Merge branch 'for-linus' of master.kernel.org:/home/rmk/linux-2.6-arm
[linux-2.6] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/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         clear_bit_unlock(BH_Lock, &bh->b_state);
80         smp_mb__after_clear_bit();
81         wake_up_bit(&bh->b_state, BH_Lock);
82 }
83
84 /*
85  * Block until a buffer comes unlocked.  This doesn't stop it
86  * from becoming locked again - you have to lock it yourself
87  * if you want to preserve its state.
88  */
89 void __wait_on_buffer(struct buffer_head * bh)
90 {
91         wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
92 }
93
94 static void
95 __clear_page_buffers(struct page *page)
96 {
97         ClearPagePrivate(page);
98         set_page_private(page, 0);
99         page_cache_release(page);
100 }
101
102 static void buffer_io_error(struct buffer_head *bh)
103 {
104         char b[BDEVNAME_SIZE];
105
106         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
107                         bdevname(bh->b_bdev, b),
108                         (unsigned long long)bh->b_blocknr);
109 }
110
111 /*
112  * End-of-IO handler helper function which does not touch the bh after
113  * unlocking it.
114  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
115  * a race there is benign: unlock_buffer() only use the bh's address for
116  * hashing after unlocking the buffer, so it doesn't actually touch the bh
117  * itself.
118  */
119 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
120 {
121         if (uptodate) {
122                 set_buffer_uptodate(bh);
123         } else {
124                 /* This happens, due to failed READA attempts. */
125                 clear_buffer_uptodate(bh);
126         }
127         unlock_buffer(bh);
128 }
129
130 /*
131  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
132  * unlock the buffer. This is what ll_rw_block uses too.
133  */
134 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
135 {
136         __end_buffer_read_notouch(bh, uptodate);
137         put_bh(bh);
138 }
139
140 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
141 {
142         char b[BDEVNAME_SIZE];
143
144         if (uptodate) {
145                 set_buffer_uptodate(bh);
146         } else {
147                 if (!buffer_eopnotsupp(bh) && printk_ratelimit()) {
148                         buffer_io_error(bh);
149                         printk(KERN_WARNING "lost page write due to "
150                                         "I/O error on %s\n",
151                                        bdevname(bh->b_bdev, b));
152                 }
153                 set_buffer_write_io_error(bh);
154                 clear_buffer_uptodate(bh);
155         }
156         unlock_buffer(bh);
157         put_bh(bh);
158 }
159
160 /*
161  * Write out and wait upon all the dirty data associated with a block
162  * device via its mapping.  Does not take the superblock lock.
163  */
164 int sync_blockdev(struct block_device *bdev)
165 {
166         int ret = 0;
167
168         if (bdev)
169                 ret = filemap_write_and_wait(bdev->bd_inode->i_mapping);
170         return ret;
171 }
172 EXPORT_SYMBOL(sync_blockdev);
173
174 /*
175  * Write out and wait upon all dirty data associated with this
176  * device.   Filesystem data as well as the underlying block
177  * device.  Takes the superblock lock.
178  */
179 int fsync_bdev(struct block_device *bdev)
180 {
181         struct super_block *sb = get_super(bdev);
182         if (sb) {
183                 int res = fsync_super(sb);
184                 drop_super(sb);
185                 return res;
186         }
187         return sync_blockdev(bdev);
188 }
189
190 /**
191  * freeze_bdev  --  lock a filesystem and force it into a consistent state
192  * @bdev:       blockdevice to lock
193  *
194  * This takes the block device bd_mount_sem to make sure no new mounts
195  * happen on bdev until thaw_bdev() is called.
196  * If a superblock is found on this device, we take the s_umount semaphore
197  * on it to make sure nobody unmounts until the snapshot creation is done.
198  */
199 struct super_block *freeze_bdev(struct block_device *bdev)
200 {
201         struct super_block *sb;
202
203         down(&bdev->bd_mount_sem);
204         sb = get_super(bdev);
205         if (sb && !(sb->s_flags & MS_RDONLY)) {
206                 sb->s_frozen = SB_FREEZE_WRITE;
207                 smp_wmb();
208
209                 __fsync_super(sb);
210
211                 sb->s_frozen = SB_FREEZE_TRANS;
212                 smp_wmb();
213
214                 sync_blockdev(sb->s_bdev);
215
216                 if (sb->s_op->write_super_lockfs)
217                         sb->s_op->write_super_lockfs(sb);
218         }
219
220         sync_blockdev(bdev);
221         return sb;      /* thaw_bdev releases s->s_umount and bd_mount_sem */
222 }
223 EXPORT_SYMBOL(freeze_bdev);
224
225 /**
226  * thaw_bdev  -- unlock filesystem
227  * @bdev:       blockdevice to unlock
228  * @sb:         associated superblock
229  *
230  * Unlocks the filesystem and marks it writeable again after freeze_bdev().
231  */
232 void thaw_bdev(struct block_device *bdev, struct super_block *sb)
233 {
234         if (sb) {
235                 BUG_ON(sb->s_bdev != bdev);
236
237                 if (sb->s_op->unlockfs)
238                         sb->s_op->unlockfs(sb);
239                 sb->s_frozen = SB_UNFROZEN;
240                 smp_wmb();
241                 wake_up(&sb->s_wait_unfrozen);
242                 drop_super(sb);
243         }
244
245         up(&bdev->bd_mount_sem);
246 }
247 EXPORT_SYMBOL(thaw_bdev);
248
249 /*
250  * Various filesystems appear to want __find_get_block to be non-blocking.
251  * But it's the page lock which protects the buffers.  To get around this,
252  * we get exclusion from try_to_free_buffers with the blockdev mapping's
253  * private_lock.
254  *
255  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
256  * may be quite high.  This code could TryLock the page, and if that
257  * succeeds, there is no need to take private_lock. (But if
258  * private_lock is contended then so is mapping->tree_lock).
259  */
260 static struct buffer_head *
261 __find_get_block_slow(struct block_device *bdev, sector_t block)
262 {
263         struct inode *bd_inode = bdev->bd_inode;
264         struct address_space *bd_mapping = bd_inode->i_mapping;
265         struct buffer_head *ret = NULL;
266         pgoff_t index;
267         struct buffer_head *bh;
268         struct buffer_head *head;
269         struct page *page;
270         int all_mapped = 1;
271
272         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
273         page = find_get_page(bd_mapping, index);
274         if (!page)
275                 goto out;
276
277         spin_lock(&bd_mapping->private_lock);
278         if (!page_has_buffers(page))
279                 goto out_unlock;
280         head = page_buffers(page);
281         bh = head;
282         do {
283                 if (bh->b_blocknr == block) {
284                         ret = bh;
285                         get_bh(bh);
286                         goto out_unlock;
287                 }
288                 if (!buffer_mapped(bh))
289                         all_mapped = 0;
290                 bh = bh->b_this_page;
291         } while (bh != head);
292
293         /* we might be here because some of the buffers on this page are
294          * not mapped.  This is due to various races between
295          * file io on the block device and getblk.  It gets dealt with
296          * elsewhere, don't buffer_error if we had some unmapped buffers
297          */
298         if (all_mapped) {
299                 printk("__find_get_block_slow() failed. "
300                         "block=%llu, b_blocknr=%llu\n",
301                         (unsigned long long)block,
302                         (unsigned long long)bh->b_blocknr);
303                 printk("b_state=0x%08lx, b_size=%zu\n",
304                         bh->b_state, bh->b_size);
305                 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
306         }
307 out_unlock:
308         spin_unlock(&bd_mapping->private_lock);
309         page_cache_release(page);
310 out:
311         return ret;
312 }
313
314 /* If invalidate_buffers() will trash dirty buffers, it means some kind
315    of fs corruption is going on. Trashing dirty data always imply losing
316    information that was supposed to be just stored on the physical layer
317    by the user.
318
319    Thus invalidate_buffers in general usage is not allwowed to trash
320    dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
321    be preserved.  These buffers are simply skipped.
322   
323    We also skip buffers which are still in use.  For example this can
324    happen if a userspace program is reading the block device.
325
326    NOTE: In the case where the user removed a removable-media-disk even if
327    there's still dirty data not synced on disk (due a bug in the device driver
328    or due an error of the user), by not destroying the dirty buffers we could
329    generate corruption also on the next media inserted, thus a parameter is
330    necessary to handle this case in the most safe way possible (trying
331    to not corrupt also the new disk inserted with the data belonging to
332    the old now corrupted disk). Also for the ramdisk the natural thing
333    to do in order to release the ramdisk memory is to destroy dirty buffers.
334
335    These are two special cases. Normal usage imply the device driver
336    to issue a sync on the device (without waiting I/O completion) and
337    then an invalidate_buffers call that doesn't trash dirty buffers.
338
339    For handling cache coherency with the blkdev pagecache the 'update' case
340    is been introduced. It is needed to re-read from disk any pinned
341    buffer. NOTE: re-reading from disk is destructive so we can do it only
342    when we assume nobody is changing the buffercache under our I/O and when
343    we think the disk contains more recent information than the buffercache.
344    The update == 1 pass marks the buffers we need to update, the update == 2
345    pass does the actual I/O. */
346 void invalidate_bdev(struct block_device *bdev)
347 {
348         struct address_space *mapping = bdev->bd_inode->i_mapping;
349
350         if (mapping->nrpages == 0)
351                 return;
352
353         invalidate_bh_lrus();
354         invalidate_mapping_pages(mapping, 0, -1);
355 }
356
357 /*
358  * Kick pdflush then try to free up some ZONE_NORMAL memory.
359  */
360 static void free_more_memory(void)
361 {
362         struct zone *zone;
363         int nid;
364
365         wakeup_pdflush(1024);
366         yield();
367
368         for_each_online_node(nid) {
369                 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
370                                                 gfp_zone(GFP_NOFS), NULL,
371                                                 &zone);
372                 if (zone)
373                         try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
374                                                 GFP_NOFS);
375         }
376 }
377
378 /*
379  * I/O completion handler for block_read_full_page() - pages
380  * which come unlocked at the end of I/O.
381  */
382 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
383 {
384         unsigned long flags;
385         struct buffer_head *first;
386         struct buffer_head *tmp;
387         struct page *page;
388         int page_uptodate = 1;
389
390         BUG_ON(!buffer_async_read(bh));
391
392         page = bh->b_page;
393         if (uptodate) {
394                 set_buffer_uptodate(bh);
395         } else {
396                 clear_buffer_uptodate(bh);
397                 if (printk_ratelimit())
398                         buffer_io_error(bh);
399                 SetPageError(page);
400         }
401
402         /*
403          * Be _very_ careful from here on. Bad things can happen if
404          * two buffer heads end IO at almost the same time and both
405          * decide that the page is now completely done.
406          */
407         first = page_buffers(page);
408         local_irq_save(flags);
409         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
410         clear_buffer_async_read(bh);
411         unlock_buffer(bh);
412         tmp = bh;
413         do {
414                 if (!buffer_uptodate(tmp))
415                         page_uptodate = 0;
416                 if (buffer_async_read(tmp)) {
417                         BUG_ON(!buffer_locked(tmp));
418                         goto still_busy;
419                 }
420                 tmp = tmp->b_this_page;
421         } while (tmp != bh);
422         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
423         local_irq_restore(flags);
424
425         /*
426          * If none of the buffers had errors and they are all
427          * uptodate then we can set the page uptodate.
428          */
429         if (page_uptodate && !PageError(page))
430                 SetPageUptodate(page);
431         unlock_page(page);
432         return;
433
434 still_busy:
435         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
436         local_irq_restore(flags);
437         return;
438 }
439
440 /*
441  * Completion handler for block_write_full_page() - pages which are unlocked
442  * during I/O, and which have PageWriteback cleared upon I/O completion.
443  */
444 static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
445 {
446         char b[BDEVNAME_SIZE];
447         unsigned long flags;
448         struct buffer_head *first;
449         struct buffer_head *tmp;
450         struct page *page;
451
452         BUG_ON(!buffer_async_write(bh));
453
454         page = bh->b_page;
455         if (uptodate) {
456                 set_buffer_uptodate(bh);
457         } else {
458                 if (printk_ratelimit()) {
459                         buffer_io_error(bh);
460                         printk(KERN_WARNING "lost page write due to "
461                                         "I/O error on %s\n",
462                                bdevname(bh->b_bdev, b));
463                 }
464                 set_bit(AS_EIO, &page->mapping->flags);
465                 set_buffer_write_io_error(bh);
466                 clear_buffer_uptodate(bh);
467                 SetPageError(page);
468         }
469
470         first = page_buffers(page);
471         local_irq_save(flags);
472         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
473
474         clear_buffer_async_write(bh);
475         unlock_buffer(bh);
476         tmp = bh->b_this_page;
477         while (tmp != bh) {
478                 if (buffer_async_write(tmp)) {
479                         BUG_ON(!buffer_locked(tmp));
480                         goto still_busy;
481                 }
482                 tmp = tmp->b_this_page;
483         }
484         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
485         local_irq_restore(flags);
486         end_page_writeback(page);
487         return;
488
489 still_busy:
490         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
491         local_irq_restore(flags);
492         return;
493 }
494
495 /*
496  * If a page's buffers are under async readin (end_buffer_async_read
497  * completion) then there is a possibility that another thread of
498  * control could lock one of the buffers after it has completed
499  * but while some of the other buffers have not completed.  This
500  * locked buffer would confuse end_buffer_async_read() into not unlocking
501  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
502  * that this buffer is not under async I/O.
