the filesystem and marks it writeable again after freeze_bdev().
221  */
222 void thaw_bdev(struct block_device *bdev, struct super_block *sb)
223 {
224         if (sb) {
225                 BUG_ON(sb->s_bdev != bdev);
226
227                 if (sb->s_op->unlockfs)
228                         sb->s_op->unlockfs(sb);
229                 sb->s_frozen = SB_UNFROZEN;
230                 smp_wmb();
231                 wake_up(&sb->s_wait_unfrozen);
232                 drop_super(sb);
233         }
234
235         up(&bdev->bd_mount_sem);
236 }
237 EXPORT_SYMBOL(thaw_bdev);
238
239 /*
240  * Various filesystems appear to want __find_get_block to be non-blocking.
241  * But it's the page lock which protects the buffers.  To get around this,
242  * we get exclusion from try_to_free_buffers with the blockdev mapping's
243  * private_lock.
244  *
245  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
246  * may be quite high.  This code could TryLock the page, and if that
247  * succeeds, there is no need to take private_lock. (But if
248  * private_lock is contended then so is mapping->tree_lock).
249  */
250 static struct buffer_head *
251 __find_get_block_slow(struct block_device *bdev, sector_t block)
252 {
253         struct inode *bd_inode = bdev->bd_inode;
254         struct address_space *bd_mapping = bd_inode->i_mapping;
255         struct buffer_head *ret = NULL;
256         pgoff_t index;
257         struct buffer_head *bh;
258         struct buffer_head *head;
259         struct page *page;
260         int all_mapped = 1;
261
262         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
263         page = find_get_page(bd_mapping, index);
264         if (!page)
265                 goto out;
266
267         spin_lock(&bd_mapping->private_lock);
268         if (!page_has_buffers(page))
269                 goto out_unlock;
270         head = page_buffers(page);
271         bh = head;
272         do {
273                 if (bh->b_blocknr == block) {
274                         ret = bh;
275                         get_bh(bh);
276                         goto out_unlock;
277                 }
278                 if (!buffer_mapped(bh))
279                         all_mapped = 0;
280                 bh = bh->b_this_page;
281         } while (bh != head);
282
283         /* we might be here because some of the buffers on this page are
284          * not mapped.  This is due to various races between
285          * file io on the block device and getblk.  It gets dealt with
286          * elsewhere, don't buffer_error if we had some unmapped buffers
287          */
288         if (all_mapped) {
289                 printk("__find_get_block_slow() failed. "
290                         "block=%llu, b_blocknr=%llu\n",
291                         (unsigned long long)block,
292                         (unsigned long long)bh->b_blocknr);
293                 printk("b_state=0x%08lx, b_size=%zu\n",
294                         bh->b_state, bh->b_size);
295                 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
296         }
297 out_unlock:
298         spin_unlock(&bd_mapping->private_lock);
299         page_cache_release(page);
300 out:
301         return ret;
302 }
303
304 /* If invalidate_buffers() will trash dirty buffers, it means some kind
305    of fs corruption is going on. Trashing dirty data always imply losing
306    information that was supposed to be just stored on the physical layer
307    by the user.
308
309    Thus invalidate_buffers in general usage is not allwowed to trash
310    dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
311    be preserved.  These buffers are simply skipped.
312   
313    We also skip buffers which are still in use.  For example this can
314    happen if a userspace program is reading the block device.
315
316    NOTE: In the case where the user removed a removable-media-disk even if
317    there's still dirty data not synced on disk (due a bug in the device driver
318    or due an error of the user), by not destroying the dirty buffers we could
319    generate corruption also on the next media inserted, thus a parameter is
320    necessary to handle this case in the most safe way possible (trying
321    to not corrupt also the new disk inserted with the data belonging to
322    the old now corrupted disk). Also for the ramdisk the natural thing
323    to do in order to release the ramdisk memory is to destroy dirty buffers.
324
325    These are two special cases. Normal usage imply the device driver
326    to issue a sync on the device (without waiting I/O completion) and
327    then an invalidate_buffers call that doesn't trash dirty buffers.
328
329    For handling cache coherency with the blkdev pagecache the 'update' case
330    is been introduced. It is needed to re-read from disk any pinned
331    buffer. NOTE: re-reading from disk is destructive so we can do it only
332    when we assume nobody is changing the buffercache under our I/O and when
333    we think the disk contains more recent information than the buffercache.
334    The update == 1 pass marks the buffers we need to update, the update == 2
335    pass does the actual I/O. */
336 void invalidate_bdev(struct block_device *bdev, int destroy_dirty_buffers)
337 {
338         struct address_space *mapping = bdev->bd_inode->i_mapping;
339
340         if (mapping->nrpages == 0)
341                 return;
342
343         invalidate_bh_lrus();
344         /*
345          * FIXME: what about destroy_dirty_buffers?
346          * We really want to use invalidate_inode_pages2() for
347          * that, but not until that's cleaned up.
348          */
349         invalidate_mapping_pages(mapping, 0, -1);
350 }
351
352 /*
353  * Kick pdflush then try to free up some ZONE_NORMAL memory.
354  */
355 static void free_more_memory(void)
356 {
357         struct zone **zones;
358         pg_data_t *pgdat;
359
360         wakeup_pdflush(1024);
361         yield();
362
363         for_each_online_pgdat(pgdat) {
364                 zones = pgdat->node_zonelists[gfp_zone(GFP_NOFS)].zones;
365                 if (*zones)
366                         try_to_free_pages(zones, GFP_NOFS);
367         }
368 }
369
370 /*
371  * I/O completion handler for block_read_full_page() - pages
372  * which come unlocked at the end of I/O.
373  */
374 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
375 {
376         unsigned long flags;
377         struct buffer_head *first;
378         struct buffer_head *tmp;
379         struct page *page;
380         int page_uptodate = 1;
381
382         BUG_ON(!buffer_async_read(bh));
383
384         page = bh->b_page;
385         if (uptodate) {
386                 set_buffer_uptodate(bh);
387         } else {
388                 clear_buffer_uptodate(bh);
389                 if (printk_ratelimit())
390                         buffer_io_error(bh);
391                 SetPageError(page);
392         }
393
394         /*
395          * Be _very_ careful from here on. Bad things can happen if
396          * two buffer heads end IO at almost the same time and both
397          * decide that the page is now completely done.
398          */
399         first = page_buffers(page);
400         local_irq_save(flags);
401         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
402         clear_buffer_async_read(bh);
403         unlock_buffer(bh);
404         tmp = bh;
405         do {
406                 if (!buffer_uptodate(tmp))
407                         page_uptodate = 0;
408                 if (buffer_async_read(tmp)) {
409                         BUG_ON(!buffer_locked(tmp));
410                         goto still_busy;
411                 }
412                 tmp = tmp->b_this_page;
413         } while (tmp != bh);
414         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
415         local_irq_restore(flags);
416
417         /*
418          * If none of the buffers had errors and they are all
419          * uptodate then we can set the page uptodate.
420          */
421         if (page_uptodate && !PageError(page))
422                 SetPageUptodate(page);
423         unlock_page(page);
424         return;
425
426 still_busy:
427         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
428         local_irq_restore(flags);
429         return;
430 }
431
432 /*
433  * Completion handler for block_write_full_page() - pages which are unlocked
434  * during I/O, and which have PageWriteback cleared upon I/O completion.
435  */
436 static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
437 {
438         char b[BDEVNAME_SIZE];
439         unsigned long flags;
440         struct buffer_head *first;
441         struct buffer_head *tmp;
442         struct page *page;
443
444         BUG_ON(!buffer_async_write(bh));
445
446         page = bh->b_page;
447         if (uptodate) {
448                 set_buffer_uptodate(bh);
449         } else {
450                 if (printk_ratelimit()) {
451                         buffer_io_error(bh);
452                         printk(KERN_WARNING "lost page write due to "
453                                         "I/O error on %s\n",
454                                bdevname(bh->b_bdev, b));
455                 }
456                 set_bit(AS_EIO, &page->mapping->flags);
457                 set_buffer_write_io_error(bh);
458                 clear_buffer_uptodate(bh);
459                 SetPageError(page);
460         }
461
462         first = page_buffers(page);
463         local_irq_save(flags);
464         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
465
466         clear_buffer_async_write(bh);
467         unlock_buffer(bh);
468         tmp = bh->b_this_page;
469         while (tmp != bh) {
470                 if (buffer_async_write(tmp)) {
471                         BUG_ON(!buffer_locked(tmp));
472                         goto still_busy;
473                 }
474                 tmp = tmp->b_this_page;
475         }
476         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
477         local_irq_restore(flags);
478         end_page_writeback(page);
479         return;
480
481 still_busy:
482         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
483         local_irq_restore(flags);
484         return;
485 }
486
487 /*
488  * If a page's buffers are under async readin (end_buffer_async_read
489  * completion) then there is a possibility that another thread of
490  * control could lock one of the buffers after it has completed
491  * but while some of the other buffers have not completed.  This
492  * locked buffer would confuse end_buffer_async_read() into not unlocking
493  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
494  * that this buffer is not under async I/O.
495  *
496  * The page comes unlocked when it has no locked buffer_async buffers
497  * left.
498  *
499  * PageLocked prevents anyone starting new async I/O reads any of
500  * the buffers.
501  *
502  * PageWriteback is used to prevent simultaneous writeout of the same
503  * page.
504  *
505  * PageLocked prevents anyone from starting writeback of a page which is
506  * under read I/O (PageWriteback is only ever set against a locked page).
507  */
508 static void mark_buffer_async_read(struct buffer_head *bh)
509 {
510         bh->b_end_io = end_buffer_async_read;
511         set_buffer_async_read(bh);
512 }
513
514 void mark_buffer_async_write(struct buffer_head *bh)
515 {
516         bh->b_end_io = end_buffer_async_write;
517         set_buffer_async_write(bh);
518 }
519 EXPORT_SYMBOL(mark_buffer_async_write);
520
521
522 /*
523  * fs/buffer.c contains helper functions for buffer-backed address space's
524  * fsync functions.  A common requirement for buffer-based filesystems is
525  * that certain data from the backing blockdev needs to be written out for
526  * a successful fsync().  For example, ext2 indirect blocks need to be
527  * written back and waited upon before fsync() returns.
528  *
529  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
530  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
531  * management of a list of dependent buffers at ->i_mapping->private_list.
532  *
533  * Locking is a little subtle: try_to_free_buffers() will remove buffers
534  * from their controlling inode's queue when they are being freed.  But
535  * try_to_free_buffers() will be operating against the *blockdev* mapping
536  * at the time, not against the S_ISREG file which depends on those buffers.
537  * So the locking for private_list is via the private_lock in the address_space
538  * which backs the buffers.  Which is different from the address_space 
539  * against which the buffers are listed.  So for a particular address_space,
540  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
541  * mapping->private_list will always be protected by the backing blockdev's
542  * ->private_lock.
543  *
544  * Which introduces a requirement: all buffers on an address_space's
545  * ->private_list must be from the same address_space: the blockdev's.
546  *
547  * address_spaces which do not place buffers at ->private_list via these
548  * utility functions are free to use private_lock and private_list for
549  * whatever they want.  The only requirement is that list_empty(private_list)
550  * be true at clear_inode() time.
551  *
552  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
553  * filesystems should do that.  invalidate_inode_buffers() should just go
554  * BUG_ON(!list_empty).
555  *
556  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
557  * take an address_space, not an inode.  And it should be called
558  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
559  * queued up.
560  *
561  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
562  * list if it is already on a list.  Because if the buffer is on a list,
563  * it *must* already be on the right one.  If not, the filesystem is being
564  * silly.  This will save a ton of locking.  But first we have to ensure
565  * that buffers are taken *off* the old inode's list when they are freed
566  * (presumably in truncate).  That requires careful auditing of all
567  * filesystems (do it inside bforget()).  It could also be done by bringing
568  * b_inode back.
