vfs: pagecache usage optimization for pagesize!=blocksize
[linux-2.6.git] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/capability.h>
28 #include <linux/blkdev.h>
29 #include <linux/file.h>
30 #include <linux/quotaops.h>
31 #include <linux/highmem.h>
32 #include <linux/module.h>
33 #include <linux/writeback.h>
34 #include <linux/hash.h>
35 #include <linux/suspend.h>
36 #include <linux/buffer_head.h>
37 #include <linux/task_io_accounting_ops.h>
38 #include <linux/bio.h>
39 #include <linux/notifier.h>
40 #include <linux/cpu.h>
41 #include <linux/bitops.h>
42 #include <linux/mpage.h>
43 #include <linux/bit_spinlock.h>
44
45 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
46
47 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
48
49 inline void
50 init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
51 {
52         bh->b_end_io = handler;
53         bh->b_private = private;
54 }
55
56 static int sync_buffer(void *word)
57 {
58         struct block_device *bd;
59         struct buffer_head *bh
60                 = container_of(word, struct buffer_head, b_state);
61
62         smp_mb();
63         bd = bh->b_bdev;
64         if (bd)
65                 blk_run_address_space(bd->bd_inode->i_mapping);
66         io_schedule();
67         return 0;
68 }
69
70 void __lock_buffer(struct buffer_head *bh)
71 {
72         wait_on_bit_lock(&bh->b_state, BH_Lock, sync_buffer,
73                                                         TASK_UNINTERRUPTIBLE);
74 }
75 EXPORT_SYMBOL(__lock_buffer);
76
77 void unlock_buffer(struct buffer_head *bh)
78 {
79         smp_mb__before_clear_bit();
80         clear_buffer_locked(bh);
81         smp_mb__after_clear_bit();
82         wake_up_bit(&bh->b_state, BH_Lock);
83 }
84
85 /*
86  * Block until a buffer comes unlocked.  This doesn't stop it
87  * from becoming locked again - you have to lock it yourself
88  * if you want to preserve its state.
89  */
90 void __wait_on_buffer(struct buffer_head * bh)
91 {
92         wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
93 }
94
95 static void
96 __clear_page_buffers(struct page *page)
97 {
98         ClearPagePrivate(page);
99         set_page_private(page, 0);
100         page_cache_release(page);
101 }
102
103 static void buffer_io_error(struct buffer_head *bh)
104 {
105         char b[BDEVNAME_SIZE];
106
107         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
108                         bdevname(bh->b_bdev, b),
109                         (unsigned long long)bh->b_blocknr);
110 }
111
112 /*
113  * End-of-IO handler helper function which does not touch the bh after
114  * unlocking it.
115  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
116  * a race there is benign: unlock_buffer() only use the bh's address for
117  * hashing after unlocking the buffer, so it doesn't actually touch the bh
118  * itself.
119  */
120 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
121 {
122         if (uptodate) {
123                 set_buffer_uptodate(bh);
124         } else {
125                 /* This happens, due to failed READA attempts. */
126                 clear_buffer_uptodate(bh);
127         }
128         unlock_buffer(bh);
129 }
130
131 /*
132  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
133  * unlock the buffer. This is what ll_rw_block uses too.
134  */
135 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
136 {
137         __end_buffer_read_notouch(bh, uptodate);
138         put_bh(bh);
139 }
140
141 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
142 {
143         char b[BDEVNAME_SIZE];
144
145         if (uptodate) {
146                 set_buffer_uptodate(bh);
147         } else {
148                 if (!buffer_eopnotsupp(bh) && printk_ratelimit()) {
149                         buffer_io_error(bh);
150                         printk(KERN_WARNING "lost page write due to "
151                                         "I/O error on %s\n",
152                                        bdevname(bh->b_bdev, b));
153                 }
154                 set_buffer_write_io_error(bh);
155                 clear_buffer_uptodate(bh);
156         }
157         unlock_buffer(bh);
158         put_bh(bh);
159 }
160
161 /*
162  * Write out and wait upon all the dirty data associated with a block
163  * device via its mapping.  Does not take the superblock lock.
164  */
165 int sync_blockdev(struct block_device *bdev)
166 {
167         int ret = 0;
168
169         if (bdev)
170                 ret = filemap_write_and_wait(bdev->bd_inode->i_mapping);
171         return ret;
172 }
173 EXPORT_SYMBOL(sync_blockdev);
174
175 /*
176  * Write out and wait upon all dirty data associated with this
177  * device.   Filesystem data as well as the underlying block
178  * device.  Takes the superblock lock.
179  */
180 int fsync_bdev(struct block_device *bdev)
181 {
182         struct super_block *sb = get_super(bdev);
183         if (sb) {
184                 int res = fsync_super(sb);
185                 drop_super(sb);
186                 return res;
187         }
188         return sync_blockdev(bdev);
189 }
190
191 /**
192  * freeze_bdev  --  lock a filesystem and force it into a consistent state
193  * @bdev:       blockdevice to lock
194  *
195  * This takes the block device bd_mount_sem to make sure no new mounts
196  * happen on bdev until thaw_bdev() is called.
197  * If a superblock is found on this device, we take the s_umount semaphore
198  * on it to make sure nobody unmounts until the snapshot creation is done.
199  */
200 struct super_block *freeze_bdev(struct block_device *bdev)
201 {
202         struct super_block *sb;
203
204         down(&bdev->bd_mount_sem);
205         sb = get_super(bdev);
206         if (sb && !(sb->s_flags & MS_RDONLY)) {
207                 sb->s_frozen = SB_FREEZE_WRITE;
208                 smp_wmb();
209
210                 __fsync_super(sb);
211
212                 sb->s_frozen = SB_FREEZE_TRANS;
213                 smp_wmb();
214
215                 sync_blockdev(sb->s_bdev);
216
217                 if (sb->s_op->write_super_lockfs)
218                         sb->s_op->write_super_lockfs(sb);
219         }
220
221         sync_blockdev(bdev);
222         return sb;      /* thaw_bdev releases s->s_umount and bd_mount_sem */
223 }
224 EXPORT_SYMBOL(freeze_bdev);
225
226 /**
227  * thaw_bdev  -- unlock filesystem
228  * @bdev:       blockdevice to unlock
229  * @sb:         associated superblock
230  *
231  * Unlocks the filesystem and marks it writeable again after freeze_bdev().
232  */
233 void thaw_bdev(struct block_device *bdev, struct super_block *sb)
234 {
235         if (sb) {
236                 BUG_ON(sb->s_bdev != bdev);
237
238                 if (sb->s_op->unlockfs)
239                         sb->s_op->unlockfs(sb);
240                 sb->s_frozen = SB_UNFROZEN;
241                 smp_wmb();
242                 wake_up(&sb->s_wait_unfrozen);
243                 drop_super(sb);
244         }
245
246         up(&bdev->bd_mount_sem);
247 }
248 EXPORT_SYMBOL(thaw_bdev);
249
250 /*
251  * Various filesystems appear to want __find_get_block to be non-blocking.
252  * But it's the page lock which protects the buffers.  To get around this,
253  * we get exclusion from try_to_free_buffers with the blockdev mapping's
254  * private_lock.
255  *
256  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
257  * may be quite high.  This code could TryLock the page, and if that
258  * succeeds, there is no need to take private_lock. (But if
259  * private_lock is contended then so is mapping->tree_lock).
260  */
261 static struct buffer_head *
262 __find_get_block_slow(struct block_device *bdev, sector_t block)
263 {
264         struct inode *bd_inode = bdev->bd_inode;
265         struct address_space *bd_mapping = bd_inode->i_mapping;
266         struct buffer_head *ret = NULL;
267         pgoff_t index;
268         struct buffer_head *bh;
269         struct buffer_head *head;
270         struct page *page;
271         int all_mapped = 1;
272
273         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
274         page = find_get_page(bd_mapping, index);
275         if (!page)
276                 goto out;
277
278         spin_lock(&bd_mapping->private_lock);
279         if (!page_has_buffers(page))
280                 goto out_unlock;
281         head = page_buffers(page);
282         bh = head;
283         do {
284                 if (bh->b_blocknr == block) {
285                         ret = bh;
286                         get_bh(bh);
287                         goto out_unlock;
288                 }
289                 if (!buffer_mapped(bh))
290                         all_mapped = 0;
291                 bh = bh->b_this_page;
292         } while (bh != head);
293
294         /* we might be here because some of the buffers on this page are
295          * not mapped.  This is due to various races between
296          * file io on the block device and getblk.  It gets dealt with
297          * elsewhere, don't buffer_error if we had some unmapped buffers
298          */
299         if (all_mapped) {
300                 printk("__find_get_block_slow() failed. "
301                         "block=%llu, b_blocknr=%llu\n",
302                         (unsigned long long)block,
303                         (unsigned long long)bh->b_blocknr);
304                 printk("b_state=0x%08lx, b_size=%zu\n",
305                         bh->b_state, bh->b_size);
306                 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
307         }
308 out_unlock:
309         spin_unlock(&bd_mapping->private_lock);
310         page_cache_release(page);
311 out:
312         return ret;
313 }
314
315 /* If invalidate_buffers() will trash dirty buffers, it means some kind
316    of fs corruption is going on. Trashing dirty data always imply losing
317    information that was supposed to be just stored on the physical layer
318    by the user.
319
320    Thus invalidate_buffers in general usage is not allwowed to trash
321    dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
322    be preserved.  These buffers are simply skipped.
323   
324    We also skip buffers which are still in use.  For example this can
325    happen if a userspace program is reading the block device.
326
327    NOTE: In the case where the user removed a removable-media-disk even if
328    there's still dirty data not synced on disk (due a bug in the device driver
329    or due an error of the user), by not destroying the dirty buffers we could
330    generate corruption also on the next media inserted, thus a parameter is
331    necessary to handle this case in the most safe way possible (trying
332    to not corrupt also the new disk inserted with the data belonging to
333    the old now corrupted disk). Also for the ramdisk the natural thing
334    to do in order to release the ramdisk memory is to destroy dirty buffers.
335
336    These are two special cases. Normal usage imply the device driver
337    to issue a sync on the device (without waiting I/O completion) and
338    then an invalidate_buffers call that doesn't trash dirty buffers.
339
340    For handling cache coherency with the blkdev pagecache the 'update' case
341    is been introduced. It is needed to re-read from disk any pinned
342    buffer. NOTE: re-reading from disk is destructive so we can do it only
343    when we assume nobody is changing the buffercache under our I/O and when
344    we think the disk contains more recent information than the buffercache.
345    The update == 1 pass marks the buffers we need to update, the update == 2
346    pass does the actual I/O. */
347 void invalidate_bdev(struct block_device *bdev)
348 {
349         struct address_space *mapping = bdev->bd_inode->i_mapping;
350
351         if (mapping->nrpages == 0)
352                 return;
353
354         invalidate_bh_lrus();
355         invalidate_mapping_pages(mapping, 0, -1);
356 }
357
358 /*
359  * Kick pdflush then try to free up some ZONE_NORMAL memory.
360  */
361 static void free_more_memory(void)
362 {
363         struct zone *zone;
364         int nid;
365
366         wakeup_pdflush(1024);
367         yield();
368
369         for_each_online_node(nid) {
370                 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
371                                                 gfp_zone(GFP_NOFS), NULL,
372                                                 &zone);
373                 if (zone)
374                         try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
375                                                 GFP_NOFS);
376         }
377 }
378
379 /*
380  * I/O completion handler for block_read_full_page() - pages
381  * which come unlocked at the end of I/O.
382  */
383 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
384 {
385         unsigned long flags;
386         struct buffer_head *first;
387         struct buffer_head *tmp;
388         struct page *page;
389         int page_uptodate = 1;
390
391         BUG_ON(!buffer_async_read(bh));
392
393         page = bh->b_page;
394         if (uptodate) {
395                 set_buffer_uptodate(bh);
396         } else {
397                 clear_buffer_uptodate(bh);
398                 if (printk_ratelimit())
399                         buffer_io_error(bh);
400                 SetPageError(page);
401         }
402
403         /*
404          * Be _very_ careful from here on. Bad things can happen if
405          * two buffer heads end IO at almost the same time and both
406          * decide that the page is now completely done.
407          */
408         first = page_buffers(page);
409         local_irq_save(flags);
410         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
411         clear_buffer_async_read(bh);
412         unlock_buffer(bh);
413         tmp = bh;
414         do {
415                 if (!buffer_uptodate(tmp))
416                         page_uptodate = 0;
417                 if (buffer_async_read(tmp)) {
418                         BUG_ON(!buffer_locked(tmp));
419                         goto still_busy;
420                 }
421                 tmp = tmp->b_this_page;
422         } while (tmp != bh);
423         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
424         local_irq_restore(flags);
425
426         /*
427          * If none of the buffers had errors and they are all
428          * uptodate then we can set the page uptodate.
429          */
430         if (page_uptodate && !PageError(page))
431                 SetPageUptodate(page);
432         unlock_page(page);
433         return;
434
435 still_busy:
436         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
437         local_irq_restore(flags);
438         return;
439 }
440
441 /*
442  * Completion handler for block_write_full_page() - pages which are unlocked
443  * during I/O, and which have PageWriteback cleared upon I/O completion.