503  *
504  * The page comes unlocked when it has no locked buffer_async buffers
505  * left.
506  *
507  * PageLocked prevents anyone starting new async I/O reads any of
508  * the buffers.
509  *
510  * PageWriteback is used to prevent simultaneous writeout of the same
511  * page.
512  *
513  * PageLocked prevents anyone from starting writeback of a page which is
514  * under read I/O (PageWriteback is only ever set against a locked page).
515  */
516 static void mark_buffer_async_read(struct buffer_head *bh)
517 {
518         bh->b_end_io = end_buffer_async_read;
519         set_buffer_async_read(bh);
520 }
521
522 void mark_buffer_async_write(struct buffer_head *bh)
523 {
524         bh->b_end_io = end_buffer_async_write;
525         set_buffer_async_write(bh);
526 }
527 EXPORT_SYMBOL(mark_buffer_async_write);
528
529
530 /*
531  * fs/buffer.c contains helper functions for buffer-backed address space's
532  * fsync functions.  A common requirement for buffer-based filesystems is
533  * that certain data from the backing blockdev needs to be written out for
534  * a successful fsync().  For example, ext2 indirect blocks need to be
535  * written back and waited upon before fsync() returns.
536  *
537  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
538  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
539  * management of a list of dependent buffers at ->i_mapping->private_list.
540  *
541  * Locking is a little subtle: try_to_free_buffers() will remove buffers
542  * from their controlling inode's queue when they are being freed.  But
543  * try_to_free_buffers() will be operating against the *blockdev* mapping
544  * at the time, not against the S_ISREG file which depends on those buffers.
545  * So the locking for private_list is via the private_lock in the address_space
546  * which backs the buffers.  Which is different from the address_space 
547  * against which the buffers are listed.  So for a particular address_space,
548  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
549  * mapping->private_list will always be protected by the backing blockdev's
550  * ->private_lock.
551  *
552  * Which introduces a requirement: all buffers on an address_space's
553  * ->private_list must be from the same address_space: the blockdev's.
554  *
555  * address_spaces which do not place buffers at ->private_list via these
556  * utility functions are free to use private_lock and private_list for
557  * whatever they want.  The only requirement is that list_empty(private_list)
558  * be true at clear_inode() time.
559  *
560  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
561  * filesystems should do that.  invalidate_inode_buffers() should just go
562  * BUG_ON(!list_empty).
563  *
564  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
565  * take an address_space, not an inode.  And it should be called
566  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
567  * queued up.
568  *
569  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
570  * list if it is already on a list.  Because if the buffer is on a list,
571  * it *must* already be on the right one.  If not, the filesystem is being
572  * silly.  This will save a ton of locking.  But first we have to ensure
573  * that buffers are taken *off* the old inode's list when they are freed
574  * (presumably in truncate).  That requires careful auditing of all
575  * filesystems (do it inside bforget()).  It could also be done by bringing
576  * b_inode back.
577  */
578
579 /*
580  * The buffer's backing address_space's private_lock must be held
581  */
582 static void __remove_assoc_queue(struct buffer_head *bh)
583 {
584         list_del_init(&bh->b_assoc_buffers);
585         WARN_ON(!bh->b_assoc_map);
586         if (buffer_write_io_error(bh))
587                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
588         bh->b_assoc_map = NULL;
589 }
590
591 int inode_has_buffers(struct inode *inode)
592 {
593         return !list_empty(&inode->i_data.private_list);
594 }
595
596 /*
597  * osync is designed to support O_SYNC io.  It waits synchronously for
598  * all already-submitted IO to complete, but does not queue any new
599  * writes to the disk.
600  *
601  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
602  * you dirty the buffers, and then use osync_inode_buffers to wait for
603  * completion.  Any other dirty buffers which are not yet queued for
604  * write will not be flushed to disk by the osync.
605  */
606 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
607 {
608         struct buffer_head *bh;
609         struct list_head *p;
610         int err = 0;
611
612         spin_lock(lock);
613 repeat:
614         list_for_each_prev(p, list) {
615                 bh = BH_ENTRY(p);
616                 if (buffer_locked(bh)) {
617                         get_bh(bh);
618                         spin_unlock(lock);
619                         wait_on_buffer(bh);
620                         if (!buffer_uptodate(bh))
621                                 err = -EIO;
622                         brelse(bh);
623                         spin_lock(lock);
624                         goto repeat;
625                 }
626         }
627         spin_unlock(lock);
628         return err;
629 }
630
631 /**
632  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
633  * @mapping: the mapping which wants those buffers written
634  *
635  * Starts I/O against the buffers at mapping->private_list, and waits upon
636  * that I/O.
637  *
638  * Basically, this is a convenience function for fsync().
639  * @mapping is a file or directory which needs those buffers to be written for
640  * a successful fsync().
641  */
642 int sync_mapping_buffers(struct address_space *mapping)
643 {
644         struct address_space *buffer_mapping = mapping->assoc_mapping;
645
646         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
647                 return 0;
648
649         return fsync_buffers_list(&buffer_mapping->private_lock,
650                                         &mapping->private_list);
651 }
652 EXPORT_SYMBOL(sync_mapping_buffers);
653
654 /*
655  * Called when we've recently written block `bblock', and it is known that
656  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
657  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
658  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
659  */
660 void write_boundary_block(struct block_device *bdev,
661                         sector_t bblock, unsigned blocksize)
662 {
663         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
664         if (bh) {
665                 if (buffer_dirty(bh))
666                         ll_rw_block(WRITE, 1, &bh);
667                 put_bh(bh);
668         }
669 }
670
671 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
672 {
673         struct address_space *mapping = inode->i_mapping;
674         struct address_space *buffer_mapping = bh->b_page->mapping;
675
676         mark_buffer_dirty(bh);
677         if (!mapping->assoc_mapping) {
678                 mapping->assoc_mapping = buffer_mapping;
679         } else {
680                 BUG_ON(mapping->assoc_mapping != buffer_mapping);
681         }
682         if (!bh->b_assoc_map) {
683                 spin_lock(&buffer_mapping->private_lock);
684                 list_move_tail(&bh->b_assoc_buffers,
685                                 &mapping->private_list);
686                 bh->b_assoc_map = mapping;
687                 spin_unlock(&buffer_mapping->private_lock);
688         }
689 }
690 EXPORT_SYMBOL(mark_buffer_dirty_inode);
691
692 /*
693  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
694  * dirty.
695  *
696  * If warn is true, then emit a warning if the page is not uptodate and has
697  * not been truncated.
698  */
699 static int __set_page_dirty(struct page *page,
700                 struct address_space *mapping, int warn)
701 {
702         if (unlikely(!mapping))
703                 return !TestSetPageDirty(page);
704
705         if (TestSetPageDirty(page))
706                 return 0;
707
708         spin_lock_irq(&mapping->tree_lock);
709         if (page->mapping) {    /* Race with truncate? */
710                 WARN_ON_ONCE(warn && !PageUptodate(page));
711
712                 if (mapping_cap_account_dirty(mapping)) {
713                         __inc_zone_page_state(page, NR_FILE_DIRTY);
714                         __inc_bdi_stat(mapping->backing_dev_info,
715                                         BDI_RECLAIMABLE);
716                         task_io_account_write(PAGE_CACHE_SIZE);
717                 }
718                 radix_tree_tag_set(&mapping->page_tree,
719                                 page_index(page), PAGECACHE_TAG_DIRTY);
720         }
721         spin_unlock_irq(&mapping->tree_lock);
722         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
723
724         return 1;
725 }
726
727 /*
728  * Add a page to the dirty page list.
729  *
730  * It is a sad fact of life that this function is called from several places
731  * deeply under spinlocking.  It may not sleep.
732  *
733  * If the page has buffers, the uptodate buffers are set dirty, to preserve
734  * dirty-state coherency between the page and the buffers.  It the page does
735  * not have buffers then when they are later attached they will all be set
736  * dirty.
737  *
738  * The buffers are dirtied before the page is dirtied.  There's a small race
739  * window in which a writepage caller may see the page cleanness but not the
740  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
741  * before the buffers, a concurrent writepage caller could clear the page dirty
742  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
743  * page on the dirty page list.
744  *
745  * We use private_lock to lock against try_to_free_buffers while using the
746  * page's buffer list.  Also use this to protect against clean buffers being
747  * added to the page after it was set dirty.
748  *
749  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
750  * address_space though.
751  */
752 int __set_page_dirty_buffers(struct page *page)
753 {
754         struct address_space *mapping = page_mapping(page);
755
756         if (unlikely(!mapping))
757                 return !TestSetPageDirty(page);
758
759         spin_lock(&mapping->private_lock);
760         if (page_has_buffers(page)) {
761                 struct buffer_head *head = page_buffers(page);
762                 struct buffer_head *bh = head;
763
764                 do {
765                         set_buffer_dirty(bh);
766                         bh = bh->b_this_page;
767                 } while (bh != head);
768         }
769         spin_unlock(&mapping->private_lock);
770
771         return __set_page_dirty(page, mapping, 1);
772 }
773 EXPORT_SYMBOL(__set_page_dirty_buffers);
774
775 /*
776  * Write out and wait upon a list of buffers.
777  *
778  * We have conflicting pressures: we want to make sure that all
779  * initially dirty buffers get waited on, but that any subsequently
780  * dirtied buffers don't.  After all, we don't want fsync to last
781  * forever if somebody is actively writing to the file.
782  *
783  * Do this in two main stages: first we copy dirty buffers to a
784  * temporary inode list, queueing the writes as we go.  Then we clean
785  * up, waiting for those writes to complete.
786  * 
787  * During this second stage, any subsequent updates to the file may end
788  * up refiling the buffer on the original inode's dirty list again, so
789  * there is a chance we will end up with a buffer queued for write but
790  * not yet completed on that list.  So, as a final cleanup we go through
791  * the osync code to catch these locked, dirty buffers without requeuing
792  * any newly dirty buffers for write.
793  */
794 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
795 {
796         struct buffer_head *bh;
797         struct list_head tmp;
798         struct address_space *mapping;
799         int err = 0, err2;
800
801         INIT_LIST_HEAD(&tmp);
802
803         spin_lock(lock);
804         while (!list_empty(list)) {
805                 bh = BH_ENTRY(list->next);
806                 mapping = bh->b_assoc_map;
807                 __remove_assoc_queue(bh);
808                 /* Avoid race with mark_buffer_dirty_inode() which does
809                  * a lockless check and we rely on seeing the dirty bit */
810                 smp_mb();
811                 if (buffer_dirty(bh) || buffer_locked(bh)) {
812                         list_add(&bh->b_assoc_buffers, &tmp);
813                         bh->b_assoc_map = mapping;
814                         if (buffer_dirty(bh)) {
815                                 get_bh(bh);
816                                 spin_unlock(lock);
817                                 /*
818                                  * Ensure any pending I/O completes so that
819                                  * ll_rw_block() actually writes the current
820                                  * contents - it is a noop if I/O is still in
821                                  * flight on potentially older contents.
822                                  */
823                                 ll_rw_block(SWRITE_SYNC, 1, &bh);
824                                 brelse(bh);
825                                 spin_lock(lock);
826                         }
827                 }
828         }
829
830         while (!list_empty(&tmp)) {
831                 bh = BH_ENTRY(tmp.prev);
832                 get_bh(bh);
833                 mapping = bh->b_assoc_map;
834                 __remove_assoc_queue(bh);
835                 /* Avoid race with mark_buffer_dirty_inode() which does
836                  * a lockless check and we rely on seeing the dirty bit */
837                 smp_mb();
838                 if (buffer_dirty(bh)) {
839                         list_add(&bh->b_assoc_buffers,
840                                  &mapping->private_list);
841                         bh->b_assoc_map = mapping;
842                 }
843                 spin_unlock(lock);
844                 wait_on_buffer(bh);
845                 if (!buffer_uptodate(bh))
846                         err = -EIO;
847                 brelse(bh);
848                 spin_lock(lock);
849         }
850         
851         spin_unlock(lock);
852         err2 = osync_buffers_list(lock, list);
853         if (err)
854                 return err;
855         else
856                 return err2;
857 }
858
859 /*
860  * Invalidate any and all dirty buffers on a given inode.  We are
861  * probably unmounting the fs, but that doesn't mean we have already
862  * done a sync().  Just drop the buffers from the inode list.
863  *
864  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
865  * assumes that all the buffers are against the blockdev.  Not true
866  * for reiserfs.
867  */
868 void invalidate_inode_buffers(struct inode *inode)
869 {
870         if (inode_has_buffers(inode)) {
871                 struct address_space *mapping = &inode->i_data;
872                 struct list_head *list = &mapping->private_list;
873                 struct address_space *buffer_mapping = mapping->assoc_mapping;
874
875                 spin_lock(&buffer_mapping->private_lock);
876                 while (!list_empty(list))
877                         __remove_assoc_queue(BH_ENTRY(list->next));
878                 spin_unlock(&buffer_mapping->private_lock);
879         }
880 }
881
882 /*
883  * Remove any clean buffers from the inode's buffer list.  This is called
884  * when we're trying to free the inode itself.  Those buffers can pin it.
885  *
886  * Returns true if all buffers were removed.
