Merge with rsync://rsync.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6.git
[linux-2.6] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/config.h>
22 #include <linux/kernel.h>
23 #include <linux/syscalls.h>
24 #include <linux/fs.h>
25 #include <linux/mm.h>
26 #include <linux/percpu.h>
27 #include <linux/slab.h>
28 #include <linux/smp_lock.h>
29 #include <linux/blkdev.h>
30 #include <linux/file.h>
31 #include <linux/quotaops.h>
32 #include <linux/highmem.h>
33 #include <linux/module.h>
34 #include <linux/writeback.h>
35 #include <linux/hash.h>
36 #include <linux/suspend.h>
37 #include <linux/buffer_head.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
44 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
45 static void invalidate_bh_lrus(void);
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 fastcall __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 fastcall unlock_buffer(struct buffer_head *bh)
78 {
79         clear_buffer_locked(bh);
80         smp_mb__after_clear_bit();
81         wake_up_bit(&bh->b_state, BH_Lock);
82 }
83
84 /*
85  * Block until a buffer comes unlocked.  This doesn't stop it
86  * from becoming locked again - you have to lock it yourself
87  * if you want to preserve its state.
88  */
89 void __wait_on_buffer(struct buffer_head * bh)
90 {
91         wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
92 }
93
94 static void
95 __clear_page_buffers(struct page *page)
96 {
97         ClearPagePrivate(page);
98         page->private = 0;
99         page_cache_release(page);
100 }
101
102 static void buffer_io_error(struct buffer_head *bh)
103 {
104         char b[BDEVNAME_SIZE];
105
106         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
107                         bdevname(bh->b_bdev, b),
108                         (unsigned long long)bh->b_blocknr);
109 }
110
111 /*
112  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
113  * unlock the buffer. This is what ll_rw_block uses too.
114  */
115 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
116 {
117         if (uptodate) {
118                 set_buffer_uptodate(bh);
119         } else {
120                 /* This happens, due to failed READA attempts. */
121                 clear_buffer_uptodate(bh);
122         }
123         unlock_buffer(bh);
124         put_bh(bh);
125 }
126
127 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
128 {
129         char b[BDEVNAME_SIZE];
130
131         if (uptodate) {
132                 set_buffer_uptodate(bh);
133         } else {
134                 if (!buffer_eopnotsupp(bh) && printk_ratelimit()) {
135                         buffer_io_error(bh);
136                         printk(KERN_WARNING "lost page write due to "
137                                         "I/O error on %s\n",
138                                        bdevname(bh->b_bdev, b));
139                 }
140                 set_buffer_write_io_error(bh);
141                 clear_buffer_uptodate(bh);
142         }
143         unlock_buffer(bh);
144         put_bh(bh);
145 }
146
147 /*
148  * Write out and wait upon all the dirty data associated with a block
149  * device via its mapping.  Does not take the superblock lock.
150  */
151 int sync_blockdev(struct block_device *bdev)
152 {
153         int ret = 0;
154
155         if (bdev) {
156                 int err;
157
158                 ret = filemap_fdatawrite(bdev->bd_inode->i_mapping);
159                 err = filemap_fdatawait(bdev->bd_inode->i_mapping);
160                 if (!ret)
161                         ret = err;
162         }
163         return ret;
164 }
165 EXPORT_SYMBOL(sync_blockdev);
166
167 /*
168  * Write out and wait upon all dirty data associated with this
169  * superblock.  Filesystem data as well as the underlying block
170  * device.  Takes the superblock lock.
171  */
172 int fsync_super(struct super_block *sb)
173 {
174         sync_inodes_sb(sb, 0);
175         DQUOT_SYNC(sb);
176         lock_super(sb);
177         if (sb->s_dirt && sb->s_op->write_super)
178                 sb->s_op->write_super(sb);
179         unlock_super(sb);
180         if (sb->s_op->sync_fs)
181                 sb->s_op->sync_fs(sb, 1);
182         sync_blockdev(sb->s_bdev);
183         sync_inodes_sb(sb, 1);
184
185         return sync_blockdev(sb->s_bdev);
186 }
187
188 /*
189  * Write out and wait upon all dirty data associated with this
190  * device.   Filesystem data as well as the underlying block
191  * device.  Takes the superblock lock.
192  */
193 int fsync_bdev(struct block_device *bdev)
194 {
195         struct super_block *sb = get_super(bdev);
196         if (sb) {
197                 int res = fsync_super(sb);
198                 drop_super(sb);
199                 return res;
200         }
201         return sync_blockdev(bdev);
202 }
203
204 /**
205  * freeze_bdev  --  lock a filesystem and force it into a consistent state
206  * @bdev:       blockdevice to lock
207  *
208  * This takes the block device bd_mount_sem to make sure no new mounts
209  * happen on bdev until thaw_bdev() is called.
210  * If a superblock is found on this device, we take the s_umount semaphore
211  * on it to make sure nobody unmounts until the snapshot creation is done.
212  */
213 struct super_block *freeze_bdev(struct block_device *bdev)
214 {
215         struct super_block *sb;
216
217         down(&bdev->bd_mount_sem);
218         sb = get_super(bdev);
219         if (sb && !(sb->s_flags & MS_RDONLY)) {
220                 sb->s_frozen = SB_FREEZE_WRITE;
221                 smp_wmb();
222
223                 sync_inodes_sb(sb, 0);
224                 DQUOT_SYNC(sb);
225
226                 lock_super(sb);
227                 if (sb->s_dirt && sb->s_op->write_super)
228                         sb->s_op->write_super(sb);
229                 unlock_super(sb);
230
231                 if (sb->s_op->sync_fs)
232                         sb->s_op->sync_fs(sb, 1);
233
234                 sync_blockdev(sb->s_bdev);
235                 sync_inodes_sb(sb, 1);
236
237                 sb->s_frozen = SB_FREEZE_TRANS;
238                 smp_wmb();
239
240                 sync_blockdev(sb->s_bdev);
241
242                 if (sb->s_op->write_super_lockfs)
243                         sb->s_op->write_super_lockfs(sb);
244         }
245
246         sync_blockdev(bdev);
247         return sb;      /* thaw_bdev releases s->s_umount and bd_mount_sem */
248 }
249 EXPORT_SYMBOL(freeze_bdev);
250
251 /**
252  * thaw_bdev  -- unlock filesystem
253  * @bdev:       blockdevice to unlock
254  * @sb:         associated superblock
255  *
256  * Unlocks the filesystem and marks it writeable again after freeze_bdev().
257  */
258 void thaw_bdev(struct block_device *bdev, struct super_block *sb)
259 {
260         if (sb) {
261                 BUG_ON(sb->s_bdev != bdev);
262
263                 if (sb->s_op->unlockfs)
264                         sb->s_op->unlockfs(sb);
265                 sb->s_frozen = SB_UNFROZEN;
266                 smp_wmb();
267                 wake_up(&sb->s_wait_unfrozen);
268                 drop_super(sb);
269         }
270
271         up(&bdev->bd_mount_sem);
272 }
273 EXPORT_SYMBOL(thaw_bdev);
274
275 /*
276  * sync everything.  Start out by waking pdflush, because that writes back
277  * all queues in parallel.