444  */
445 static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
446 {
447         char b[BDEVNAME_SIZE];
448         unsigned long flags;
449         struct buffer_head *first;
450         struct buffer_head *tmp;
451         struct page *page;
452
453         BUG_ON(!buffer_async_write(bh));
454
455         page = bh->b_page;
456         if (uptodate) {
457                 set_buffer_uptodate(bh);
458         } else {
459                 if (printk_ratelimit()) {
460                         buffer_io_error(bh);
461                         printk(KERN_WARNING "lost page write due to "
462                                         "I/O error on %s\n",
463                                bdevname(bh->b_bdev, b));
464                 }
465                 set_bit(AS_EIO, &page->mapping->flags);
466                 set_buffer_write_io_error(bh);
467                 clear_buffer_uptodate(bh);
468                 SetPageError(page);
469         }
470
471         first = page_buffers(page);
472         local_irq_save(flags);
473         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
474
475         clear_buffer_async_write(bh);
476         unlock_buffer(bh);
477         tmp = bh->b_this_page;
478         while (tmp != bh) {
479                 if (buffer_async_write(tmp)) {
480                         BUG_ON(!buffer_locked(tmp));
481                         goto still_busy;
482                 }
483                 tmp = tmp->b_this_page;
484         }
485         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
486         local_irq_restore(flags);
487         end_page_writeback(page);
488         return;
489
490 still_busy:
491         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
492         local_irq_restore(flags);
493         return;
494 }
495
496 /*
497  * If a page's buffers are under async readin (end_buffer_async_read
498  * completion) then there is a possibility that another thread of
499  * control could lock one of the buffers after it has completed
500  * but while some of the other buffers have not completed.  This
501  * locked buffer would confuse end_buffer_async_read() into not unlocking
502  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
503  * that this buffer is not under async I/O.
504  *
505  * The page comes unlocked when it has no locked buffer_async buffers
506  * left.
507  *
508  * PageLocked prevents anyone starting new async I/O reads any of
509  * the buffers.
510  *
511  * PageWriteback is used to prevent simultaneous writeout of the same
512  * page.
513  *
514  * PageLocked prevents anyone from starting writeback of a page which is
515  * under read I/O (PageWriteback is only ever set against a locked page).
516  */
517 static void mark_buffer_async_read(struct buffer_head *bh)
518 {
519         bh->b_end_io = end_buffer_async_read;
520         set_buffer_async_read(bh);
521 }
522
523 void mark_buffer_async_write(struct buffer_head *bh)
524 {
525         bh->b_end_io = end_buffer_async_write;
526         set_buffer_async_write(bh);
527 }
528 EXPORT_SYMBOL(mark_buffer_async_write);
529
530
531 /*
532  * fs/buffer.c contains helper functions for buffer-backed address space's
533  * fsync functions.  A common requirement for buffer-based filesystems is
534  * that certain data from the backing blockdev needs to be written out for
535  * a successful fsync().  For example, ext2 indirect blocks need to be
536  * written back and waited upon before fsync() returns.
537  *
538  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
539  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
540  * management of a list of dependent buffers at ->i_mapping->private_list.
541  *
542  * Locking is a little subtle: try_to_free_buffers() will remove buffers
543  * from their controlling inode's queue when they are being freed.  But
544  * try_to_free_buffers() will be operating against the *blockdev* mapping
545  * at the time, not against the S_ISREG file which depends on those buffers.
546  * So the locking for private_list is via the private_lock in the address_space
547  * which backs the buffers.  Which is different from the address_space 
548  * against which the buffers are listed.  So for a particular address_space,
549  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
550  * mapping->private_list will always be protected by the backing blockdev's
551  * ->private_lock.
552  *
553  * Which introduces a requirement: all buffers on an address_space's
554  * ->private_list must be from the same address_space: the blockdev's.
555  *
556  * address_spaces which do not place buffers at ->private_list via these
557  * utility functions are free to use private_lock and private_list for
558  * whatever they want.  The only requirement is that list_empty(private_list)
559  * be true at clear_inode() time.
560  *
561  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
562  * filesystems should do that.  invalidate_inode_buffers() should just go
563  * BUG_ON(!list_empty).
564  *
565  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
566  * take an address_space, not an inode.  And it should be called
567  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
568  * queued up.
569  *
570  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
571  * list if it is already on a list.  Because if the buffer is on a list,
572  * it *must* already be on the right one.  If not, the filesystem is being
573  * silly.  This will save a ton of locking.  But first we have to ensure
574  * that buffers are taken *off* the old inode's list when they are freed
575  * (presumably in truncate).  That requires careful auditing of all
576  * filesystems (do it inside bforget()).  It could also be done by bringing
577  * b_inode back.
578  */
579
580 /*
581  * The buffer's backing address_space's private_lock must be held
582  */
583 static inline void __remove_assoc_queue(struct buffer_head *bh)
584 {
585         list_del_init(&bh->b_assoc_buffers);
586         WARN_ON(!bh->b_assoc_map);
587         if (buffer_write_io_error(bh))
588                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
589         bh->b_assoc_map = NULL;
590 }
591
592 int inode_has_buffers(struct inode *inode)
593 {
594         return !list_empty(&inode->i_data.private_list);
595 }
596
597 /*
598  * osync is designed to support O_SYNC io.  It waits synchronously for
599  * all already-submitted IO to complete, but does not queue any new
600  * writes to the disk.
601  *
602  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
603  * you dirty the buffers, and then use osync_inode_buffers to wait for
604  * completion.  Any other dirty buffers which are not yet queued for
605  * write will not be flushed to disk by the osync.
606  */
607 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
608 {
609         struct buffer_head *bh;
610         struct list_head *p;
611         int err = 0;
612
613         spin_lock(lock);
614 repeat:
615         list_for_each_prev(p, list) {
616                 bh = BH_ENTRY(p);
617                 if (buffer_locked(bh)) {
618                         get_bh(bh);
619                         spin_unlock(lock);
620                         wait_on_buffer(bh);
621                         if (!buffer_uptodate(bh))
622                                 err = -EIO;
623                         brelse(bh);
624                         spin_lock(lock);
625                         goto repeat;
626                 }
627         }
628         spin_unlock(lock);
629         return err;
630 }
631
632 /**
633  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
634  * @mapping: the mapping which wants those buffers written
635  *
636  * Starts I/O against the buffers at mapping->private_list, and waits upon
637  * that I/O.
638  *
639  * Basically, this is a convenience function for fsync().
640  * @mapping is a file or directory which needs those buffers to be written for
641  * a successful fsync().
642  */
643 int sync_mapping_buffers(struct address_space *mapping)
644 {
645         struct address_space *buffer_mapping = mapping->assoc_mapping;
646
647         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
648                 return 0;
649
650         return fsync_buffers_list(&buffer_mapping->private_lock,
651                                         &mapping->private_list);
652 }
653 EXPORT_SYMBOL(sync_mapping_buffers);
654
655 /*
656  * Called when we've recently written block `bblock', and it is known that
657  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
658  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
659  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
660  */
661 void write_boundary_block(struct block_device *bdev,
662                         sector_t bblock, unsigned blocksize)
663 {
664         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
665         if (bh) {
666                 if (buffer_dirty(bh))
667                         ll_rw_block(WRITE, 1, &bh);
668                 put_bh(bh);
669         }
670 }
671
672 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
673 {
674         struct address_space *mapping = inode->i_mapping;
675         struct address_space *buffer_mapping = bh->b_page->mapping;
676
677         mark_buffer_dirty(bh);
678         if (!mapping->assoc_mapping) {
679                 mapping->assoc_mapping = buffer_mapping;
680         } else {
681                 BUG_ON(mapping->assoc_mapping != buffer_mapping);
682         }
683         if (!bh->b_assoc_map) {
684                 spin_lock(&buffer_mapping->private_lock);
685                 list_move_tail(&bh->b_assoc_buffers,
686                                 &mapping->private_list);
687                 bh->b_assoc_map = mapping;
688                 spin_unlock(&buffer_mapping->private_lock);
689         }
690 }
691 EXPORT_SYMBOL(mark_buffer_dirty_inode);
692
693 /*
694  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
695  * dirty.
696  *
697  * If warn is true, then emit a warning if the page is not uptodate and has
698  * not been truncated.
699  */
700 static int __set_page_dirty(struct page *page,
701                 struct address_space *mapping, int warn)
702 {
703         if (unlikely(!mapping))
704                 return !TestSetPageDirty(page);
705
706         if (TestSetPageDirty(page))
707                 return 0;
708
709         spin_lock_irq(&mapping->tree_lock);
710         if (page->mapping) {    /* Race with truncate? */
711                 WARN_ON_ONCE(warn && !PageUptodate(page));
712
713                 if (mapping_cap_account_dirty(mapping)) {
714                         __inc_zone_page_state(page, NR_FILE_DIRTY);
715                         __inc_bdi_stat(mapping->backing_dev_info,
716                                         BDI_RECLAIMABLE);
717                         task_io_account_write(PAGE_CACHE_SIZE);
718                 }
719                 radix_tree_tag_set(&mapping->page_tree,
720                                 page_index(page), PAGECACHE_TAG_DIRTY);
721         }
722         spin_unlock_irq(&mapping->tree_lock);
723         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
724
725         return 1;
726 }
727
728 /*
729  * Add a page to the dirty page list.
730  *
731  * It is a sad fact of life that this function is called from several places
732  * deeply under spinlocking.  It may not sleep.
733  *
734  * If the page has buffers, the uptodate buffers are set dirty, to preserve
735  * dirty-state coherency between the page and the buffers.  It the page does
736  * not have buffers then when they are later attached they will all be set
737  * dirty.
738  *
739  * The buffers are dirtied before the page is dirtied.  There's a small race
740  * window in which a writepage caller may see the page cleanness but not the
741  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
742  * before the buffers, a concurrent writepage caller could clear the page dirty
743  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
744  * page on the dirty page list.
745  *
746  * We use private_lock to lock against try_to_free_buffers while using the
747  * page's buffer list.  Also use this to protect against clean buffers being
748  * added to the page after it was set dirty.
749  *
750  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
751  * address_space though.
752  */
753 int __set_page_dirty_buffers(struct page *page)
754 {
755         struct address_space *mapping = page_mapping(page);
756
757         if (unlikely(!mapping))
758                 return !TestSetPageDirty(page);
759
760         spin_lock(&mapping->private_lock);
761         if (page_has_buffers(page)) {
762                 struct buffer_head *head = page_buffers(page);
763                 struct buffer_head *bh = head;
764
765                 do {
766                         set_buffer_dirty(bh);
767                         bh = bh->b_this_page;
768                 } while (bh != head);
769         }
770         spin_unlock(&mapping->private_lock);
771
772         return __set_page_dirty(page, mapping, 1);
773 }
774 EXPORT_SYMBOL(__set_page_dirty_buffers);
775
776 /*
777  * Write out and wait upon a list of buffers.
778  *
779  * We have conflicting pressures: we want to make sure that all
780  * initially dirty buffers get waited on, but that any subsequently
781  * dirtied buffers don't.  After all, we don't want fsync to last
782  * forever if somebody is actively writing to the file.
783  *
784  * Do this in two main stages: first we copy dirty buffers to a
785  * temporary inode list, queueing the writes as we go.  Then we clean
786  * up, waiting for those writes to complete.
787  * 
788  * During this second stage, any subsequent updates to the file may end
789  * up refiling the buffer on the original inode's dirty list again, so
790  * there is a chance we will end up with a buffer queued for write but
791  * not yet completed on that list.  So, as a final cleanup we go through
792  * the osync code to catch these locked, dirty buffers without requeuing
793  * any newly dirty buffers for write.
794  */
795 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
796 {
797         struct buffer_head *bh;
798         struct list_head tmp;
799         struct address_space *mapping;
800         int err = 0, err2;
801
802         INIT_LIST_HEAD(&tmp);
803
804         spin_lock(lock);
805         while (!list_empty(list)) {
806                 bh = BH_ENTRY(list->next);
807                 mapping = bh->b_assoc_map;
808                 __remove_assoc_queue(bh);
809                 /* Avoid race with mark_buffer_dirty_inode() which does
810                  * a lockless check and we rely on seeing the dirty bit */
811                 smp_mb();
812                 if (buffer_dirty(bh) || buffer_locked(bh)) {
813                         list_add(&bh->b_assoc_buffers, &tmp);
814                         bh->b_assoc_map = mapping;
815                         if (buffer_dirty(bh)) {
816                                 get_bh(bh);
817                                 spin_unlock(lock);
818                                 /*
819                                  * Ensure any pending I/O completes so that
820                                  * ll_rw_block() actually writes the current
821                                  * contents - it is a noop if I/O is still in
822                                  * flight on potentially older contents.
823                                  */
824                                 ll_rw_block(SWRITE_SYNC, 1, &bh);
825                                 brelse(bh);
826                                 spin_lock(lock);
827                         }
828                 }
829         }
830
831         while (!list_empty(&tmp)) {
832                 bh = BH_ENTRY(tmp.prev);
833                 get_bh(bh);
834                 mapping = bh->b_assoc_map;
835                 __remove_assoc_queue(bh);
836                 /* Avoid race with mark_buffer_dirty_inode() which does
837                  * a lockless check and we rely on seeing the dirty bit */
838                 smp_mb();
839                 if (buffer_dirty(bh)) {
840                         list_add(&bh->b_assoc_buffers,
841                                  &mapping->private_list);
842                         bh->b_assoc_map = mapping;
843                 }
844                 spin_unlock(lock);
845                 wait_on_buffer(bh);
846                 if (!buffer_uptodate(bh))
847                         err = -EIO;
848                 brelse(bh);
849                 spin_lock(lock);
850         }
851         
852         spin_unlock(lock);
853         err2 = osync_buffers_list(lock, list);
854         if (err)
855                 return err;
856         else
857                 return err2;
858 }
859
860 /*
861  * Invalidate any and all dirty buffers on a given inode.  We are
862  * probably unmounting the fs, but that doesn't mean we have already
863  * done a sync().  Just drop the buffers from the inode list.