569  */
570
571 /*
572  * The buffer's backing address_space's private_lock must be held
573  */
574 static inline void __remove_assoc_queue(struct buffer_head *bh)
575 {
576         list_del_init(&bh->b_assoc_buffers);
577         WARN_ON(!bh->b_assoc_map);
578         if (buffer_write_io_error(bh))
579                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
580         bh->b_assoc_map = NULL;
581 }
582
583 int inode_has_buffers(struct inode *inode)
584 {
585         return !list_empty(&inode->i_data.private_list);
586 }
587
588 /*
589  * osync is designed to support O_SYNC io.  It waits synchronously for
590  * all already-submitted IO to complete, but does not queue any new
591  * writes to the disk.
592  *
593  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
594  * you dirty the buffers, and then use osync_inode_buffers to wait for
595  * completion.  Any other dirty buffers which are not yet queued for
596  * write will not be flushed to disk by the osync.
597  */
598 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
599 {
600         struct buffer_head *bh;
601         struct list_head *p;
602         int err = 0;
603
604         spin_lock(lock);
605 repeat:
606         list_for_each_prev(p, list) {
607                 bh = BH_ENTRY(p);
608                 if (buffer_locked(bh)) {
609                         get_bh(bh);
610                         spin_unlock(lock);
611                         wait_on_buffer(bh);
612                         if (!buffer_uptodate(bh))
613                                 err = -EIO;
614                         brelse(bh);
615                         spin_lock(lock);
616                         goto repeat;
617                 }
618         }
619         spin_unlock(lock);
620         return err;
621 }
622
623 /**
624  * sync_mapping_buffers - write out and wait upon a mapping's "associated"
625  *                        buffers
626  * @mapping: the mapping which wants those buffers written
627  *
628  * Starts I/O against the buffers at mapping->private_list, and waits upon
629  * that I/O.
630  *
631  * Basically, this is a convenience function for fsync().
632  * @mapping is a file or directory which needs those buffers to be written for
633  * a successful fsync().
634  */
635 int sync_mapping_buffers(struct address_space *mapping)
636 {
637         struct address_space *buffer_mapping = mapping->assoc_mapping;
638
639         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
640                 return 0;
641
642         return fsync_buffers_list(&buffer_mapping->private_lock,
643                                         &mapping->private_list);
644 }
645 EXPORT_SYMBOL(sync_mapping_buffers);
646
647 /*
648  * Called when we've recently written block `bblock', and it is known that
649  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
650  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
651  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
652  */
653 void write_boundary_block(struct block_device *bdev,
654                         sector_t bblock, unsigned blocksize)
655 {
656         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
657         if (bh) {
658                 if (buffer_dirty(bh))
659                         ll_rw_block(WRITE, 1, &bh);
660                 put_bh(bh);
661         }
662 }
663
664 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
665 {
666         struct address_space *mapping = inode->i_mapping;
667         struct address_space *buffer_mapping = bh->b_page->mapping;
668
669         mark_buffer_dirty(bh);
670         if (!mapping->assoc_mapping) {
671                 mapping->assoc_mapping = buffer_mapping;
672         } else {
673                 BUG_ON(mapping->assoc_mapping != buffer_mapping);
674         }
675         if (list_empty(&bh->b_assoc_buffers)) {
676                 spin_lock(&buffer_mapping->private_lock);
677                 list_move_tail(&bh->b_assoc_buffers,
678                                 &mapping->private_list);
679                 bh->b_assoc_map = mapping;
680                 spin_unlock(&buffer_mapping->private_lock);
681         }
682 }
683 EXPORT_SYMBOL(mark_buffer_dirty_inode);
684
685 /*
686  * Add a page to the dirty page list.
687  *
688  * It is a sad fact of life that this function is called from several places
689  * deeply under spinlocking.  It may not sleep.
690  *
691  * If the page has buffers, the uptodate buffers are set dirty, to preserve
692  * dirty-state coherency between the page and the buffers.  It the page does
693  * not have buffers then when they are later attached they will all be set
694  * dirty.
695  *
696  * The buffers are dirtied before the page is dirtied.  There's a small race
697  * window in which a writepage caller may see the page cleanness but not the
698  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
699  * before the buffers, a concurrent writepage caller could clear the page dirty
700  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
701  * page on the dirty page list.
702  *
703  * We use private_lock to lock against try_to_free_buffers while using the
704  * page's buffer list.  Also use this to protect against clean buffers being
705  * added to the page after it was set dirty.
706  *
707  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
708  * address_space though.
709  */
710 int __set_page_dirty_buffers(struct page *page)
711 {
712         struct address_space * const mapping = page_mapping(page);
713
714         if (unlikely(!mapping))
715                 return !TestSetPageDirty(page);
716
717         spin_lock(&mapping->private_lock);
718         if (page_has_buffers(page)) {
719                 struct buffer_head *head = page_buffers(page);
720                 struct buffer_head *bh = head;
721
722                 do {
723                         set_buffer_dirty(bh);
724                         bh = bh->b_this_page;
725                 } while (bh != head);
726         }
727         spin_unlock(&mapping->private_lock);
728
729         if (TestSetPageDirty(page))
730                 return 0;
731
732         write_lock_irq(&mapping->tree_lock);
733         if (page->mapping) {    /* Race with truncate? */
734                 if (mapping_cap_account_dirty(mapping)) {
735                         __inc_zone_page_state(page, NR_FILE_DIRTY);
736                         task_io_account_write(PAGE_CACHE_SIZE);
737                 }
738                 radix_tree_tag_set(&mapping->page_tree,
739                                 page_index(page), PAGECACHE_TAG_DIRTY);
740         }
741         write_unlock_irq(&mapping->tree_lock);
742         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
743         return 1;
744 }
745 EXPORT_SYMBOL(__set_page_dirty_buffers);
746
747 /*
748  * Write out and wait upon a list of buffers.
749  *
750  * We have conflicting pressures: we want to make sure that all
751  * initially dirty buffers get waited on, but that any subsequently
752  * dirtied buffers don't.  After all, we don't want fsync to last
753  * forever if somebody is actively writing to the file.
754  *
755  * Do this in two main stages: first we copy dirty buffers to a
756  * temporary inode list, queueing the writes as we go.  Then we clean
757  * up, waiting for those writes to complete.
758  * 
759  * During this second stage, any subsequent updates to the file may end
760  * up refiling the buffer on the original inode's dirty list again, so
761  * there is a chance we will end up with a buffer queued for write but
762  * not yet completed on that list.  So, as a final cleanup we go through
763  * the osync code to catch these locked, dirty buffers without requeuing
764  * any newly dirty buffers for write.
765  */
766 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
767 {
768         struct buffer_head *bh;
769         struct list_head tmp;
770         int err = 0, err2;
771
772         INIT_LIST_HEAD(&tmp);
773
774         spin_lock(lock);
775         while (!list_empty(list)) {
776                 bh = BH_ENTRY(list->next);
777                 __remove_assoc_queue(bh);
778                 if (buffer_dirty(bh) || buffer_locked(bh)) {
779                         list_add(&bh->b_assoc_buffers, &tmp);
780                         if (buffer_dirty(bh)) {
781                                 get_bh(bh);
782                                 spin_unlock(lock);
783                                 /*
784                                  * Ensure any pending I/O completes so that
785                                  * ll_rw_block() actually writes the current
786                                  * contents - it is a noop if I/O is still in
787                                  * flight on potentially older contents.
788                                  */
789                                 ll_rw_block(SWRITE, 1, &bh);
790                                 brelse(bh);
791                                 spin_lock(lock);
792                         }
793                 }
794         }
795
796         while (!list_empty(&tmp)) {
797                 bh = BH_ENTRY(tmp.prev);
798                 list_del_init(&bh->b_assoc_buffers);
799                 get_bh(bh);
800                 spin_unlock(lock);
801                 wait_on_buffer(bh);
802                 if (!buffer_uptodate(bh))
803                         err = -EIO;
804                 brelse(bh);
805                 spin_lock(lock);
806         }
807         
808         spin_unlock(lock);
809         err2 = osync_buffers_list(lock, list);
810         if (err)
811                 return err;
812         else
813                 return err2;
814 }
815
816 /*
817  * Invalidate any and all dirty buffers on a given inode.  We are
818  * probably unmounting the fs, but that doesn't mean we have already
819  * done a sync().  Just drop the buffers from the inode list.
820  *
821  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
822  * assumes that all the buffers are against the blockdev.  Not true
823  * for reiserfs.
824  */
825 void invalidate_inode_buffers(struct inode *inode)
826 {
827         if (inode_has_buffers(inode)) {
828                 struct address_space *mapping = &inode->i_data;
829                 struct list_head *list = &mapping->private_list;
830                 struct address_space *buffer_mapping = mapping->assoc_mapping;
831
832                 spin_lock(&buffer_mapping->private_lock);
833                 while (!list_empty(list))
834                         __remove_assoc_queue(BH_ENTRY(list->next));
835                 spin_unlock(&buffer_mapping->private_lock);
836         }
837 }
838
839 /*
840  * Remove any clean buffers from the inode's buffer list.  This is called
841  * when we're trying to free the inode itself.  Those buffers can pin it.
842  *
843  * Returns true if all buffers were removed.
844  */
845 int remove_inode_buffers(struct inode *inode)
846 {
847         int ret = 1;
848
849         if (inode_has_buffers(inode)) {
850                 struct address_space *mapping = &inode->i_data;
851                 struct list_head *list = &mapping->private_list;
852                 struct address_space *buffer_mapping = mapping->assoc_mapping;
853
854                 spin_lock(&buffer_mapping->private_lock);
855                 while (!list_empty(list)) {
856                         struct buffer_head *bh = BH_ENTRY(list->next);
857                         if (buffer_dirty(bh)) {
858                                 ret = 0;
859                                 break;
860                         }
861                         __remove_assoc_queue(bh);
862                 }
863                 spin_unlock(&buffer_mapping->private_lock);
864         }
865         return ret;
866 }
867
868 /*
869  * Create the appropriate buffers when given a page for data area and
870  * the size of each buffer.. Use the bh->b_this_page linked list to
871  * follow the buffers created.  Return NULL if unable to create more
872  * buffers.
873  *
874  * The retry flag is used to differentiate async IO (paging, swapping)
875  * which may not fail from ordinary buffer allocations.
876  */
877 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
878                 int retry)
879 {
880         struct buffer_head *bh, *head;
881         long offset;
882
883 try_again:
884         head = NULL;
885         offset = PAGE_SIZE;
886         while ((offset -= size) >= 0) {
887                 bh = alloc_buffer_head(GFP_NOFS);
888                 if (!bh)
889                         goto no_grow;
890
891                 bh->b_bdev = NULL;
892                 bh->b_this_page = head;
893                 bh->b_blocknr = -1;
894                 head = bh;
895
896                 bh->b_state = 0;
897                 atomic_set(&bh->b_count, 0);
898                 bh->b_private = NULL;
899                 bh->b_size = size;
900
901                 /* Link the buffer to its page */
902                 set_bh_page(bh, page, offset);
903
904                 init_buffer(bh, NULL, NULL);
905         }
906         return head;
907 /*
908  * In case anything failed, we just free everything we got.
909  */
910 no_grow:
911         if (head) {
912                 do {
913                         bh = head;
914                         head = head->b_this_page;
915                         free_buffer_head(bh);
916                 } while (head);
917         }
918
919         /*
920          * Return failure for non-async IO requests.  Async IO requests
921          * are not allowed to fail, so we have to wait until buffer heads
922          * become available.  But we don't want tasks sleeping with 
923          * partially complete buffers, so all were released above.
924          */
925         if (!retry)
926                 return NULL;
927
928         /* We're _really_ low on memory. Now we just
929          * wait for old buffer heads to become free due to
930          * finishing IO.  Since this is an async request and
931          * the reserve list is empty, we're sure there are 
932          * async buffer heads in use.
