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