864  *
865  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
866  * assumes that all the buffers are against the blockdev.  Not true
867  * for reiserfs.
868  */
869 void invalidate_inode_buffers(struct inode *inode)
870 {
871         if (inode_has_buffers(inode)) {
872                 struct address_space *mapping = &inode->i_data;
873                 struct list_head *list = &mapping->private_list;
874                 struct address_space *buffer_mapping = mapping->assoc_mapping;
875
876                 spin_lock(&buffer_mapping->private_lock);
877                 while (!list_empty(list))
878                         __remove_assoc_queue(BH_ENTRY(list->next));
879                 spin_unlock(&buffer_mapping->private_lock);
880         }
881 }
882
883 /*
884  * Remove any clean buffers from the inode's buffer list.  This is called
885  * when we're trying to free the inode itself.  Those buffers can pin it.
886  *
887  * Returns true if all buffers were removed.
888  */
889 int remove_inode_buffers(struct inode *inode)
890 {
891         int ret = 1;
892
893         if (inode_has_buffers(inode)) {
894                 struct address_space *mapping = &inode->i_data;
895                 struct list_head *list = &mapping->private_list;
896                 struct address_space *buffer_mapping = mapping->assoc_mapping;
897
898                 spin_lock(&buffer_mapping->private_lock);
899                 while (!list_empty(list)) {
900                         struct buffer_head *bh = BH_ENTRY(list->next);
901                         if (buffer_dirty(bh)) {
902                                 ret = 0;
903                                 break;
904                         }
905                         __remove_assoc_queue(bh);
906                 }
907                 spin_unlock(&buffer_mapping->private_lock);
908         }
909         return ret;
910 }
911
912 /*
913  * Create the appropriate buffers when given a page for data area and
914  * the size of each buffer.. Use the bh->b_this_page linked list to
915  * follow the buffers created.  Return NULL if unable to create more
916  * buffers.
917  *
918  * The retry flag is used to differentiate async IO (paging, swapping)
919  * which may not fail from ordinary buffer allocations.
920  */
921 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
922                 int retry)
923 {
924         struct buffer_head *bh, *head;
925         long offset;
926
927 try_again:
928         head = NULL;
929         offset = PAGE_SIZE;
930         while ((offset -= size) >= 0) {
931                 bh = alloc_buffer_head(GFP_NOFS);
932                 if (!bh)
933                         goto no_grow;
934
935                 bh->b_bdev = NULL;
936                 bh->b_this_page = head;
937                 bh->b_blocknr = -1;
938                 head = bh;
939
940                 bh->b_state = 0;
941                 atomic_set(&bh->b_count, 0);
942                 bh->b_private = NULL;
943                 bh->b_size = size;
944
945                 /* Link the buffer to its page */
946                 set_bh_page(bh, page, offset);
947
948                 init_buffer(bh, NULL, NULL);
949         }
950         return head;
951 /*
952  * In case anything failed, we just free everything we got.
953  */
954 no_grow:
955         if (head) {
956                 do {
957                         bh = head;
958                         head = head->b_this_page;
959                         free_buffer_head(bh);
960                 } while (head);
961         }
962
963         /*
964          * Return failure for non-async IO requests.  Async IO requests
965          * are not allowed to fail, so we have to wait until buffer heads
966          * become available.  But we don't want tasks sleeping with 
967          * partially complete buffers, so all were released above.
968          */
969         if (!retry)
970                 return NULL;
971
972         /* We're _really_ low on memory. Now we just
973          * wait for old buffer heads to become free due to
974          * finishing IO.  Since this is an async request and
975          * the reserve list is empty, we're sure there are 
976          * async buffer heads in use.
977          */
978         free_more_memory();
979         goto try_again;
980 }
981 EXPORT_SYMBOL_GPL(alloc_page_buffers);
982
983 static inline void
984 link_dev_buffers(struct page *page, struct buffer_head *head)
985 {
986         struct buffer_head *bh, *tail;
987
988         bh = head;
989         do {
990                 tail = bh;
991                 bh = bh->b_this_page;
992         } while (bh);
993         tail->b_this_page = head;
994         attach_page_buffers(page, head);
995 }
996
997 /*
998  * Initialise the state of a blockdev page's buffers.
999  */ 
1000 static void
1001 init_page_buffers(struct page *page, struct block_device *bdev,
1002                         sector_t block, int size)
1003 {
1004         struct buffer_head *head = page_buffers(page);
1005         struct buffer_head *bh = head;
1006         int uptodate = PageUptodate(page);
1007
1008         do {
1009                 if (!buffer_mapped(bh)) {
1010                         init_buffer(bh, NULL, NULL);
1011                         bh->b_bdev = bdev;
1012                         bh->b_blocknr = block;
1013                         if (uptodate)
1014                                 set_buffer_uptodate(bh);
1015                         set_buffer_mapped(bh);
1016                 }
1017                 block++;
1018                 bh = bh->b_this_page;
1019         } while (bh != head);
1020 }
1021
1022 /*
1023  * Create the page-cache page that contains the requested block.
1024  *
1025  * This is user purely for blockdev mappings.
1026  */
1027 static struct page *
1028 grow_dev_page(struct block_device *bdev, sector_t block,
1029                 pgoff_t index, int size)
1030 {
1031         struct inode *inode = bdev->bd_inode;
1032         struct page *page;
1033         struct buffer_head *bh;
1034
1035         page = find_or_create_page(inode->i_mapping, index,
1036                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
1037         if (!page)
1038                 return NULL;
1039
1040         BUG_ON(!PageLocked(page));
1041
1042         if (page_has_buffers(page)) {
1043                 bh = page_buffers(page);
1044                 if (bh->b_size == size) {
1045                         init_page_buffers(page, bdev, block, size);
1046                         return page;
1047                 }
1048                 if (!try_to_free_buffers(page))
1049                         goto failed;
1050         }
1051
1052         /*
1053          * Allocate some buffers for this page
1054          */
1055         bh = alloc_page_buffers(page, size, 0);
1056         if (!bh)
1057                 goto failed;
1058
1059         /*
1060          * Link the page to the buffers and initialise them.  Take the
1061          * lock to be atomic wrt __find_get_block(), which does not
1062          * run under the page lock.
1063          */
1064         spin_lock(&inode->i_mapping->private_lock);
1065         link_dev_buffers(page, bh);
1066         init_page_buffers(page, bdev, block, size);
1067         spin_unlock(&inode->i_mapping->private_lock);
1068         return page;
1069
1070 failed:
1071         BUG();
1072         unlock_page(page);
1073         page_cache_release(page);
1074         return NULL;
1075 }
1076
1077 /*
1078  * Create buffers for the specified block device block's page.  If
1079  * that page was dirty, the buffers are set dirty also.
1080  */
1081 static int
1082 grow_buffers(struct block_device *bdev, sector_t block, int size)
1083 {
1084         struct page *page;
1085         pgoff_t index;
1086         int sizebits;
1087
1088         sizebits = -1;
1089         do {
1090                 sizebits++;
1091         } while ((size << sizebits) < PAGE_SIZE);
1092
1093         index = block >> sizebits;
1094
1095         /*
1096          * Check for a block which wants to lie outside our maximum possible
1097          * pagecache index.  (this comparison is done using sector_t types).
1098          */
1099         if (unlikely(index != block >> sizebits)) {
1100                 char b[BDEVNAME_SIZE];
1101
1102                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1103                         "device %s\n",
1104                         __func__, (unsigned long long)block,
1105                         bdevname(bdev, b));
1106                 return -EIO;
1107         }
1108         block = index << sizebits;
1109         /* Create a page with the proper size buffers.. */
1110         page = grow_dev_page(bdev, block, index, size);
1111         if (!page)
1112                 return 0;
1113         unlock_page(page);
1114         page_cache_release(page);
1115         return 1;
1116 }
1117
1118 static struct buffer_head *
1119 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1120 {
1121         /* Size must be multiple of hard sectorsize */
1122         if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1123                         (size < 512 || size > PAGE_SIZE))) {
1124                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1125                                         size);
1126                 printk(KERN_ERR "hardsect size: %d\n",
1127                                         bdev_hardsect_size(bdev));
1128
1129                 dump_stack();
1130                 return NULL;
1131         }
1132
1133         for (;;) {
1134                 struct buffer_head * bh;
1135                 int ret;
1136
1137                 bh = __find_get_block(bdev, block, size);
1138                 if (bh)
1139                         return bh;
1140
1141                 ret = grow_buffers(bdev, block, size);
1142                 if (ret < 0)
1143                         return NULL;
1144                 if (ret == 0)
1145                         free_more_memory();
1146         }
1147 }
1148
1149 /*
1150  * The relationship between dirty buffers and dirty pages:
1151  *
1152  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1153  * the page is tagged dirty in its radix tree.
1154  *
1155  * At all times, the dirtiness of the buffers represents the dirtiness of
1156  * subsections of the page.  If the page has buffers, the page dirty bit is
1157  * merely a hint about the true dirty state.
1158  *
1159  * When a page is set dirty in its entirety, all its buffers are marked dirty
1160  * (if the page has buffers).
1161  *
1162  * When a buffer is marked dirty, its page is dirtied, but the page's other
1163  * buffers are not.
1164  *
1165  * Also.  When blockdev buffers are explicitly read with bread(), they
1166  * individually become uptodate.  But their backing page remains not
1167  * uptodate - even if all of its buffers are uptodate.  A subsequent
1168  * block_read_full_page() against that page will discover all the uptodate
1169  * buffers, will set the page uptodate and will perform no I/O.
1170  */
1171
1172 /**
1173  * mark_buffer_dirty - mark a buffer_head as needing writeout
1174  * @bh: the buffer_head to mark dirty
1175  *
1176  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1177  * backing page dirty, then tag the page as dirty in its address_space's radix
1178  * tree and then attach the address_space's inode to its superblock's dirty
1179  * inode list.
1180  *
1181  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1182  * mapping->tree_lock and the global inode_lock.
1183  */
1184 void mark_buffer_dirty(struct buffer_head *bh)
1185 {
1186         WARN_ON_ONCE(!buffer_uptodate(bh));
1187
1188         /*
1189          * Very *carefully* optimize the it-is-already-dirty case.
1190          *
1191          * Don't let the final "is it dirty" escape to before we
1192          * perhaps modified the buffer.
1193          */
1194         if (buffer_dirty(bh)) {
1195                 smp_mb();
1196                 if (buffer_dirty(bh))
1197                         return;
1198         }
1199
1200         if (!test_set_buffer_dirty(bh))
1201                 __set_page_dirty(bh->b_page, page_mapping(bh->b_page), 0);
1202 }
1203
1204 /*
1205  * Decrement a buffer_head's reference count.  If all buffers against a page
1206  * have zero reference count, are clean and unlocked, and if the page is clean
1207  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1208  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1209  * a page but it ends up not being freed, and buffers may later be reattached).
1210  */
1211 void __brelse(struct buffer_head * buf)
1212 {
1213         if (atomic_read(&buf->b_count)) {
1214                 put_bh(buf);
1215                 return;
1216         }
1217         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1218 }
1219
1220 /*
1221  * bforget() is like brelse(), except it discards any
1222  * potentially dirty data.
1223  */
1224 void __bforget(struct buffer_head *bh)
1225 {
1226         clear_buffer_dirty(bh);
1227         if (bh->b_assoc_map) {
1228                 struct address_space *buffer_mapping = bh->b_page->mapping;
1229
1230                 spin_lock(&buffer_mapping->private_lock);
1231                 list_del_init(&bh->b_assoc_buffers);
1232                 bh->b_assoc_map = NULL;
1233                 spin_unlock(&buffer_mapping->private_lock);
1234         }
1235         __brelse(bh);
1236 }
1237
1238 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1239 {
1240         lock_buffer(bh);
1241         if (buffer_uptodate(bh)) {
1242                 unlock_buffer(bh);
1243                 return bh;
1244         } else {
1245                 get_bh(bh);
1246                 bh->b_end_io = end_buffer_read_sync;
1247                 submit_bh(READ, bh);
1248                 wait_on_buffer(bh);
1249                 if (buffer_uptodate(bh))
1250                         return bh;
1251         }
1252         brelse(bh);
1253         return NULL;
1254 }
1255
1256 /*
1257  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1258  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1259  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1260  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1261  * CPU's LRUs at the same time.
1262  *
1263  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1264  * sb_find_get_block().
1265  *
1266  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1267  * a local interrupt disable for that.
1268  */
1269
1270 #define BH_LRU_SIZE     8
1271
1272 struct bh_lru {
1273         struct buffer_head *bhs[BH_LRU_SIZE];
1274 };
1275
1276 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1277
1278 #ifdef CONFIG_SMP
1279 #define bh_lru_lock()   local_irq_disable()
1280 #define bh_lru_unlock() local_irq_enable()
1281 #else
1282 #define bh_lru_lock()   preempt_disable()
1283 #define bh_lru_unlock() preempt_enable()
1284 #endif
1285
1286 static inline void check_irqs_on(void)
1287 {
1288 #ifdef irqs_disabled
1289         BUG_ON(irqs_disabled());
1290 #endif
1291 }
1292
1293 /*
1294  * The LRU management algorithm is dopey-but-simple.  Sorry.