278  */
279 static void do_sync(unsigned long wait)
280 {
281         wakeup_pdflush(0);
282         sync_inodes(0);         /* All mappings, inodes and their blockdevs */
283         DQUOT_SYNC(NULL);
284         sync_supers();          /* Write the superblocks */
285         sync_filesystems(0);    /* Start syncing the filesystems */
286         sync_filesystems(wait); /* Waitingly sync the filesystems */
287         sync_inodes(wait);      /* Mappings, inodes and blockdevs, again. */
288         if (!wait)
289                 printk("Emergency Sync complete\n");
290         if (unlikely(laptop_mode))
291                 laptop_sync_completion();
292 }
293
294 asmlinkage long sys_sync(void)
295 {
296         do_sync(1);
297         return 0;
298 }
299
300 void emergency_sync(void)
301 {
302         pdflush_operation(do_sync, 0);
303 }
304
305 /*
306  * Generic function to fsync a file.
307  *
308  * filp may be NULL if called via the msync of a vma.
309  */
310  
311 int file_fsync(struct file *filp, struct dentry *dentry, int datasync)
312 {
313         struct inode * inode = dentry->d_inode;
314         struct super_block * sb;
315         int ret, err;
316
317         /* sync the inode to buffers */
318         ret = write_inode_now(inode, 0);
319
320         /* sync the superblock to buffers */
321         sb = inode->i_sb;
322         lock_super(sb);
323         if (sb->s_op->write_super)
324                 sb->s_op->write_super(sb);
325         unlock_super(sb);
326
327         /* .. finally sync the buffers to disk */
328         err = sync_blockdev(sb->s_bdev);
329         if (!ret)
330                 ret = err;
331         return ret;
332 }
333
334 static long do_fsync(unsigned int fd, int datasync)
335 {
336         struct file * file;
337         struct address_space *mapping;
338         int ret, err;
339
340         ret = -EBADF;
341         file = fget(fd);
342         if (!file)
343                 goto out;
344
345         ret = -EINVAL;
346         if (!file->f_op || !file->f_op->fsync) {
347                 /* Why?  We can still call filemap_fdatawrite */
348                 goto out_putf;
349         }
350
351         mapping = file->f_mapping;
352
353         current->flags |= PF_SYNCWRITE;
354         ret = filemap_fdatawrite(mapping);
355
356         /*
357          * We need to protect against concurrent writers,
358          * which could cause livelocks in fsync_buffers_list
359          */
360         down(&mapping->host->i_sem);
361         err = file->f_op->fsync(file, file->f_dentry, datasync);
362         if (!ret)
363                 ret = err;
364         up(&mapping->host->i_sem);
365         err = filemap_fdatawait(mapping);
366         if (!ret)
367                 ret = err;
368         current->flags &= ~PF_SYNCWRITE;
369
370 out_putf:
371         fput(file);
372 out:
373         return ret;
374 }
375
376 asmlinkage long sys_fsync(unsigned int fd)
377 {
378         return do_fsync(fd, 0);
379 }
380
381 asmlinkage long sys_fdatasync(unsigned int fd)
382 {
383         return do_fsync(fd, 1);
384 }
385
386 /*
387  * Various filesystems appear to want __find_get_block to be non-blocking.
388  * But it's the page lock which protects the buffers.  To get around this,
389  * we get exclusion from try_to_free_buffers with the blockdev mapping's
390  * private_lock.
391  *
392  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
393  * may be quite high.  This code could TryLock the page, and if that
394  * succeeds, there is no need to take private_lock. (But if
395  * private_lock is contended then so is mapping->tree_lock).
396  */
397 static struct buffer_head *
398 __find_get_block_slow(struct block_device *bdev, sector_t block, int unused)
399 {
400         struct inode *bd_inode = bdev->bd_inode;
401         struct address_space *bd_mapping = bd_inode->i_mapping;
402         struct buffer_head *ret = NULL;
403         pgoff_t index;
404         struct buffer_head *bh;
405         struct buffer_head *head;
406         struct page *page;
407         int all_mapped = 1;
408
409         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
410         page = find_get_page(bd_mapping, index);
411         if (!page)
412                 goto out;
413
414         spin_lock(&bd_mapping->private_lock);
415         if (!page_has_buffers(page))
416                 goto out_unlock;
417         head = page_buffers(page);
418         bh = head;
419         do {
420                 if (bh->b_blocknr == block) {
421                         ret = bh;
422                         get_bh(bh);
423                         goto out_unlock;
424                 }
425                 if (!buffer_mapped(bh))
426                         all_mapped = 0;
427                 bh = bh->b_this_page;
428         } while (bh != head);
429
430         /* we might be here because some of the buffers on this page are
431          * not mapped.  This is due to various races between
432          * file io on the block device and getblk.  It gets dealt with
433          * elsewhere, don't buffer_error if we had some unmapped buffers
434          */
435         if (all_mapped) {
436                 printk("__find_get_block_slow() failed. "
437                         "block=%llu, b_blocknr=%llu\n",
438                         (unsigned long long)block, (unsigned long long)bh->b_blocknr);
439                 printk("b_state=0x%08lx, b_size=%u\n", bh->b_state, bh->b_size);
440                 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
441         }
442 out_unlock:
443         spin_unlock(&bd_mapping->private_lock);
444         page_cache_release(page);
445 out:
446         return ret;
447 }
448
449 /* If invalidate_buffers() will trash dirty buffers, it means some kind
450    of fs corruption is going on. Trashing dirty data always imply losing
451    information that was supposed to be just stored on the physical layer
452    by the user.
453
454    Thus invalidate_buffers in general usage is not allwowed to trash
455    dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
456    be preserved.  These buffers are simply skipped.
457   
458    We also skip buffers which are still in use.  For example this can
459    happen if a userspace program is reading the block device.
460
461    NOTE: In the case where the user removed a removable-media-disk even if
462    there's still dirty data not synced on disk (due a bug in the device driver
463    or due an error of the user), by not destroying the dirty buffers we could
464    generate corruption also on the next media inserted, thus a parameter is
465    necessary to handle this case in the most safe way possible (trying
466    to not corrupt also the new disk inserted with the data belonging to
467    the old now corrupted disk). Also for the ramdisk the natural thing
468    to do in order to release the ramdisk memory is to destroy dirty buffers.
469
470    These are two special cases. Normal usage imply the device driver
471    to issue a sync on the device (without waiting I/O completion) and
472    then an invalidate_buffers call that doesn't trash dirty buffers.