933          */
934         free_more_memory();
935         goto try_again;
936 }
937 EXPORT_SYMBOL_GPL(alloc_page_buffers);
938
939 static inline void
940 link_dev_buffers(struct page *page, struct buffer_head *head)
941 {
942         struct buffer_head *bh, *tail;
943
944         bh = head;
945         do {
946                 tail = bh;
947                 bh = bh->b_this_page;
948         } while (bh);
949         tail->b_this_page = head;
950         attach_page_buffers(page, head);
951 }
952
953 /*
954  * Initialise the state of a blockdev page's buffers.
955  */ 
956 static void
957 init_page_buffers(struct page *page, struct block_device *bdev,
958                         sector_t block, int size)
959 {
960         struct buffer_head *head = page_buffers(page);
961         struct buffer_head *bh = head;
962         int uptodate = PageUptodate(page);
963
964         do {
965                 if (!buffer_mapped(bh)) {
966                         init_buffer(bh, NULL, NULL);
967                         bh->b_bdev = bdev;
968                         bh->b_blocknr = block;
969                         if (uptodate)
970                                 set_buffer_uptodate(bh);
971                         set_buffer_mapped(bh);
972                 }
973                 block++;
974                 bh = bh->b_this_page;
975         } while (bh != head);
976 }
977
978 /*
979  * Create the page-cache page that contains the requested block.
980  *
981  * This is user purely for blockdev mappings.
982  */
983 static struct page *
984 grow_dev_page(struct block_device *bdev, sector_t block,
985                 pgoff_t index, int size)
986 {
987         struct inode *inode = bdev->bd_inode;
988         struct page *page;
989         struct buffer_head *bh;
990
991         page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
992         if (!page)
993                 return NULL;
994
995         BUG_ON(!PageLocked(page));
996
997         if (page_has_buffers(page)) {
998                 bh = page_buffers(page);
999                 if (bh->b_size == size) {
1000                         init_page_buffers(page, bdev, block, size);
1001                         return page;
1002                 }
1003                 if (!try_to_free_buffers(page))
1004                         goto failed;
1005         }
1006
1007         /*
1008          * Allocate some buffers for this page
1009          */
1010         bh = alloc_page_buffers(page, size, 0);
1011         if (!bh)
1012                 goto failed;
1013
1014         /*
1015          * Link the page to the buffers and initialise them.  Take the
1016          * lock to be atomic wrt __find_get_block(), which does not
1017          * run under the page lock.
1018          */
1019         spin_lock(&inode->i_mapping->private_lock);
1020         link_dev_buffers(page, bh);
1021         init_page_buffers(page, bdev, block, size);
1022         spin_unlock(&inode->i_mapping->private_lock);
1023         return page;
1024
1025 failed:
1026         BUG();
1027         unlock_page(page);
1028         page_cache_release(page);
1029         return NULL;
1030 }
1031
1032 /*
1033  * Create buffers for the specified block device block's page.  If
1034  * that page was dirty, the buffers are set dirty also.
1035  *
1036  * Except that's a bug.  Attaching dirty buffers to a dirty
1037  * blockdev's page can result in filesystem corruption, because
1038  * some of those buffers may be aliases of filesystem data.
1039  * grow_dev_page() will go BUG() if this happens.
1040  */
1041 static int
1042 grow_buffers(struct block_device *bdev, sector_t block, int size)
1043 {
1044         struct page *page;
1045         pgoff_t index;
1046         int sizebits;
1047
1048         sizebits = -1;
1049         do {
1050                 sizebits++;
1051         } while ((size << sizebits) < PAGE_SIZE);
1052
1053         index = block >> sizebits;
1054
1055         /*
1056          * Check for a block which wants to lie outside our maximum possible
1057          * pagecache index.  (this comparison is done using sector_t types).
1058          */
1059         if (unlikely(index != block >> sizebits)) {
1060                 char b[BDEVNAME_SIZE];
1061
1062                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1063                         "device %s\n",
1064                         __FUNCTION__, (unsigned long long)block,
1065                         bdevname(bdev, b));
1066                 return -EIO;
1067         }
1068         block = index << sizebits;
1069         /* Create a page with the proper size buffers.. */
1070         page = grow_dev_page(bdev, block, index, size);
1071         if (!page)
1072                 return 0;
1073         unlock_page(page);
1074         page_cache_release(page);
1075         return 1;
1076 }
1077
1078 static struct buffer_head *
1079 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1080 {
1081         /* Size must be multiple of hard sectorsize */
1082         if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1083                         (size < 512 || size > PAGE_SIZE))) {
1084                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1085                                         size);
1086                 printk(KERN_ERR "hardsect size: %d\n",
1087                                         bdev_hardsect_size(bdev));
1088
1089                 dump_stack();
1090                 return NULL;
1091         }
1092
1093         for (;;) {
1094                 struct buffer_head * bh;
1095                 int ret;
1096
1097                 bh = __find_get_block(bdev, block, size);
1098                 if (bh)
1099                         return bh;
1100
1101                 ret = grow_buffers(bdev, block, size);
1102                 if (ret < 0)
1103                         return NULL;
1104                 if (ret == 0)
1105                         free_more_memory();
1106         }
1107 }
1108
1109 /*
1110  * The relationship between dirty buffers and dirty pages:
1111  *
1112  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1113  * the page is tagged dirty in its radix tree.
1114  *
1115  * At all times, the dirtiness of the buffers represents the dirtiness of
1116  * subsections of the page.  If the page has buffers, the page dirty bit is
1117  * merely a hint about the true dirty state.
1118  *
1119  * When a page is set dirty in its entirety, all its buffers are marked dirty
1120  * (if the page has buffers).
1121  *
1122  * When a buffer is marked dirty, its page is dirtied, but the page's other
1123  * buffers are not.
1124  *
1125  * Also.  When blockdev buffers are explicitly read with bread(), they
1126  * individually become uptodate.  But their backing page remains not
1127  * uptodate - even if all of its buffers are uptodate.  A subsequent
1128  * block_read_full_page() against that page will discover all the uptodate
1129  * buffers, will set the page uptodate and will perform no I/O.
1130  */
1131
1132 /**
1133  * mark_buffer_dirty - mark a buffer_head as needing writeout
1134  * @bh: the buffer_head to mark dirty
1135  *
1136  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1137  * backing page dirty, then tag the page as dirty in its address_space's radix
1138  * tree and then attach the address_space's inode to its superblock's dirty
1139  * inode list.
1140  *
1141  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1142  * mapping->tree_lock and the global inode_lock.
1143  */
1144 void fastcall mark_buffer_dirty(struct buffer_head *bh)
1145 {
1146         if (!buffer_dirty(bh) && !test_set_buffer_dirty(bh))
1147                 __set_page_dirty_nobuffers(bh->b_page);
1148 }
1149
1150 /*
1151  * Decrement a buffer_head's reference count.  If all buffers against a page
1152  * have zero reference count, are clean and unlocked, and if the page is clean
1153  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1154  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1155  * a page but it ends up not being freed, and buffers may later be reattached).
1156  */
1157 void __brelse(struct buffer_head * buf)
1158 {
1159         if (atomic_read(&buf->b_count)) {
1160                 put_bh(buf);
1161                 return;
1162         }
1163         printk(KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1164         WARN_ON(1);
1165 }
1166
1167 /*
1168  * bforget() is like brelse(), except it discards any
1169  * potentially dirty data.
1170  */
1171 void __bforget(struct buffer_head *bh)
1172 {
1173         clear_buffer_dirty(bh);
1174         if (!list_empty(&bh->b_assoc_buffers)) {
1175                 struct address_space *buffer_mapping = bh->b_page->mapping;
1176
1177                 spin_lock(&buffer_mapping->private_lock);
1178                 list_del_init(&bh->b_assoc_buffers);
1179                 bh->b_assoc_map = NULL;
1180                 spin_unlock(&buffer_mapping->private_lock);
1181         }
1182         __brelse(bh);
1183 }
1184
1185 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1186 {
1187         lock_buffer(bh);
1188         if (buffer_uptodate(bh)) {
1189                 unlock_buffer(bh);
1190                 return bh;
1191         } else {
1192                 get_bh(bh);
1193                 bh->b_end_io = end_buffer_read_sync;
1194                 submit_bh(READ, bh);
1195                 wait_on_buffer(bh);
1196                 if (buffer_uptodate(bh))
1197                         return bh;
1198         }
1199         brelse(bh);
1200         return NULL;
1201 }
1202
1203 /*
1204  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1205  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1206  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1207  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1208  * CPU's LRUs at the same time.
1209  *
1210  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1211  * sb_find_get_block().
1212  *
1213  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1214  * a local interrupt disable for that.
1215  */
1216
1217 #define BH_LRU_SIZE     8
1218
1219 struct bh_lru {
1220         struct buffer_head *bhs[BH_LRU_SIZE];
1221 };
1222
1223 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1224
1225 #ifdef CONFIG_SMP
1226 #define bh_lru_lock()   local_irq_disable()
1227 #define bh_lru_unlock() local_irq_enable()
1228 #else
1229 #define bh_lru_lock()   preempt_disable()
1230 #define bh_lru_unlock() preempt_enable()
1231 #endif
1232
1233 static inline void check_irqs_on(void)
1234 {
1235 #ifdef irqs_disabled
1236         BUG_ON(irqs_disabled());
1237 #endif
1238 }
1239
1240 /*
1241  * The LRU management algorithm is dopey-but-simple.  Sorry.
1242  */
1243 static void bh_lru_install(struct buffer_head *bh)
1244 {
1245         struct buffer_head *evictee = NULL;
1246         struct bh_lru *lru;
1247
1248         check_irqs_on();
1249         bh_lru_lock();
1250         lru = &__get_cpu_var(bh_lrus);
1251         if (lru->bhs[0] != bh) {
1252                 struct buffer_head *bhs[BH_LRU_SIZE];
1253                 int in;
1254                 int out = 0;
1255
1256                 get_bh(bh);
1257                 bhs[out++] = bh;
1258                 for (in = 0; in < BH_LRU_SIZE; in++) {
1259                         struct buffer_head *bh2 = lru->bhs[in];
1260
1261                         if (bh2 == bh) {
1262                                 __brelse(bh2);
1263                         } else {
1264                                 if (out >= BH_LRU_SIZE) {
1265                                         BUG_ON(evictee != NULL);
1266                                         evictee = bh2;
1267                                 } else {
1268                                         bhs[out++] = bh2;
1269                                 }
1270                         }
1271                 }
1272                 while (out < BH_LRU_SIZE)
1273                         bhs[out++] = NULL;
1274                 memcpy(lru->bhs, bhs, sizeof(bhs));
1275         }
1276         bh_lru_unlock();
1277
1278         if (evictee)
1279                 __brelse(evictee);
1280 }
1281
1282 /*
1283  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1284  */
1285 static struct buffer_head *
1286 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1287 {
1288         struct buffer_head *ret = NULL;
1289         struct bh_lru *lru;
1290         unsigned int i;
1291
1292         check_irqs_on();
1293         bh_lru_lock();
1294         lru = &__get_cpu_var(bh_lrus);
1295         for (i = 0; i < BH_LRU_SIZE; i++) {
1296                 struct buffer_head *bh = lru->bhs[i];
1297
1298                 if (bh && bh->b_bdev == bdev &&
1299                                 bh->b_blocknr == block && bh->b_size == size) {
1300                         if (i) {
1301                                 while (i) {
1302                                         lru->bhs[i] = lru->bhs[i - 1];
1303                                         i--;
1304                                 }
1305                                 lru->bhs[0] = bh;
1306                         }
1307                         get_bh(bh);
1308                         ret = bh;
1309                         break;
1310                 }
1311         }
1312         bh_lru_unlock();
1313         return ret;
1314 }
1315
1316 /*
1317  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1318  * it in the LRU and mark it as accessed.  If it is not present then return
1319  * NULL
1320  */
1321 struct buffer_head *
1322 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1323 {
1324         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1325
1326         if (bh == NULL) {
1327                 bh = __find_get_block_slow(bdev, block);
1328                 if (bh)
1329                         bh_lru_install(bh);
1330         }
1331         if (bh)
1332                 touch_buffer(bh);
1333         return bh;
1334 }
1335 EXPORT_SYMBOL(__find_get_block);
1336
1337 /*
1338  * __getblk will locate (and, if necessary, create) the buffer_head
1339  * which corresponds to the passed block_device, block and size. The
1340  * returned buffer has its reference count incremented.