1295  */
1296 static void bh_lru_install(struct buffer_head *bh)
1297 {
1298         struct buffer_head *evictee = NULL;
1299         struct bh_lru *lru;
1300
1301         check_irqs_on();
1302         bh_lru_lock();
1303         lru = &__get_cpu_var(bh_lrus);
1304         if (lru->bhs[0] != bh) {
1305                 struct buffer_head *bhs[BH_LRU_SIZE];
1306                 int in;
1307                 int out = 0;
1308
1309                 get_bh(bh);
1310                 bhs[out++] = bh;
1311                 for (in = 0; in < BH_LRU_SIZE; in++) {
1312                         struct buffer_head *bh2 = lru->bhs[in];
1313
1314                         if (bh2 == bh) {
1315                                 __brelse(bh2);
1316                         } else {
1317                                 if (out >= BH_LRU_SIZE) {
1318                                         BUG_ON(evictee != NULL);
1319                                         evictee = bh2;
1320                                 } else {
1321                                         bhs[out++] = bh2;
1322                                 }
1323                         }
1324                 }
1325                 while (out < BH_LRU_SIZE)
1326                         bhs[out++] = NULL;
1327                 memcpy(lru->bhs, bhs, sizeof(bhs));
1328         }
1329         bh_lru_unlock();
1330
1331         if (evictee)
1332                 __brelse(evictee);
1333 }
1334
1335 /*
1336  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1337  */
1338 static struct buffer_head *
1339 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1340 {
1341         struct buffer_head *ret = NULL;
1342         struct bh_lru *lru;
1343         unsigned int i;
1344
1345         check_irqs_on();
1346         bh_lru_lock();
1347         lru = &__get_cpu_var(bh_lrus);
1348         for (i = 0; i < BH_LRU_SIZE; i++) {
1349                 struct buffer_head *bh = lru->bhs[i];
1350
1351                 if (bh && bh->b_bdev == bdev &&
1352                                 bh->b_blocknr == block && bh->b_size == size) {
1353                         if (i) {
1354                                 while (i) {
1355                                         lru->bhs[i] = lru->bhs[i - 1];
1356                                         i--;
1357                                 }
1358                                 lru->bhs[0] = bh;
1359                         }
1360                         get_bh(bh);
1361                         ret = bh;
1362                         break;
1363                 }
1364         }
1365         bh_lru_unlock();
1366         return ret;
1367 }
1368
1369 /*
1370  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1371  * it in the LRU and mark it as accessed.  If it is not present then return
1372  * NULL
1373  */
1374 struct buffer_head *
1375 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1376 {
1377         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1378
1379         if (bh == NULL) {
1380                 bh = __find_get_block_slow(bdev, block);
1381                 if (bh)
1382                         bh_lru_install(bh);
1383         }
1384         if (bh)
1385                 touch_buffer(bh);
1386         return bh;
1387 }
1388 EXPORT_SYMBOL(__find_get_block);
1389
1390 /*
1391  * __getblk will locate (and, if necessary, create) the buffer_head
1392  * which corresponds to the passed block_device, block and size. The
1393  * returned buffer has its reference count incremented.
1394  *
1395  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1396  * illegal block number, __getblk() will happily return a buffer_head
1397  * which represents the non-existent block.  Very weird.
1398  *
1399  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1400  * attempt is failing.  FIXME, perhaps?
1401  */
1402 struct buffer_head *
1403 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1404 {
1405         struct buffer_head *bh = __find_get_block(bdev, block, size);
1406
1407         might_sleep();
1408         if (bh == NULL)
1409                 bh = __getblk_slow(bdev, block, size);
1410         return bh;
1411 }
1412 EXPORT_SYMBOL(__getblk);
1413
1414 /*
1415  * Do async read-ahead on a buffer..
1416  */
1417 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1418 {
1419         struct buffer_head *bh = __getblk(bdev, block, size);
1420         if (likely(bh)) {
1421                 ll_rw_block(READA, 1, &bh);
1422                 brelse(bh);
1423         }
1424 }
1425 EXPORT_SYMBOL(__breadahead);
1426
1427 /**
1428  *  __bread() - reads a specified block and returns the bh
1429  *  @bdev: the block_device to read from
1430  *  @block: number of block
1431  *  @size: size (in bytes) to read
1432  * 
1433  *  Reads a specified block, and returns buffer head that contains it.
1434  *  It returns NULL if the block was unreadable.
1435  */
1436 struct buffer_head *
1437 __bread(struct block_device *bdev, sector_t block, unsigned size)
1438 {
1439         struct buffer_head *bh = __getblk(bdev, block, size);
1440
1441         if (likely(bh) && !buffer_uptodate(bh))
1442                 bh = __bread_slow(bh);
1443         return bh;
1444 }
1445 EXPORT_SYMBOL(__bread);
1446
1447 /*
1448  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1449  * This doesn't race because it runs in each cpu either in irq
1450  * or with preempt disabled.
1451  */
1452 static void invalidate_bh_lru(void *arg)
1453 {
1454         struct bh_lru *b = &get_cpu_var(bh_lrus);
1455         int i;
1456
1457         for (i = 0; i < BH_LRU_SIZE; i++) {
1458                 brelse(b->bhs[i]);
1459                 b->bhs[i] = NULL;
1460         }
1461         put_cpu_var(bh_lrus);
1462 }
1463         
1464 void invalidate_bh_lrus(void)
1465 {
1466         on_each_cpu(invalidate_bh_lru, NULL, 1);
1467 }
1468 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1469
1470 void set_bh_page(struct buffer_head *bh,
1471                 struct page *page, unsigned long offset)
1472 {
1473         bh->b_page = page;
1474         BUG_ON(offset >= PAGE_SIZE);
1475         if (PageHighMem(page))
1476                 /*
1477                  * This catches illegal uses and preserves the offset:
1478                  */
1479                 bh->b_data = (char *)(0 + offset);
1480         else
1481                 bh->b_data = page_address(page) + offset;
1482 }
1483 EXPORT_SYMBOL(set_bh_page);
1484
1485 /*
1486  * Called when truncating a buffer on a page completely.
1487  */
1488 static void discard_buffer(struct buffer_head * bh)
1489 {
1490         lock_buffer(bh);
1491         clear_buffer_dirty(bh);
1492         bh->b_bdev = NULL;
1493         clear_buffer_mapped(bh);
1494         clear_buffer_req(bh);
1495         clear_buffer_new(bh);
1496         clear_buffer_delay(bh);
1497         clear_buffer_unwritten(bh);
1498         unlock_buffer(bh);
1499 }
1500
1501 /**
1502  * block_invalidatepage - invalidate part of all of a buffer-backed page
1503  *
1504  * @page: the page which is affected
1505  * @offset: the index of the truncation point
1506  *
1507  * block_invalidatepage() is called when all or part of the page has become
1508  * invalidatedby a truncate operation.
1509  *
1510  * block_invalidatepage() does not have to release all buffers, but it must
1511  * ensure that no dirty buffer is left outside @offset and that no I/O
1512  * is underway against any of the blocks which are outside the truncation
1513  * point.  Because the caller is about to free (and possibly reuse) those
1514  * blocks on-disk.
1515  */
1516 void block_invalidatepage(struct page *page, unsigned long offset)
1517 {
1518         struct buffer_head *head, *bh, *next;
1519         unsigned int curr_off = 0;
1520
1521         BUG_ON(!PageLocked(page));
1522         if (!page_has_buffers(page))
1523                 goto out;
1524
1525         head = page_buffers(page);
1526         bh = head;
1527         do {
1528                 unsigned int next_off = curr_off + bh->b_size;
1529                 next = bh->b_this_page;
1530
1531                 /*
1532                  * is this block fully invalidated?
1533                  */
1534                 if (offset <= curr_off)
1535                         discard_buffer(bh);
1536                 curr_off = next_off;
1537                 bh = next;
1538         } while (bh != head);
1539
1540         /*
1541          * We release buffers only if the entire page is being invalidated.
1542          * The get_block cached value has been unconditionally invalidated,
1543          * so real IO is not possible anymore.
1544          */
1545         if (offset == 0)
1546                 try_to_release_page(page, 0);
1547 out:
1548         return;
1549 }
1550 EXPORT_SYMBOL(block_invalidatepage);
1551
1552 /*
1553  * We attach and possibly dirty the buffers atomically wrt
1554  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1555  * is already excluded via the page lock.
1556  */
1557 void create_empty_buffers(struct page *page,
1558                         unsigned long blocksize, unsigned long b_state)
1559 {
1560         struct buffer_head *bh, *head, *tail;
1561
1562         head = alloc_page_buffers(page, blocksize, 1);
1563         bh = head;
1564         do {
1565                 bh->b_state |= b_state;
1566                 tail = bh;
1567                 bh = bh->b_this_page;
1568         } while (bh);
1569         tail->b_this_page = head;
1570
1571         spin_lock(&page->mapping->private_lock);
1572         if (PageUptodate(page) || PageDirty(page)) {
1573                 bh = head;
1574                 do {
1575                         if (PageDirty(page))
1576                                 set_buffer_dirty(bh);
1577                         if (PageUptodate(page))
1578                                 set_buffer_uptodate(bh);
1579                         bh = bh->b_this_page;
1580                 } while (bh != head);
1581         }
1582         attach_page_buffers(page, head);
1583         spin_unlock(&page->mapping->private_lock);
1584 }
1585 EXPORT_SYMBOL(create_empty_buffers);
1586
1587 /*
1588  * We are taking a block for data and we don't want any output from any
1589  * buffer-cache aliases starting from return from that function and
1590  * until the moment when something will explicitly mark the buffer
1591  * dirty (hopefully that will not happen until we will free that block ;-)
1592  * We don't even need to mark it not-uptodate - nobody can expect
1593  * anything from a newly allocated buffer anyway. We used to used
1594  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1595  * don't want to mark the alias unmapped, for example - it would confuse
1596  * anyone who might pick it with bread() afterwards...
1597  *
1598  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1599  * be writeout I/O going on against recently-freed buffers.  We don't
1600  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1601  * only if we really need to.  That happens here.
1602  */
1603 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1604 {
1605         struct buffer_head *old_bh;
1606
1607         might_sleep();
1608
1609         old_bh = __find_get_block_slow(bdev, block);
1610         if (old_bh) {
1611                 clear_buffer_dirty(old_bh);
1612                 wait_on_buffer(old_bh);
1613                 clear_buffer_req(old_bh);
1614                 __brelse(old_bh);
1615         }
1616 }
1617 EXPORT_SYMBOL(unmap_underlying_metadata);
1618
1619 /*
1620  * NOTE! All mapped/uptodate combinations are valid:
1621  *
1622  *      Mapped  Uptodate        Meaning
1623  *
1624  *      No      No              "unknown" - must do get_block()
1625  *      No      Yes             "hole" - zero-filled
1626  *      Yes     No              "allocated" - allocated on disk, not read in
1627  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1628  *
1629  * "Dirty" is valid only with the last case (mapped+uptodate).
1630  */
1631
1632 /*
1633  * While block_write_full_page is writing back the dirty buffers under
1634  * the page lock, whoever dirtied the buffers may decide to clean them
1635  * again at any time.  We handle that by only looking at the buffer
1636  * state inside lock_buffer().
1637  *
1638  * If block_write_full_page() is called for regular writeback
1639  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1640  * locked buffer.   This only can happen if someone has written the buffer
1641  * directly, with submit_bh().  At the address_space level PageWriteback
1642  * prevents this contention from occurring.
1643  */
1644 static int __block_write_full_page(struct inode *inode, struct page *page,
1645                         get_block_t *get_block, struct writeback_control *wbc)
1646 {
1647         int err;
1648         sector_t block;
1649         sector_t last_block;
1650         struct buffer_head *bh, *head;
1651         const unsigned blocksize = 1 << inode->i_blkbits;
1652         int nr_underway = 0;
1653
1654         BUG_ON(!PageLocked(page));
1655
1656         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1657
1658         if (!page_has_buffers(page)) {
1659                 create_empty_buffers(page, blocksize,
1660                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1661         }
1662
1663         /*
1664          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1665          * here, and the (potentially unmapped) buffers may become dirty at
1666          * any time.  If a buffer becomes dirty here after we've inspected it
1667          * then we just miss that fact, and the page stays dirty.
1668          *
1669          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1670          * handle that here by just cleaning them.
1671          */
1672
1673         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1674         head = page_buffers(page);
1675         bh = head;
1676
1677         /*
1678          * Get all the dirty buffers mapped to disk addresses and
1679          * handle any aliases from the underlying blockdev's mapping.
1680          */
1681         do {
1682                 if (block > last_block) {
1683                         /*
1684                          * mapped buffers outside i_size will occur, because
1685                          * this page can be outside i_size when there is a
1686                          * truncate in progress.
1687                          */
1688                         /*
1689                          * The buffer was zeroed by block_write_full_page()
1690                          */
1691                         clear_buffer_dirty(bh);
1692                         set_buffer_uptodate(bh);
1693                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1694                            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                         clear_buffer_delay(bh);
1700                         if (buffer_new(bh)) {
1701                                 /* blockdev mappings never come here */
1702                                 clear_buffer_new(bh);
1703                                 unmap_underlying_metadata(bh->b_bdev,
1704                                                         bh->b_blocknr);
1705                         }
1706                 }
1707                 bh = bh->b_this_page;
1708                 block++;
1709         } while (bh != head);
1710
1711         do {
1712                 if (!buffer_mapped(bh))
1713                         continue;
1714                 /*
1715                  * If it's a fully non-blocking write attempt and we cannot
1716                  * lock the buffer then redirty the page.  Note that this can
1717                  * potentially cause a busy-wait loop from pdflush and kswapd
1718                  * activity, but those code paths have their own higher-level
1719                  * throttling.