473
474    For handling cache coherency with the blkdev pagecache the 'update' case
475    is been introduced. It is needed to re-read from disk any pinned
476    buffer. NOTE: re-reading from disk is destructive so we can do it only
477    when we assume nobody is changing the buffercache under our I/O and when
478    we think the disk contains more recent information than the buffercache.
479    The update == 1 pass marks the buffers we need to update, the update == 2
480    pass does the actual I/O. */
481 void invalidate_bdev(struct block_device *bdev, int destroy_dirty_buffers)
482 {
483         invalidate_bh_lrus();
484         /*
485          * FIXME: what about destroy_dirty_buffers?
486          * We really want to use invalidate_inode_pages2() for
487          * that, but not until that's cleaned up.
488          */
489         invalidate_inode_pages(bdev->bd_inode->i_mapping);
490 }
491
492 /*
493  * Kick pdflush then try to free up some ZONE_NORMAL memory.
494  */
495 static void free_more_memory(void)
496 {
497         struct zone **zones;
498         pg_data_t *pgdat;
499
500         wakeup_pdflush(1024);
501         yield();
502
503         for_each_pgdat(pgdat) {
504                 zones = pgdat->node_zonelists[GFP_NOFS&GFP_ZONEMASK].zones;
505                 if (*zones)
506                         try_to_free_pages(zones, GFP_NOFS);
507         }
508 }
509
510 /*
511  * I/O completion handler for block_read_full_page() - pages
512  * which come unlocked at the end of I/O.
513  */
514 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
515 {
516         static DEFINE_SPINLOCK(page_uptodate_lock);
517         unsigned long flags;
518         struct buffer_head *tmp;
519         struct page *page;
520         int page_uptodate = 1;
521
522         BUG_ON(!buffer_async_read(bh));
523
524         page = bh->b_page;
525         if (uptodate) {
526                 set_buffer_uptodate(bh);
527         } else {
528                 clear_buffer_uptodate(bh);
529                 if (printk_ratelimit())
530                         buffer_io_error(bh);
531                 SetPageError(page);
532         }
533
534         /*
535          * Be _very_ careful from here on. Bad things can happen if
536          * two buffer heads end IO at almost the same time and both
537          * decide that the page is now completely done.
538          */
539         spin_lock_irqsave(&page_uptodate_lock, flags);
540         clear_buffer_async_read(bh);
541         unlock_buffer(bh);
542         tmp = bh;
543         do {
544                 if (!buffer_uptodate(tmp))
545                         page_uptodate = 0;
546                 if (buffer_async_read(tmp)) {
547                         BUG_ON(!buffer_locked(tmp));
548                         goto still_busy;
549                 }
550                 tmp = tmp->b_this_page;
551         } while (tmp != bh);
552         spin_unlock_irqrestore(&page_uptodate_lock, flags);
553
554         /*
555          * If none of the buffers had errors and they are all
556          * uptodate then we can set the page uptodate.
557          */
558         if (page_uptodate && !PageError(page))
559                 SetPageUptodate(page);
560         unlock_page(page);
561         return;
562
563 still_busy:
564         spin_unlock_irqrestore(&page_uptodate_lock, flags);
565         return;
566 }
567
568 /*
569  * Completion handler for block_write_full_page() - pages which are unlocked
570  * during I/O, and which have PageWriteback cleared upon I/O completion.
571  */
572 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
573 {
574         char b[BDEVNAME_SIZE];
575         static DEFINE_SPINLOCK(page_uptodate_lock);
576         unsigned long flags;
577         struct buffer_head *tmp;
578         struct page *page;
579
580         BUG_ON(!buffer_async_write(bh));
581
582         page = bh->b_page;
583         if (uptodate) {
584                 set_buffer_uptodate(bh);
585         } else {
586                 if (printk_ratelimit()) {
587                         buffer_io_error(bh);
588                         printk(KERN_WARNING "lost page write due to "
589                                         "I/O error on %s\n",
590                                bdevname(bh->b_bdev, b));
591                 }
592                 set_bit(AS_EIO, &page->mapping->flags);
593                 clear_buffer_uptodate(bh);
594                 SetPageError(page);
595         }
596
597         spin_lock_irqsave(&page_uptodate_lock, flags);
598         clear_buffer_async_write(bh);
599         unlock_buffer(bh);
600         tmp = bh->b_this_page;
601         while (tmp != bh) {
602                 if (buffer_async_write(tmp)) {
603                         BUG_ON(!buffer_locked(tmp));
604                         goto still_busy;
605                 }
606                 tmp = tmp->b_this_page;
607         }
608         spin_unlock_irqrestore(&page_uptodate_lock, flags);
609         end_page_writeback(page);
610         return;
611
612 still_busy:
613         spin_unlock_irqrestore(&page_uptodate_lock, flags);
614         return;
615 }
616
617 /*
618  * If a page's buffers are under async readin (end_buffer_async_read
619  * completion) then there is a possibility that another thread of
620  * control could lock one of the buffers after it has completed
621  * but while some of the other buffers have not completed.  This
622  * locked buffer would confuse end_buffer_async_read() into not unlocking
623  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
624  * that this buffer is not under async I/O.
625  *
626  * The page comes unlocked when it has no locked buffer_async buffers
627  * left.
628  *
629  * PageLocked prevents anyone starting new async I/O reads any of
630  * the buffers.
631  *
632  * PageWriteback is used to prevent simultaneous writeout of the same
633  * page.
634  *
635  * PageLocked prevents anyone from starting writeback of a page which is
636  * under read I/O (PageWriteback is only ever set against a locked page).
637  */
638 static void mark_buffer_async_read(struct buffer_head *bh)
639 {
640         bh->b_end_io = end_buffer_async_read;
641         set_buffer_async_read(bh);
642 }
643
644 void mark_buffer_async_write(struct buffer_head *bh)
645 {
646         bh->b_end_io = end_buffer_async_write;
647         set_buffer_async_write(bh);
648 }
649 EXPORT_SYMBOL(mark_buffer_async_write);
650
651
652 /*
653  * fs/buffer.c contains helper functions for buffer-backed address space's
654  * fsync functions.  A common requirement for buffer-based filesystems is
655  * that certain data from the backing blockdev needs to be written out for
656  * a successful fsync().  For example, ext2 indirect blocks need to be
657  * written back and waited upon before fsync() returns.
658  *
659  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
660  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
661  * management of a list of dependent buffers at ->i_mapping->private_list.
662  *
663  * Locking is a little subtle: try_to_free_buffers() will remove buffers
664  * from their controlling inode's queue when they are being freed.  But
665  * try_to_free_buffers() will be operating against the *blockdev* mapping
666  * at the time, not against the S_ISREG file which depends on those buffers.
667  * So the locking for private_list is via the private_lock in the address_space
668  * which backs the buffers.  Which is different from the address_space 
669  * against which the buffers are listed.  So for a particular address_space,
670  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
671  * mapping->private_list will always be protected by the backing blockdev's
672  * ->private_lock.