1341  *
1342  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1343  * illegal block number, __getblk() will happily return a buffer_head
1344  * which represents the non-existent block.  Very weird.
1345  *
1346  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1347  * attempt is failing.  FIXME, perhaps?
1348  */
1349 struct buffer_head *
1350 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1351 {
1352         struct buffer_head *bh = __find_get_block(bdev, block, size);
1353
1354         might_sleep();
1355         if (bh == NULL)
1356                 bh = __getblk_slow(bdev, block, size);
1357         return bh;
1358 }
1359 EXPORT_SYMBOL(__getblk);
1360
1361 /*
1362  * Do async read-ahead on a buffer..
1363  */
1364 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1365 {
1366         struct buffer_head *bh = __getblk(bdev, block, size);
1367         if (likely(bh)) {
1368                 ll_rw_block(READA, 1, &bh);
1369                 brelse(bh);
1370         }
1371 }
1372 EXPORT_SYMBOL(__breadahead);
1373
1374 /**
1375  *  __bread() - reads a specified block and returns the bh
1376  *  @bdev: the block_device to read from
1377  *  @block: number of block
1378  *  @size: size (in bytes) to read
1379  * 
1380  *  Reads a specified block, and returns buffer head that contains it.
1381  *  It returns NULL if the block was unreadable.
1382  */
1383 struct buffer_head *
1384 __bread(struct block_device *bdev, sector_t block, unsigned size)
1385 {
1386         struct buffer_head *bh = __getblk(bdev, block, size);
1387
1388         if (likely(bh) && !buffer_uptodate(bh))
1389                 bh = __bread_slow(bh);
1390         return bh;
1391 }
1392 EXPORT_SYMBOL(__bread);
1393
1394 /*
1395  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1396  * This doesn't race because it runs in each cpu either in irq
1397  * or with preempt disabled.
1398  */
1399 static void invalidate_bh_lru(void *arg)
1400 {
1401         struct bh_lru *b = &get_cpu_var(bh_lrus);
1402         int i;
1403
1404         for (i = 0; i < BH_LRU_SIZE; i++) {
1405                 brelse(b->bhs[i]);
1406                 b->bhs[i] = NULL;
1407         }
1408         put_cpu_var(bh_lrus);
1409 }
1410         
1411 static void invalidate_bh_lrus(void)
1412 {
1413         on_each_cpu(invalidate_bh_lru, NULL, 1, 1);
1414 }
1415
1416 void set_bh_page(struct buffer_head *bh,
1417                 struct page *page, unsigned long offset)
1418 {
1419         bh->b_page = page;
1420         BUG_ON(offset >= PAGE_SIZE);
1421         if (PageHighMem(page))
1422                 /*
1423                  * This catches illegal uses and preserves the offset:
1424                  */
1425                 bh->b_data = (char *)(0 + offset);
1426         else
1427                 bh->b_data = page_address(page) + offset;
1428 }
1429 EXPORT_SYMBOL(set_bh_page);
1430
1431 /*
1432  * Called when truncating a buffer on a page completely.
1433  */
1434 static void discard_buffer(struct buffer_head * bh)
1435 {
1436         lock_buffer(bh);
1437         clear_buffer_dirty(bh);
1438         bh->b_bdev = NULL;
1439         clear_buffer_mapped(bh);
1440         clear_buffer_req(bh);
1441         clear_buffer_new(bh);
1442         clear_buffer_delay(bh);
1443         clear_buffer_unwritten(bh);
1444         unlock_buffer(bh);
1445 }
1446
1447 /**
1448  * block_invalidatepage - invalidate part of all of a buffer-backed page
1449  *
1450  * @page: the page which is affected
1451  * @offset: the index of the truncation point
1452  *
1453  * block_invalidatepage() is called when all or part of the page has become
1454  * invalidatedby a truncate operation.
1455  *
1456  * block_invalidatepage() does not have to release all buffers, but it must
1457  * ensure that no dirty buffer is left outside @offset and that no I/O
1458  * is underway against any of the blocks which are outside the truncation
1459  * point.  Because the caller is about to free (and possibly reuse) those
1460  * blocks on-disk.
1461  */
1462 void block_invalidatepage(struct page *page, unsigned long offset)
1463 {
1464         struct buffer_head *head, *bh, *next;
1465         unsigned int curr_off = 0;
1466
1467         BUG_ON(!PageLocked(page));
1468         if (!page_has_buffers(page))
1469                 goto out;
1470
1471         head = page_buffers(page);
1472         bh = head;
1473         do {
1474                 unsigned int next_off = curr_off + bh->b_size;
1475                 next = bh->b_this_page;
1476
1477                 /*
1478                  * is this block fully invalidated?
1479                  */
1480                 if (offset <= curr_off)
1481                         discard_buffer(bh);
1482                 curr_off = next_off;
1483                 bh = next;
1484         } while (bh != head);
1485
1486         /*
1487          * We release buffers only if the entire page is being invalidated.
1488          * The get_block cached value has been unconditionally invalidated,
1489          * so real IO is not possible anymore.
1490          */
1491         if (offset == 0)
1492                 try_to_release_page(page, 0);
1493 out:
1494         return;
1495 }
1496 EXPORT_SYMBOL(block_invalidatepage);
1497
1498 /*
1499  * We attach and possibly dirty the buffers atomically wrt
1500  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1501  * is already excluded via the page lock.
1502  */
1503 void create_empty_buffers(struct page *page,
1504                         unsigned long blocksize, unsigned long b_state)
1505 {
1506         struct buffer_head *bh, *head, *tail;
1507
1508         head = alloc_page_buffers(page, blocksize, 1);
1509         bh = head;
1510         do {
1511                 bh->b_state |= b_state;
1512                 tail = bh;
1513                 bh = bh->b_this_page;
1514         } while (bh);
1515         tail->b_this_page = head;
1516
1517         spin_lock(&page->mapping->private_lock);
1518         if (PageUptodate(page) || PageDirty(page)) {
1519                 bh = head;
1520                 do {
1521                         if (PageDirty(page))
1522                                 set_buffer_dirty(bh);
1523                         if (PageUptodate(page))
1524                                 set_buffer_uptodate(bh);
1525                         bh = bh->b_this_page;
1526                 } while (bh != head);
1527         }
1528         attach_page_buffers(page, head);
1529         spin_unlock(&page->mapping->private_lock);
1530 }
1531 EXPORT_SYMBOL(create_empty_buffers);
1532
1533 /*
1534  * We are taking a block for data and we don't want any output from any
1535  * buffer-cache aliases starting from return from that function and
1536  * until the moment when something will explicitly mark the buffer
1537  * dirty (hopefully that will not happen until we will free that block ;-)
1538  * We don't even need to mark it not-uptodate - nobody can expect
1539  * anything from a newly allocated buffer anyway. We used to used
1540  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1541  * don't want to mark the alias unmapped, for example - it would confuse
1542  * anyone who might pick it with bread() afterwards...
1543  *
1544  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1545  * be writeout I/O going on against recently-freed buffers.  We don't
1546  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1547  * only if we really need to.  That happens here.
1548  */
1549 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1550 {
1551         struct buffer_head *old_bh;
1552
1553         might_sleep();
1554
1555         old_bh = __find_get_block_slow(bdev, block);
1556         if (old_bh) {
1557                 clear_buffer_dirty(old_bh);
1558                 wait_on_buffer(old_bh);
1559                 clear_buffer_req(old_bh);
1560                 __brelse(old_bh);
1561         }
1562 }
1563 EXPORT_SYMBOL(unmap_underlying_metadata);
1564
1565 /*
1566  * NOTE! All mapped/uptodate combinations are valid:
1567  *
1568  *      Mapped  Uptodate        Meaning
1569  *
1570  *      No      No              "unknown" - must do get_block()
1571  *      No      Yes             "hole" - zero-filled
1572  *      Yes     No              "allocated" - allocated on disk, not read in
1573  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1574  *
1575  * "Dirty" is valid only with the last case (mapped+uptodate).
1576  */
1577
1578 /*
1579  * While block_write_full_page is writing back the dirty buffers under
1580  * the page lock, whoever dirtied the buffers may decide to clean them
1581  * again at any time.  We handle that by only looking at the buffer
1582  * state inside lock_buffer().
1583  *
1584  * If block_write_full_page() is called for regular writeback
1585  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1586  * locked buffer.   This only can happen if someone has written the buffer
1587  * directly, with submit_bh().  At the address_space level PageWriteback
1588  * prevents this contention from occurring.
1589  */
1590 static int __block_write_full_page(struct inode *inode, struct page *page,
1591                         get_block_t *get_block, struct writeback_control *wbc)
1592 {
1593         int err;
1594         sector_t block;
1595         sector_t last_block;
1596         struct buffer_head *bh, *head;
1597         const unsigned blocksize = 1 << inode->i_blkbits;
1598         int nr_underway = 0;
1599
1600         BUG_ON(!PageLocked(page));
1601
1602         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1603
1604         if (!page_has_buffers(page)) {
1605                 create_empty_buffers(page, blocksize,
1606                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1607         }
1608
1609         /*
1610          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1611          * here, and the (potentially unmapped) buffers may become dirty at
1612          * any time.  If a buffer becomes dirty here after we've inspected it
1613          * then we just miss that fact, and the page stays dirty.
1614          *
1615          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1616          * handle that here by just cleaning them.
1617          */
1618
1619         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1620         head = page_buffers(page);
1621         bh = head;
1622
1623         /*
1624          * Get all the dirty buffers mapped to disk addresses and
1625          * handle any aliases from the underlying blockdev's mapping.
1626          */
1627         do {
1628                 if (block > last_block) {
1629                         /*
1630                          * mapped buffers outside i_size will occur, because
1631                          * this page can be outside i_size when there is a
1632                          * truncate in progress.
1633                          */
1634                         /*
1635                          * The buffer was zeroed by block_write_full_page()
1636                          */
1637                         clear_buffer_dirty(bh);
1638                         set_buffer_uptodate(bh);
1639                 } else if (!buffer_mapped(bh) && buffer_dirty(bh)) {
1640                         WARN_ON(bh->b_size != blocksize);
1641                         err = get_block(inode, block, bh, 1);
1642                         if (err)
1643                                 goto recover;
1644                         if (buffer_new(bh)) {
1645                                 /* blockdev mappings never come here */
1646                                 clear_buffer_new(bh);
1647                                 unmap_underlying_metadata(bh->b_bdev,
1648                                                         bh->b_blocknr);
1649                         }
1650                 }
1651                 bh = bh->b_this_page;
1652                 block++;
1653         } while (bh != head);
1654
1655         do {
1656                 if (!buffer_mapped(bh))
1657                         continue;
1658                 /*
1659                  * If it's a fully non-blocking write attempt and we cannot
1660                  * lock the buffer then redirty the page.  Note that this can
1661                  * potentially cause a busy-wait loop from pdflush and kswapd
1662                  * activity, but those code paths have their own higher-level
1663                  * throttling.
1664                  */
1665                 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1666                         lock_buffer(bh);
1667                 } else if (test_set_buffer_locked(bh)) {
1668                         redirty_page_for_writepage(wbc, page);
1669                         continue;
1670                 }
1671                 if (test_clear_buffer_dirty(bh)) {
1672                         mark_buffer_async_write(bh);
1673                 } else {
1674                         unlock_buffer(bh);
1675                 }
1676         } while ((bh = bh->b_this_page) != head);
1677
1678         /*
1679          * The page and its buffers are protected by PageWriteback(), so we can
1680          * drop the bh refcounts early.