1720                  */
1721                 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1722                         lock_buffer(bh);
1723                 } else if (test_set_buffer_locked(bh)) {
1724                         redirty_page_for_writepage(wbc, page);
1725                         continue;
1726                 }
1727                 if (test_clear_buffer_dirty(bh)) {
1728                         mark_buffer_async_write(bh);
1729                 } else {
1730                         unlock_buffer(bh);
1731                 }
1732         } while ((bh = bh->b_this_page) != head);
1733
1734         /*
1735          * The page and its buffers are protected by PageWriteback(), so we can
1736          * drop the bh refcounts early.
1737          */
1738         BUG_ON(PageWriteback(page));
1739         set_page_writeback(page);
1740
1741         do {
1742                 struct buffer_head *next = bh->b_this_page;
1743                 if (buffer_async_write(bh)) {
1744                         submit_bh(WRITE, bh);
1745                         nr_underway++;
1746                 }
1747                 bh = next;
1748         } while (bh != head);
1749         unlock_page(page);
1750
1751         err = 0;
1752 done:
1753         if (nr_underway == 0) {
1754                 /*
1755                  * The page was marked dirty, but the buffers were
1756                  * clean.  Someone wrote them back by hand with
1757                  * ll_rw_block/submit_bh.  A rare case.
1758                  */
1759                 end_page_writeback(page);
1760
1761                 /*
1762                  * The page and buffer_heads can be released at any time from
1763                  * here on.
1764                  */
1765         }
1766         return err;
1767
1768 recover:
1769         /*
1770          * ENOSPC, or some other error.  We may already have added some
1771          * blocks to the file, so we need to write these out to avoid
1772          * exposing stale data.
1773          * The page is currently locked and not marked for writeback
1774          */
1775         bh = head;
1776         /* Recovery: lock and submit the mapped buffers */
1777         do {
1778                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1779                     !buffer_delay(bh)) {
1780                         lock_buffer(bh);
1781                         mark_buffer_async_write(bh);
1782                 } else {
1783                         /*
1784                          * The buffer may have been set dirty during
1785                          * attachment to a dirty page.
1786                          */
1787                         clear_buffer_dirty(bh);
1788                 }
1789         } while ((bh = bh->b_this_page) != head);
1790         SetPageError(page);
1791         BUG_ON(PageWriteback(page));
1792         mapping_set_error(page->mapping, err);
1793         set_page_writeback(page);
1794         do {
1795                 struct buffer_head *next = bh->b_this_page;
1796                 if (buffer_async_write(bh)) {
1797                         clear_buffer_dirty(bh);
1798                         submit_bh(WRITE, bh);
1799                         nr_underway++;
1800                 }
1801                 bh = next;
1802         } while (bh != head);
1803         unlock_page(page);
1804         goto done;
1805 }
1806
1807 /*
1808  * If a page has any new buffers, zero them out here, and mark them uptodate
1809  * and dirty so they'll be written out (in order to prevent uninitialised
1810  * block data from leaking). And clear the new bit.
1811  */
1812 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1813 {
1814         unsigned int block_start, block_end;
1815         struct buffer_head *head, *bh;
1816
1817         BUG_ON(!PageLocked(page));
1818         if (!page_has_buffers(page))
1819                 return;
1820
1821         bh = head = page_buffers(page);
1822         block_start = 0;
1823         do {
1824                 block_end = block_start + bh->b_size;
1825
1826                 if (buffer_new(bh)) {
1827                         if (block_end > from && block_start < to) {
1828                                 if (!PageUptodate(page)) {
1829                                         unsigned start, size;
1830
1831                                         start = max(from, block_start);
1832                                         size = min(to, block_end) - start;
1833
1834                                         zero_user(page, start, size);
1835                                         set_buffer_uptodate(bh);
1836                                 }
1837
1838                                 clear_buffer_new(bh);
1839                                 mark_buffer_dirty(bh);
1840                         }
1841                 }
1842
1843                 block_start = block_end;
1844                 bh = bh->b_this_page;
1845         } while (bh != head);
1846 }
1847 EXPORT_SYMBOL(page_zero_new_buffers);
1848
1849 static int __block_prepare_write(struct inode *inode, struct page *page,
1850                 unsigned from, unsigned to, get_block_t *get_block)
1851 {
1852         unsigned block_start, block_end;
1853         sector_t block;
1854         int err = 0;
1855         unsigned blocksize, bbits;
1856         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1857
1858         BUG_ON(!PageLocked(page));
1859         BUG_ON(from > PAGE_CACHE_SIZE);
1860         BUG_ON(to > PAGE_CACHE_SIZE);
1861         BUG_ON(from > to);
1862
1863         blocksize = 1 << inode->i_blkbits;
1864         if (!page_has_buffers(page))
1865                 create_empty_buffers(page, blocksize, 0);
1866         head = page_buffers(page);
1867
1868         bbits = inode->i_blkbits;
1869         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1870
1871         for(bh = head, block_start = 0; bh != head || !block_start;
1872             block++, block_start=block_end, bh = bh->b_this_page) {
1873                 block_end = block_start + blocksize;
1874                 if (block_end <= from || block_start >= to) {
1875                         if (PageUptodate(page)) {
1876                                 if (!buffer_uptodate(bh))
1877                                         set_buffer_uptodate(bh);
1878                         }
1879                         continue;
1880                 }
1881                 if (buffer_new(bh))
1882                         clear_buffer_new(bh);
1883                 if (!buffer_mapped(bh)) {
1884                         WARN_ON(bh->b_size != blocksize);
1885                         err = get_block(inode, block, bh, 1);
1886                         if (err)
1887                                 break;
1888                         if (buffer_new(bh)) {
1889                                 unmap_underlying_metadata(bh->b_bdev,
1890                                                         bh->b_blocknr);
1891                                 if (PageUptodate(page)) {
1892                                         clear_buffer_new(bh);
1893                                         set_buffer_uptodate(bh);
1894                                         mark_buffer_dirty(bh);
1895                                         continue;
1896                                 }
1897                                 if (block_end > to || block_start < from)
1898                                         zero_user_segments(page,
1899                                                 to, block_end,
1900                                                 block_start, from);
1901                                 continue;
1902                         }
1903                 }
1904                 if (PageUptodate(page)) {
1905                         if (!buffer_uptodate(bh))
1906                                 set_buffer_uptodate(bh);
1907                         continue; 
1908                 }
1909                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1910                     !buffer_unwritten(bh) &&
1911                      (block_start < from || block_end > to)) {
1912                         ll_rw_block(READ, 1, &bh);
1913                         *wait_bh++=bh;
1914                 }
1915         }
1916         /*
1917          * If we issued read requests - let them complete.
1918          */
1919         while(wait_bh > wait) {
1920                 wait_on_buffer(*--wait_bh);
1921                 if (!buffer_uptodate(*wait_bh))
1922                         err = -EIO;
1923         }
1924         if (unlikely(err))
1925                 page_zero_new_buffers(page, from, to);
1926         return err;
1927 }
1928
1929 static int __block_commit_write(struct inode *inode, struct page *page,
1930                 unsigned from, unsigned to)
1931 {
1932         unsigned block_start, block_end;
1933         int partial = 0;
1934         unsigned blocksize;
1935         struct buffer_head *bh, *head;
1936
1937         blocksize = 1 << inode->i_blkbits;
1938
1939         for(bh = head = page_buffers(page), block_start = 0;
1940             bh != head || !block_start;
1941             block_start=block_end, bh = bh->b_this_page) {
1942                 block_end = block_start + blocksize;
1943                 if (block_end <= from || block_start >= to) {
1944                         if (!buffer_uptodate(bh))
1945                                 partial = 1;
1946                 } else {
1947                         set_buffer_uptodate(bh);
1948                         mark_buffer_dirty(bh);
1949                 }
1950                 clear_buffer_new(bh);
1951         }
1952
1953         /*
1954          * If this is a partial write which happened to make all buffers
1955          * uptodate then we can optimize away a bogus readpage() for
1956          * the next read(). Here we 'discover' whether the page went
1957          * uptodate as a result of this (potentially partial) write.
1958          */
1959         if (!partial)
1960                 SetPageUptodate(page);
1961         return 0;
1962 }
1963
1964 /*
1965  * block_write_begin takes care of the basic task of block allocation and
1966  * bringing partial write blocks uptodate first.
1967  *
1968  * If *pagep is not NULL, then block_write_begin uses the locked page
1969  * at *pagep rather than allocating its own. In this case, the page will
1970  * not be unlocked or deallocated on failure.
1971  */
1972 int block_write_begin(struct file *file, struct address_space *mapping,
1973                         loff_t pos, unsigned len, unsigned flags,
1974                         struct page **pagep, void **fsdata,
1975                         get_block_t *get_block)
1976 {
1977         struct inode *inode = mapping->host;
1978         int status = 0;
1979         struct page *page;
1980         pgoff_t index;
1981         unsigned start, end;
1982         int ownpage = 0;
1983
1984         index = pos >> PAGE_CACHE_SHIFT;
1985         start = pos & (PAGE_CACHE_SIZE - 1);
1986         end = start + len;
1987
1988         page = *pagep;
1989         if (page == NULL) {
1990                 ownpage = 1;
1991                 page = __grab_cache_page(mapping, index);
1992                 if (!page) {
1993                         status = -ENOMEM;
1994                         goto out;
1995                 }
1996                 *pagep = page;
1997         } else
1998                 BUG_ON(!PageLocked(page));
1999
2000         status = __block_prepare_write(inode, page, start, end, get_block);
2001         if (unlikely(status)) {
2002                 ClearPageUptodate(page);
2003
2004                 if (ownpage) {
2005                         unlock_page(page);
2006                         page_cache_release(page);
2007                         *pagep = NULL;
2008
2009                         /*
2010                          * prepare_write() may have instantiated a few blocks
2011                          * outside i_size.  Trim these off again. Don't need
2012                          * i_size_read because we hold i_mutex.
2013                          */
2014                         if (pos + len > inode->i_size)
2015                                 vmtruncate(inode, inode->i_size);
2016                 }
2017                 goto out;
2018         }
2019
2020 out:
2021         return status;
2022 }
2023 EXPORT_SYMBOL(block_write_begin);
2024
2025 int block_write_end(struct file *file, struct address_space *mapping,
2026                         loff_t pos, unsigned len, unsigned copied,
2027                         struct page *page, void *fsdata)
2028 {
2029         struct inode *inode = mapping->host;
2030         unsigned start;
2031
2032         start = pos & (PAGE_CACHE_SIZE - 1);
2033
2034         if (unlikely(copied < len)) {
2035                 /*
2036                  * The buffers that were written will now be uptodate, so we
2037                  * don't have to worry about a readpage reading them and
2038                  * overwriting a partial write. However if we have encountered
2039                  * a short write and only partially written into a buffer, it
2040                  * will not be marked uptodate, so a readpage might come in and
2041                  * destroy our partial write.
2042                  *
2043                  * Do the simplest thing, and just treat any short write to a
2044                  * non uptodate page as a zero-length write, and force the
2045                  * caller to redo the whole thing.
2046                  */
2047                 if (!PageUptodate(page))
2048                         copied = 0;
2049
2050                 page_zero_new_buffers(page, start+copied, start+len);
2051         }
2052         flush_dcache_page(page);
2053
2054         /* This could be a short (even 0-length) commit */
2055         __block_commit_write(inode, page, start, start+copied);
2056
2057         return copied;
2058 }
2059 EXPORT_SYMBOL(block_write_end);
2060
2061 int generic_write_end(struct file *file, struct address_space *mapping,
2062                         loff_t pos, unsigned len, unsigned copied,
2063                         struct page *page, void *fsdata)
2064 {
2065         struct inode *inode = mapping->host;
2066         int i_size_changed = 0;
2067
2068         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2069
2070         /*
2071          * No need to use i_size_read() here, the i_size
2072          * cannot change under us because we hold i_mutex.
2073          *
2074          * But it's important to update i_size while still holding page lock:
2075          * page writeout could otherwise come in and zero beyond i_size.
2076          */
2077         if (pos+copied > inode->i_size) {
2078                 i_size_write(inode, pos+copied);
2079                 i_size_changed = 1;
2080         }
2081
2082         unlock_page(page);
2083         page_cache_release(page);
2084
2085         /*
2086          * Don't mark the inode dirty under page lock. First, it unnecessarily
2087          * makes the holding time of page lock longer. Second, it forces lock
2088          * ordering of page lock and transaction start for journaling
2089          * filesystems.
2090          */
2091         if (i_size_changed)
2092                 mark_inode_dirty(inode);
2093
2094         return copied;
2095 }
2096 EXPORT_SYMBOL(generic_write_end);
2097
2098 /*
2099  * block_is_partially_uptodate checks whether buffers within a page are
2100  * uptodate or not.
2101  *
2102  * Returns true if all buffers which correspond to a file portion
2103  * we want to read are uptodate.
2104  */
2105 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2106                                         unsigned long from)
2107 {
2108         struct inode *inode = page->mapping->host;
2109         unsigned block_start, block_end, blocksize;
2110         unsigned to;
2111         struct buffer_head *bh, *head;
2112         int ret = 1;
2113
2114         if (!page_has_buffers(page))
2115                 return 0;
2116
2117         blocksize = 1 << inode->i_blkbits;
2118         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2119         to = from + to;
2120         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2121                 return 0;
2122
2123         head = page_buffers(page);
2124         bh = head;
2125         block_start = 0;
2126         do {
2127                 block_end = block_start + blocksize;
2128                 if (block_end > from && block_start < to) {
2129                         if (!buffer_uptodate(bh)) {
2130                                 ret = 0;
2131                                 break;
2132                         }
2133                         if (block_end >= to)
2134                                 break;
2135                 }
2136                 block_start = block_end;
2137                 bh = bh->b_this_page;
2138         } while (bh != head);
2139
2140         return ret;
2141 }
2142 EXPORT_SYMBOL(block_is_partially_uptodate);
2143
2144 /*
2145  * Generic "read page" function for block devices that have the normal
2146  * get_block functionality. This is most of the block device filesystems.