673  *
674  * Which introduces a requirement: all buffers on an address_space's
675  * ->private_list must be from the same address_space: the blockdev's.
676  *
677  * address_spaces which do not place buffers at ->private_list via these
678  * utility functions are free to use private_lock and private_list for
679  * whatever they want.  The only requirement is that list_empty(private_list)
680  * be true at clear_inode() time.
681  *
682  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
683  * filesystems should do that.  invalidate_inode_buffers() should just go
684  * BUG_ON(!list_empty).
685  *
686  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
687  * take an address_space, not an inode.  And it should be called
688  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
689  * queued up.
690  *
691  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
692  * list if it is already on a list.  Because if the buffer is on a list,
693  * it *must* already be on the right one.  If not, the filesystem is being
694  * silly.  This will save a ton of locking.  But first we have to ensure
695  * that buffers are taken *off* the old inode's list when they are freed
696  * (presumably in truncate).  That requires careful auditing of all
697  * filesystems (do it inside bforget()).  It could also be done by bringing
698  * b_inode back.
699  */
700
701 /*
702  * The buffer's backing address_space's private_lock must be held
703  */
704 static inline void __remove_assoc_queue(struct buffer_head *bh)
705 {
706         list_del_init(&bh->b_assoc_buffers);
707 }
708
709 int inode_has_buffers(struct inode *inode)
710 {
711         return !list_empty(&inode->i_data.private_list);
712 }
713
714 /*
715  * osync is designed to support O_SYNC io.  It waits synchronously for
716  * all already-submitted IO to complete, but does not queue any new
717  * writes to the disk.
718  *
719  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
720  * you dirty the buffers, and then use osync_inode_buffers to wait for
721  * completion.  Any other dirty buffers which are not yet queued for
722  * write will not be flushed to disk by the osync.
723  */
724 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
725 {
726         struct buffer_head *bh;
727         struct list_head *p;
728         int err = 0;
729
730         spin_lock(lock);
731 repeat:
732         list_for_each_prev(p, list) {
733                 bh = BH_ENTRY(p);
734                 if (buffer_locked(bh)) {
735                         get_bh(bh);
736                         spin_unlock(lock);
737                         wait_on_buffer(bh);
738                         if (!buffer_uptodate(bh))
739                                 err = -EIO;
740                         brelse(bh);
741                         spin_lock(lock);
742                         goto repeat;
743                 }
744         }
745         spin_unlock(lock);
746         return err;
747 }
748
749 /**
750  * sync_mapping_buffers - write out and wait upon a mapping's "associated"
751  *                        buffers
752  * @mapping: the mapping which wants those buffers written
753  *
754  * Starts I/O against the buffers at mapping->private_list, and waits upon
755  * that I/O.
756  *
757  * Basically, this is a convenience function for fsync().
758  * @mapping is a file or directory which needs those buffers to be written for
759  * a successful fsync().
760  */
761 int sync_mapping_buffers(struct address_space *mapping)
762 {
763         struct address_space *buffer_mapping = mapping->assoc_mapping;
764
765         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
766                 return 0;
767
768         return fsync_buffers_list(&buffer_mapping->private_lock,
769                                         &mapping->private_list);
770 }
771 EXPORT_SYMBOL(sync_mapping_buffers);
772
773 /*
774  * Called when we've recently written block `bblock', and it is known that
775  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
776  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
777  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
778  */
779 void write_boundary_block(struct block_device *bdev,
780                         sector_t bblock, unsigned blocksize)
781 {
782         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
783         if (bh) {
784                 if (buffer_dirty(bh))
785                         ll_rw_block(WRITE, 1, &bh);
786                 put_bh(bh);
787         }
788 }
789
790 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
791 {
792         struct address_space *mapping = inode->i_mapping;
793         struct address_space *buffer_mapping = bh->b_page->mapping;
794
795         mark_buffer_dirty(bh);
796         if (!mapping->assoc_mapping) {
797                 mapping->assoc_mapping = buffer_mapping;
798         } else {
799                 if (mapping->assoc_mapping != buffer_mapping)
800                         BUG();
801         }
802         if (list_empty(&bh->b_assoc_buffers)) {
803                 spin_lock(&buffer_mapping->private_lock);
804                 list_move_tail(&bh->b_assoc_buffers,
805                                 &mapping->private_list);
806                 spin_unlock(&buffer_mapping->private_lock);
807         }
808 }
809 EXPORT_SYMBOL(mark_buffer_dirty_inode);
810
811 /*
812  * Add a page to the dirty page list.
813  *
814  * It is a sad fact of life that this function is called from several places
815  * deeply under spinlocking.  It may not sleep.
816  *
817  * If the page has buffers, the uptodate buffers are set dirty, to preserve
818  * dirty-state coherency between the page and the buffers.  It the page does
819  * not have buffers then when they are later attached they will all be set
820  * dirty.
821  *
822  * The buffers are dirtied before the page is dirtied.  There's a small race
823  * window in which a writepage caller may see the page cleanness but not the
824  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
825  * before the buffers, a concurrent writepage caller could clear the page dirty
826  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
827  * page on the dirty page list.
828  *
829  * We use private_lock to lock against try_to_free_buffers while using the
830  * page's buffer list.  Also use this to protect against clean buffers being
831  * added to the page after it was set dirty.
832  *
833  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
834  * address_space though.
835  */
836 int __set_page_dirty_buffers(struct page *page)
837 {
838         struct address_space * const mapping = page->mapping;
839
840         spin_lock(&mapping->private_lock);
841         if (page_has_buffers(page)) {
842                 struct buffer_head *head = page_buffers(page);
843                 struct buffer_head *bh = head;
844
845                 do {
846                         set_buffer_dirty(bh);
847                         bh = bh->b_this_page;
848                 } while (bh != head);
849         }
850         spin_unlock(&mapping->private_lock);
851
852         if (!TestSetPageDirty(page)) {
853                 write_lock_irq(&mapping->tree_lock);
854                 if (page->mapping) {    /* Race with truncate? */
855                         if (mapping_cap_account_dirty(mapping))
856                                 inc_page_state(nr_dirty);
857                         radix_tree_tag_set(&mapping->page_tree,
858                                                 page_index(page),
859                                                 PAGECACHE_TAG_DIRTY);
860                 }
861                 write_unlock_irq(&mapping->tree_lock);
862                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
863         }
864         
865         return 0;
866 }
867 EXPORT_SYMBOL(__set_page_dirty_buffers);
868
869 /*
870  * Write out and wait upon a list of buffers.
871  *
872  * We have conflicting pressures: we want to make sure that all
873  * initially dirty buffers get waited on, but that any subsequently
874  * dirtied buffers don't.  After all, we don't want fsync to last
875  * forever if somebody is actively writing to the file.
876  *
877  * Do this in two main stages: first we copy dirty buffers to a
878  * temporary inode list, queueing the writes as we go.  Then we clean
879  * up, waiting for those writes to complete.