1681          */
1682         BUG_ON(PageWriteback(page));
1683         set_page_writeback(page);
1684
1685         do {
1686                 struct buffer_head *next = bh->b_this_page;
1687                 if (buffer_async_write(bh)) {
1688                         submit_bh(WRITE, bh);
1689                         nr_underway++;
1690                 }
1691                 bh = next;
1692         } while (bh != head);
1693         unlock_page(page);
1694
1695         err = 0;
1696 done:
1697         if (nr_underway == 0) {
1698                 /*
1699                  * The page was marked dirty, but the buffers were
1700                  * clean.  Someone wrote them back by hand with
1701                  * ll_rw_block/submit_bh.  A rare case.
1702                  */
1703                 int uptodate = 1;
1704                 do {
1705                         if (!buffer_uptodate(bh)) {
1706                                 uptodate = 0;
1707                                 break;
1708                         }
1709                         bh = bh->b_this_page;
1710                 } while (bh != head);
1711                 if (uptodate)
1712                         SetPageUptodate(page);
1713                 end_page_writeback(page);
1714                 /*
1715                  * The page and buffer_heads can be released at any time from
1716                  * here on.
1717                  */
1718                 wbc->pages_skipped++;   /* We didn't write this page */
1719         }
1720         return err;
1721
1722 recover:
1723         /*
1724          * ENOSPC, or some other error.  We may already have added some
1725          * blocks to the file, so we need to write these out to avoid
1726          * exposing stale data.
1727          * The page is currently locked and not marked for writeback
1728          */
1729         bh = head;
1730         /* Recovery: lock and submit the mapped buffers */
1731         do {
1732                 if (buffer_mapped(bh) && buffer_dirty(bh)) {
1733                         lock_buffer(bh);
1734                         mark_buffer_async_write(bh);
1735                 } else {
1736                         /*
1737                          * The buffer may have been set dirty during
1738                          * attachment to a dirty page.
1739                          */
1740                         clear_buffer_dirty(bh);
1741                 }
1742         } while ((bh = bh->b_this_page) != head);
1743         SetPageError(page);
1744         BUG_ON(PageWriteback(page));
1745         set_page_writeback(page);
1746         do {
1747                 struct buffer_head *next = bh->b_this_page;
1748                 if (buffer_async_write(bh)) {
1749                         clear_buffer_dirty(bh);
1750                         submit_bh(WRITE, bh);
1751                         nr_underway++;
1752                 }
1753                 bh = next;
1754         } while (bh != head);
1755         unlock_page(page);
1756         goto done;
1757 }
1758
1759 static int __block_prepare_write(struct inode *inode, struct page *page,
1760                 unsigned from, unsigned to, get_block_t *get_block)
1761 {
1762         unsigned block_start, block_end;
1763         sector_t block;
1764         int err = 0;
1765         unsigned blocksize, bbits;
1766         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1767
1768         BUG_ON(!PageLocked(page));
1769         BUG_ON(from > PAGE_CACHE_SIZE);
1770         BUG_ON(to > PAGE_CACHE_SIZE);
1771         BUG_ON(from > to);
1772
1773         blocksize = 1 << inode->i_blkbits;
1774         if (!page_has_buffers(page))
1775                 create_empty_buffers(page, blocksize, 0);
1776         head = page_buffers(page);
1777
1778         bbits = inode->i_blkbits;
1779         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1780
1781         for(bh = head, block_start = 0; bh != head || !block_start;
1782             block++, block_start=block_end, bh = bh->b_this_page) {
1783                 block_end = block_start + blocksize;
1784                 if (block_end <= from || block_start >= to) {
1785                         if (PageUptodate(page)) {
1786                                 if (!buffer_uptodate(bh))
1787                                         set_buffer_uptodate(bh);
1788                         }
1789                         continue;
1790                 }
1791                 if (buffer_new(bh))
1792                         clear_buffer_new(bh);
1793                 if (!buffer_mapped(bh)) {
1794                         WARN_ON(bh->b_size != blocksize);
1795                         err = get_block(inode, block, bh, 1);
1796                         if (err)
1797                                 break;
1798                         if (buffer_new(bh)) {
1799                                 unmap_underlying_metadata(bh->b_bdev,
1800                                                         bh->b_blocknr);
1801                                 if (PageUptodate(page)) {
1802                                         set_buffer_uptodate(bh);
1803                                         continue;
1804                                 }
1805                                 if (block_end > to || block_start < from) {
1806                                         void *kaddr;
1807
1808                                         kaddr = kmap_atomic(page, KM_USER0);
1809                                         if (block_end > to)
1810                                                 memset(kaddr+to, 0,
1811                                                         block_end-to);
1812                                         if (block_start < from)
1813                                                 memset(kaddr+block_start,
1814                                                         0, from-block_start);
1815                                         flush_dcache_page(page);
1816                                         kunmap_atomic(kaddr, KM_USER0);
1817                                 }
1818                                 continue;
1819                         }
1820                 }
1821                 if (PageUptodate(page)) {
1822                         if (!buffer_uptodate(bh))
1823                                 set_buffer_uptodate(bh);
1824                         continue; 
1825                 }
1826                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1827                     !buffer_unwritten(bh) &&
1828                      (block_start < from || block_end > to)) {
1829                         ll_rw_block(READ, 1, &bh);
1830                         *wait_bh++=bh;
1831                 }
1832         }
1833         /*
1834          * If we issued read requests - let them complete.
1835          */
1836         while(wait_bh > wait) {
1837                 wait_on_buffer(*--wait_bh);
1838                 if (!buffer_uptodate(*wait_bh))
1839                         err = -EIO;
1840         }
1841         if (!err) {
1842                 bh = head;
1843                 do {
1844                         if (buffer_new(bh))
1845                                 clear_buffer_new(bh);
1846                 } while ((bh = bh->b_this_page) != head);
1847                 return 0;
1848         }
1849         /* Error case: */
1850         /*
1851          * Zero out any newly allocated blocks to avoid exposing stale
1852          * data.  If BH_New is set, we know that the block was newly
1853          * allocated in the above loop.
1854          */
1855         bh = head;
1856         block_start = 0;
1857         do {
1858                 block_end = block_start+blocksize;
1859                 if (block_end <= from)
1860                         goto next_bh;
1861                 if (block_start >= to)
1862                         break;
1863                 if (buffer_new(bh)) {
1864                         void *kaddr;
1865
1866                         clear_buffer_new(bh);
1867                         kaddr = kmap_atomic(page, KM_USER0);
1868                         memset(kaddr+block_start, 0, bh->b_size);
1869                         flush_dcache_page(page);
1870                         kunmap_atomic(kaddr, KM_USER0);
1871                         set_buffer_uptodate(bh);
1872                         mark_buffer_dirty(bh);
1873                 }
1874 next_bh:
1875                 block_start = block_end;
1876                 bh = bh->b_this_page;
1877         } while (bh != head);
1878         return err;
1879 }
1880
1881 static int __block_commit_write(struct inode *inode, struct page *page,
1882                 unsigned from, unsigned to)
1883 {
1884         unsigned block_start, block_end;
1885         int partial = 0;
1886         unsigned blocksize;
1887         struct buffer_head *bh, *head;
1888
1889         blocksize = 1 << inode->i_blkbits;
1890
1891         for(bh = head = page_buffers(page), block_start = 0;
1892             bh != head || !block_start;
1893             block_start=block_end, bh = bh->b_this_page) {
1894                 block_end = block_start + blocksize;
1895                 if (block_end <= from || block_start >= to) {
1896                         if (!buffer_uptodate(bh))
1897                                 partial = 1;
1898                 } else {
1899                         set_buffer_uptodate(bh);
1900                         mark_buffer_dirty(bh);
1901                 }
1902         }
1903
1904         /*
1905          * If this is a partial write which happened to make all buffers
1906          * uptodate then we can optimize away a bogus readpage() for
1907          * the next read(). Here we 'discover' whether the page went
1908          * uptodate as a result of this (potentially partial) write.
1909          */
1910         if (!partial)
1911                 SetPageUptodate(page);
1912         return 0;
1913 }
1914
1915 /*
1916  * Generic "read page" function for block devices that have the normal
1917  * get_block functionality. This is most of the block device filesystems.
1918  * Reads the page asynchronously --- the unlock_buffer() and
1919  * set/clear_buffer_uptodate() functions propagate buffer state into the
1920  * page struct once IO has completed.
1921  */
1922 int block_read_full_page(struct page *page, get_block_t *get_block)
1923 {
1924         struct inode *inode = page->mapping->host;
1925         sector_t iblock, lblock;
1926         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
1927         unsigned int blocksize;
1928         int nr, i;
1929         int fully_mapped = 1;
1930
1931         BUG_ON(!PageLocked(page));
1932         blocksize = 1 << inode->i_blkbits;
1933         if (!page_has_buffers(page))
1934                 create_empty_buffers(page, blocksize, 0);
1935         head = page_buffers(page);
1936
1937         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1938         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
1939         bh = head;
1940         nr = 0;
1941         i = 0;
1942
1943         do {
1944                 if (buffer_uptodate(bh))
1945                         continue;
1946
1947                 if (!buffer_mapped(bh)) {
1948                         int err = 0;
1949
1950                         fully_mapped = 0;
1951                         if (iblock < lblock) {
1952                                 WARN_ON(bh->b_size != blocksize);
1953                                 err = get_block(inode, iblock, bh, 0);
1954                                 if (err)
1955                                         SetPageError(page);
1956                         }
1957                         if (!buffer_mapped(bh)) {
1958                                 void *kaddr = kmap_atomic(page, KM_USER0);
1959                                 memset(kaddr + i * blocksize, 0, blocksize);
1960                                 flush_dcache_page(page);
1961                                 kunmap_atomic(kaddr, KM_USER0);
1962                                 if (!err)
1963                                         set_buffer_uptodate(bh);
1964                                 continue;
1965                         }
1966                         /*
1967                          * get_block() might have updated the buffer
1968                          * synchronously
1969                          */
1970                         if (buffer_uptodate(bh))
1971                                 continue;
1972                 }
1973                 arr[nr++] = bh;
1974         } while (i++, iblock++, (bh = bh->b_this_page) != head);
1975
1976         if (fully_mapped)
1977                 SetPageMappedToDisk(page);
1978
1979         if (!nr) {
1980                 /*
1981                  * All buffers are uptodate - we can set the page uptodate
1982                  * as well. But not if get_block() returned an error.
1983                  */
1984                 if (!PageError(page))
1985                         SetPageUptodate(page);
1986                 unlock_page(page);
1987                 return 0;
1988         }
1989
1990         /* Stage two: lock the buffers */
1991         for (i = 0; i < nr; i++) {
1992                 bh = arr[i];
1993                 lock_buffer(bh);
1994                 mark_buffer_async_read(bh);
1995         }
1996
1997         /*
1998          * Stage 3: start the IO.  Check for uptodateness
1999          * inside the buffer lock in case another process reading
2000          * the underlying blockdev brought it uptodate (the sct fix).
2001          */
2002         for (i = 0; i < nr; i++) {
2003                 bh = arr[i];
2004                 if (buffer_uptodate(bh))
2005                         end_buffer_async_read(bh, 1);
2006                 else
2007                         submit_bh(READ, bh);
2008         }
2009         return 0;
2010 }
2011
2012 /* utility function for filesystems that need to do work on expanding
2013  * truncates.  Uses prepare/commit_write to allow the filesystem to
2014  * deal with the hole.  
2015  */
2016 static int __generic_cont_expand(struct inode *inode, loff_t size,
2017                                  pgoff_t index, unsigned int offset)
2018 {
2019         struct address_space *mapping = inode->i_mapping;
2020         struct page *page;
2021         unsigned long limit;
2022         int err;
2023
2024         err = -EFBIG;
2025         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2026         if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2027                 send_sig(SIGXFSZ, current, 0);
2028                 goto out;
2029         }
2030         if (size > inode->i_sb->s_maxbytes)
2031                 goto out;
2032
2033         err = -ENOMEM;
2034         page = grab_cache_page(mapping, index);
2035         if (!page)
2036                 goto out;
2037         err = mapping->a_ops->prepare_write(NULL, page, offset, offset);
2038         if (err) {
2039                 /*
2040                  * ->prepare_write() may have instantiated a few blocks
2041                  * outside i_size.  Trim these off again.