2147  * Reads the page asynchronously --- the unlock_buffer() and
2148  * set/clear_buffer_uptodate() functions propagate buffer state into the
2149  * page struct once IO has completed.
2150  */
2151 int block_read_full_page(struct page *page, get_block_t *get_block)
2152 {
2153         struct inode *inode = page->mapping->host;
2154         sector_t iblock, lblock;
2155         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2156         unsigned int blocksize;
2157         int nr, i;
2158         int fully_mapped = 1;
2159
2160         BUG_ON(!PageLocked(page));
2161         blocksize = 1 << inode->i_blkbits;
2162         if (!page_has_buffers(page))
2163                 create_empty_buffers(page, blocksize, 0);
2164         head = page_buffers(page);
2165
2166         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2167         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2168         bh = head;
2169         nr = 0;
2170         i = 0;
2171
2172         do {
2173                 if (buffer_uptodate(bh))
2174                         continue;
2175
2176                 if (!buffer_mapped(bh)) {
2177                         int err = 0;
2178
2179                         fully_mapped = 0;
2180                         if (iblock < lblock) {
2181                                 WARN_ON(bh->b_size != blocksize);
2182                                 err = get_block(inode, iblock, bh, 0);
2183                                 if (err)
2184                                         SetPageError(page);
2185                         }
2186                         if (!buffer_mapped(bh)) {
2187                                 zero_user(page, i * blocksize, blocksize);
2188                                 if (!err)
2189                                         set_buffer_uptodate(bh);
2190                                 continue;
2191                         }
2192                         /*
2193                          * get_block() might have updated the buffer
2194                          * synchronously
2195                          */
2196                         if (buffer_uptodate(bh))
2197                                 continue;
2198                 }
2199                 arr[nr++] = bh;
2200         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2201
2202         if (fully_mapped)
2203                 SetPageMappedToDisk(page);
2204
2205         if (!nr) {
2206                 /*
2207                  * All buffers are uptodate - we can set the page uptodate
2208                  * as well. But not if get_block() returned an error.
2209                  */
2210                 if (!PageError(page))
2211                         SetPageUptodate(page);
2212                 unlock_page(page);
2213                 return 0;
2214         }
2215
2216         /* Stage two: lock the buffers */
2217         for (i = 0; i < nr; i++) {
2218                 bh = arr[i];
2219                 lock_buffer(bh);
2220                 mark_buffer_async_read(bh);
2221         }
2222
2223         /*
2224          * Stage 3: start the IO.  Check for uptodateness
2225          * inside the buffer lock in case another process reading
2226          * the underlying blockdev brought it uptodate (the sct fix).
2227          */
2228         for (i = 0; i < nr; i++) {
2229                 bh = arr[i];
2230                 if (buffer_uptodate(bh))
2231                         end_buffer_async_read(bh, 1);
2232                 else
2233                         submit_bh(READ, bh);
2234         }
2235         return 0;
2236 }
2237
2238 /* utility function for filesystems that need to do work on expanding
2239  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2240  * deal with the hole.  
2241  */
2242 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2243 {
2244         struct address_space *mapping = inode->i_mapping;
2245         struct page *page;
2246         void *fsdata;
2247         unsigned long limit;
2248         int err;
2249
2250         err = -EFBIG;
2251         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2252         if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2253                 send_sig(SIGXFSZ, current, 0);
2254                 goto out;
2255         }
2256         if (size > inode->i_sb->s_maxbytes)
2257                 goto out;
2258
2259         err = pagecache_write_begin(NULL, mapping, size, 0,
2260                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2261                                 &page, &fsdata);
2262         if (err)
2263                 goto out;
2264
2265         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2266         BUG_ON(err > 0);
2267
2268 out:
2269         return err;
2270 }
2271
2272 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2273                             loff_t pos, loff_t *bytes)
2274 {
2275         struct inode *inode = mapping->host;
2276         unsigned blocksize = 1 << inode->i_blkbits;
2277         struct page *page;
2278         void *fsdata;
2279         pgoff_t index, curidx;
2280         loff_t curpos;
2281         unsigned zerofrom, offset, len;
2282         int err = 0;
2283
2284         index = pos >> PAGE_CACHE_SHIFT;
2285         offset = pos & ~PAGE_CACHE_MASK;
2286
2287         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2288                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2289                 if (zerofrom & (blocksize-1)) {
2290                         *bytes |= (blocksize-1);
2291                         (*bytes)++;
2292                 }
2293                 len = PAGE_CACHE_SIZE - zerofrom;
2294
2295                 err = pagecache_write_begin(file, mapping, curpos, len,
2296                                                 AOP_FLAG_UNINTERRUPTIBLE,
2297                                                 &page, &fsdata);
2298                 if (err)
2299                         goto out;
2300                 zero_user(page, zerofrom, len);
2301                 err = pagecache_write_end(file, mapping, curpos, len, len,
2302                                                 page, fsdata);
2303                 if (err < 0)
2304                         goto out;
2305                 BUG_ON(err != len);
2306                 err = 0;
2307
2308                 balance_dirty_pages_ratelimited(mapping);
2309         }
2310
2311         /* page covers the boundary, find the boundary offset */
2312         if (index == curidx) {
2313                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2314                 /* if we will expand the thing last block will be filled */
2315                 if (offset <= zerofrom) {
2316                         goto out;
2317                 }
2318                 if (zerofrom & (blocksize-1)) {
2319                         *bytes |= (blocksize-1);
2320                         (*bytes)++;
2321                 }
2322                 len = offset - zerofrom;
2323
2324                 err = pagecache_write_begin(file, mapping, curpos, len,
2325                                                 AOP_FLAG_UNINTERRUPTIBLE,
2326                                                 &page, &fsdata);
2327                 if (err)
2328                         goto out;
2329                 zero_user(page, zerofrom, len);
2330                 err = pagecache_write_end(file, mapping, curpos, len, len,
2331                                                 page, fsdata);
2332                 if (err < 0)
2333                         goto out;
2334                 BUG_ON(err != len);
2335                 err = 0;
2336         }
2337 out:
2338         return err;
2339 }
2340
2341 /*
2342  * For moronic filesystems that do not allow holes in file.
2343  * We may have to extend the file.
2344  */
2345 int cont_write_begin(struct file *file, struct address_space *mapping,
2346                         loff_t pos, unsigned len, unsigned flags,
2347                         struct page **pagep, void **fsdata,
2348                         get_block_t *get_block, loff_t *bytes)
2349 {
2350         struct inode *inode = mapping->host;
2351         unsigned blocksize = 1 << inode->i_blkbits;
2352         unsigned zerofrom;
2353         int err;
2354
2355         err = cont_expand_zero(file, mapping, pos, bytes);
2356         if (err)
2357                 goto out;
2358
2359         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2360         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2361                 *bytes |= (blocksize-1);
2362                 (*bytes)++;
2363         }
2364
2365         *pagep = NULL;
2366         err = block_write_begin(file, mapping, pos, len,
2367                                 flags, pagep, fsdata, get_block);
2368 out:
2369         return err;
2370 }
2371
2372 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2373                         get_block_t *get_block)
2374 {
2375         struct inode *inode = page->mapping->host;
2376         int err = __block_prepare_write(inode, page, from, to, get_block);
2377         if (err)
2378                 ClearPageUptodate(page);
2379         return err;
2380 }
2381
2382 int block_commit_write(struct page *page, unsigned from, unsigned to)
2383 {
2384         struct inode *inode = page->mapping->host;
2385         __block_commit_write(inode,page,from,to);
2386         return 0;
2387 }
2388
2389 /*
2390  * block_page_mkwrite() is not allowed to change the file size as it gets
2391  * called from a page fault handler when a page is first dirtied. Hence we must
2392  * be careful to check for EOF conditions here. We set the page up correctly
2393  * for a written page which means we get ENOSPC checking when writing into
2394  * holes and correct delalloc and unwritten extent mapping on filesystems that
2395  * support these features.
2396  *
2397  * We are not allowed to take the i_mutex here so we have to play games to
2398  * protect against truncate races as the page could now be beyond EOF.  Because
2399  * vmtruncate() writes the inode size before removing pages, once we have the
2400  * page lock we can determine safely if the page is beyond EOF. If it is not
2401  * beyond EOF, then the page is guaranteed safe against truncation until we
2402  * unlock the page.
2403  */
2404 int
2405 block_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2406                    get_block_t get_block)
2407 {
2408         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2409         unsigned long end;
2410         loff_t size;
2411         int ret = -EINVAL;
2412
2413         lock_page(page);
2414         size = i_size_read(inode);
2415         if ((page->mapping != inode->i_mapping) ||
2416             (page_offset(page) > size)) {
2417                 /* page got truncated out from underneath us */
2418                 goto out_unlock;
2419         }
2420
2421         /* page is wholly or partially inside EOF */
2422         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2423                 end = size & ~PAGE_CACHE_MASK;
2424         else
2425                 end = PAGE_CACHE_SIZE;
2426
2427         ret = block_prepare_write(page, 0, end, get_block);
2428         if (!ret)
2429                 ret = block_commit_write(page, 0, end);
2430
2431 out_unlock:
2432         unlock_page(page);
2433         return ret;
2434 }
2435
2436 /*
2437  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2438  * immediately, while under the page lock.  So it needs a special end_io
2439  * handler which does not touch the bh after unlocking it.
2440  */
2441 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2442 {
2443         __end_buffer_read_notouch(bh, uptodate);
2444 }
2445
2446 /*
2447  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2448  * the page (converting it to circular linked list and taking care of page
2449  * dirty races).
2450  */
2451 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2452 {
2453         struct buffer_head *bh;
2454
2455         BUG_ON(!PageLocked(page));
2456
2457         spin_lock(&page->mapping->private_lock);
2458         bh = head;
2459         do {
2460                 if (PageDirty(page))
2461                         set_buffer_dirty(bh);
2462                 if (!bh->b_this_page)
2463                         bh->b_this_page = head;
2464                 bh = bh->b_this_page;
2465         } while (bh != head);
2466         attach_page_buffers(page, head);
2467         spin_unlock(&page->mapping->private_lock);
2468 }
2469
2470 /*
2471  * On entry, the page is fully not uptodate.
2472  * On exit the page is fully uptodate in the areas outside (from,to)
2473  */
2474 int nobh_write_begin(struct file *file, struct address_space *mapping,
2475                         loff_t pos, unsigned len, unsigned flags,
2476                         struct page **pagep, void **fsdata,
2477                         get_block_t *get_block)
2478 {
2479         struct inode *inode = mapping->host;
2480         const unsigned blkbits = inode->i_blkbits;
2481         const unsigned blocksize = 1 << blkbits;
2482         struct buffer_head *head, *bh;
2483         struct page *page;
2484         pgoff_t index;
2485         unsigned from, to;
2486         unsigned block_in_page;
2487         unsigned block_start, block_end;
2488         sector_t block_in_file;
2489         int nr_reads = 0;
2490         int ret = 0;
2491         int is_mapped_to_disk = 1;
2492
2493         index = pos >> PAGE_CACHE_SHIFT;
2494         from = pos & (PAGE_CACHE_SIZE - 1);
2495         to = from + len;
2496
2497         page = __grab_cache_page(mapping, index);
2498         if (!page)
2499                 return -ENOMEM;
2500         *pagep = page;
2501         *fsdata = NULL;
2502
2503         if (page_has_buffers(page)) {
2504                 unlock_page(page);
2505                 page_cache_release(page);
2506                 *pagep = NULL;
2507                 return block_write_begin(file, mapping, pos, len, flags, pagep,
2508                                         fsdata, get_block);
2509         }
2510
2511         if (PageMappedToDisk(page))
2512                 return 0;
2513
2514         /*
2515          * Allocate buffers so that we can keep track of state, and potentially
2516          * attach them to the page if an error occurs. In the common case of
2517          * no error, they will just be freed again without ever being attached
2518          * to the page (which is all OK, because we're under the page lock).
2519          *
2520          * Be careful: the buffer linked list is a NULL terminated one, rather
2521          * than the circular one we're used to.
2522          */
2523         head = alloc_page_buffers(page, blocksize, 0);
2524         if (!head) {
2525                 ret = -ENOMEM;
2526                 goto out_release;
2527         }
2528
2529         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2530
2531         /*
2532          * We loop across all blocks in the page, whether or not they are
2533          * part of the affected region.  This is so we can discover if the
2534          * page is fully mapped-to-disk.
2535          */
2536         for (block_start = 0, block_in_page = 0, bh = head;
2537                   block_start < PAGE_CACHE_SIZE;
2538                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2539                 int create;
2540
2541                 block_end = block_start + blocksize;
2542                 bh->b_state = 0;
2543                 create = 1;
2544                 if (block_start >= to)
2545                         create = 0;
2546                 ret = get_block(inode, block_in_file + block_in_page,
2547                                         bh, create);
2548                 if (ret)
2549                         goto failed;
2550                 if (!buffer_mapped(bh))
2551                         is_mapped_to_disk = 0;
2552                 if (buffer_new(bh))
2553                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2554                 if (PageUptodate(page)) {
2555                         set_buffer_uptodate(bh);
2556                         continue;
2557                 }
2558                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2559                         zero_user_segments(page, block_start, from,
2560                                                         to, block_end);
2561                         continue;
2562                 }
2563                 if (buffer_uptodate(bh))
2564                         continue;       /* reiserfs does this */
2565                 if (block_start < from || block_end > to) {
2566                         lock_buffer(bh);
2567                         bh->b_end_io = end_buffer_read_nobh;
2568                         submit_bh(READ, bh);
2569                         nr_reads++;
2570                 }
2571         }
2572
2573         if (nr_reads) {
2574                 /*
2575                  * The page is locked, so these buffers are protected from
2576                  * any VM or truncate activity.  Hence we don't need to care
2577                  * for the buffer_head refcounts.