880  * 
881  * During this second stage, any subsequent updates to the file may end
882  * up refiling the buffer on the original inode's dirty list again, so
883  * there is a chance we will end up with a buffer queued for write but
884  * not yet completed on that list.  So, as a final cleanup we go through
885  * the osync code to catch these locked, dirty buffers without requeuing
886  * any newly dirty buffers for write.
887  */
888 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
889 {
890         struct buffer_head *bh;
891         struct list_head tmp;
892         int err = 0, err2;
893
894         INIT_LIST_HEAD(&tmp);
895
896         spin_lock(lock);
897         while (!list_empty(list)) {
898                 bh = BH_ENTRY(list->next);
899                 list_del_init(&bh->b_assoc_buffers);
900                 if (buffer_dirty(bh) || buffer_locked(bh)) {
901                         list_add(&bh->b_assoc_buffers, &tmp);
902                         if (buffer_dirty(bh)) {
903                                 get_bh(bh);
904                                 spin_unlock(lock);
905                                 /*
906                                  * Ensure any pending I/O completes so that
907                                  * ll_rw_block() actually writes the current
908                                  * contents - it is a noop if I/O is still in
909                                  * flight on potentially older contents.
910                                  */
911                                 wait_on_buffer(bh);
912                                 ll_rw_block(WRITE, 1, &bh);
913                                 brelse(bh);
914                                 spin_lock(lock);
915                         }
916                 }
917         }
918
919         while (!list_empty(&tmp)) {
920                 bh = BH_ENTRY(tmp.prev);
921                 __remove_assoc_queue(bh);
922                 get_bh(bh);
923                 spin_unlock(lock);
924                 wait_on_buffer(bh);
925                 if (!buffer_uptodate(bh))
926                         err = -EIO;
927                 brelse(bh);
928                 spin_lock(lock);
929         }
930         
931         spin_unlock(lock);
932         err2 = osync_buffers_list(lock, list);
933         if (err)
934                 return err;
935         else
936                 return err2;
937 }
938
939 /*
940  * Invalidate any and all dirty buffers on a given inode.  We are
941  * probably unmounting the fs, but that doesn't mean we have already
942  * done a sync().  Just drop the buffers from the inode list.
943  *
944  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
945  * assumes that all the buffers are against the blockdev.  Not true
946  * for reiserfs.
947  */
948 void invalidate_inode_buffers(struct inode *inode)
949 {
950         if (inode_has_buffers(inode)) {
951                 struct address_space *mapping = &inode->i_data;
952                 struct list_head *list = &mapping->private_list;
953                 struct address_space *buffer_mapping = mapping->assoc_mapping;
954
955                 spin_lock(&buffer_mapping->private_lock);
956                 while (!list_empty(list))
957                         __remove_assoc_queue(BH_ENTRY(list->next));
958                 spin_unlock(&buffer_mapping->private_lock);
959         }
960 }
961
962 /*
963  * Remove any clean buffers from the inode's buffer list.  This is called
964  * when we're trying to free the inode itself.  Those buffers can pin it.
965  *
966  * Returns true if all buffers were removed.
967  */
968 int remove_inode_buffers(struct inode *inode)
969 {
970         int ret = 1;
971
972         if (inode_has_buffers(inode)) {
973                 struct address_space *mapping = &inode->i_data;
974                 struct list_head *list = &mapping->private_list;
975                 struct address_space *buffer_mapping = mapping->assoc_mapping;
976
977                 spin_lock(&buffer_mapping->private_lock);
978                 while (!list_empty(list)) {
979                         struct buffer_head *bh = BH_ENTRY(list->next);
980                         if (buffer_dirty(bh)) {
981                                 ret = 0;
982                                 break;
983                         }
984                         __remove_assoc_queue(bh);
985                 }
986                 spin_unlock(&buffer_mapping->private_lock);
987         }
988         return ret;
989 }
990
991 /*
992  * Create the appropriate buffers when given a page for data area and
993  * the size of each buffer.. Use the bh->b_this_page linked list to
994  * follow the buffers created.  Return NULL if unable to create more
995  * buffers.
996  *
997  * The retry flag is used to differentiate async IO (paging, swapping)
998  * which may not fail from ordinary buffer allocations.
999  */
1000 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
1001                 int retry)
1002 {
1003         struct buffer_head *bh, *head;
1004         long offset;
1005
1006 try_again:
1007         head = NULL;
1008         offset = PAGE_SIZE;
1009         while ((offset -= size) >= 0) {
1010                 bh = alloc_buffer_head(GFP_NOFS);
1011                 if (!bh)
1012                         goto no_grow;
1013
1014                 bh->b_bdev = NULL;
1015                 bh->b_this_page = head;
1016                 bh->b_blocknr = -1;
1017                 head = bh;
1018
1019                 bh->b_state = 0;
1020                 atomic_set(&bh->b_count, 0);
1021                 bh->b_size = size;
1022
1023                 /* Link the buffer to its page */
1024                 set_bh_page(bh, page, offset);
1025
1026                 bh->b_end_io = NULL;
1027         }
1028         return head;
1029 /*
1030  * In case anything failed, we just free everything we got.
1031  */
1032 no_grow:
1033         if (head) {
1034                 do {
1035                         bh = head;
1036                         head = head->b_this_page;
1037                         free_buffer_head(bh);
1038                 } while (head);
1039         }
1040
1041         /*
1042          * Return failure for non-async IO requests.  Async IO requests
1043          * are not allowed to fail, so we have to wait until buffer heads
1044          * become available.  But we don't want tasks sleeping with 
1045          * partially complete buffers, so all were released above.
1046          */
1047         if (!retry)
1048                 return NULL;
1049
1050         /* We're _really_ low on memory. Now we just
1051          * wait for old buffer heads to become free due to
1052          * finishing IO.  Since this is an async request and
1053          * the reserve list is empty, we're sure there are 
1054          * async buffer heads in use.
1055          */
1056         free_more_memory();
1057         goto try_again;
1058 }
1059 EXPORT_SYMBOL_GPL(alloc_page_buffers);
1060
1061 static inline void
1062 link_dev_buffers(struct page *page, struct buffer_head *head)
1063 {
1064         struct buffer_head *bh, *tail;
1065
1066         bh = head;
1067         do {
1068                 tail = bh;
1069                 bh = bh->b_this_page;
1070         } while (bh);
1071         tail->b_this_page = head;
1072         attach_page_buffers(page, head);
1073 }
1074
1075 /*
1076  * Initialise the state of a blockdev page's buffers.