887  */
888 int remove_inode_buffers(struct inode *inode)
889 {
890         int ret = 1;
891
892         if (inode_has_buffers(inode)) {
893                 struct address_space *mapping = &inode->i_data;
894                 struct list_head *list = &mapping->private_list;
895                 struct address_space *buffer_mapping = mapping->assoc_mapping;
896
897                 spin_lock(&buffer_mapping->private_lock);
898                 while (!list_empty(list)) {
899                         struct buffer_head *bh = BH_ENTRY(list->next);
900                         if (buffer_dirty(bh)) {
901                                 ret = 0;
902                                 break;
903                         }
904                         __remove_assoc_queue(bh);
905                 }
906                 spin_unlock(&buffer_mapping->private_lock);
907         }
908         return ret;
909 }
910
911 /*
912  * Create the appropriate buffers when given a page for data area and
913  * the size of each buffer.. Use the bh->b_this_page linked list to
914  * follow the buffers created.  Return NULL if unable to create more
915  * buffers.
916  *
917  * The retry flag is used to differentiate async IO (paging, swapping)
918  * which may not fail from ordinary buffer allocations.
919  */
920 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
921                 int retry)
922 {
923         struct buffer_head *bh, *head;
924         long offset;
925
926 try_again:
927         head = NULL;
928         offset = PAGE_SIZE;
929         while ((offset -= size) >= 0) {
930                 bh = alloc_buffer_head(GFP_NOFS);
931                 if (!bh)
932                         goto no_grow;
933
934                 bh->b_bdev = NULL;
935                 bh->b_this_page = head;
936                 bh->b_blocknr = -1;
937                 head = bh;
938
939                 bh->b_state = 0;
940                 atomic_set(&bh->b_count, 0);
941                 bh->b_private = NULL;
942                 bh->b_size = size;
943
944                 /* Link the buffer to its page */
945                 set_bh_page(bh, page, offset);
946
947                 init_buffer(bh, NULL, NULL);
948         }
949         return head;
950 /*
951  * In case anything failed, we just free everything we got.
952  */
953 no_grow:
954         if (head) {
955                 do {
956                         bh = head;
957                         head = head->b_this_page;
958                         free_buffer_head(bh);
959                 } while (head);
960         }
961
962         /*
963          * Return failure for non-async IO requests.  Async IO requests
964          * are not allowed to fail, so we have to wait until buffer heads
965          * become available.  But we don't want tasks sleeping with 
966          * partially complete buffers, so all were released above.
967          */
968         if (!retry)
969                 return NULL;
970
971         /* We're _really_ low on memory. Now we just
972          * wait for old buffer heads to become free due to
973          * finishing IO.  Since this is an async request and
974          * the reserve list is empty, we're sure there are 
975          * async buffer heads in use.
976          */
977         free_more_memory();
978         goto try_again;
979 }
980 EXPORT_SYMBOL_GPL(alloc_page_buffers);
981
982 static inline void
983 link_dev_buffers(struct page *page, struct buffer_head *head)
984 {
985         struct buffer_head *bh, *tail;
986
987         bh = head;
988         do {
989                 tail = bh;
990                 bh = bh->b_this_page;
991         } while (bh);
992         tail->b_this_page = head;
993         attach_page_buffers(page, head);
994 }
995
996 /*
997  * Initialise the state of a blockdev page's buffers.
998  */ 
999 static void
1000 init_page_buffers(struct page *page, struct block_device *bdev,
1001                         sector_t block, int size)
1002 {
1003         struct buffer_head *head = page_buffers(page);
1004         struct buffer_head *bh = head;
1005         int uptodate = PageUptodate(page);
1006
1007         do {
1008                 if (!buffer_mapped(bh)) {
1009                         init_buffer(bh, NULL, NULL);
1010                         bh->b_bdev = bdev;
1011                         bh->b_blocknr = block;
1012                         if (uptodate)
1013                                 set_buffer_uptodate(bh);
1014                         set_buffer_mapped(bh);
1015                 }
1016                 block++;
1017                 bh = bh->b_this_page;
1018         } while (bh != head);
1019 }
1020
1021 /*
1022  * Create the page-cache page that contains the requested block.
1023  *
1024  * This is user purely for blockdev mappings.
1025  */
1026 static struct page *
1027 grow_dev_page(struct block_device *bdev, sector_t block,
1028                 pgoff_t index, int size)
1029 {
1030         struct inode *inode = bdev->bd_inode;
1031         struct page *page;
1032         struct buffer_head *bh;
1033
1034         page = find_or_create_page(inode->i_mapping, index,
1035                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
1036         if (!page)
1037                 return NULL;
1038
1039         BUG_ON(!PageLocked(page));
1040
1041         if (page_has_buffers(page)) {
1042                 bh = page_buffers(page);
1043                 if (bh->b_size == size) {
1044                         init_page_buffers(page, bdev, block, size);
1045                         return page;
1046                 }
1047                 if (!try_to_free_buffers(page))
1048                         goto failed;
1049         }
1050
1051         /*
1052          * Allocate some buffers for this page
1053          */
1054         bh = alloc_page_buffers(page, size, 0);
1055         if (!bh)
1056                 goto failed;
1057
1058         /*
1059          * Link the page to the buffers and initialise them.  Take the
1060          * lock to be atomic wrt __find_get_block(), which does not
1061          * run under the page lock.
1062          */
1063         spin_lock(&inode->i_mapping->private_lock);
1064         link_dev_buffers(page, bh);
1065         init_page_buffers(page, bdev, block, size);
1066         spin_unlock(&inode->i_mapping->private_lock);
1067         return page;
1068
1069 failed:
1070         BUG();
1071         unlock_page(page);
1072         page_cache_release(page);
1073         return NULL;
1074 }
1075
1076 /*
1077  * Create buffers for the specified block device block's page.  If
1078  * that page was dirty, the buffers are set dirty also.
1079  */
1080 static int
1081 grow_buffers(struct block_device *bdev, sector_t block, int size)
1082 {
1083         struct page *page;
1084         pgoff_t index;
1085         int sizebits;
1086
1087         sizebits = -1;
1088         do {
1089                 sizebits++;
1090         } while ((size << sizebits) < PAGE_SIZE);
1091
1092         index = block >> sizebits;
1093
1094         /*
1095          * Check for a block which wants to lie outside our maximum possible
1096          * pagecache index.  (this comparison is done using sector_t types).
1097          */
1098         if (unlikely(index != block >> sizebits)) {
1099                 char b[BDEVNAME_SIZE];
1100
1101                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1102                         "device %s\n",
1103                         __func__, (unsigned long long)block,
1104                         bdevname(bdev, b));
1105                 return -EIO;
1106         }
1107         block = index << sizebits;
1108         /* Create a page with the proper size buffers.. */
1109         page = grow_dev_page(bdev, block, index, size);
1110         if (!page)
1111                 return 0;
1112         unlock_page(page);
1113         page_cache_release(page);
1114         return 1;
1115 }
1116
1117 static struct buffer_head *
1118 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1119 {
1120         /* Size must be multiple of hard sectorsize */
1121         if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1122                         (size < 512 || size > PAGE_SIZE))) {
1123                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1124                                         size);
1125                 printk(KERN_ERR "hardsect size: %d\n",
1126                                         bdev_hardsect_size(bdev));
1127
1128                 dump_stack();
1129                 return NULL;
1130         }
1131
1132         for (;;) {
1133                 struct buffer_head * bh;
1134                 int ret;
1135
1136                 bh = __find_get_block(bdev, block, size);
1137                 if (bh)
1138                         return bh;
1139
1140                 ret = grow_buffers(bdev, block, size);
1141                 if (ret < 0)
1142                         return NULL;
1143                 if (ret == 0)
1144                         free_more_memory();
1145         }
1146 }
1147
1148 /*
1149  * The relationship between dirty buffers and dirty pages:
1150  *
1151  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1152  * the page is tagged dirty in its radix tree.
1153  *
1154  * At all times, the dirtiness of the buffers represents the dirtiness of
1155  * subsections of the page.  If the page has buffers, the page dirty bit is
1156  * merely a hint about the true dirty state.
1157  *
1158  * When a page is set dirty in its entirety, all its buffers are marked dirty
1159  * (if the page has buffers).
1160  *
1161  * When a buffer is marked dirty, its page is dirtied, but the page's other
1162  * buffers are not.
1163  *
1164  * Also.  When blockdev buffers are explicitly read with bread(), they
1165  * individually become uptodate.  But their backing page remains not
1166  * uptodate - even if all of its buffers are uptodate.  A subsequent
1167  * block_read_full_page() against that page will discover all the uptodate
1168  * buffers, will set the page uptodate and will perform no I/O.
1169  */
1170
1171 /**
1172  * mark_buffer_dirty - mark a buffer_head as needing writeout
1173  * @bh: the buffer_head to mark dirty
1174  *
1175  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1176  * backing page dirty, then tag the page as dirty in its address_space's radix
1177  * tree and then attach the address_space's inode to its superblock's dirty
1178  * inode list.
1179  *
1180  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1181  * mapping->tree_lock and the global inode_lock.
1182  */
1183 void mark_buffer_dirty(struct buffer_head *bh)
1184 {
1185         WARN_ON_ONCE(!buffer_uptodate(bh));
1186
1187         /*
1188          * Very *carefully* optimize the it-is-already-dirty case.
1189          *
1190          * Don't let the final "is it dirty" escape to before we
1191          * perhaps modified the buffer.
1192          */
1193         if (buffer_dirty(bh)) {
1194                 smp_mb();
1195                 if (buffer_dirty(bh))
1196                         return;
1197         }
1198
1199         if (!test_set_buffer_dirty(bh))
1200                 __set_page_dirty(bh->b_page, page_mapping(bh->b_page), 0);
1201 }
1202
1203 /*
1204  * Decrement a buffer_head's reference count.  If all buffers against a page
1205  * have zero reference count, are clean and unlocked, and if the page is clean
1206  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1207  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1208  * a page but it ends up not being freed, and buffers may later be reattached).
1209  */
1210 void __brelse(struct buffer_head * buf)
1211 {
1212         if (atomic_read(&buf->b_count)) {
1213                 put_bh(buf);
1214                 return;
1215         }
1216         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1217 }
1218
1219 /*
1220  * bforget() is like brelse(), except it discards any
1221  * potentially dirty data.
1222  */
1223 void __bforget(struct buffer_head *bh)
1224 {
1225         clear_buffer_dirty(bh);
1226         if (bh->b_assoc_map) {
1227                 struct address_space *buffer_mapping = bh->b_page->mapping;
1228
1229                 spin_lock(&buffer_mapping->private_lock);
1230                 list_del_init(&bh->b_assoc_buffers);
1231                 bh->b_assoc_map = NULL;
1232                 spin_unlock(&buffer_mapping->private_lock);
1233         }
1234         __brelse(bh);
1235 }
1236
1237 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1238 {
1239         lock_buffer(bh);
1240         if (buffer_uptodate(bh)) {
1241                 unlock_buffer(bh);
1242                 return bh;
1243         } else {
1244                 get_bh(bh);
1245                 bh->b_end_io = end_buffer_read_sync;
1246                 submit_bh(READ, bh);
1247                 wait_on_buffer(bh);
1248                 if (buffer_uptodate(bh))
1249                         return bh;
1250         }
1251         brelse(bh);
1252         return NULL;
1253 }
1254
1255 /*
1256  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1257  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1258  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1259  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1260  * CPU's LRUs at the same time.
1261  *
1262  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1263  * sb_find_get_block().
1264  *
1265  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1266  * a local interrupt disable for that.
1267  */
1268
1269 #define BH_LRU_SIZE     8
1270
1271 struct bh_lru {
1272         struct buffer_head *bhs[BH_LRU_SIZE];
1273 };
1274
1275 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1276
1277 #ifdef CONFIG_SMP
1278 #define bh_lru_lock()   local_irq_disable()
1279 #define bh_lru_unlock() local_irq_enable()
1280 #else
1281 #define bh_lru_lock()   preempt_disable()
1282 #define bh_lru_unlock() preempt_enable()
1283 #endif
1284
1285 static inline void check_irqs_on(void)
1286 {
1287 #ifdef irqs_disabled
1288         BUG_ON(irqs_disabled());
1289 #endif
1290 }
1291
1292 /*
1293  * The LRU management algorithm is dopey-but-simple.  Sorry.