2578                  */
2579                 for (bh = head; bh; bh = bh->b_this_page) {
2580                         wait_on_buffer(bh);
2581                         if (!buffer_uptodate(bh))
2582                                 ret = -EIO;
2583                 }
2584                 if (ret)
2585                         goto failed;
2586         }
2587
2588         if (is_mapped_to_disk)
2589                 SetPageMappedToDisk(page);
2590
2591         *fsdata = head; /* to be released by nobh_write_end */
2592
2593         return 0;
2594
2595 failed:
2596         BUG_ON(!ret);
2597         /*
2598          * Error recovery is a bit difficult. We need to zero out blocks that
2599          * were newly allocated, and dirty them to ensure they get written out.
2600          * Buffers need to be attached to the page at this point, otherwise
2601          * the handling of potential IO errors during writeout would be hard
2602          * (could try doing synchronous writeout, but what if that fails too?)
2603          */
2604         attach_nobh_buffers(page, head);
2605         page_zero_new_buffers(page, from, to);
2606
2607 out_release:
2608         unlock_page(page);
2609         page_cache_release(page);
2610         *pagep = NULL;
2611
2612         if (pos + len > inode->i_size)
2613                 vmtruncate(inode, inode->i_size);
2614
2615         return ret;
2616 }
2617 EXPORT_SYMBOL(nobh_write_begin);
2618
2619 int nobh_write_end(struct file *file, struct address_space *mapping,
2620                         loff_t pos, unsigned len, unsigned copied,
2621                         struct page *page, void *fsdata)
2622 {
2623         struct inode *inode = page->mapping->host;
2624         struct buffer_head *head = fsdata;
2625         struct buffer_head *bh;
2626         BUG_ON(fsdata != NULL && page_has_buffers(page));
2627
2628         if (unlikely(copied < len) && !page_has_buffers(page))
2629                 attach_nobh_buffers(page, head);
2630         if (page_has_buffers(page))
2631                 return generic_write_end(file, mapping, pos, len,
2632                                         copied, page, fsdata);
2633
2634         SetPageUptodate(page);
2635         set_page_dirty(page);
2636         if (pos+copied > inode->i_size) {
2637                 i_size_write(inode, pos+copied);
2638                 mark_inode_dirty(inode);
2639         }
2640
2641         unlock_page(page);
2642         page_cache_release(page);
2643
2644         while (head) {
2645                 bh = head;
2646                 head = head->b_this_page;
2647                 free_buffer_head(bh);
2648         }
2649
2650         return copied;
2651 }
2652 EXPORT_SYMBOL(nobh_write_end);
2653
2654 /*
2655  * nobh_writepage() - based on block_full_write_page() except
2656  * that it tries to operate without attaching bufferheads to
2657  * the page.
2658  */
2659 int nobh_writepage(struct page *page, get_block_t *get_block,
2660                         struct writeback_control *wbc)
2661 {
2662         struct inode * const inode = page->mapping->host;
2663         loff_t i_size = i_size_read(inode);
2664         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2665         unsigned offset;
2666         int ret;
2667
2668         /* Is the page fully inside i_size? */
2669         if (page->index < end_index)
2670                 goto out;
2671
2672         /* Is the page fully outside i_size? (truncate in progress) */
2673         offset = i_size & (PAGE_CACHE_SIZE-1);
2674         if (page->index >= end_index+1 || !offset) {
2675                 /*
2676                  * The page may have dirty, unmapped buffers.  For example,
2677                  * they may have been added in ext3_writepage().  Make them
2678                  * freeable here, so the page does not leak.
2679                  */
2680 #if 0
2681                 /* Not really sure about this  - do we need this ? */
2682                 if (page->mapping->a_ops->invalidatepage)
2683                         page->mapping->a_ops->invalidatepage(page, offset);
2684 #endif
2685                 unlock_page(page);
2686                 return 0; /* don't care */
2687         }
2688
2689         /*
2690          * The page straddles i_size.  It must be zeroed out on each and every
2691          * writepage invocation because it may be mmapped.  "A file is mapped
2692          * in multiples of the page size.  For a file that is not a multiple of
2693          * the  page size, the remaining memory is zeroed when mapped, and
2694          * writes to that region are not written out to the file."
2695          */
2696         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2697 out:
2698         ret = mpage_writepage(page, get_block, wbc);
2699         if (ret == -EAGAIN)
2700                 ret = __block_write_full_page(inode, page, get_block, wbc);
2701         return ret;
2702 }
2703 EXPORT_SYMBOL(nobh_writepage);
2704
2705 int nobh_truncate_page(struct address_space *mapping,
2706                         loff_t from, get_block_t *get_block)
2707 {
2708         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2709         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2710         unsigned blocksize;
2711         sector_t iblock;
2712         unsigned length, pos;
2713         struct inode *inode = mapping->host;
2714         struct page *page;
2715         struct buffer_head map_bh;
2716         int err;
2717
2718         blocksize = 1 << inode->i_blkbits;
2719         length = offset & (blocksize - 1);
2720
2721         /* Block boundary? Nothing to do */
2722         if (!length)
2723                 return 0;
2724
2725         length = blocksize - length;
2726         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2727
2728         page = grab_cache_page(mapping, index);
2729         err = -ENOMEM;
2730         if (!page)
2731                 goto out;
2732
2733         if (page_has_buffers(page)) {
2734 has_buffers:
2735                 unlock_page(page);
2736                 page_cache_release(page);
2737                 return block_truncate_page(mapping, from, get_block);
2738         }
2739
2740         /* Find the buffer that contains "offset" */
2741         pos = blocksize;
2742         while (offset >= pos) {
2743                 iblock++;
2744                 pos += blocksize;
2745         }
2746
2747         err = get_block(inode, iblock, &map_bh, 0);
2748         if (err)
2749                 goto unlock;
2750         /* unmapped? It's a hole - nothing to do */
2751         if (!buffer_mapped(&map_bh))
2752                 goto unlock;
2753
2754         /* Ok, it's mapped. Make sure it's up-to-date */
2755         if (!PageUptodate(page)) {
2756                 err = mapping->a_ops->readpage(NULL, page);
2757                 if (err) {
2758                         page_cache_release(page);
2759                         goto out;
2760                 }
2761                 lock_page(page);
2762                 if (!PageUptodate(page)) {
2763                         err = -EIO;
2764                         goto unlock;
2765                 }
2766                 if (page_has_buffers(page))
2767                         goto has_buffers;
2768         }
2769         zero_user(page, offset, length);
2770         set_page_dirty(page);
2771         err = 0;
2772
2773 unlock:
2774         unlock_page(page);
2775         page_cache_release(page);
2776 out:
2777         return err;
2778 }
2779 EXPORT_SYMBOL(nobh_truncate_page);
2780
2781 int block_truncate_page(struct address_space *mapping,
2782                         loff_t from, get_block_t *get_block)
2783 {
2784         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2785         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2786         unsigned blocksize;
2787         sector_t iblock;
2788         unsigned length, pos;
2789         struct inode *inode = mapping->host;
2790         struct page *page;
2791         struct buffer_head *bh;
2792         int err;
2793
2794         blocksize = 1 << inode->i_blkbits;
2795         length = offset & (blocksize - 1);
2796
2797         /* Block boundary? Nothing to do */
2798         if (!length)
2799                 return 0;
2800
2801         length = blocksize - length;
2802         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2803         
2804         page = grab_cache_page(mapping, index);
2805         err = -ENOMEM;
2806         if (!page)
2807                 goto out;
2808
2809         if (!page_has_buffers(page))
2810                 create_empty_buffers(page, blocksize, 0);
2811
2812         /* Find the buffer that contains "offset" */
2813         bh = page_buffers(page);
2814         pos = blocksize;
2815         while (offset >= pos) {
2816                 bh = bh->b_this_page;
2817                 iblock++;
2818                 pos += blocksize;
2819         }
2820
2821         err = 0;
2822         if (!buffer_mapped(bh)) {
2823                 WARN_ON(bh->b_size != blocksize);
2824                 err = get_block(inode, iblock, bh, 0);
2825                 if (err)
2826                         goto unlock;
2827                 /* unmapped? It's a hole - nothing to do */
2828                 if (!buffer_mapped(bh))
2829                         goto unlock;
2830         }
2831
2832         /* Ok, it's mapped. Make sure it's up-to-date */
2833         if (PageUptodate(page))
2834                 set_buffer_uptodate(bh);
2835
2836         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2837                 err = -EIO;
2838                 ll_rw_block(READ, 1, &bh);
2839                 wait_on_buffer(bh);
2840                 /* Uhhuh. Read error. Complain and punt. */
2841                 if (!buffer_uptodate(bh))
2842                         goto unlock;
2843         }
2844
2845         zero_user(page, offset, length);
2846         mark_buffer_dirty(bh);
2847         err = 0;
2848
2849 unlock:
2850         unlock_page(page);
2851         page_cache_release(page);
2852 out:
2853         return err;
2854 }
2855
2856 /*
2857  * The generic ->writepage function for buffer-backed address_spaces
2858  */
2859 int block_write_full_page(struct page *page, get_block_t *get_block,
2860                         struct writeback_control *wbc)
2861 {
2862         struct inode * const inode = page->mapping->host;
2863         loff_t i_size = i_size_read(inode);
2864         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2865         unsigned offset;
2866
2867         /* Is the page fully inside i_size? */
2868         if (page->index < end_index)
2869                 return __block_write_full_page(inode, page, get_block, wbc);
2870
2871         /* Is the page fully outside i_size? (truncate in progress) */
2872         offset = i_size & (PAGE_CACHE_SIZE-1);
2873         if (page->index >= end_index+1 || !offset) {
2874                 /*
2875                  * The page may have dirty, unmapped buffers.  For example,
2876                  * they may have been added in ext3_writepage().  Make them
2877                  * freeable here, so the page does not leak.
2878                  */
2879                 do_invalidatepage(page, 0);
2880                 unlock_page(page);
2881                 return 0; /* don't care */
2882         }
2883
2884         /*
2885          * The page straddles i_size.  It must be zeroed out on each and every
2886          * writepage invokation because it may be mmapped.  "A file is mapped
2887          * in multiples of the page size.  For a file that is not a multiple of
2888          * the  page size, the remaining memory is zeroed when mapped, and
2889          * writes to that region are not written out to the file."
2890          */
2891         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2892         return __block_write_full_page(inode, page, get_block, wbc);
2893 }
2894
2895 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2896                             get_block_t *get_block)
2897 {
2898         struct buffer_head tmp;
2899         struct inode *inode = mapping->host;
2900         tmp.b_state = 0;
2901         tmp.b_blocknr = 0;
2902         tmp.b_size = 1 << inode->i_blkbits;
2903         get_block(inode, block, &tmp, 0);
2904         return tmp.b_blocknr;
2905 }
2906
2907 static void end_bio_bh_io_sync(struct bio *bio, int err)
2908 {
2909         struct buffer_head *bh = bio->bi_private;
2910
2911         if (err == -EOPNOTSUPP) {
2912                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2913                 set_bit(BH_Eopnotsupp, &bh->b_state);
2914         }
2915
2916         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2917         bio_put(bio);
2918 }
2919
2920 int submit_bh(int rw, struct buffer_head * bh)
2921 {
2922         struct bio *bio;
2923         int ret = 0;
2924
2925         BUG_ON(!buffer_locked(bh));
2926         BUG_ON(!buffer_mapped(bh));
2927         BUG_ON(!bh->b_end_io);
2928
2929         if (buffer_ordered(bh) && (rw == WRITE))
2930                 rw = WRITE_BARRIER;
2931
2932         /*
2933          * Only clear out a write error when rewriting, should this
2934          * include WRITE_SYNC as well?
2935          */
2936         if (test_set_buffer_req(bh) && (rw == WRITE || rw == WRITE_BARRIER))
2937                 clear_buffer_write_io_error(bh);
2938
2939         /*
2940          * from here on down, it's all bio -- do the initial mapping,
2941          * submit_bio -> generic_make_request may further map this bio around
2942          */
2943         bio = bio_alloc(GFP_NOIO, 1);
2944
2945         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2946         bio->bi_bdev = bh->b_bdev;
2947         bio->bi_io_vec[0].bv_page = bh->b_page;
2948         bio->bi_io_vec[0].bv_len = bh->b_size;
2949         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2950
2951         bio->bi_vcnt = 1;
2952         bio->bi_idx = 0;
2953         bio->bi_size = bh->b_size;
2954
2955         bio->bi_end_io = end_bio_bh_io_sync;
2956         bio->bi_private = bh;
2957
2958         bio_get(bio);
2959         submit_bio(rw, bio);
2960
2961         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2962                 ret = -EOPNOTSUPP;
2963
2964         bio_put(bio);
2965         return ret;
2966 }
2967
2968 /**
2969  * ll_rw_block: low-level access to block devices (DEPRECATED)
2970  * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
2971  * @nr: number of &struct buffer_heads in the array
2972  * @bhs: array of pointers to &struct buffer_head
2973  *
2974  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2975  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2976  * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2977  * are sent to disk. The fourth %READA option is described in the documentation
2978  * for generic_make_request() which ll_rw_block() calls.