1077  */ 
1078 static void
1079 init_page_buffers(struct page *page, struct block_device *bdev,
1080                         sector_t block, int size)
1081 {
1082         struct buffer_head *head = page_buffers(page);
1083         struct buffer_head *bh = head;
1084         int uptodate = PageUptodate(page);
1085
1086         do {
1087                 if (!buffer_mapped(bh)) {
1088                         init_buffer(bh, NULL, NULL);
1089                         bh->b_bdev = bdev;
1090                         bh->b_blocknr = block;
1091                         if (uptodate)
1092                                 set_buffer_uptodate(bh);
1093                         set_buffer_mapped(bh);
1094                 }
1095                 block++;
1096                 bh = bh->b_this_page;
1097         } while (bh != head);
1098 }
1099
1100 /*
1101  * Create the page-cache page that contains the requested block.
1102  *
1103  * This is user purely for blockdev mappings.
1104  */
1105 static struct page *
1106 grow_dev_page(struct block_device *bdev, sector_t block,
1107                 pgoff_t index, int size)
1108 {
1109         struct inode *inode = bdev->bd_inode;
1110         struct page *page;
1111         struct buffer_head *bh;
1112
1113         page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
1114         if (!page)
1115                 return NULL;
1116
1117         if (!PageLocked(page))
1118                 BUG();
1119
1120         if (page_has_buffers(page)) {
1121                 bh = page_buffers(page);
1122                 if (bh->b_size == size) {
1123                         init_page_buffers(page, bdev, block, size);
1124                         return page;
1125                 }
1126                 if (!try_to_free_buffers(page))
1127                         goto failed;
1128         }
1129
1130         /*
1131          * Allocate some buffers for this page
1132          */
1133         bh = alloc_page_buffers(page, size, 0);
1134         if (!bh)
1135                 goto failed;
1136
1137         /*
1138          * Link the page to the buffers and initialise them.  Take the
1139          * lock to be atomic wrt __find_get_block(), which does not
1140          * run under the page lock.
1141          */
1142         spin_lock(&inode->i_mapping->private_lock);
1143         link_dev_buffers(page, bh);
1144         init_page_buffers(page, bdev, block, size);
1145         spin_unlock(&inode->i_mapping->private_lock);
1146         return page;
1147
1148 failed:
1149         BUG();
1150         unlock_page(page);
1151         page_cache_release(page);
1152         return NULL;
1153 }
1154
1155 /*
1156  * Create buffers for the specified block device block's page.  If
1157  * that page was dirty, the buffers are set dirty also.
1158  *
1159  * Except that's a bug.  Attaching dirty buffers to a dirty
1160  * blockdev's page can result in filesystem corruption, because
1161  * some of those buffers may be aliases of filesystem data.
1162  * grow_dev_page() will go BUG() if this happens.
1163  */
1164 static inline int
1165 grow_buffers(struct block_device *bdev, sector_t block, int size)
1166 {
1167         struct page *page;
1168         pgoff_t index;
1169         int sizebits;
1170
1171         sizebits = -1;
1172         do {
1173                 sizebits++;
1174         } while ((size << sizebits) < PAGE_SIZE);
1175
1176         index = block >> sizebits;
1177         block = index << sizebits;
1178
1179         /* Create a page with the proper size buffers.. */
1180         page = grow_dev_page(bdev, block, index, size);
1181         if (!page)
1182                 return 0;
1183         unlock_page(page);
1184         page_cache_release(page);
1185         return 1;
1186 }
1187
1188 static struct buffer_head *
1189 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1190 {
1191         /* Size must be multiple of hard sectorsize */
1192         if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1193                         (size < 512 || size > PAGE_SIZE))) {
1194                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1195                                         size);
1196                 printk(KERN_ERR "hardsect size: %d\n",
1197                                         bdev_hardsect_size(bdev));
1198
1199                 dump_stack();
1200                 return NULL;
1201         }
1202
1203         for (;;) {
1204                 struct buffer_head * bh;
1205
1206                 bh = __find_get_block(bdev, block, size);
1207                 if (bh)
1208                         return bh;
1209
1210                 if (!grow_buffers(bdev, block, size))
1211                         free_more_memory();
1212         }
1213 }
1214
1215 /*
1216  * The relationship between dirty buffers and dirty pages:
1217  *
1218  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1219  * the page is tagged dirty in its radix tree.
1220  *
1221  * At all times, the dirtiness of the buffers represents the dirtiness of
1222  * subsections of the page.  If the page has buffers, the page dirty bit is
1223  * merely a hint about the true dirty state.
1224  *
1225  * When a page is set dirty in its entirety, all its buffers are marked dirty
1226  * (if the page has buffers).
1227  *
1228  * When a buffer is marked dirty, its page is dirtied, but the page's other
1229  * buffers are not.
1230  *
1231  * Also.  When blockdev buffers are explicitly read with bread(), they
1232  * individually become uptodate.  But their backing page remains not
1233  * uptodate - even if all of its buffers are uptodate.  A subsequent
1234  * block_read_full_page() against that page will discover all the uptodate
1235  * buffers, will set the page uptodate and will perform no I/O.
1236  */
1237
1238 /**
1239  * mark_buffer_dirty - mark a buffer_head as needing writeout
1240  * @bh: the buffer_head to mark dirty
1241  *
1242  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1243  * backing page dirty, then tag the page as dirty in its address_space's radix
1244  * tree and then attach the address_space's inode to its superblock's dirty
1245  * inode list.
1246  *
1247  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1248  * mapping->tree_lock and the global inode_lock.
1249  */
1250 void fastcall mark_buffer_dirty(struct buffer_head *bh)
1251 {
1252         if (!buffer_dirty(bh) && !test_set_buffer_dirty(bh))
1253                 __set_page_dirty_nobuffers(bh->b_page);
1254 }
1255
1256 /*
1257  * Decrement a buffer_head's reference count.  If all buffers against a page
1258  * have zero reference count, are clean and unlocked, and if the page is clean
1259  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1260  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1261  * a page but it ends up not being freed, and buffers may later be reattached).
1262  */
1263 void __brelse(struct buffer_head * buf)
1264 {
1265         if (atomic_read(&buf->b_count)) {
1266                 put_bh(buf);
1267                 return;
1268         }
1269         printk(KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1270         WARN_ON(1);
1271 }
1272
1273 /*
1274  * bforget() is like brelse(), except it discards any
1275  * potentially dirty data.
1276  */
1277 void __bforget(struct buffer_head *bh)
1278 {
1279         clear_buffer_dirty(bh);
1280         if (!list_empty(&bh->b_assoc_buffers)) {
1281                 struct address_space *buffer_mapping = bh->b_page->mapping;
1282
1283                 spin_lock(&buffer_mapping->private_lock);
1284                 list_del_init(&bh->b_assoc_buffers);
1285                 spin_unlock(&buffer_mapping->private_lock);
1286         }
1287         __brelse(bh);
1288 }
1289
1290 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1291 {
1292         lock_buffer(bh);
1293         if (buffer_uptodate(bh)) {
1294                 unlock_buffer(bh);
1295                 return bh;
1296         } else {
1297                 get_bh(bh);
1298                 bh->b_end_io = end_buffer_read_sync;
1299                 submit_bh(READ, bh);
1300                 wait_on_buffer(bh);
1301                 if (buffer_uptodate(bh))
1302                         return bh;
1303         }
1304         brelse(bh);
1305         return NULL;
1306 }
1307
1308 /*
1309  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1310  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1311  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1312  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1313  * CPU's LRUs at the same time.
