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