1294  */
1295 static void bh_lru_install(struct buffer_head *bh)
1296 {
1297         struct buffer_head *evictee = NULL;
1298         struct bh_lru *lru;
1299
1300         check_irqs_on();
1301         bh_lru_lock();
1302         lru = &__get_cpu_var(bh_lrus);
1303         if (lru->bhs[0] != bh) {
1304                 struct buffer_head *bhs[BH_LRU_SIZE];
1305                 int in;
1306                 int out = 0;
1307
1308                 get_bh(bh);
1309                 bhs[out++] = bh;
1310                 for (in = 0; in < BH_LRU_SIZE; in++) {
1311                         struct buffer_head *bh2 = lru->bhs[in];
1312
1313                         if (bh2 == bh) {
1314                                 __brelse(bh2);
1315                         } else {
1316                                 if (out >= BH_LRU_SIZE) {
1317                                         BUG_ON(evictee != NULL);
1318                                         evictee = bh2;
1319                                 } else {
1320                                         bhs[out++] = bh2;
1321                                 }
1322                         }
1323                 }
1324                 while (out < BH_LRU_SIZE)
1325                         bhs[out++] = NULL;
1326                 memcpy(lru->bhs, bhs, sizeof(bhs));
1327         }
1328         bh_lru_unlock();
1329
1330         if (evictee)
1331                 __brelse(evictee);
1332 }
1333
1334 /*
1335  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1336  */
1337 static struct buffer_head *
1338 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1339 {
1340         struct buffer_head *ret = NULL;
1341         struct bh_lru *lru;
1342         unsigned int i;
1343
1344         check_irqs_on();
1345         bh_lru_lock();
1346         lru = &__get_cpu_var(bh_lrus);
1347         for (i = 0; i < BH_LRU_SIZE; i++) {
1348                 struct buffer_head *bh = lru->bhs[i];
1349
1350                 if (bh && bh->b_bdev == bdev &&
1351                                 bh->b_blocknr == block && bh->b_size == size) {
1352                         if (i) {
1353                                 while (i) {
1354                                         lru->bhs[i] = lru->bhs[i - 1];
1355                                         i--;
1356                                 }
1357                                 lru->bhs[0] = bh;
1358                         }
1359                         get_bh(bh);
1360                         ret = bh;
1361                         break;
1362                 }
1363         }
1364         bh_lru_unlock();
1365         return ret;
1366 }
1367
1368 /*
1369  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1370  * it in the LRU and mark it as accessed.  If it is not present then return
1371  * NULL
1372  */
1373 struct buffer_head *
1374 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1375 {
1376         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1377
1378         if (bh == NULL) {
1379                 bh = __find_get_block_slow(bdev, block);
1380                 if (bh)
1381                         bh_lru_install(bh);
1382         }
1383         if (bh)
1384                 touch_buffer(bh);
1385         return bh;
1386 }
1387 EXPORT_SYMBOL(__find_get_block);
1388
1389 /*
1390  * __getblk will locate (and, if necessary, create) the buffer_head
1391  * which corresponds to the passed block_device, block and size. The
1392  * returned buffer has its reference count incremented.
1393  *
1394  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1395  * illegal block number, __getblk() will happily return a buffer_head
1396  * which represents the non-existent block.  Very weird.
1397  *
1398  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1399  * attempt is failing.  FIXME, perhaps?
1400  */
1401 struct buffer_head *
1402 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1403 {
1404         struct buffer_head *bh = __find_get_block(bdev, block, size);
1405
1406         might_sleep();
1407         if (bh == NULL)
1408                 bh = __getblk_slow(bdev, block, size);
1409         return bh;
1410 }
1411 EXPORT_SYMBOL(__getblk);
1412
1413 /*
1414  * Do async read-ahead on a buffer..
1415  */
1416 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1417 {
1418         struct buffer_head *bh = __getblk(bdev, block, size);
1419         if (likely(bh)) {
1420                 ll_rw_block(READA, 1, &bh);
1421                 brelse(bh);
1422         }
1423 }
1424 EXPORT_SYMBOL(__breadahead);
1425
1426 /**
1427  *  __bread() - reads a specified block and returns the bh
1428  *  @bdev: the block_device to read from
1429  *  @block: number of block
1430  *  @size: size (in bytes) to read
1431  * 
1432  *  Reads a specified block, and returns buffer head that contains it.
1433  *  It returns NULL if the block was unreadable.
1434  */
1435 struct buffer_head *
1436 __bread(struct block_device *bdev, sector_t block, unsigned size)
1437 {
1438         struct buffer_head *bh = __getblk(bdev, block, size);
1439
1440         if (likely(bh) && !buffer_uptodate(bh))
1441                 bh = __bread_slow(bh);
1442         return bh;
1443 }
1444 EXPORT_SYMBOL(__bread);
1445
1446 /*
1447  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1448  * This doesn't race because it runs in each cpu either in irq
1449  * or with preempt disabled.
1450  */
1451 static void invalidate_bh_lru(void *arg)
1452 {
1453         struct bh_lru *b = &get_cpu_var(bh_lrus);
1454         int i;
1455
1456         for (i = 0; i < BH_LRU_SIZE; i++) {
1457                 brelse(b->bhs[i]);
1458                 b->bhs[i] = NULL;
1459         }
1460         put_cpu_var(bh_lrus);
1461 }
1462         
1463 void invalidate_bh_lrus(void)
1464 {
1465         on_each_cpu(invalidate_bh_lru, NULL, 1);
1466 }
1467 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1468
1469 void set_bh_page(struct buffer_head *bh,
1470                 struct page *page, unsigned long offset)
1471 {
1472         bh->b_page = page;
1473         BUG_ON(offset >= PAGE_SIZE);
1474         if (PageHighMem(page))
1475                 /*
1476                  * This catches illegal uses and preserves the offset:
1477                  */
1478                 bh->b_data = (char *)(0 + offset);
1479         else
1480                 bh->b_data = page_address(page) + offset;
1481 }
1482 EXPORT_SYMBOL(set_bh_page);
1483
1484 /*
1485  * Called when truncating a buffer on a page completely.
1486  */
1487 static void discard_buffer(struct buffer_head * bh)
1488 {
1489         lock_buffer(bh);
1490         clear_buffer_dirty(bh);
1491         bh->b_bdev = NULL;
1492         clear_buffer_mapped(bh);
1493         clear_buffer_req(bh);
1494         clear_buffer_new(bh);
1495         clear_buffer_delay(bh);
1496         clear_buffer_unwritten(bh);
1497         unlock_buffer(bh);
1498 }
1499
1500 /**
1501  * block_invalidatepage - invalidate part of all of a buffer-backed page
1502  *
1503  * @page: the page which is affected
1504  * @offset: the index of the truncation point
1505  *
1506  * block_invalidatepage() is called when all or part of the page has become
1507  * invalidatedby a truncate operation.
1508  *
1509  * block_invalidatepage() does not have to release all buffers, but it must
1510  * ensure that no dirty buffer is left outside @offset and that no I/O
1511  * is underway against any of the blocks which are outside the truncation
1512  * point.  Because the caller is about to free (and possibly reuse) those
1513  * blocks on-disk.
1514  */
1515 void block_invalidatepage(struct page *page, unsigned long offset)
1516 {
1517         struct buffer_head *head, *bh, *next;
1518         unsigned int curr_off = 0;
1519
1520         BUG_ON(!PageLocked(page));
1521         if (!page_has_buffers(page))
1522                 goto out;
1523
1524         head = page_buffers(page);
1525         bh = head;
1526         do {
1527                 unsigned int next_off = curr_off + bh->b_size;
1528                 next = bh->b_this_page;
1529
1530                 /*
1531                  * is this block fully invalidated?
1532                  */
1533                 if (offset <= curr_off)
1534                         discard_buffer(bh);
1535                 curr_off = next_off;
1536                 bh = next;
1537         } while (bh != head);
1538
1539         /*
1540          * We release buffers only if the entire page is being invalidated.
1541          * The get_block cached value has been unconditionally invalidated,
1542          * so real IO is not possible anymore.
1543          */
1544         if (offset == 0)
1545                 try_to_release_page(page, 0);
1546 out:
1547         return;
1548 }
1549 EXPORT_SYMBOL(block_invalidatepage);
1550
1551 /*
1552  * We attach and possibly dirty the buffers atomically wrt
1553  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1554  * is already excluded via the page lock.
1555  */
1556 void create_empty_buffers(struct page *page,
1557                         unsigned long blocksize, unsigned long b_state)
1558 {
1559         struct buffer_head *bh, *head, *tail;
1560
1561         head = alloc_page_buffers(page, blocksize, 1);
1562         bh = head;
1563         do {
1564                 bh->b_state |= b_state;
1565                 tail = bh;
1566                 bh = bh->b_this_page;
1567         } while (bh);
1568         tail->b_this_page = head;
1569
1570         spin_lock(&page->mapping->private_lock);
1571         if (PageUptodate(page) || PageDirty(page)) {
1572                 bh = head;
1573                 do {
1574                         if (PageDirty(page))
1575                                 set_buffer_dirty(bh);
1576                         if (PageUptodate(page))
1577                                 set_buffer_uptodate(bh);
1578                         bh = bh->b_this_page;
1579                 } while (bh != head);
1580         }
1581         attach_page_buffers(page, head);
1582         spin_unlock(&page->mapping->private_lock);
1583 }
1584 EXPORT_SYMBOL(create_empty_buffers);
1585
1586 /*
1587  * We are taking a block for data and we don't want any output from any
1588  * buffer-cache aliases starting from return from that function and
1589  * until the moment when something will explicitly mark the buffer
1590  * dirty (hopefully that will not happen until we will free that block ;-)
1591  * We don't even need to mark it not-uptodate - nobody can expect
1592  * anything from a newly allocated buffer anyway. We used to used
1593  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1594  * don't want to mark the alias unmapped, for example - it would confuse
1595  * anyone who might pick it with bread() afterwards...
1596  *
1597  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1598  * be writeout I/O going on against recently-freed buffers.  We don't
1599  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1600  * only if we really need to.  That happens here.
1601  */
1602 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1603 {
1604         struct buffer_head *old_bh;
1605
1606         might_sleep();
1607
1608         old_bh = __find_get_block_slow(bdev, block);
1609         if (old_bh) {
1610                 clear_buffer_dirty(old_bh);
1611                 wait_on_buffer(old_bh);
1612                 clear_buffer_req(old_bh);
1613                 __brelse(old_bh);
1614         }
1615 }
1616 EXPORT_SYMBOL(unmap_underlying_metadata);
1617
1618 /*
1619  * NOTE! All mapped/uptodate combinations are valid:
1620  *
1621  *      Mapped  Uptodate        Meaning
1622  *
1623  *      No      No              "unknown" - must do get_block()
1624  *      No      Yes             "hole" - zero-filled
1625  *      Yes     No              "allocated" - allocated on disk, not read in
1626  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1627  *
1628  * "Dirty" is valid only with the last case (mapped+uptodate).
1629  */
1630
1631 /*
1632  * While block_write_full_page is writing back the dirty buffers under
1633  * the page lock, whoever dirtied the buffers may decide to clean them
1634  * again at any time.  We handle that by only looking at the buffer
1635  * state inside lock_buffer().
1636  *
1637  * If block_write_full_page() is called for regular writeback
1638  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1639  * locked buffer.   This only can happen if someone has written the buffer
1640  * directly, with submit_bh().  At the address_space level PageWriteback
1641  * prevents this contention from occurring.
1642  */
1643 static int __block_write_full_page(struct inode *inode, struct page *page,
1644                         get_block_t *get_block, struct writeback_control *wbc)
1645 {
1646         int err;
1647         sector_t block;
1648         sector_t last_block;
1649         struct buffer_head *bh, *head;
1650         const unsigned blocksize = 1 << inode->i_blkbits;
1651         int nr_underway = 0;
1652
1653         BUG_ON(!PageLocked(page));
1654
1655         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1656
1657         if (!page_has_buffers(page)) {
1658                 create_empty_buffers(page, blocksize,
1659                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1660         }
1661
1662         /*
1663          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1664          * here, and the (potentially unmapped) buffers may become dirty at
1665          * any time.  If a buffer becomes dirty here after we've inspected it
1666          * then we just miss that fact, and the page stays dirty.
1667          *
1668          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1669          * handle that here by just cleaning them.
1670          */
1671
1672         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1673         head = page_buffers(page);
1674         bh = head;
1675
1676         /*
1677          * Get all the dirty buffers mapped to disk addresses and
1678          * handle any aliases from the underlying blockdev's mapping.
1679          */
1680         do {
1681                 if (block > last_block) {
1682                         /*
1683                          * mapped buffers outside i_size will occur, because
1684                          * this page can be outside i_size when there is a
1685                          * truncate in progress.
1686                          */
1687                         /*
1688                          * The buffer was zeroed by block_write_full_page()
1689                          */
1690                         clear_buffer_dirty(bh);
1691                         set_buffer_uptodate(bh);
1692                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1693                            buffer_dirty(bh)) {
1694                         WARN_ON(bh->b_size != blocksize);
1695                         err = get_block(inode, block, bh, 1);
1696                         if (err)
1697                                 goto recover;
1698                         clear_buffer_delay(bh);
1699                         if (buffer_new(bh)) {
1700                                 /* blockdev mappings never come here */
1701                                 clear_buffer_new(bh);
1702                                 unmap_underlying_metadata(bh->b_bdev,
1703                                                         bh->b_blocknr);
1704                         }
1705                 }
1706                 bh = bh->b_this_page;
1707                 block++;
1708         } while (bh != head);
1709
1710         do {
1711                 if (!buffer_mapped(bh))
1712                         continue;
1713                 /*
1714                  * If it's a fully non-blocking write attempt and we cannot
1715                  * lock the buffer then redirty the page.  Note that this can
1716                  * potentially cause a busy-wait loop from pdflush and kswapd
1717                  * activity, but those code paths have their own higher-level
1718                  * throttling.
1719                  */
1720                 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1721                         lock_buffer(bh);
1722                 } else if (!trylock_buffer(bh)) {
1723                         redirty_page_for_writepage(wbc, page);
1724                         continue;
1725                 }
1726                 if (test_clear_buffer_dirty(bh)) {
1727                         mark_buffer_async_write(bh);
1728                 } else {
1729                         unlock_buffer(bh);
1730                 }
1731         } while ((bh = bh->b_this_page) != head);
1732
1733         /*
1734          * The page and its buffers are protected by PageWriteback(), so we can
1735          * drop the bh refcounts early.