1314  *
1315  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1316  * sb_find_get_block().
1317  *
1318  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1319  * a local interrupt disable for that.
1320  */
1321
1322 #define BH_LRU_SIZE     8
1323
1324 struct bh_lru {
1325         struct buffer_head *bhs[BH_LRU_SIZE];
1326 };
1327
1328 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1329
1330 #ifdef CONFIG_SMP
1331 #define bh_lru_lock()   local_irq_disable()
1332 #define bh_lru_unlock() local_irq_enable()
1333 #else
1334 #define bh_lru_lock()   preempt_disable()
1335 #define bh_lru_unlock() preempt_enable()
1336 #endif
1337
1338 static inline void check_irqs_on(void)
1339 {
1340 #ifdef irqs_disabled
1341         BUG_ON(irqs_disabled());
1342 #endif
1343 }
1344
1345 /*
1346  * The LRU management algorithm is dopey-but-simple.  Sorry.
1347  */
1348 static void bh_lru_install(struct buffer_head *bh)
1349 {
1350         struct buffer_head *evictee = NULL;
1351         struct bh_lru *lru;
1352
1353         check_irqs_on();
1354         bh_lru_lock();
1355         lru = &__get_cpu_var(bh_lrus);
1356         if (lru->bhs[0] != bh) {
1357                 struct buffer_head *bhs[BH_LRU_SIZE];
1358                 int in;
1359                 int out = 0;
1360
1361                 get_bh(bh);
1362                 bhs[out++] = bh;
1363                 for (in = 0; in < BH_LRU_SIZE; in++) {
1364                         struct buffer_head *bh2 = lru->bhs[in];
1365
1366                         if (bh2 == bh) {
1367                                 __brelse(bh2);
1368                         } else {
1369                                 if (out >= BH_LRU_SIZE) {
1370                                         BUG_ON(evictee != NULL);
1371                                         evictee = bh2;
1372                                 } else {
1373                                         bhs[out++] = bh2;
1374                                 }
1375                         }
1376                 }
1377                 while (out < BH_LRU_SIZE)
1378                         bhs[out++] = NULL;
1379                 memcpy(lru->bhs, bhs, sizeof(bhs));
1380         }
1381         bh_lru_unlock();
1382
1383         if (evictee)
1384                 __brelse(evictee);
1385 }
1386
1387 /*
1388  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1389  */
1390 static inline struct buffer_head *
1391 lookup_bh_lru(struct block_device *bdev, sector_t block, int size)
1392 {
1393         struct buffer_head *ret = NULL;
1394         struct bh_lru *lru;
1395         int i;
1396
1397         check_irqs_on();
1398         bh_lru_lock();
1399         lru = &__get_cpu_var(bh_lrus);
1400         for (i = 0; i < BH_LRU_SIZE; i++) {
1401                 struct buffer_head *bh = lru->bhs[i];
1402
1403                 if (bh && bh->b_bdev == bdev &&
1404                                 bh->b_blocknr == block && bh->b_size == size) {
1405                         if (i) {
1406                                 while (i) {
1407                                         lru->bhs[i] = lru->bhs[i - 1];
1408                                         i--;
1409                                 }
1410                                 lru->bhs[0] = bh;
1411                         }
1412                         get_bh(bh);
1413                         ret = bh;
1414                         break;
1415                 }
1416         }
1417         bh_lru_unlock();
1418         return ret;
1419 }
1420
1421 /*
1422  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1423  * it in the LRU and mark it as accessed.  If it is not present then return
1424  * NULL
1425  */
1426 struct buffer_head *
1427 __find_get_block(struct block_device *bdev, sector_t block, int size)
1428 {
1429         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1430
1431         if (bh == NULL) {
1432                 bh = __find_get_block_slow(bdev, block, size);
1433                 if (bh)
1434                         bh_lru_install(bh);
1435         }
1436         if (bh)
1437                 touch_buffer(bh);
1438         return bh;
1439 }
1440 EXPORT_SYMBOL(__find_get_block);
1441
1442 /*
1443  * __getblk will locate (and, if necessary, create) the buffer_head
1444  * which corresponds to the passed block_device, block and size. The
1445  * returned buffer has its reference count incremented.
1446  *
1447  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1448  * illegal block number, __getblk() will happily return a buffer_head
1449  * which represents the non-existent block.  Very weird.
1450  *
1451  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1452  * attempt is failing.  FIXME, perhaps?
1453  */
1454 struct buffer_head *
1455 __getblk(struct block_device *bdev, sector_t block, int size)
1456 {
1457         struct buffer_head *bh = __find_get_block(bdev, block, size);
1458
1459         might_sleep();
1460         if (bh == NULL)
1461                 bh = __getblk_slow(bdev, block, size);
1462         return bh;
1463 }
1464 EXPORT_SYMBOL(__getblk);
1465
1466 /*
1467  * Do async read-ahead on a buffer..
1468  */
1469 void __breadahead(struct block_device *bdev, sector_t block, int size)
1470 {
1471         struct buffer_head *bh = __getblk(bdev, block, size);
1472         ll_rw_block(READA, 1, &bh);
1473         brelse(bh);
1474 }
1475 EXPORT_SYMBOL(__breadahead);
1476
1477 /**
1478  *  __bread() - reads a specified block and returns the bh
1479  *  @bdev: the block_device to read from
1480  *  @block: number of block
1481  *  @size: size (in bytes) to read
1482  * 
1483  *  Reads a specified block, and returns buffer head that contains it.
1484  *  It returns NULL if the block was unreadable.
1485  */
1486 struct buffer_head *
1487 __bread(struct block_device *bdev, sector_t block, int size)
1488 {
1489         struct buffer_head *bh = __getblk(bdev, block, size);
1490
1491         if (!buffer_uptodate(bh))
1492                 bh = __bread_slow(bh);
1493         return bh;
1494 }
1495 EXPORT_SYMBOL(__bread);
1496
1497 /*
1498  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1499  * This doesn't race because it runs in each cpu either in irq
1500  * or with preempt disabled.