2979  *
2980  * This function drops any buffer that it cannot get a lock on (with the
2981  * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
2982  * clean when doing a write request, and any buffer that appears to be
2983  * up-to-date when doing read request.  Further it marks as clean buffers that
2984  * are processed for writing (the buffer cache won't assume that they are
2985  * actually clean until the buffer gets unlocked).
2986  *
2987  * ll_rw_block sets b_end_io to simple completion handler that marks
2988  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2989  * any waiters. 
2990  *
2991  * All of the buffers must be for the same device, and must also be a
2992  * multiple of the current approved size for the device.
2993  */
2994 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2995 {
2996         int i;
2997
2998         for (i = 0; i < nr; i++) {
2999                 struct buffer_head *bh = bhs[i];
3000
3001                 if (rw == SWRITE || rw == SWRITE_SYNC)
3002                         lock_buffer(bh);
3003                 else if (test_set_buffer_locked(bh))
3004                         continue;
3005
3006                 if (rw == WRITE || rw == SWRITE || rw == SWRITE_SYNC) {
3007                         if (test_clear_buffer_dirty(bh)) {
3008                                 bh->b_end_io = end_buffer_write_sync;
3009                                 get_bh(bh);
3010                                 if (rw == SWRITE_SYNC)
3011                                         submit_bh(WRITE_SYNC, bh);
3012                                 else
3013                                         submit_bh(WRITE, bh);
3014                                 continue;
3015                         }
3016                 } else {
3017                         if (!buffer_uptodate(bh)) {
3018                                 bh->b_end_io = end_buffer_read_sync;
3019                                 get_bh(bh);
3020                                 submit_bh(rw, bh);
3021                                 continue;
3022                         }
3023                 }
3024                 unlock_buffer(bh);
3025         }
3026 }
3027
3028 /*
3029  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3030  * and then start new I/O and then wait upon it.  The caller must have a ref on
3031  * the buffer_head.
3032  */
3033 int sync_dirty_buffer(struct buffer_head *bh)
3034 {
3035         int ret = 0;
3036
3037         WARN_ON(atomic_read(&bh->b_count) < 1);
3038         lock_buffer(bh);
3039         if (test_clear_buffer_dirty(bh)) {
3040                 get_bh(bh);
3041                 bh->b_end_io = end_buffer_write_sync;
3042                 ret = submit_bh(WRITE_SYNC, bh);
3043                 wait_on_buffer(bh);
3044                 if (buffer_eopnotsupp(bh)) {
3045                         clear_buffer_eopnotsupp(bh);
3046                         ret = -EOPNOTSUPP;
3047                 }
3048                 if (!ret && !buffer_uptodate(bh))
3049                         ret = -EIO;
3050         } else {
3051                 unlock_buffer(bh);
3052         }
3053         return ret;
3054 }
3055
3056 /*
3057  * try_to_free_buffers() checks if all the buffers on this particular page
3058  * are unused, and releases them if so.
3059  *
3060  * Exclusion against try_to_free_buffers may be obtained by either
3061  * locking the page or by holding its mapping's private_lock.
3062  *
3063  * If the page is dirty but all the buffers are clean then we need to
3064  * be sure to mark the page clean as well.  This is because the page
3065  * may be against a block device, and a later reattachment of buffers
3066  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3067  * filesystem data on the same device.
3068  *
3069  * The same applies to regular filesystem pages: if all the buffers are
3070  * clean then we set the page clean and proceed.  To do that, we require
3071  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3072  * private_lock.
3073  *
3074  * try_to_free_buffers() is non-blocking.
3075  */
3076 static inline int buffer_busy(struct buffer_head *bh)
3077 {
3078         return atomic_read(&bh->b_count) |
3079                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3080 }
3081
3082 static int
3083 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3084 {
3085         struct buffer_head *head = page_buffers(page);
3086         struct buffer_head *bh;
3087
3088         bh = head;
3089         do {
3090                 if (buffer_write_io_error(bh) && page->mapping)
3091                         set_bit(AS_EIO, &page->mapping->flags);
3092                 if (buffer_busy(bh))
3093                         goto failed;
3094                 bh = bh->b_this_page;
3095         } while (bh != head);
3096
3097         do {
3098                 struct buffer_head *next = bh->b_this_page;
3099
3100                 if (bh->b_assoc_map)
3101                         __remove_assoc_queue(bh);
3102                 bh = next;
3103         } while (bh != head);
3104         *buffers_to_free = head;
3105         __clear_page_buffers(page);
3106         return 1;
3107 failed:
3108         return 0;
3109 }
3110
3111 int try_to_free_buffers(struct page *page)
3112 {
3113         struct address_space * const mapping = page->mapping;
3114         struct buffer_head *buffers_to_free = NULL;
3115         int ret = 0;
3116
3117         BUG_ON(!PageLocked(page));
3118         if (PageWriteback(page))
3119                 return 0;
3120
3121         if (mapping == NULL) {          /* can this still happen? */
3122                 ret = drop_buffers(page, &buffers_to_free);
3123                 goto out;
3124         }
3125
3126         spin_lock(&mapping->private_lock);
3127         ret = drop_buffers(page, &buffers_to_free);
3128
3129         /*
3130          * If the filesystem writes its buffers by hand (eg ext3)
3131          * then we can have clean buffers against a dirty page.  We
3132          * clean the page here; otherwise the VM will never notice
3133          * that the filesystem did any IO at all.
3134          *
3135          * Also, during truncate, discard_buffer will have marked all
3136          * the page's buffers clean.  We discover that here and clean
3137          * the page also.
3138          *
3139          * private_lock must be held over this entire operation in order
3140          * to synchronise against __set_page_dirty_buffers and prevent the
3141          * dirty bit from being lost.
3142          */
3143         if (ret)
3144                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3145         spin_unlock(&mapping->private_lock);
3146 out:
3147         if (buffers_to_free) {
3148                 struct buffer_head *bh = buffers_to_free;
3149
3150                 do {
3151                         struct buffer_head *next = bh->b_this_page;
3152                         free_buffer_head(bh);
3153                         bh = next;
3154                 } while (bh != buffers_to_free);
3155         }
3156         return ret;
3157 }
3158 EXPORT_SYMBOL(try_to_free_buffers);
3159
3160 void block_sync_page(struct page *page)
3161 {
3162         struct address_space *mapping;
3163
3164         smp_mb();
3165         mapping = page_mapping(page);
3166         if (mapping)
3167                 blk_run_backing_dev(mapping->backing_dev_info, page);
3168 }
3169
3170 /*
3171  * There are no bdflush tunables left.  But distributions are
3172  * still running obsolete flush daemons, so we terminate them here.
3173  *
3174  * Use of bdflush() is deprecated and will be removed in a future kernel.
3175  * The `pdflush' kernel threads fully replace bdflush daemons and this call.
3176  */
3177 asmlinkage long sys_bdflush(int func, long data)
3178 {
3179         static int msg_count;
3180
3181         if (!capable(CAP_SYS_ADMIN))
3182                 return -EPERM;
3183
3184         if (msg_count < 5) {
3185                 msg_count++;
3186                 printk(KERN_INFO
3187                         "warning: process `%s' used the obsolete bdflush"
3188                         " system call\n", current->comm);
3189                 printk(KERN_INFO "Fix your initscripts?\n");
3190         }
3191
3192         if (func == 1)
3193                 do_exit(0);
3194         return 0;
3195 }
3196
3197 /*
3198  * Buffer-head allocation
3199  */
3200 static struct kmem_cache *bh_cachep;
3201
3202 /*
3203  * Once the number of bh's in the machine exceeds this level, we start
3204  * stripping them in writeback.
3205  */
3206 static int max_buffer_heads;
3207
3208 int buffer_heads_over_limit;
3209
3210 struct bh_accounting {
3211         int nr;                 /* Number of live bh's */
3212         int ratelimit;          /* Limit cacheline bouncing */
3213 };
3214
3215 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3216
3217 static void recalc_bh_state(void)
3218 {
3219         int i;
3220         int tot = 0;
3221
3222         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3223                 return;
3224         __get_cpu_var(bh_accounting).ratelimit = 0;
3225         for_each_online_cpu(i)
3226                 tot += per_cpu(bh_accounting, i).nr;
3227         buffer_heads_over_limit = (tot > max_buffer_heads);
3228 }
3229         
3230 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3231 {
3232         struct buffer_head *ret = kmem_cache_alloc(bh_cachep, gfp_flags);
3233         if (ret) {
3234                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3235                 get_cpu_var(bh_accounting).nr++;
3236                 recalc_bh_state();
3237                 put_cpu_var(bh_accounting);
3238         }
3239         return ret;
3240 }
3241 EXPORT_SYMBOL(alloc_buffer_head);
3242
3243 void free_buffer_head(struct buffer_head *bh)
3244 {
3245         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3246         kmem_cache_free(bh_cachep, bh);
3247         get_cpu_var(bh_accounting).nr--;
3248         recalc_bh_state();
3249         put_cpu_var(bh_accounting);
3250 }
3251 EXPORT_SYMBOL(free_buffer_head);
3252
3253 static void buffer_exit_cpu(int cpu)
3254 {
3255         int i;
3256         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3257
3258         for (i = 0; i < BH_LRU_SIZE; i++) {
3259                 brelse(b->bhs[i]);
3260                 b->bhs[i] = NULL;
3261         }
3262         get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3263         per_cpu(bh_accounting, cpu).nr = 0;
3264         put_cpu_var(bh_accounting);
3265 }
3266
3267 static int buffer_cpu_notify(struct notifier_block *self,
3268                               unsigned long action, void *hcpu)
3269 {
3270         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3271                 buffer_exit_cpu((unsigned long)hcpu);
3272         return NOTIFY_OK;
3273 }
3274
3275 /**
3276  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3277  * @bh: struct buffer_head
3278  *
3279  * Return true if the buffer is up-to-date and false,
3280  * with the buffer locked, if not.
3281  */
3282 int bh_uptodate_or_lock(struct buffer_head *bh)
3283 {
3284         if (!buffer_uptodate(bh)) {
3285                 lock_buffer(bh);
3286                 if (!buffer_uptodate(bh))
3287                         return 0;
3288                 unlock_buffer(bh);
3289         }
3290         return 1;
3291 }
3292 EXPORT_SYMBOL(bh_uptodate_or_lock);
3293
3294 /**
3295  * bh_submit_read - Submit a locked buffer for reading
3296  * @bh: struct buffer_head
3297  *
3298  * Returns zero on success and -EIO on error.
3299  */
3300 int bh_submit_read(struct buffer_head *bh)
3301 {
3302         BUG_ON(!buffer_locked(bh));
3303
3304         if (buffer_uptodate(bh)) {
3305                 unlock_buffer(bh);
3306                 return 0;
3307         }
3308
3309         get_bh(bh);
3310         bh->b_end_io = end_buffer_read_sync;
3311         submit_bh(READ, bh);
3312         wait_on_buffer(bh);
3313         if (buffer_uptodate(bh))
3314                 return 0;
3315         return -EIO;
3316 }
3317 EXPORT_SYMBOL(bh_submit_read);
3318
3319 static void
3320 init_buffer_head(void *data)
3321 {
3322         struct buffer_head *bh = data;
3323
3324         memset(bh, 0, sizeof(*bh));
3325         INIT_LIST_HEAD(&bh->b_assoc_buffers);
3326 }
3327
3328 void __init buffer_init(void)
3329 {
3330         int nrpages;
3331
3332         bh_cachep = kmem_cache_create("buffer_head",
3333                         sizeof(struct buffer_head), 0,
3334                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3335                                 SLAB_MEM_SPREAD),
3336                                 init_buffer_head);
3337
3338         /*
3339          * Limit the bh occupancy to 10% of ZONE_NORMAL
3340          */
3341         nrpages = (nr_free_buffer_pages() * 10) / 100;
3342         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3343         hotcpu_notifier(buffer_cpu_notify, 0);
3344 }
3345
3346 EXPORT_SYMBOL(__bforget);
3347 EXPORT_SYMBOL(__brelse);
3348 EXPORT_SYMBOL(__wait_on_buffer);
3349 EXPORT_SYMBOL(block_commit_write);
3350 EXPORT_SYMBOL(block_prepare_write);
3351 EXPORT_SYMBOL(block_page_mkwrite);
3352 EXPORT_SYMBOL(block_read_full_page);
3353 EXPORT_SYMBOL(block_sync_page);
3354 EXPORT_SYMBOL(block_truncate_page);
3355 EXPORT_SYMBOL(block_write_full_page);
3356 EXPORT_SYMBOL(cont_write_begin);
3357 EXPORT_SYMBOL(end_buffer_read_sync);
3358 EXPORT_SYMBOL(end_buffer_write_sync);
3359 EXPORT_SYMBOL(file_fsync);
3360 EXPORT_SYMBOL(fsync_bdev);
3361 EXPORT_SYMBOL(generic_block_bmap);
3362 EXPORT_SYMBOL(generic_cont_expand_simple);
3363 EXPORT_SYMBOL(init_buffer);
3364 EXPORT_SYMBOL(invalidate_bdev);
3365 EXPORT_SYMBOL(ll_rw_block);
3366 EXPORT_SYMBOL(mark_buffer_dirty);
3367 EXPORT_SYMBOL(submit_bh);
3368 EXPORT_SYMBOL(sync_dirty_buffer);
3369 EXPORT_SYMBOL(unlock_buffer);