1736          */
1737         BUG_ON(PageWriteback(page));
1738         set_page_writeback(page);
1739
1740         do {
1741                 struct buffer_head *next = bh->b_this_page;
1742                 if (buffer_async_write(bh)) {
1743                         submit_bh(WRITE, bh);
1744                         nr_underway++;
1745                 }
1746                 bh = next;
1747         } while (bh != head);
1748         unlock_page(page);
1749
1750         err = 0;
1751 done:
1752         if (nr_underway == 0) {
1753                 /*
1754                  * The page was marked dirty, but the buffers were
1755                  * clean.  Someone wrote them back by hand with
1756                  * ll_rw_block/submit_bh.  A rare case.
1757                  */
1758                 end_page_writeback(page);
1759
1760                 /*
1761                  * The page and buffer_heads can be released at any time from
1762                  * here on.
1763                  */
1764         }
1765         return err;
1766
1767 recover:
1768         /*
1769          * ENOSPC, or some other error.  We may already have added some
1770          * blocks to the file, so we need to write these out to avoid
1771          * exposing stale data.
1772          * The page is currently locked and not marked for writeback
1773          */
1774         bh = head;
1775         /* Recovery: lock and submit the mapped buffers */
1776         do {
1777                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1778                     !buffer_delay(bh)) {
1779                         lock_buffer(bh);
1780                         mark_buffer_async_write(bh);
1781                 } else {
1782                         /*
1783                          * The buffer may have been set dirty during
1784                          * attachment to a dirty page.
1785                          */
1786                         clear_buffer_dirty(bh);
1787                 }
1788         } while ((bh = bh->b_this_page) != head);
1789         SetPageError(page);
1790         BUG_ON(PageWriteback(page));
1791         mapping_set_error(page->mapping, err);
1792         set_page_writeback(page);
1793         do {
1794                 struct buffer_head *next = bh->b_this_page;
1795                 if (buffer_async_write(bh)) {
1796                         clear_buffer_dirty(bh);
1797                         submit_bh(WRITE, bh);
1798                         nr_underway++;
1799                 }
1800                 bh = next;
1801         } while (bh != head);
1802         unlock_page(page);
1803         goto done;
1804 }
1805
1806 /*
1807  * If a page has any new buffers, zero them out here, and mark them uptodate
1808  * and dirty so they'll be written out (in order to prevent uninitialised
1809  * block data from leaking). And clear the new bit.
1810  */
1811 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1812 {
1813         unsigned int block_start, block_end;
1814         struct buffer_head *head, *bh;
1815
1816         BUG_ON(!PageLocked(page));
1817         if (!page_has_buffers(page))
1818                 return;
1819
1820         bh = head = page_buffers(page);
1821         block_start = 0;
1822         do {
1823                 block_end = block_start + bh->b_size;
1824
1825                 if (buffer_new(bh)) {
1826                         if (block_end > from && block_start < to) {
1827                                 if (!PageUptodate(page)) {
1828                                         unsigned start, size;
1829
1830                                         start = max(from, block_start);
1831                                         size = min(to, block_end) - start;
1832
1833                                         zero_user(page, start, size);
1834                                         set_buffer_uptodate(bh);
1835                                 }
1836
1837                                 clear_buffer_new(bh);
1838                                 mark_buffer_dirty(bh);
1839                         }
1840                 }
1841
1842                 block_start = block_end;
1843                 bh = bh->b_this_page;
1844         } while (bh != head);
1845 }
1846 EXPORT_SYMBOL(page_zero_new_buffers);
1847
1848 static int __block_prepare_write(struct inode *inode, struct page *page,
1849                 unsigned from, unsigned to, get_block_t *get_block)
1850 {
1851         unsigned block_start, block_end;
1852         sector_t block;
1853         int err = 0;
1854         unsigned blocksize, bbits;
1855         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1856
1857         BUG_ON(!PageLocked(page));
1858         BUG_ON(from > PAGE_CACHE_SIZE);
1859         BUG_ON(to > PAGE_CACHE_SIZE);
1860         BUG_ON(from > to);
1861
1862         blocksize = 1 << inode->i_blkbits;
1863         if (!page_has_buffers(page))
1864                 create_empty_buffers(page, blocksize, 0);
1865         head = page_buffers(page);
1866
1867         bbits = inode->i_blkbits;
1868         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1869
1870         for(bh = head, block_start = 0; bh != head || !block_start;
1871             block++, block_start=block_end, bh = bh->b_this_page) {
1872                 block_end = block_start + blocksize;
1873                 if (block_end <= from || block_start >= to) {
1874                         if (PageUptodate(page)) {
1875                                 if (!buffer_uptodate(bh))
1876                                         set_buffer_uptodate(bh);
1877                         }
1878                         continue;
1879                 }
1880                 if (buffer_new(bh))
1881                         clear_buffer_new(bh);
1882                 if (!buffer_mapped(bh)) {
1883                         WARN_ON(bh->b_size != blocksize);
1884                         err = get_block(inode, block, bh, 1);
1885                         if (err)
1886                                 break;
1887                         if (buffer_new(bh)) {
1888                                 unmap_underlying_metadata(bh->b_bdev,
1889                                                         bh->b_blocknr);
1890                                 if (PageUptodate(page)) {
1891                                         clear_buffer_new(bh);
1892                                         set_buffer_uptodate(bh);
1893                                         mark_buffer_dirty(bh);
1894                                         continue;
1895                                 }
1896                                 if (block_end > to || block_start < from)
1897                                         zero_user_segments(page,
1898                                                 to, block_end,
1899                                                 block_start, from);
1900                                 continue;
1901                         }
1902                 }
1903                 if (PageUptodate(page)) {
1904                         if (!buffer_uptodate(bh))
1905                                 set_buffer_uptodate(bh);
1906                         continue; 
1907                 }
1908                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1909                     !buffer_unwritten(bh) &&
1910                      (block_start < from || block_end > to)) {
1911                         ll_rw_block(READ, 1, &bh);
1912                         *wait_bh++=bh;
1913                 }
1914         }
1915         /*
1916          * If we issued read requests - let them complete.
1917          */
1918         while(wait_bh > wait) {
1919                 wait_on_buffer(*--wait_bh);
1920                 if (!buffer_uptodate(*wait_bh))
1921                         err = -EIO;
1922         }
1923         if (unlikely(err))
1924                 page_zero_new_buffers(page, from, to);
1925         return err;
1926 }
1927
1928 static int __block_commit_write(struct inode *inode, struct page *page,
1929                 unsigned from, unsigned to)
1930 {
1931         unsigned block_start, block_end;
1932         int partial = 0;
1933         unsigned blocksize;
1934         struct buffer_head *bh, *head;
1935
1936         blocksize = 1 << inode->i_blkbits;
1937
1938         for(bh = head = page_buffers(page), block_start = 0;
1939             bh != head || !block_start;
1940             block_start=block_end, bh = bh->b_this_page) {
1941                 block_end = block_start + blocksize;
1942                 if (block_end <= from || block_start >= to) {
1943                         if (!buffer_uptodate(bh))
1944                                 partial = 1;
1945                 } else {
1946                         set_buffer_uptodate(bh);
1947                         mark_buffer_dirty(bh);
1948                 }
1949                 clear_buffer_new(bh);
1950         }
1951
1952         /*
1953          * If this is a partial write which happened to make all buffers
1954          * uptodate then we can optimize away a bogus readpage() for
1955          * the next read(). Here we 'discover' whether the page went
1956          * uptodate as a result of this (potentially partial) write.
1957          */
1958         if (!partial)
1959                 SetPageUptodate(page);
1960         return 0;
1961 }
1962
1963 /*
1964  * block_write_begin takes care of the basic task of block allocation and
1965  * bringing partial write blocks uptodate first.
1966  *
1967  * If *pagep is not NULL, then block_write_begin uses the locked page
1968  * at *pagep rather than allocating its own. In this case, the page will
1969  * not be unlocked or deallocated on failure.
1970  */
1971 int block_write_begin(struct file *file, struct address_space *mapping,
1972                         loff_t pos, unsigned len, unsigned flags,
1973                         struct page **pagep, void **fsdata,
1974                         get_block_t *get_block)
1975 {
1976         struct inode *inode = mapping->host;
1977         int status = 0;
1978         struct page *page;
1979         pgoff_t index;
1980         unsigned start, end;
1981         int ownpage = 0;
1982
1983         index = pos >> PAGE_CACHE_SHIFT;
1984         start = pos & (PAGE_CACHE_SIZE - 1);
1985         end = start + len;
1986
1987         page = *pagep;
1988         if (page == NULL) {
1989                 ownpage = 1;
1990                 page = __grab_cache_page(mapping, index);
1991                 if (!page) {
1992                         status = -ENOMEM;
1993                         goto out;
1994                 }
1995                 *pagep = page;
1996         } else
1997                 BUG_ON(!PageLocked(page));
1998
1999         status = __block_prepare_write(inode, page, start, end, get_block);
2000         if (unlikely(status)) {
2001                 ClearPageUptodate(page);
2002
2003                 if (ownpage) {
2004                         unlock_page(page);
2005                         page_cache_release(page);
2006                         *pagep = NULL;
2007
2008                         /*
2009                          * prepare_write() may have instantiated a few blocks
2010                          * outside i_size.  Trim these off again. Don't need
2011                          * i_size_read because we hold i_mutex.
2012                          */
2013                         if (pos + len > inode->i_size)
2014                                 vmtruncate(inode, inode->i_size);
2015                 }
2016                 goto out;
2017         }
2018
2019 out:
2020         return status;
2021 }
2022 EXPORT_SYMBOL(block_write_begin);
2023
2024 int block_write_end(struct file *file, struct address_space *mapping,
2025                         loff_t pos, unsigned len, unsigned copied,
2026                         struct page *page, void *fsdata)
2027 {
2028         struct inode *inode = mapping->host;
2029         unsigned start;
2030
2031         start = pos & (PAGE_CACHE_SIZE - 1);
2032
2033         if (unlikely(copied < len)) {
2034                 /*
2035                  * The buffers that were written will now be uptodate, so we
2036                  * don't have to worry about a readpage reading them and
2037                  * overwriting a partial write. However if we have encountered
2038                  * a short write and only partially written into a buffer, it
2039                  * will not be marked uptodate, so a readpage might come in and
2040                  * destroy our partial write.
2041                  *
2042                  * Do the simplest thing, and just treat any short write to a
2043                  * non uptodate page as a zero-length write, and force the
2044                  * caller to redo the whole thing.
2045                  */
2046                 if (!PageUptodate(page))
2047                         copied = 0;
2048
2049                 page_zero_new_buffers(page, start+copied, start+len);
2050         }
2051         flush_dcache_page(page);
2052
2053         /* This could be a short (even 0-length) commit */
2054         __block_commit_write(inode, page, start, start+copied);
2055
2056         return copied;
2057 }
2058 EXPORT_SYMBOL(block_write_end);
2059
2060 int generic_write_end(struct file *file, struct address_space *mapping,
2061                         loff_t pos, unsigned len, unsigned copied,
2062                         struct page *page, void *fsdata)
2063 {
2064         struct inode *inode = mapping->host;
2065         int i_size_changed = 0;
2066
2067         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2068
2069         /*
2070          * No need to use i_size_read() here, the i_size
2071          * cannot change under us because we hold i_mutex.
2072          *
2073          * But it's important to update i_size while still holding page lock:
2074          * page writeout could otherwise come in and zero beyond i_size.
2075          */
2076         if (pos+copied > inode->i_size) {
2077                 i_size_write(inode, pos+copied);
2078                 i_size_changed = 1;
2079         }
2080
2081         unlock_page(page);
2082         page_cache_release(page);
2083
2084         /*
2085          * Don't mark the inode dirty under page lock. First, it unnecessarily
2086          * makes the holding time of page lock longer. Second, it forces lock
2087          * ordering of page lock and transaction start for journaling
2088          * filesystems.
2089          */
2090         if (i_size_changed)
2091                 mark_inode_dirty(inode);
2092
2093         return copied;
2094 }
2095 EXPORT_SYMBOL(generic_write_end);
2096
2097 /*
2098  * block_is_partially_uptodate checks whether buffers within a page are
2099  * uptodate or not.
2100  *
2101  * Returns true if all buffers which correspond to a file portion
2102  * we want to read are uptodate.
2103  */
2104 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2105                                         unsigned long from)
2106 {
2107         struct inode *inode = page->mapping->host;
2108         unsigned block_start, block_end, blocksize;
2109         unsigned to;
2110         struct buffer_head *bh, *head;
2111         int ret = 1;
2112
2113         if (!page_has_buffers(page))
2114                 return 0;
2115
2116         blocksize = 1 << inode->i_blkbits;
2117         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2118         to = from + to;
2119         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2120                 return 0;
2121
2122         head = page_buffers(page);
2123         bh = head;
2124         block_start = 0;
2125         do {
2126                 block_end = block_start + blocksize;
2127                 if (block_end > from && block_start < to) {
2128                         if (!buffer_uptodate(bh)) {
2129                                 ret = 0;
2130                                 break;
2131                         }
2132                         if (block_end >= to)
2133                                 break;
2134                 }
2135                 block_start = block_end;
2136                 bh = bh->b_this_page;
2137         } while (bh != head);
2138
2139         return ret;
2140 }
2141 EXPORT_SYMBOL(block_is_partially_uptodate);
2142
2143 /*
2144  * Generic "read page" function for block devices that have the normal
2145  * get_block functionality. This is most of the block device filesystems.
2146  * Reads the page asynchronously --- the unlock_buffer() and
2147  * set/clear_buffer_uptodate() functions propagate buffer state into the
2148  * page struct once IO has completed.