2042                  */
2043                 unlock_page(page);
2044                 page_cache_release(page);
2045                 vmtruncate(inode, inode->i_size);
2046                 goto out;
2047         }
2048
2049         err = mapping->a_ops->commit_write(NULL, page, offset, offset);
2050
2051         unlock_page(page);
2052         page_cache_release(page);
2053         if (err > 0)
2054                 err = 0;
2055 out:
2056         return err;
2057 }
2058
2059 int generic_cont_expand(struct inode *inode, loff_t size)
2060 {
2061         pgoff_t index;
2062         unsigned int offset;
2063
2064         offset = (size & (PAGE_CACHE_SIZE - 1)); /* Within page */
2065
2066         /* ugh.  in prepare/commit_write, if from==to==start of block, we
2067         ** skip the prepare.  make sure we never send an offset for the start
2068         ** of a block
2069         */
2070         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
2071                 /* caller must handle this extra byte. */
2072                 offset++;
2073         }
2074         index = size >> PAGE_CACHE_SHIFT;
2075
2076         return __generic_cont_expand(inode, size, index, offset);
2077 }
2078
2079 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2080 {
2081         loff_t pos = size - 1;
2082         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
2083         unsigned int offset = (pos & (PAGE_CACHE_SIZE - 1)) + 1;
2084
2085         /* prepare/commit_write can handle even if from==to==start of block. */
2086         return __generic_cont_expand(inode, size, index, offset);
2087 }
2088
2089 /*
2090  * For moronic filesystems that do not allow holes in file.
2091  * We may have to extend the file.
2092  */
2093
2094 int cont_prepare_write(struct page *page, unsigned offset,
2095                 unsigned to, get_block_t *get_block, loff_t *bytes)
2096 {
2097         struct address_space *mapping = page->mapping;
2098         struct inode *inode = mapping->host;
2099         struct page *new_page;
2100         pgoff_t pgpos;
2101         long status;
2102         unsigned zerofrom;
2103         unsigned blocksize = 1 << inode->i_blkbits;
2104         void *kaddr;
2105
2106         while(page->index > (pgpos = *bytes>>PAGE_CACHE_SHIFT)) {
2107                 status = -ENOMEM;
2108                 new_page = grab_cache_page(mapping, pgpos);
2109                 if (!new_page)
2110                         goto out;
2111                 /* we might sleep */
2112                 if (*bytes>>PAGE_CACHE_SHIFT != pgpos) {
2113                         unlock_page(new_page);
2114                         page_cache_release(new_page);
2115                         continue;
2116                 }
2117                 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2118                 if (zerofrom & (blocksize-1)) {
2119                         *bytes |= (blocksize-1);
2120                         (*bytes)++;
2121                 }
2122                 status = __block_prepare_write(inode, new_page, zerofrom,
2123                                                 PAGE_CACHE_SIZE, get_block);
2124                 if (status)
2125                         goto out_unmap;
2126                 kaddr = kmap_atomic(new_page, KM_USER0);
2127                 memset(kaddr+zerofrom, 0, PAGE_CACHE_SIZE-zerofrom);
2128                 flush_dcache_page(new_page);
2129                 kunmap_atomic(kaddr, KM_USER0);
2130                 generic_commit_write(NULL, new_page, zerofrom, PAGE_CACHE_SIZE);
2131                 unlock_page(new_page);
2132                 page_cache_release(new_page);
2133         }
2134
2135         if (page->index < pgpos) {
2136                 /* completely inside the area */
2137                 zerofrom = offset;
2138         } else {
2139                 /* page covers the boundary, find the boundary offset */
2140                 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2141
2142                 /* if we will expand the thing last block will be filled */
2143                 if (to > zerofrom && (zerofrom & (blocksize-1))) {
2144                         *bytes |= (blocksize-1);
2145                         (*bytes)++;
2146                 }
2147
2148                 /* starting below the boundary? Nothing to zero out */
2149                 if (offset <= zerofrom)
2150                         zerofrom = offset;
2151         }
2152         status = __block_prepare_write(inode, page, zerofrom, to, get_block);
2153         if (status)
2154                 goto out1;
2155         if (zerofrom < offset) {
2156                 kaddr = kmap_atomic(page, KM_USER0);
2157                 memset(kaddr+zerofrom, 0, offset-zerofrom);
2158                 flush_dcache_page(page);
2159                 kunmap_atomic(kaddr, KM_USER0);
2160                 __block_commit_write(inode, page, zerofrom, offset);
2161         }
2162         return 0;
2163 out1:
2164         ClearPageUptodate(page);
2165         return status;
2166
2167 out_unmap:
2168         ClearPageUptodate(new_page);
2169         unlock_page(new_page);
2170         page_cache_release(new_page);
2171 out:
2172         return status;
2173 }
2174
2175 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2176                         get_block_t *get_block)
2177 {
2178         struct inode *inode = page->mapping->host;
2179         int err = __block_prepare_write(inode, page, from, to, get_block);
2180         if (err)
2181                 ClearPageUptodate(page);
2182         return err;
2183 }
2184
2185 int block_commit_write(struct page *page, unsigned from, unsigned to)
2186 {
2187         struct inode *inode = page->mapping->host;
2188         __block_commit_write(inode,page,from,to);
2189         return 0;
2190 }
2191
2192 int generic_commit_write(struct file *file, struct page *page,
2193                 unsigned from, unsigned to)
2194 {
2195         struct inode *inode = page->mapping->host;
2196         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2197         __block_commit_write(inode,page,from,to);
2198         /*
2199          * No need to use i_size_read() here, the i_size
2200          * cannot change under us because we hold i_mutex.
2201          */
2202         if (pos > inode->i_size) {
2203                 i_size_write(inode, pos);
2204                 mark_inode_dirty(inode);
2205         }
2206         return 0;
2207 }
2208
2209
2210 /*
2211  * nobh_prepare_write()'s prereads are special: the buffer_heads are freed
2212  * immediately, while under the page lock.  So it needs a special end_io
2213  * handler which does not touch the bh after unlocking it.
2214  *
2215  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
2216  * a race there is benign: unlock_buffer() only use the bh's address for
2217  * hashing after unlocking the buffer, so it doesn't actually touch the bh
2218  * itself.
2219  */
2220 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2221 {
2222         if (uptodate) {
2223                 set_buffer_uptodate(bh);
2224         } else {
2225                 /* This happens, due to failed READA attempts. */
2226                 clear_buffer_uptodate(bh);
2227         }
2228         unlock_buffer(bh);
2229 }
2230
2231 /*
2232  * On entry, the page is fully not uptodate.
2233  * On exit the page is fully uptodate in the areas outside (from,to)
2234  */
2235 int nobh_prepare_write(struct page *page, unsigned from, unsigned to,
2236                         get_block_t *get_block)
2237 {
2238         struct inode *inode = page->mapping->host;
2239         const unsigned blkbits = inode->i_blkbits;
2240         const unsigned blocksize = 1 << blkbits;
2241         struct buffer_head map_bh;
2242         struct buffer_head *read_bh[MAX_BUF_PER_PAGE];
2243         unsigned block_in_page;
2244         unsigned block_start;
2245         sector_t block_in_file;
2246         char *kaddr;
2247         int nr_reads = 0;
2248         int i;
2249         int ret = 0;
2250         int is_mapped_to_disk = 1;
2251
2252         if (PageMappedToDisk(page))
2253                 return 0;
2254
2255         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2256         map_bh.b_page = page;
2257
2258         /*
2259          * We loop across all blocks in the page, whether or not they are
2260          * part of the affected region.  This is so we can discover if the
2261          * page is fully mapped-to-disk.
2262          */
2263         for (block_start = 0, block_in_page = 0;
2264                   block_start < PAGE_CACHE_SIZE;
2265                   block_in_page++, block_start += blocksize) {
2266                 unsigned block_end = block_start + blocksize;
2267                 int create;
2268
2269                 map_bh.b_state = 0;
2270                 create = 1;
2271                 if (block_start >= to)
2272                         create = 0;
2273                 map_bh.b_size = blocksize;
2274                 ret = get_block(inode, block_in_file + block_in_page,
2275                                         &map_bh, create);
2276                 if (ret)
2277                         goto failed;
2278                 if (!buffer_mapped(&map_bh))
2279                         is_mapped_to_disk = 0;
2280                 if (buffer_new(&map_bh))
2281                         unmap_underlying_metadata(map_bh.b_bdev,
2282                                                         map_bh.b_blocknr);
2283                 if (PageUptodate(page))
2284                         continue;
2285                 if (buffer_new(&map_bh) || !buffer_mapped(&map_bh)) {
2286                         kaddr = kmap_atomic(page, KM_USER0);
2287                         if (block_start < from)
2288                                 memset(kaddr+block_start, 0, from-block_start);
2289                         if (block_end > to)
2290                                 memset(kaddr + to, 0, block_end - to);
2291                         flush_dcache_page(page);
2292                         kunmap_atomic(kaddr, KM_USER0);
2293                         continue;
2294                 }
2295                 if (buffer_uptodate(&map_bh))
2296                         continue;       /* reiserfs does this */
2297                 if (block_start < from || block_end > to) {
2298                         struct buffer_head *bh = alloc_buffer_head(GFP_NOFS);
2299
2300                         if (!bh) {
2301                                 ret = -ENOMEM;
2302                                 goto failed;
2303                         }
2304                         bh->b_state = map_bh.b_state;
2305                         atomic_set(&bh->b_count, 0);
2306                         bh->b_this_page = NULL;
2307                         bh->b_page = page;
2308                         bh->b_blocknr = map_bh.b_blocknr;
2309                         bh->b_size = blocksize;
2310                         bh->b_data = (char *)(long)block_start;
2311                         bh->b_bdev = map_bh.b_bdev;
2312                         bh->b_private = NULL;
2313                         read_bh[nr_reads++] = bh;
2314                 }
2315         }
2316
2317         if (nr_reads) {
2318                 struct buffer_head *bh;
2319
2320                 /*
2321                  * The page is locked, so these buffers are protected from
2322                  * any VM or truncate activity.  Hence we don't need to care
2323                  * for the buffer_head refcounts.
2324                  */
2325                 for (i = 0; i < nr_reads; i++) {
2326                         bh = read_bh[i];
2327                         lock_buffer(bh);
2328                         bh->b_end_io = end_buffer_read_nobh;
2329                         submit_bh(READ, bh);
2330                 }
2331                 for (i = 0; i < nr_reads; i++) {
2332                         bh = read_bh[i];
2333                         wait_on_buffer(bh);
2334                         if (!buffer_uptodate(bh))
2335                                 ret = -EIO;
2336                         free_buffer_head(bh);
2337                         read_bh[i] = NULL;
2338                 }
2339                 if (ret)
2340                         goto failed;
2341         }
2342
2343         if (is_mapped_to_disk)
2344                 SetPageMappedToDisk(page);
2345
2346         return 0;
2347
2348 failed:
2349         for (i = 0; i < nr_reads; i++) {
2350                 if (read_bh[i])
2351                         free_buffer_head(read_bh[i]);
2352         }
2353
2354         /*
2355          * Error recovery is pretty slack.  Clear the page and mark it dirty
2356          * so we'll later zero out any blocks which _were_ allocated.
2357          */
2358         kaddr = kmap_atomic(page, KM_USER0);
2359         memset(kaddr, 0, PAGE_CACHE_SIZE);
2360         flush_dcache_page(page);
2361         kunmap_atomic(kaddr, KM_USER0);
2362         SetPageUptodate(page);
2363         set_page_dirty(page);
2364         return ret;
2365 }
2366 EXPORT_SYMBOL(nobh_prepare_write);
2367
2368 /*
2369  * Make sure any changes to nobh_commit_write() are reflected in
2370  * nobh_truncate_page(), since it doesn't call commit_write().
