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