2149  */
2150 int block_read_full_page(struct page *page, get_block_t *get_block)
2151 {
2152         struct inode *inode = page->mapping->host;
2153         sector_t iblock, lblock;
2154         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2155         unsigned int blocksize;
2156         int nr, i;
2157         int fully_mapped = 1;
2158
2159         BUG_ON(!PageLocked(page));
2160         blocksize = 1 << inode->i_blkbits;
2161         if (!page_has_buffers(page))
2162                 create_empty_buffers(page, blocksize, 0);
2163         head = page_buffers(page);
2164
2165         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2166         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2167         bh = head;
2168         nr = 0;
2169         i = 0;
2170
2171         do {
2172                 if (buffer_uptodate(bh))
2173                         continue;
2174
2175                 if (!buffer_mapped(bh)) {
2176                         int err = 0;
2177
2178                         fully_mapped = 0;
2179                         if (iblock < lblock) {
2180                                 WARN_ON(bh->b_size != blocksize);
2181                                 err = get_block(inode, iblock, bh, 0);
2182                                 if (err)
2183                                         SetPageError(page);
2184                         }
2185                         if (!buffer_mapped(bh)) {
2186                                 zero_user(page, i * blocksize, blocksize);
2187                                 if (!err)
2188                                         set_buffer_uptodate(bh);
2189                                 continue;
2190                         }
2191                         /*
2192                          * get_block() might have updated the buffer
2193                          * synchronously
2194                          */
2195                         if (buffer_uptodate(bh))
2196                                 continue;
2197                 }
2198                 arr[nr++] = bh;
2199         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2200
2201         if (fully_mapped)
2202                 SetPageMappedToDisk(page);
2203
2204         if (!nr) {
2205                 /*
2206                  * All buffers are uptodate - we can set the page uptodate
2207                  * as well. But not if get_block() returned an error.
2208                  */
2209                 if (!PageError(page))
2210                         SetPageUptodate(page);
2211                 unlock_page(page);
2212                 return 0;
2213         }
2214
2215         /* Stage two: lock the buffers */
2216         for (i = 0; i < nr; i++) {
2217                 bh = arr[i];
2218                 lock_buffer(bh);
2219                 mark_buffer_async_read(bh);
2220         }
2221
2222         /*
2223          * Stage 3: start the IO.  Check for uptodateness
2224          * inside the buffer lock in case another process reading
2225          * the underlying blockdev brought it uptodate (the sct fix).
2226          */
2227         for (i = 0; i < nr; i++) {
2228                 bh = arr[i];
2229                 if (buffer_uptodate(bh))
2230                         end_buffer_async_read(bh, 1);
2231                 else
2232                         submit_bh(READ, bh);
2233         }
2234         return 0;
2235 }
2236
2237 /* utility function for filesystems that need to do work on expanding
2238  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2239  * deal with the hole.  
2240  */
2241 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2242 {
2243         struct address_space *mapping = inode->i_mapping;
2244         struct page *page;
2245         void *fsdata;
2246         unsigned long limit;
2247         int err;
2248
2249         err = -EFBIG;
2250         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2251         if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2252                 send_sig(SIGXFSZ, current, 0);
2253                 goto out;
2254         }
2255         if (size > inode->i_sb->s_maxbytes)
2256                 goto out;
2257
2258         err = pagecache_write_begin(NULL, mapping, size, 0,
2259                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2260                                 &page, &fsdata);
2261         if (err)
2262                 goto out;
2263
2264         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2265         BUG_ON(err > 0);
2266
2267 out:
2268         return err;
2269 }
2270
2271 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2272                             loff_t pos, loff_t *bytes)
2273 {
2274         struct inode *inode = mapping->host;
2275         unsigned blocksize = 1 << inode->i_blkbits;
2276         struct page *page;
2277         void *fsdata;
2278         pgoff_t index, curidx;
2279         loff_t curpos;
2280         unsigned zerofrom, offset, len;
2281         int err = 0;
2282
2283         index = pos >> PAGE_CACHE_SHIFT;
2284         offset = pos & ~PAGE_CACHE_MASK;
2285
2286         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2287                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2288                 if (zerofrom & (blocksize-1)) {
2289                         *bytes |= (blocksize-1);
2290                         (*bytes)++;
2291                 }
2292                 len = PAGE_CACHE_SIZE - zerofrom;
2293
2294                 err = pagecache_write_begin(file, mapping, curpos, len,
2295                                                 AOP_FLAG_UNINTERRUPTIBLE,
2296                                                 &page, &fsdata);
2297                 if (err)
2298                         goto out;
2299                 zero_user(page, zerofrom, len);
2300                 err = pagecache_write_end(file, mapping, curpos, len, len,
2301                                                 page, fsdata);
2302                 if (err < 0)
2303                         goto out;
2304                 BUG_ON(err != len);
2305                 err = 0;
2306
2307                 balance_dirty_pages_ratelimited(mapping);
2308         }
2309
2310         /* page covers the boundary, find the boundary offset */
2311         if (index == curidx) {
2312                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2313                 /* if we will expand the thing last block will be filled */
2314                 if (offset <= zerofrom) {
2315                         goto out;
2316                 }
2317                 if (zerofrom & (blocksize-1)) {
2318                         *bytes |= (blocksize-1);
2319                         (*bytes)++;
2320                 }
2321                 len = offset - zerofrom;
2322
2323                 err = pagecache_write_begin(file, mapping, curpos, len,
2324                                                 AOP_FLAG_UNINTERRUPTIBLE,
2325                                                 &page, &fsdata);
2326                 if (err)
2327                         goto out;
2328                 zero_user(page, zerofrom, len);
2329                 err = pagecache_write_end(file, mapping, curpos, len, len,
2330                                                 page, fsdata);
2331                 if (err < 0)
2332                         goto out;
2333                 BUG_ON(err != len);
2334                 err = 0;
2335         }
2336 out:
2337         return err;
2338 }
2339
2340 /*
2341  * For moronic filesystems that do not allow holes in file.
2342  * We may have to extend the file.
2343  */
2344 int cont_write_begin(struct file *file, struct address_space *mapping,
2345                         loff_t pos, unsigned len, unsigned flags,
2346                         struct page **pagep, void **fsdata,
2347                         get_block_t *get_block, loff_t *bytes)
2348 {
2349         struct inode *inode = mapping->host;
2350         unsigned blocksize = 1 << inode->i_blkbits;
2351         unsigned zerofrom;
2352         int err;
2353
2354         err = cont_expand_zero(file, mapping, pos, bytes);
2355         if (err)
2356                 goto out;
2357
2358         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2359         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2360                 *bytes |= (blocksize-1);
2361                 (*bytes)++;
2362         }
2363
2364         *pagep = NULL;
2365         err = block_write_begin(file, mapping, pos, len,
2366                                 flags, pagep, fsdata, get_block);
2367 out:
2368         return err;
2369 }
2370
2371 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2372                         get_block_t *get_block)
2373 {
2374         struct inode *inode = page->mapping->host;
2375         int err = __block_prepare_write(inode, page, from, to, get_block);
2376         if (err)
2377                 ClearPageUptodate(page);
2378         return err;
2379 }
2380
2381 int block_commit_write(struct page *page, unsigned from, unsigned to)
2382 {
2383         struct inode *inode = page->mapping->host;
2384         __block_commit_write(inode,page,from,to);
2385         return 0;
2386 }
2387
2388 /*
2389  * block_page_mkwrite() is not allowed to change the file size as it gets
2390  * called from a page fault handler when a page is first dirtied. Hence we must
2391  * be careful to check for EOF conditions here. We set the page up correctly
2392  * for a written page which means we get ENOSPC checking when writing into
2393  * holes and correct delalloc and unwritten extent mapping on filesystems that
2394  * support these features.
2395  *
2396  * We are not allowed to take the i_mutex here so we have to play games to
2397  * protect against truncate races as the page could now be beyond EOF.  Because
2398  * vmtruncate() writes the inode size before removing pages, once we have the
2399  * page lock we can determine safely if the page is beyond EOF. If it is not
2400  * beyond EOF, then the page is guaranteed safe against truncation until we
2401  * unlock the page.
2402  */
2403 int
2404 block_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2405                    get_block_t get_block)
2406 {
2407         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2408         unsigned long end;
2409         loff_t size;
2410         int ret = -EINVAL;
2411
2412         lock_page(page);
2413         size = i_size_read(inode);
2414         if ((page->mapping != inode->i_mapping) ||
2415             (page_offset(page) > size)) {
2416                 /* page got truncated out from underneath us */
2417                 goto out_unlock;
2418         }
2419
2420         /* page is wholly or partially inside EOF */
2421         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2422                 end = size & ~PAGE_CACHE_MASK;
2423         else
2424                 end = PAGE_CACHE_SIZE;
2425
2426         ret = block_prepare_write(page, 0, end, get_block);
2427         if (!ret)
2428                 ret = block_commit_write(page, 0, end);
2429
2430 out_unlock:
2431         unlock_page(page);
2432         return ret;
2433 }
2434
2435 /*
2436  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2437  * immediately, while under the page lock.  So it needs a special end_io
2438  * handler which does not touch the bh after unlocking it.
2439  */
2440 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2441 {
2442         __end_buffer_read_notouch(bh, uptodate);
2443 }
2444
2445 /*
2446  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2447  * the page (converting it to circular linked list and taking care of page
2448  * dirty races).
2449  */
2450 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2451 {
2452         struct buffer_head *bh;
2453
2454         BUG_ON(!PageLocked(page));
2455
2456         spin_lock(&page->mapping->private_lock);
2457         bh = head;
2458         do {
2459                 if (PageDirty(page))
2460                         set_buffer_dirty(bh);
2461                 if (!bh->b_this_page)
2462                         bh->b_this_page = head;
2463                 bh = bh->b_this_page;
2464         } while (bh != head);
2465         attach_page_buffers(page, head);
2466         spin_unlock(&page->mapping->private_lock);
2467 }
2468
2469 /*
2470  * On entry, the page is fully not uptodate.
2471  * On exit the page is fully uptodate in the areas outside (from,to)
2472  */
2473 int nobh_write_begin(struct file *file, struct address_space *mapping,
2474                         loff_t pos, unsigned len, unsigned flags,
2475                         struct page **pagep, void **fsdata,
2476                         get_block_t *get_block)
2477 {
2478         struct inode *inode = mapping->host;
2479         const unsigned blkbits = inode->i_blkbits;
2480         const unsigned blocksize = 1 << blkbits;
2481         struct buffer_head *head, *bh;
2482         struct page *page;
2483         pgoff_t index;
2484         unsigned from, to;
2485         unsigned block_in_page;
2486         unsigned block_start, block_end;
2487         sector_t block_in_file;
2488         int nr_reads = 0;
2489         int ret = 0;
2490         int is_mapped_to_disk = 1;
2491
2492         index = pos >> PAGE_CACHE_SHIFT;
2493         from = pos & (PAGE_CACHE_SIZE - 1);
2494         to = from + len;
2495
2496         page = __grab_cache_page(mapping, index);
2497         if (!page)
2498                 return -ENOMEM;
2499         *pagep = page;
2500         *fsdata = NULL;
2501
2502         if (page_has_buffers(page)) {
2503                 unlock_page(page);
2504                 page_cache_release(page);
2505                 *pagep = NULL;
2506                 return block_write_begin(file, mapping, pos, len, flags, pagep,
2507                                         fsdata, get_block);
2508         }
2509
2510         if (PageMappedToDisk(page))
2511                 return 0;
2512
2513         /*
2514          * Allocate buffers so that we can keep track of state, and potentially
2515          * attach them to the page if an error occurs. In the common case of
2516          * no error, they will just be freed again without ever being attached
2517          * to the page (which is all OK, because we're under the page lock).
2518          *
2519          * Be careful: the buffer linked list is a NULL terminated one, rather
2520          * than the circular one we're used to.
2521          */
2522         head = alloc_page_buffers(page, blocksize, 0);
2523         if (!head) {
2524                 ret = -ENOMEM;
2525                 goto out_release;
2526         }
2527
2528         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2529
2530         /*
2531          * We loop across all blocks in the page, whether or not they are
2532          * part of the affected region.  This is so we can discover if the
2533          * page is fully mapped-to-disk.
2534          */
2535         for (block_start = 0, block_in_page = 0, bh = head;
2536                   block_start < PAGE_CACHE_SIZE;
2537                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2538                 int create;
2539
2540                 block_end = block_start + blocksize;
2541                 bh->b_state = 0;
2542                 create = 1;
2543                 if (block_start >= to)
2544                         create = 0;
2545                 ret = get_block(inode, block_in_file + block_in_page,
2546                                         bh, create);
2547                 if (ret)
2548                         goto failed;
2549                 if (!buffer_mapped(bh))
2550                         is_mapped_to_disk = 0;
2551                 if (buffer_new(bh))
2552                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2553                 if (PageUptodate(page)) {
2554                         set_buffer_uptodate(bh);
2555                         continue;
2556                 }
2557                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2558                         zero_user_segments(page, block_start, from,
2559                                                         to, block_end);
2560                         continue;
2561                 }
2562                 if (buffer_uptodate(bh))
2563                         continue;       /* reiserfs does this */
2564                 if (block_start < from || block_end > to) {
2565                         lock_buffer(bh);
2566                         bh->b_end_io = end_buffer_read_nobh;
2567                         submit_bh(READ, bh);
2568                         nr_reads++;
2569                 }
2570         }
2571
2572         if (nr_reads) {
2573                 /*
2574                  * The page is locked, so these buffers are protected from
2575                  * any VM or truncate activity.  Hence we don't need to care
2576                  * for the buffer_head refcounts.