2371  */
2372 int nobh_commit_write(struct file *file, struct page *page,
2373                 unsigned from, unsigned to)
2374 {
2375         struct inode *inode = page->mapping->host;
2376         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2377
2378         SetPageUptodate(page);
2379         set_page_dirty(page);
2380         if (pos > inode->i_size) {
2381                 i_size_write(inode, pos);
2382                 mark_inode_dirty(inode);
2383         }
2384         return 0;
2385 }
2386 EXPORT_SYMBOL(nobh_commit_write);
2387
2388 /*
2389  * nobh_writepage() - based on block_full_write_page() except
2390  * that it tries to operate without attaching bufferheads to
2391  * the page.
2392  */
2393 int nobh_writepage(struct page *page, get_block_t *get_block,
2394                         struct writeback_control *wbc)
2395 {
2396         struct inode * const inode = page->mapping->host;
2397         loff_t i_size = i_size_read(inode);
2398         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2399         unsigned offset;
2400         void *kaddr;
2401         int ret;
2402
2403         /* Is the page fully inside i_size? */
2404         if (page->index < end_index)
2405                 goto out;
2406
2407         /* Is the page fully outside i_size? (truncate in progress) */
2408         offset = i_size & (PAGE_CACHE_SIZE-1);
2409         if (page->index >= end_index+1 || !offset) {
2410                 /*
2411                  * The page may have dirty, unmapped buffers.  For example,
2412                  * they may have been added in ext3_writepage().  Make them
2413                  * freeable here, so the page does not leak.
2414                  */
2415 #if 0
2416                 /* Not really sure about this  - do we need this ? */
2417                 if (page->mapping->a_ops->invalidatepage)
2418                         page->mapping->a_ops->invalidatepage(page, offset);
2419 #endif
2420                 unlock_page(page);
2421                 return 0; /* don't care */
2422         }
2423
2424         /*
2425          * The page straddles i_size.  It must be zeroed out on each and every
2426          * writepage invocation because it may be mmapped.  "A file is mapped
2427          * in multiples of the page size.  For a file that is not a multiple of
2428          * the  page size, the remaining memory is zeroed when mapped, and
2429          * writes to that region are not written out to the file."
2430          */
2431         kaddr = kmap_atomic(page, KM_USER0);
2432         memset(kaddr + offset, 0, PAGE_CACHE_SIZE - offset);
2433         flush_dcache_page(page);
2434         kunmap_atomic(kaddr, KM_USER0);
2435 out:
2436         ret = mpage_writepage(page, get_block, wbc);
2437         if (ret == -EAGAIN)
2438                 ret = __block_write_full_page(inode, page, get_block, wbc);
2439         return ret;
2440 }
2441 EXPORT_SYMBOL(nobh_writepage);
2442
2443 /*
2444  * This function assumes that ->prepare_write() uses nobh_prepare_write().
2445  */
2446 int nobh_truncate_page(struct address_space *mapping, loff_t from)
2447 {
2448         struct inode *inode = mapping->host;
2449         unsigned blocksize = 1 << inode->i_blkbits;
2450         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2451         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2452         unsigned to;
2453         struct page *page;
2454         const struct address_space_operations *a_ops = mapping->a_ops;
2455         char *kaddr;
2456         int ret = 0;
2457
2458         if ((offset & (blocksize - 1)) == 0)
2459                 goto out;
2460
2461         ret = -ENOMEM;
2462         page = grab_cache_page(mapping, index);
2463         if (!page)
2464                 goto out;
2465
2466         to = (offset + blocksize) & ~(blocksize - 1);
2467         ret = a_ops->prepare_write(NULL, page, offset, to);
2468         if (ret == 0) {
2469                 kaddr = kmap_atomic(page, KM_USER0);
2470                 memset(kaddr + offset, 0, PAGE_CACHE_SIZE - offset);
2471                 flush_dcache_page(page);
2472                 kunmap_atomic(kaddr, KM_USER0);
2473                 /*
2474                  * It would be more correct to call aops->commit_write()
2475                  * here, but this is more efficient.
2476                  */
2477                 SetPageUptodate(page);
2478                 set_page_dirty(page);
2479         }
2480         unlock_page(page);
2481         page_cache_release(page);
2482 out:
2483         return ret;
2484 }
2485 EXPORT_SYMBOL(nobh_truncate_page);
2486
2487 int block_truncate_page(struct address_space *mapping,
2488                         loff_t from, get_block_t *get_block)
2489 {
2490         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2491         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2492         unsigned blocksize;
2493         sector_t iblock;
2494         unsigned length, pos;
2495         struct inode *inode = mapping->host;
2496         struct page *page;
2497         struct buffer_head *bh;
2498         void *kaddr;
2499         int err;
2500
2501         blocksize = 1 << inode->i_blkbits;
2502         length = offset & (blocksize - 1);
2503
2504         /* Block boundary? Nothing to do */
2505         if (!length)
2506                 return 0;
2507
2508         length = blocksize - length;
2509         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2510         
2511         page = grab_cache_page(mapping, index);
2512         err = -ENOMEM;
2513         if (!page)
2514                 goto out;
2515
2516         if (!page_has_buffers(page))
2517                 create_empty_buffers(page, blocksize, 0);
2518
2519         /* Find the buffer that contains "offset" */
2520         bh = page_buffers(page);
2521         pos = blocksize;
2522         while (offset >= pos) {
2523                 bh = bh->b_this_page;
2524                 iblock++;
2525                 pos += blocksize;
2526         }
2527
2528         err = 0;
2529         if (!buffer_mapped(bh)) {
2530                 WARN_ON(bh->b_size != blocksize);
2531                 err = get_block(inode, iblock, bh, 0);
2532                 if (err)
2533                         goto unlock;
2534                 /* unmapped? It's a hole - nothing to do */
2535                 if (!buffer_mapped(bh))
2536                         goto unlock;
2537         }
2538
2539         /* Ok, it's mapped. Make sure it's up-to-date */
2540         if (PageUptodate(page))
2541                 set_buffer_uptodate(bh);
2542
2543         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2544                 err = -EIO;
2545                 ll_rw_block(READ, 1, &bh);
2546                 wait_on_buffer(bh);
2547                 /* Uhhuh. Read error. Complain and punt. */
2548                 if (!buffer_uptodate(bh))
2549                         goto unlock;
2550         }
2551
2552         kaddr = kmap_atomic(page, KM_USER0);
2553         memset(kaddr + offset, 0, length);
2554         flush_dcache_page(page);
2555         kunmap_atomic(kaddr, KM_USER0);
2556
2557         mark_buffer_dirty(bh);
2558         err = 0;
2559
2560 unlock:
2561         unlock_page(page);
2562         page_cache_release(page);
2563 out:
2564         return err;
2565 }
2566
2567 /*
2568  * The generic ->writepage function for buffer-backed address_spaces
2569  */
2570 int block_write_full_page(struct page *page, get_block_t *get_block,
2571                         struct writeback_control *wbc)
2572 {
2573         struct inode * const inode = page->mapping->host;
2574         loff_t i_size = i_size_read(inode);
2575         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2576         unsigned offset;
2577         void *kaddr;
2578
2579         /* Is the page fully inside i_size? */
2580         if (page->index < end_index)
2581                 return __block_write_full_page(inode, page, get_block, wbc);
2582
2583         /* Is the page fully outside i_size? (truncate in progress) */
2584         offset = i_size & (PAGE_CACHE_SIZE-1);
2585         if (page->index >= end_index+1 || !offset) {
2586                 /*
2587                  * The page may have dirty, unmapped buffers.  For example,
2588                  * they may have been added in ext3_writepage().  Make them
2589                  * freeable here, so the page does not leak.
2590                  */
2591                 do_invalidatepage(page, 0);
2592                 unlock_page(page);
2593                 return 0; /* don't care */
2594         }
2595
2596         /*
2597          * The page straddles i_size.  It must be zeroed out on each and every
2598          * writepage invokation because it may be mmapped.  "A file is mapped
2599          * in multiples of the page size.  For a file that is not a multiple of
2600          * the  page size, the remaining memory is zeroed when mapped, and
2601          * writes to that region are not written out to the file."
2602          */
2603         kaddr = kmap_atomic(page, KM_USER0);
2604         memset(kaddr + offset, 0, PAGE_CACHE_SIZE - offset);
2605         flush_dcache_page(page);
2606         kunmap_atomic(kaddr, KM_USER0);
2607         return __block_write_full_page(inode, page, get_block, wbc);
2608 }
2609
2610 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2611                             get_block_t *get_block)
2612 {
2613         struct buffer_head tmp;
2614         struct inode *inode = mapping->host;
2615         tmp.b_state = 0;
2616         tmp.b_blocknr = 0;
2617         tmp.b_size = 1 << inode->i_blkbits;
2618         get_block(inode, block, &tmp, 0);
2619         return tmp.b_blocknr;
2620 }
2621
2622 static int end_bio_bh_io_sync(struct bio *bio, unsigned int bytes_done, int err)
2623 {
2624         struct buffer_head *bh = bio->bi_private;
2625
2626         if (bio->bi_size)
2627                 return 1;
2628
2629         if (err == -EOPNOTSUPP) {
2630                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2631                 set_bit(BH_Eopnotsupp, &bh->b_state);
2632         }
2633
2634         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2635         bio_put(bio);
2636         return 0;
2637 }
2638
2639 int submit_bh(int rw, struct buffer_head * bh)
2640 {
2641         struct bio *bio;
2642         int ret = 0;
2643
2644         BUG_ON(!buffer_locked(bh));
2645         BUG_ON(!buffer_mapped(bh));
2646         BUG_ON(!bh->b_end_io);
2647
2648         if (buffer_ordered(bh) && (rw == WRITE))
2649                 rw = WRITE_BARRIER;
2650
2651         /*
2652          * Only clear out a write error when rewriting, should this
2653          * include WRITE_SYNC as well?
2654          */
2655         if (test_set_buffer_req(bh) && (rw == WRITE || rw == WRITE_BARRIER))
2656                 clear_buffer_write_io_error(bh);
2657
2658         /*
2659          * from here on down, it's all bio -- do the initial mapping,
2660          * submit_bio -> generic_make_request may further map this bio around
2661          */
2662         bio = bio_alloc(GFP_NOIO, 1);
2663
2664         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2665         bio->bi_bdev = bh->b_bdev;
2666         bio->bi_io_vec[0].bv_page = bh->b_page;
2667         bio->bi_io_vec[0].bv_len = bh->b_size;
2668         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2669
2670         bio->bi_vcnt = 1;
2671         bio->bi_idx = 0;
2672         bio->bi_size = bh->b_size;
2673
2674         bio->bi_end_io = end_bio_bh_io_sync;
2675         bio->bi_private = bh;
2676
2677         bio_get(bio);
2678         submit_bio(rw, bio);
2679
2680         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2681                 ret = -EOPNOTSUPP;
2682
2683         bio_put(bio);
2684         return ret;
2685 }
2686
2687 /**
2688  * ll_rw_block: low-level access to block devices (DEPRECATED)
2689  * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
2690  * @nr: number of &struct buffer_heads in the array
2691  * @bhs: array of pointers to &struct buffer_head
2692  *
2693  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2694  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2695  * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2696  * are sent to disk. The fourth %READA option is described in the documentation
2697  * for generic_make_request() which ll_rw_block() calls.
2698  *
2699  * This function drops any buffer that it cannot get a lock on (with the
2700  * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
2701  * clean when doing a write request, and any buffer that appears to be
2702  * up-to-date when doing read request.  Further it marks as clean buffers that
2703  * are processed for writing (the buffer cache won't assume that they are
2704  * actually clean until the buffer gets unlocked).
2705  *
2706  * ll_rw_block sets b_end_io to simple completion handler that marks
2707  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2708  * any waiters. 
