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