2577                  */
2578                 for (bh = head; bh; bh = bh->b_this_page) {
2579                         wait_on_buffer(bh);
2580                         if (!buffer_uptodate(bh))
2581                                 ret = -EIO;
2582                 }
2583                 if (ret)
2584                         goto failed;
2585         }
2586
2587         if (is_mapped_to_disk)
2588                 SetPageMappedToDisk(page);
2589
2590         *fsdata = head; /* to be released by nobh_write_end */
2591
2592         return 0;
2593
2594 failed:
2595         BUG_ON(!ret);
2596         /*
2597          * Error recovery is a bit difficult. We need to zero out blocks that
2598          * were newly allocated, and dirty them to ensure they get written out.
2599          * Buffers need to be attached to the page at this point, otherwise
2600          * the handling of potential IO errors during writeout would be hard
2601          * (could try doing synchronous writeout, but what if that fails too?)
2602          */
2603         attach_nobh_buffers(page, head);
2604         page_zero_new_buffers(page, from, to);
2605
2606 out_release:
2607         unlock_page(page);
2608         page_cache_release(page);
2609         *pagep = NULL;
2610
2611         if (pos + len > inode->i_size)
2612                 vmtruncate(inode, inode->i_size);
2613
2614         return ret;
2615 }
2616 EXPORT_SYMBOL(nobh_write_begin);
2617
2618 int nobh_write_end(struct file *file, struct address_space *mapping,
2619                         loff_t pos, unsigned len, unsigned copied,
2620                         struct page *page, void *fsdata)
2621 {
2622         struct inode *inode = page->mapping->host;
2623         struct buffer_head *head = fsdata;
2624         struct buffer_head *bh;
2625         BUG_ON(fsdata != NULL && page_has_buffers(page));
2626
2627         if (unlikely(copied < len) && !page_has_buffers(page))
2628                 attach_nobh_buffers(page, head);
2629         if (page_has_buffers(page))
2630                 return generic_write_end(file, mapping, pos, len,
2631                                         copied, page, fsdata);
2632
2633         SetPageUptodate(page);
2634         set_page_dirty(page);
2635         if (pos+copied > inode->i_size) {
2636                 i_size_write(inode, pos+copied);
2637                 mark_inode_dirty(inode);
2638         }
2639
2640         unlock_page(page);
2641         page_cache_release(page);
2642
2643         while (head) {
2644                 bh = head;
2645                 head = head->b_this_page;
2646                 free_buffer_head(bh);
2647         }
2648
2649         return copied;
2650 }
2651 EXPORT_SYMBOL(nobh_write_end);
2652
2653 /*
2654  * nobh_writepage() - based on block_full_write_page() except
2655  * that it tries to operate without attaching bufferheads to
2656  * the page.
2657  */
2658 int nobh_writepage(struct page *page, get_block_t *get_block,
2659                         struct writeback_control *wbc)
2660 {
2661         struct inode * const inode = page->mapping->host;
2662         loff_t i_size = i_size_read(inode);
2663         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2664         unsigned offset;
2665         int ret;
2666
2667         /* Is the page fully inside i_size? */
2668         if (page->index < end_index)
2669                 goto out;
2670
2671         /* Is the page fully outside i_size? (truncate in progress) */
2672         offset = i_size & (PAGE_CACHE_SIZE-1);
2673         if (page->index >= end_index+1 || !offset) {
2674                 /*
2675                  * The page may have dirty, unmapped buffers.  For example,
2676                  * they may have been added in ext3_writepage().  Make them
2677                  * freeable here, so the page does not leak.
2678                  */
2679 #if 0
2680                 /* Not really sure about this  - do we need this ? */
2681                 if (page->mapping->a_ops->invalidatepage)
2682                         page->mapping->a_ops->invalidatepage(page, offset);
2683 #endif
2684                 unlock_page(page);
2685                 return 0; /* don't care */
2686         }
2687
2688         /*
2689          * The page straddles i_size.  It must be zeroed out on each and every
2690          * writepage invocation because it may be mmapped.  "A file is mapped
2691          * in multiples of the page size.  For a file that is not a multiple of
2692          * the  page size, the remaining memory is zeroed when mapped, and
2693          * writes to that region are not written out to the file."
2694          */
2695         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2696 out:
2697         ret = mpage_writepage(page, get_block, wbc);
2698         if (ret == -EAGAIN)
2699                 ret = __block_write_full_page(inode, page, get_block, wbc);
2700         return ret;
2701 }
2702 EXPORT_SYMBOL(nobh_writepage);
2703
2704 int nobh_truncate_page(struct address_space *mapping,
2705                         loff_t from, get_block_t *get_block)
2706 {
2707         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2708         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2709         unsigned blocksize;
2710         sector_t iblock;
2711         unsigned length, pos;
2712         struct inode *inode = mapping->host;
2713         struct page *page;
2714         struct buffer_head map_bh;
2715         int err;
2716
2717         blocksize = 1 << inode->i_blkbits;
2718         length = offset & (blocksize - 1);
2719
2720         /* Block boundary? Nothing to do */
2721         if (!length)
2722                 return 0;
2723
2724         length = blocksize - length;
2725         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2726
2727         page = grab_cache_page(mapping, index);
2728         err = -ENOMEM;
2729         if (!page)
2730                 goto out;
2731
2732         if (page_has_buffers(page)) {
2733 has_buffers:
2734                 unlock_page(page);
2735                 page_cache_release(page);
2736                 return block_truncate_page(mapping, from, get_block);
2737         }
2738
2739         /* Find the buffer that contains "offset" */
2740         pos = blocksize;
2741         while (offset >= pos) {
2742                 iblock++;
2743                 pos += blocksize;
2744         }
2745
2746         err = get_block(inode, iblock, &map_bh, 0);
2747         if (err)
2748                 goto unlock;
2749         /* unmapped? It's a hole - nothing to do */
2750         if (!buffer_mapped(&map_bh))
2751                 goto unlock;
2752
2753         /* Ok, it's mapped. Make sure it's up-to-date */
2754         if (!PageUptodate(page)) {
2755                 err = mapping->a_ops->readpage(NULL, page);
2756                 if (err) {
2757                         page_cache_release(page);
2758                         goto out;
2759                 }
2760                 lock_page(page);
2761                 if (!PageUptodate(page)) {
2762                         err = -EIO;
2763                         goto unlock;
2764                 }
2765                 if (page_has_buffers(page))
2766                         goto has_buffers;
2767         }
2768         zero_user(page, offset, length);
2769         set_page_dirty(page);
2770         err = 0;
2771
2772 unlock:
2773         unlock_page(page);
2774         page_cache_release(page);
2775 out:
2776         return err;
2777 }
2778 EXPORT_SYMBOL(nobh_truncate_page);
2779
2780 int block_truncate_page(struct address_space *mapping,
2781                         loff_t from, get_block_t *get_block)
2782 {
2783         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2784         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2785         unsigned blocksize;
2786         sector_t iblock;
2787         unsigned length, pos;
2788         struct inode *inode = mapping->host;
2789         struct page *page;
2790         struct buffer_head *bh;
2791         int err;
2792
2793         blocksize = 1 << inode->i_blkbits;
2794         length = offset & (blocksize - 1);
2795
2796         /* Block boundary? Nothing to do */
2797         if (!length)
2798                 return 0;
2799
2800         length = blocksize - length;
2801         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2802         
2803         page = grab_cache_page(mapping, index);
2804         err = -ENOMEM;
2805         if (!page)
2806                 goto out;
2807
2808         if (!page_has_buffers(page))
2809                 create_empty_buffers(page, blocksize, 0);
2810
2811         /* Find the buffer that contains "offset" */
2812         bh = page_buffers(page);
2813         pos = blocksize;
2814         while (offset >= pos) {
2815                 bh = bh->b_this_page;
2816                 iblock++;
2817                 pos += blocksize;
2818         }
2819
2820         err = 0;
2821         if (!buffer_mapped(bh)) {
2822                 WARN_ON(bh->b_size != blocksize);
2823                 err = get_block(inode, iblock, bh, 0);
2824                 if (err)
2825                         goto unlock;
2826                 /* unmapped? It's a hole - nothing to do */
2827                 if (!buffer_mapped(bh))
2828                         goto unlock;
2829         }
2830
2831         /* Ok, it's mapped. Make sure it's up-to-date */
2832         if (PageUptodate(page))
2833                 set_buffer_uptodate(bh);
2834
2835         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2836                 err = -EIO;
2837                 ll_rw_block(READ, 1, &bh);
2838                 wait_on_buffer(bh);
2839                 /* Uhhuh. Read error. Complain and punt. */
2840                 if (!buffer_uptodate(bh))
2841                         goto unlock;
2842         }
2843
2844         zero_user(page, offset, length);
2845         mark_buffer_dirty(bh);
2846         err = 0;
2847
2848 unlock:
2849         unlock_page(page);
2850         page_cache_release(page);
2851 out:
2852         return err;
2853 }
2854
2855 /*
2856  * The generic ->writepage function for buffer-backed address_spaces
2857  */
2858 int block_write_full_page(struct page *page, get_block_t *get_block,
2859                         struct writeback_control *wbc)
2860 {
2861         struct inode * const inode = page->mapping->host;
2862         loff_t i_size = i_size_read(inode);
2863         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2864         unsigned offset;
2865
2866         /* Is the page fully inside i_size? */
2867         if (page->index < end_index)
2868                 return __block_write_full_page(inode, page, get_block, wbc);
2869
2870         /* Is the page fully outside i_size? (truncate in progress) */
2871         offset = i_size & (PAGE_CACHE_SIZE-1);
2872         if (page->index >= end_index+1 || !offset) {
2873                 /*
2874                  * The page may have dirty, unmapped buffers.  For example,
2875                  * they may have been added in ext3_writepage().  Make them
2876                  * freeable here, so the page does not leak.
2877                  */
2878                 do_invalidatepage(page, 0);
2879                 unlock_page(page);
2880                 return 0; /* don't care */
2881         }
2882
2883         /*
2884          * The page straddles i_size.  It must be zeroed out on each and every
2885          * writepage invokation because it may be mmapped.  "A file is mapped
2886          * in multiples of the page size.  For a file that is not a multiple of
2887          * the  page size, the remaining memory is zeroed when mapped, and
2888          * writes to that region are not written out to the file."
2889          */
2890         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2891         return __block_write_full_page(inode, page, get_block, wbc);
2892 }
2893
2894 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2895                             get_block_t *get_block)
2896 {
2897         struct buffer_head tmp;
2898         struct inode *inode = mapping->host;
2899         tmp.b_state = 0;
2900         tmp.b_blocknr = 0;
2901         tmp.b_size = 1 << inode->i_blkbits;
2902         get_block(inode, block, &tmp, 0);
2903         return tmp.b_blocknr;
2904 }
2905
2906 static void end_bio_bh_io_sync(struct bio *bio, int err)
2907 {
2908         struct buffer_head *bh = bio->bi_private;
2909
2910         if (err == -EOPNOTSUPP) {
2911                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2912                 set_bit(BH_Eopnotsupp, &bh->b_state);
2913         }
2914
2915         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2916         bio_put(bio);
2917 }
2918
2919 int submit_bh(int rw, struct buffer_head * bh)
2920 {
2921         struct bio *bio;
2922         int ret = 0;
2923
2924         BUG_ON(!buffer_locked(bh));
2925         BUG_ON(!buffer_mapped(bh));
2926         BUG_ON(!bh->b_end_io);
2927
2928         /*
2929          * Mask in barrier bit for a write (could be either a WRITE or a
2930          * WRITE_SYNC
2931          */
2932         if (buffer_ordered(bh) && (rw & WRITE))
2933                 rw |= WRITE_BARRIER;
2934
2935         /*
2936          * Only clear out a write error when rewriting
2937          */
2938         if (test_set_buffer_req(bh) && (rw & WRITE))
2939                 clear_buffer_write_io_error(bh);
2940
2941         /*
2942          * from here on down, it's all bio -- do the initial mapping,
2943          * submit_bio -> generic_make_request may further map this bio around
2944          */
2945         bio = bio_alloc(GFP_NOIO, 1);
2946
2947         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2948         bio->bi_bdev = bh->b_bdev;
2949         bio->bi_io_vec[0].bv_page = bh->b_page;
2950         bio->bi_io_vec[0].bv_len = bh->b_size;
2951         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2952
2953         bio->bi_vcnt = 1;
2954         bio->bi_idx = 0;
2955         bio->bi_size = bh->b_size;
2956
2957         bio->bi_end_io = end_bio_bh_io_sync;
2958         bio->bi_private = bh;
2959
2960         bio_get(bio);
2961         submit_bio(rw, bio);
2962
2963         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2964                 ret = -EOPNOTSUPP;
2965
2966         bio_put(bio);
2967         return ret;
2968 }
2969
2970 /**
2971  * ll_rw_block: low-level access to block devices (DEPRECATED)
2972  * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
2973  * @nr: number of &struct buffer_heads in the array
2974  * @bhs: array of pointers to &struct buffer_head
2975  *
2976  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2977  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2978  * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2979  * are sent to disk. The fourth %READA option is described in the documentation
2980  * for generic_make_request() which ll_rw_block() calls.