2709  *
2710  * All of the buffers must be for the same device, and must also be a
2711  * multiple of the current approved size for the device.
2712  */
2713 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2714 {
2715         int i;
2716
2717         for (i = 0; i < nr; i++) {
2718                 struct buffer_head *bh = bhs[i];
2719
2720                 if (rw == SWRITE)
2721                         lock_buffer(bh);
2722                 else if (test_set_buffer_locked(bh))
2723                         continue;
2724
2725                 if (rw == WRITE || rw == SWRITE) {
2726                         if (test_clear_buffer_dirty(bh)) {
2727                                 bh->b_end_io = end_buffer_write_sync;
2728                                 get_bh(bh);
2729                                 submit_bh(WRITE, bh);
2730                                 continue;
2731                         }
2732                 } else {
2733                         if (!buffer_uptodate(bh)) {
2734                                 bh->b_end_io = end_buffer_read_sync;
2735                                 get_bh(bh);
2736                                 submit_bh(rw, bh);
2737                                 continue;
2738                         }
2739                 }
2740                 unlock_buffer(bh);
2741         }
2742 }
2743
2744 /*
2745  * For a data-integrity writeout, we need to wait upon any in-progress I/O
2746  * and then start new I/O and then wait upon it.  The caller must have a ref on
2747  * the buffer_head.
2748  */
2749 int sync_dirty_buffer(struct buffer_head *bh)
2750 {
2751         int ret = 0;
2752
2753         WARN_ON(atomic_read(&bh->b_count) < 1);
2754         lock_buffer(bh);
2755         if (test_clear_buffer_dirty(bh)) {
2756                 get_bh(bh);
2757                 bh->b_end_io = end_buffer_write_sync;
2758                 ret = submit_bh(WRITE, bh);
2759                 wait_on_buffer(bh);
2760                 if (buffer_eopnotsupp(bh)) {
2761                         clear_buffer_eopnotsupp(bh);
2762                         ret = -EOPNOTSUPP;
2763                 }
2764                 if (!ret && !buffer_uptodate(bh))
2765                         ret = -EIO;
2766         } else {
2767                 unlock_buffer(bh);
2768         }
2769         return ret;
2770 }
2771
2772 /*
2773  * try_to_free_buffers() checks if all the buffers on this particular page
2774  * are unused, and releases them if so.
2775  *
2776  * Exclusion against try_to_free_buffers may be obtained by either
2777  * locking the page or by holding its mapping's private_lock.
2778  *
2779  * If the page is dirty but all the buffers are clean then we need to
2780  * be sure to mark the page clean as well.  This is because the page
2781  * may be against a block device, and a later reattachment of buffers
2782  * to a dirty page will set *all* buffers dirty.  Which would corrupt
2783  * filesystem data on the same device.
2784  *
2785  * The same applies to regular filesystem pages: if all the buffers are
2786  * clean then we set the page clean and proceed.  To do that, we require
2787  * total exclusion from __set_page_dirty_buffers().  That is obtained with
2788  * private_lock.
2789  *
2790  * try_to_free_buffers() is non-blocking.
2791  */
2792 static inline int buffer_busy(struct buffer_head *bh)
2793 {
2794         return atomic_read(&bh->b_count) |
2795                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
2796 }
2797
2798 static int
2799 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
2800 {
2801         struct buffer_head *head = page_buffers(page);
2802         struct buffer_head *bh;
2803
2804         bh = head;
2805         do {
2806                 if (buffer_write_io_error(bh) && page->mapping)
2807                         set_bit(AS_EIO, &page->mapping->flags);
2808                 if (buffer_busy(bh))
2809                         goto failed;
2810                 bh = bh->b_this_page;
2811         } while (bh != head);
2812
2813         do {
2814                 struct buffer_head *next = bh->b_this_page;
2815
2816                 if (!list_empty(&bh->b_assoc_buffers))
2817                         __remove_assoc_queue(bh);
2818                 bh = next;
2819         } while (bh != head);
2820         *buffers_to_free = head;
2821         __clear_page_buffers(page);
2822         return 1;
2823 failed:
2824         return 0;
2825 }
2826
2827 int try_to_free_buffers(struct page *page)
2828 {
2829         struct address_space * const mapping = page->mapping;
2830         struct buffer_head *buffers_to_free = NULL;
2831         int ret = 0;
2832
2833         BUG_ON(!PageLocked(page));
2834         if (PageWriteback(page))
2835                 return 0;
2836
2837         if (mapping == NULL) {          /* can this still happen? */
2838                 ret = drop_buffers(page, &buffers_to_free);
2839                 goto out;
2840         }
2841
2842         spin_lock(&mapping->private_lock);
2843         ret = drop_buffers(page, &buffers_to_free);
2844
2845         /*
2846          * If the filesystem writes its buffers by hand (eg ext3)
2847          * then we can have clean buffers against a dirty page.  We
2848          * clean the page here; otherwise the VM will never notice
2849          * that the filesystem did any IO at all.
2850          *
2851          * Also, during truncate, discard_buffer will have marked all
2852          * the page's buffers clean.  We discover that here and clean
2853          * the page also.
2854          *
2855          * private_lock must be held over this entire operation in order
2856          * to synchronise against __set_page_dirty_buffers and prevent the
2857          * dirty bit from being lost.
2858          */
2859         if (ret)
2860                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
2861         spin_unlock(&mapping->private_lock);
2862 out:
2863         if (buffers_to_free) {
2864                 struct buffer_head *bh = buffers_to_free;
2865
2866                 do {
2867                         struct buffer_head *next = bh->b_this_page;
2868                         free_buffer_head(bh);
2869                         bh = next;
2870                 } while (bh != buffers_to_free);
2871         }
2872         return ret;
2873 }
2874 EXPORT_SYMBOL(try_to_free_buffers);
2875
2876 void block_sync_page(struct page *page)
2877 {
2878         struct address_space *mapping;
2879
2880         smp_mb();
2881         mapping = page_mapping(page);
2882         if (mapping)
2883                 blk_run_backing_dev(mapping->backing_dev_info, page);
2884 }
2885
2886 /*
2887  * There are no bdflush tunables left.  But distributions are
2888  * still running obsolete flush daemons, so we terminate them here.
2889  *
2890  * Use of bdflush() is deprecated and will be removed in a future kernel.
2891  * The `pdflush' kernel threads fully replace bdflush daemons and this call.
2892  */
2893 asmlinkage long sys_bdflush(int func, long data)
2894 {
2895         static int msg_count;
2896
2897         if (!capable(CAP_SYS_ADMIN))
2898                 return -EPERM;
2899
2900         if (msg_count < 5) {
2901                 msg_count++;
2902                 printk(KERN_INFO
2903                         "warning: process `%s' used the obsolete bdflush"
2904                         " system call\n", current->comm);
2905                 printk(KERN_INFO "Fix your initscripts?\n");
2906         }
2907
2908         if (func == 1)
2909                 do_exit(0);
2910         return 0;
2911 }
2912
2913 /*
2914  * Buffer-head allocation
2915  */
2916 static struct kmem_cache *bh_cachep;
2917
2918 /*
2919  * Once the number of bh's in the machine exceeds this level, we start
2920  * stripping them in writeback.
2921  */
2922 static int max_buffer_heads;
2923
2924 int buffer_heads_over_limit;
2925
2926 struct bh_accounting {
2927         int nr;                 /* Number of live bh's */
2928         int ratelimit;          /* Limit cacheline bouncing */
2929 };
2930
2931 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
2932
2933 static void recalc_bh_state(void)
2934 {
2935         int i;
2936         int tot = 0;
2937
2938         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
2939                 return;
2940         __get_cpu_var(bh_accounting).ratelimit = 0;
2941         for_each_online_cpu(i)
2942                 tot += per_cpu(bh_accounting, i).nr;
2943         buffer_heads_over_limit = (tot > max_buffer_heads);
2944 }
2945         
2946 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
2947 {
2948         struct buffer_head *ret = kmem_cache_alloc(bh_cachep, gfp_flags);
2949         if (ret) {
2950                 get_cpu_var(bh_accounting).nr++;
2951                 recalc_bh_state();
2952                 put_cpu_var(bh_accounting);
2953         }
2954         return ret;
2955 }
2956 EXPORT_SYMBOL(alloc_buffer_head);
2957
2958 void free_buffer_head(struct buffer_head *bh)
2959 {
2960         BUG_ON(!list_empty(&bh->b_assoc_buffers));
2961         kmem_cache_free(bh_cachep, bh);
2962         get_cpu_var(bh_accounting).nr--;
2963         recalc_bh_state();
2964         put_cpu_var(bh_accounting);
2965 }
2966 EXPORT_SYMBOL(free_buffer_head);
2967
2968 static void
2969 init_buffer_head(void *data, struct kmem_cache *cachep, unsigned long flags)
2970 {
2971         if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) ==
2972                             SLAB_CTOR_CONSTRUCTOR) {
2973                 struct buffer_head * bh = (struct buffer_head *)data;
2974
2975                 memset(bh, 0, sizeof(*bh));
2976                 INIT_LIST_HEAD(&bh->b_assoc_buffers);
2977         }
2978 }
2979
2980 static void buffer_exit_cpu(int cpu)
2981 {
2982         int i;
2983         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
2984
2985         for (i = 0; i < BH_LRU_SIZE; i++) {
2986                 brelse(b->bhs[i]);
2987                 b->bhs[i] = NULL;
2988         }
2989         get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
2990         per_cpu(bh_accounting, cpu).nr = 0;
2991         put_cpu_var(bh_accounting);
2992 }
2993
2994 static int buffer_cpu_notify(struct notifier_block *self,
2995                               unsigned long action, void *hcpu)
2996 {
2997         if (action == CPU_DEAD)
2998                 buffer_exit_cpu((unsigned long)hcpu);
2999         return NOTIFY_OK;
3000 }
3001
3002 void __init buffer_init(void)
3003 {
3004         int nrpages;
3005
3006         bh_cachep = kmem_cache_create("buffer_head",
3007                                         sizeof(struct buffer_head), 0,
3008                                         (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3009                                         SLAB_MEM_SPREAD),
3010                                         init_buffer_head,
3011                                         NULL);
3012
3013         /*
3014          * Limit the bh occupancy to 10% of ZONE_NORMAL
3015          */
3016         nrpages = (nr_free_buffer_pages() * 10) / 100;
3017         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3018         hotcpu_notifier(buffer_cpu_notify, 0);
3019 }
3020
3021 EXPORT_SYMBOL(__bforget);
3022 EXPORT_SYMBOL(__brelse);
3023 EXPORT_SYMBOL(__wait_on_buffer);
3024 EXPORT_SYMBOL(block_commit_write);
3025 EXPORT_SYMBOL(block_prepare_write);
3026 EXPORT_SYMBOL(block_read_full_page);
3027 EXPORT_SYMBOL(block_sync_page);
3028 EXPORT_SYMBOL(block_truncate_page);
3029 EXPORT_SYMBOL(block_write_full_page);
3030 EXPORT_SYMBOL(cont_prepare_write);
3031 EXPORT_SYMBOL(end_buffer_read_sync);
3032 EXPORT_SYMBOL(end_buffer_write_sync);
3033 EXPORT_SYMBOL(file_fsync);
3034 EXPORT_SYMBOL(fsync_bdev);
3035 EXPORT_SYMBOL(generic_block_bmap);
3036 EXPORT_SYMBOL(generic_commit_write);
3037 EXPORT_SYMBOL(generic_cont_expand);
3038 EXPORT_SYMBOL(generic_cont_expand_simple);
3039 EXPORT_SYMBOL(init_buffer);
3040 EXPORT_SYMBOL(invalidate_bdev);
3041 EXPORT_SYMBOL(ll_rw_block);
3042 EXPORT_SYMBOL(mark_buffer_dirty);
3043 EXPORT_SYMBOL(submit_bh);
3044 EXPORT_SYMBOL(sync_dirty_buffer);
3045 EXPORT_SYMBOL(unlock_buffer);