2981  *
2982  * This function drops any buffer that it cannot get a lock on (with the
2983  * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
2984  * clean when doing a write request, and any buffer that appears to be
2985  * up-to-date when doing read request.  Further it marks as clean buffers that
2986  * are processed for writing (the buffer cache won't assume that they are
2987  * actually clean until the buffer gets unlocked).
2988  *
2989  * ll_rw_block sets b_end_io to simple completion handler that marks
2990  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2991  * any waiters. 
2992  *
2993  * All of the buffers must be for the same device, and must also be a
2994  * multiple of the current approved size for the device.
2995  */
2996 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2997 {
2998         int i;
2999
3000         for (i = 0; i < nr; i++) {
3001                 struct buffer_head *bh = bhs[i];
3002
3003                 if (rw == SWRITE || rw == SWRITE_SYNC)
3004                         lock_buffer(bh);
3005                 else if (!trylock_buffer(bh))
3006                         continue;
3007
3008                 if (rw == WRITE || rw == SWRITE || rw == SWRITE_SYNC) {
3009                         if (test_clear_buffer_dirty(bh)) {
3010                                 bh->b_end_io = end_buffer_write_sync;
3011                                 get_bh(bh);
3012                                 if (rw == SWRITE_SYNC)
3013                                         submit_bh(WRITE_SYNC, bh);
3014                                 else
3015                                         submit_bh(WRITE, bh);
3016                                 continue;
3017                         }
3018                 } else {
3019                         if (!buffer_uptodate(bh)) {
3020                                 bh->b_end_io = end_buffer_read_sync;
3021                                 get_bh(bh);
3022                                 submit_bh(rw, bh);
3023                                 continue;
3024                         }
3025                 }
3026                 unlock_buffer(bh);
3027         }
3028 }
3029
3030 /*
3031  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3032  * and then start new I/O and then wait upon it.  The caller must have a ref on
3033  * the buffer_head.
3034  */
3035 int sync_dirty_buffer(struct buffer_head *bh)
3036 {
3037         int ret = 0;
3038
3039         WARN_ON(atomic_read(&bh->b_count) < 1);
3040         lock_buffer(bh);
3041         if (test_clear_buffer_dirty(bh)) {
3042                 get_bh(bh);
3043                 bh->b_end_io = end_buffer_write_sync;
3044                 ret = submit_bh(WRITE_SYNC, bh);
3045                 wait_on_buffer(bh);
3046                 if (buffer_eopnotsupp(bh)) {
3047                         clear_buffer_eopnotsupp(bh);
3048                         ret = -EOPNOTSUPP;
3049                 }
3050                 if (!ret && !buffer_uptodate(bh))
3051                         ret = -EIO;
3052         } else {
3053                 unlock_buffer(bh);
3054         }
3055         return ret;
3056 }
3057
3058 /*
3059  * try_to_free_buffers() checks if all the buffers on this particular page
3060  * are unused, and releases them if so.
3061  *
3062  * Exclusion against try_to_free_buffers may be obtained by either
3063  * locking the page or by holding its mapping's private_lock.
3064  *
3065  * If the page is dirty but all the buffers are clean then we need to
3066  * be sure to mark the page clean as well.  This is because the page
3067  * may be against a block device, and a later reattachment of buffers
3068  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3069  * filesystem data on the same device.
3070  *
3071  * The same applies to regular filesystem pages: if all the buffers are
3072  * clean then we set the page clean and proceed.  To do that, we require
3073  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3074  * private_lock.
3075  *
3076  * try_to_free_buffers() is non-blocking.
3077  */
3078 static inline int buffer_busy(struct buffer_head *bh)
3079 {
3080         return atomic_read(&bh->b_count) |
3081                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3082 }
3083
3084 static int
3085 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3086 {
3087         struct buffer_head *head = page_buffers(page);
3088         struct buffer_head *bh;
3089
3090         bh = head;
3091         do {
3092                 if (buffer_write_io_error(bh) && page->mapping)
3093                         set_bit(AS_EIO, &page->mapping->flags);
3094                 if (buffer_busy(bh))
3095                         goto failed;
3096                 bh = bh->b_this_page;
3097         } while (bh != head);
3098
3099         do {
3100                 struct buffer_head *next = bh->b_this_page;
3101
3102                 if (bh->b_assoc_map)
3103                         __remove_assoc_queue(bh);
3104                 bh = next;
3105         } while (bh != head);
3106         *buffers_to_free = head;
3107         __clear_page_buffers(page);
3108         return 1;
3109 failed:
3110         return 0;
3111 }
3112
3113 int try_to_free_buffers(struct page *page)
3114 {
3115         struct address_space * const mapping = page->mapping;
3116         struct buffer_head *buffers_to_free = NULL;
3117         int ret = 0;
3118
3119         BUG_ON(!PageLocked(page));
3120         if (PageWriteback(page))
3121                 return 0;
3122
3123         if (mapping == NULL) {          /* can this still happen? */
3124                 ret = drop_buffers(page, &buffers_to_free);
3125                 goto out;
3126         }
3127
3128         spin_lock(&mapping->private_lock);
3129         ret = drop_buffers(page, &buffers_to_free);
3130
3131         /*
3132          * If the filesystem writes its buffers by hand (eg ext3)
3133          * then we can have clean buffers against a dirty page.  We
3134          * clean the page here; otherwise the VM will never notice
3135          * that the filesystem did any IO at all.
3136          *
3137          * Also, during truncate, discard_buffer will have marked all
3138          * the page's buffers clean.  We discover that here and clean
3139          * the page also.
3140          *
3141          * private_lock must be held over this entire operation in order
3142          * to synchronise against __set_page_dirty_buffers and prevent the
3143          * dirty bit from being lost.
3144          */
3145         if (ret)
3146                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3147         spin_unlock(&mapping->private_lock);
3148 out:
3149         if (buffers_to_free) {
3150                 struct buffer_head *bh = buffers_to_free;
3151
3152                 do {
3153                         struct buffer_head *next = bh->b_this_page;
3154                         free_buffer_head(bh);
3155                         bh = next;
3156                 } while (bh != buffers_to_free);
3157         }
3158         return ret;
3159 }
3160 EXPORT_SYMBOL(try_to_free_buffers);
3161
3162 void block_sync_page(struct page *page)
3163 {
3164         struct address_space *mapping;
3165
3166         smp_mb();
3167         mapping = page_mapping(page);
3168         if (mapping)
3169                 blk_run_backing_dev(mapping->backing_dev_info, page);
3170 }
3171
3172 /*
3173  * There are no bdflush tunables left.  But distributions are
3174  * still running obsolete flush daemons, so we terminate them here.
3175  *
3176  * Use of bdflush() is deprecated and will be removed in a future kernel.
3177  * The `pdflush' kernel threads fully replace bdflush daemons and this call.
3178  */
3179 asmlinkage long sys_bdflush(int func, long data)
3180 {
3181         static int msg_count;
3182
3183         if (!capable(CAP_SYS_ADMIN))
3184                 return -EPERM;
3185
3186         if (msg_count < 5) {
3187                 msg_count++;
3188                 printk(KERN_INFO
3189                         "warning: process `%s' used the obsolete bdflush"
3190                         " system call\n", current->comm);
3191                 printk(KERN_INFO "Fix your initscripts?\n");
3192         }
3193
3194         if (func == 1)
3195                 do_exit(0);
3196         return 0;
3197 }
3198
3199 /*
3200  * Buffer-head allocation
3201  */
3202 static struct kmem_cache *bh_cachep;
3203
3204 /*
3205  * Once the number of bh's in the machine exceeds this level, we start
3206  * stripping them in writeback.
3207  */
3208 static int max_buffer_heads;
3209
3210 int buffer_heads_over_limit;
3211
3212 struct bh_accounting {
3213         int nr;                 /* Number of live bh's */
3214         int ratelimit;          /* Limit cacheline bouncing */
3215 };
3216
3217 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3218
3219 static void recalc_bh_state(void)
3220 {
3221         int i;
3222         int tot = 0;
3223
3224         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3225                 return;
3226         __get_cpu_var(bh_accounting).ratelimit = 0;
3227         for_each_online_cpu(i)
3228                 tot += per_cpu(bh_accounting, i).nr;
3229         buffer_heads_over_limit = (tot > max_buffer_heads);
3230 }
3231         
3232 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3233 {
3234         struct buffer_head *ret = kmem_cache_alloc(bh_cachep, gfp_flags);
3235         if (ret) {
3236                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3237                 get_cpu_var(bh_accounting).nr++;
3238                 recalc_bh_state();
3239                 put_cpu_var(bh_accounting);
3240         }
3241         return ret;
3242 }
3243 EXPORT_SYMBOL(alloc_buffer_head);
3244
3245 void free_buffer_head(struct buffer_head *bh)
3246 {
3247         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3248         kmem_cache_free(bh_cachep, bh);
3249         get_cpu_var(bh_accounting).nr--;
3250         recalc_bh_state();
3251         put_cpu_var(bh_accounting);
3252 }
3253 EXPORT_SYMBOL(free_buffer_head);
3254
3255 static void buffer_exit_cpu(int cpu)
3256 {
3257         int i;
3258         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3259
3260         for (i = 0; i < BH_LRU_SIZE; i++) {
3261                 brelse(b->bhs[i]);
3262                 b->bhs[i] = NULL;
3263         }
3264         get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3265         per_cpu(bh_accounting, cpu).nr = 0;
3266         put_cpu_var(bh_accounting);
3267 }
3268
3269 static int buffer_cpu_notify(struct notifier_block *self,
3270                               unsigned long action, void *hcpu)
3271 {
3272         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3273                 buffer_exit_cpu((unsigned long)hcpu);
3274         return NOTIFY_OK;
3275 }
3276
3277 /**
3278  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3279  * @bh: struct buffer_head
3280  *
3281  * Return true if the buffer is up-to-date and false,
3282  * with the buffer locked, if not.
3283  */
3284 int bh_uptodate_or_lock(struct buffer_head *bh)
3285 {
3286         if (!buffer_uptodate(bh)) {
3287                 lock_buffer(bh);
3288                 if (!buffer_uptodate(bh))
3289                         return 0;
3290                 unlock_buffer(bh);
3291         }
3292         return 1;
3293 }
3294 EXPORT_SYMBOL(bh_uptodate_or_lock);
3295
3296 /**
3297  * bh_submit_read - Submit a locked buffer for reading
3298  * @bh: struct buffer_head
3299  *
3300  * Returns zero on success and -EIO on error.
3301  */
3302 int bh_submit_read(struct buffer_head *bh)
3303 {
3304         BUG_ON(!buffer_locked(bh));
3305
3306         if (buffer_uptodate(bh)) {
3307                 unlock_buffer(bh);
3308                 return 0;
3309         }
3310
3311         get_bh(bh);
3312         bh->b_end_io = end_buffer_read_sync;
3313         submit_bh(READ, bh);
3314         wait_on_buffer(bh);
3315         if (buffer_uptodate(bh))
3316                 return 0;
3317         return -EIO;
3318 }
3319 EXPORT_SYMBOL(bh_submit_read);
3320
3321 static void
3322 init_buffer_head(void *data)
3323 {
3324         struct buffer_head *bh = data;
3325
3326         memset(bh, 0, sizeof(*bh));
3327         INIT_LIST_HEAD(&bh->b_assoc_buffers);
3328 }
3329
3330 void __init buffer_init(void)
3331 {
3332         int nrpages;
3333
3334         bh_cachep = kmem_cache_create("buffer_head",
3335                         sizeof(struct buffer_head), 0,
3336                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3337                                 SLAB_MEM_SPREAD),
3338                                 init_buffer_head);
3339
3340         /*
3341          * Limit the bh occupancy to 10% of ZONE_NORMAL
3342          */
3343         nrpages = (nr_free_buffer_pages() * 10) / 100;
3344         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3345         hotcpu_notifier(buffer_cpu_notify, 0);
3346 }
3347
3348 EXPORT_SYMBOL(__bforget);
3349 EXPORT_SYMBOL(__brelse);
3350 EXPORT_SYMBOL(__wait_on_buffer);
3351 EXPORT_SYMBOL(block_commit_write);
3352 EXPORT_SYMBOL(block_prepare_write);
3353 EXPORT_SYMBOL(block_page_mkwrite);
3354 EXPORT_SYMBOL(block_read_full_page);
3355 EXPORT_SYMBOL(block_sync_page);
3356 EXPORT_SYMBOL(block_truncate_page);
3357 EXPORT_SYMBOL(block_write_full_page);
3358 EXPORT_SYMBOL(cont_write_begin);
3359 EXPORT_SYMBOL(end_buffer_read_sync);
3360 EXPORT_SYMBOL(end_buffer_write_sync);
3361 EXPORT_SYMBOL(file_fsync);
3362 EXPORT_SYMBOL(fsync_bdev);
3363 EXPORT_SYMBOL(generic_block_bmap);
3364 EXPORT_SYMBOL(generic_cont_expand_simple);
3365 EXPORT_SYMBOL(init_buffer);
3366 EXPORT_SYMBOL(invalidate_bdev);
3367 EXPORT_SYMBOL(ll_rw_block);
3368 EXPORT_SYMBOL(mark_buffer_dirty);
3369 EXPORT_SYMBOL(submit_bh);
3370 EXPORT_SYMBOL(sync_dirty_buffer);
3371 EXPORT_SYMBOL(unlock_buffer);