1501  */
1502 static void invalidate_bh_lru(void *arg)
1503 {
1504         struct bh_lru *b = &get_cpu_var(bh_lrus);
1505         int i;
1506
1507         for (i = 0; i < BH_LRU_SIZE; i++) {
1508                 brelse(b->bhs[i]);
1509                 b->bhs[i] = NULL;
1510         }
1511         put_cpu_var(bh_lrus);
1512 }
1513         
1514 static void invalidate_bh_lrus(void)
1515 {
1516         on_each_cpu(invalidate_bh_lru, NULL, 1, 1);
1517 }
1518
1519 void set_bh_page(struct buffer_head *bh,
1520                 struct page *page, unsigned long offset)
1521 {
1522         bh->b_page = page;
1523         if (offset >= PAGE_SIZE)
1524                 BUG();
1525         if (PageHighMem(page))
1526                 /*
1527                  * This catches illegal uses and preserves the offset:
1528                  */
1529                 bh->b_data = (char *)(0 + offset);
1530         else
1531                 bh->b_data = page_address(page) + offset;
1532 }
1533 EXPORT_SYMBOL(set_bh_page);
1534
1535 /*
1536  * Called when truncating a buffer on a page completely.
1537  */
1538 static inline void discard_buffer(struct buffer_head * bh)
1539 {
1540         lock_buffer(bh);
1541         clear_buffer_dirty(bh);
1542         bh->b_bdev = NULL;
1543         clear_buffer_mapped(bh);
1544         clear_buffer_req(bh);
1545         clear_buffer_new(bh);
1546         clear_buffer_delay(bh);
1547         unlock_buffer(bh);
1548 }
1549
1550 /**
1551  * try_to_release_page() - release old fs-specific metadata on a page
1552  *
1553  * @page: the page which the kernel is trying to free
1554  * @gfp_mask: memory allocation flags (and I/O mode)
1555  *
1556  * The address_space is to try to release any data against the page
1557  * (presumably at page->private).  If the release was successful, return `1'.
1558  * Otherwise return zero.
1559  *
1560  * The @gfp_mask argument specifies whether I/O may be performed to release
1561  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
1562  *
1563  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
1564  */
1565 int try_to_release_page(struct page *page, int gfp_mask)
1566 {
1567         struct address_space * const mapping = page->mapping;
1568
1569         BUG_ON(!PageLocked(page));
1570         if (PageWriteback(page))
1571                 return 0;
1572         
1573         if (mapping && mapping->a_ops->releasepage)
1574                 return mapping->a_ops->releasepage(page, gfp_mask);
1575         return try_to_free_buffers(page);
1576 }
1577 EXPORT_SYMBOL(try_to_release_page);
1578
1579 /**
1580  * block_invalidatepage - invalidate part of all of a buffer-backed page
1581  *
1582  * @page: the page which is affected
1583  * @offset: the index of the truncation point
1584  *
1585  * block_invalidatepage() is called when all or part of the page has become
1586  * invalidatedby a truncate operation.
1587  *
1588  * block_invalidatepage() does not have to release all buffers, but it must
1589  * ensure that no dirty buffer is left outside @offset and that no I/O
1590  * is underway against any of the blocks which are outside the truncation
1591  * point.  Because the caller is about to free (and possibly reuse) those
1592  * blocks on-disk.
1593  */
1594 int block_invalidatepage(struct page *page, unsigned long offset)
1595 {
1596         struct buffer_head *head, *bh, *next;
1597         unsigned int curr_off = 0;
1598         int ret = 1;
1599
1600         BUG_ON(!PageLocked(page));
1601         if (!page_has_buffers(page))
1602                 goto out;
1603
1604         head = page_buffers(page);
1605         bh = head;
1606         do {
1607                 unsigned int next_off = curr_off + bh->b_size;
1608                 next = bh->b_this_page;
1609
1610                 /*
1611                  * is this block fully invalidated?
1612                  */
1613                 if (offset <= curr_off)
1614                         discard_buffer(bh);
1615                 curr_off = next_off;
1616                 bh = next;
1617         } while (bh != head);
1618
1619         /*
1620          * We release buffers only if the entire page is being invalidated.
1621          * The get_block cached value has been unconditionally invalidated,
1622          * so real IO is not possible anymore.
1623          */
1624         if (offset == 0)
1625                 ret = try_to_release_page(page, 0);
1626 out:
1627         return ret;
1628 }
1629 EXPORT_SYMBOL(block_invalidatepage);
1630
1631 /*
1632  * We attach and possibly dirty the buffers atomically wrt
1633  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1634  * is already excluded via the page lock.
1635  */
1636 void create_empty_buffers(struct page *page,
1637                         unsigned long blocksize, unsigned long b_state)
1638 {
1639         struct buffer_head *bh, *head, *tail;
1640
1641         head = alloc_page_buffers(page, blocksize, 1);
1642         bh = head;
1643         do {
1644                 bh->b_state |= b_state;
1645                 tail = bh;
1646                 bh = bh->b_this_page;
1647         } while (bh);
1648         tail->b_this_page = head;
1649
1650         spin_lock(&page->mapping->private_lock);
1651         if (PageUptodate(page) || PageDirty(page)) {
1652                 bh = head;
1653                 do {
1654                         if (PageDirty(page))
1655                                 set_buffer_dirty(bh);
1656                         if (PageUptodate(page))
1657                                 set_buffer_uptodate(bh);
1658                         bh = bh->b_this_page;
1659                 } while (bh != head);
1660         }
1661         attach_page_buffers(page, head);
1662         spin_unlock(&page->mapping->private_lock);
1663 }
1664 EXPORT_SYMBOL(create_empty_buffers);
1665
1666 /*
1667  * We are taking a block for data and we don't want any output from any
1668  * buffer-cache aliases starting from return from that function and
1669  * until the moment when something will explicitly mark the buffer
1670  * dirty (hopefully that will not happen until we will free that block ;-)
1671  * We don't even need to mark it not-uptodate - nobody can expect
1672  * anything from a newly allocated buffer anyway. We used to used
1673  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1674  * don't want to mark the alias unmapped, for example - it would confuse
1675  * anyone who might pick it with bread() afterwards...
1676  *
1677  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1678  * be writeout I/O going on against recently-freed buffers.  We don't
1679  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1680  * only if we really need to.  That happens here.
1681  */
1682 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1683 {
1684         struct buffer_head *old_bh;
1685
1686         might_sleep();
1687
1688         old_bh = __find_get_block_slow(bdev, block, 0);
1689         if (old_bh) {
1690                 clear_buffer_dirty(old_bh);
1691                 wait_on_buffer(old_bh);
1692                 clear_buffer_req(old_bh);
1693                 __brelse(old_bh);
1694         }
1695 }
1696 EXPORT_SYMBOL(unmap_underlying_metadata);
1697
1698 /*
1699  * NOTE! All mapped/uptodate combinations are valid:
1700  *
1701  *      Mapped  Uptodate        Meaning
1702  *
1703  *      No      No              "unknown" - must do get_block()
1704  *      No      Yes             "hole" - zero-filled
1705  *      Yes     No              "allocated" - allocated on disk, not read in
1706  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1707  *
1708  * "Dirty" is valid only with the last case (mapped+uptodate).
1709  */
1710
1711 /*
1712  * While block_write_full_page is writing back the dirty buffers under
1713  * the page lock, whoever dirtied the buffers may decide to clean them
1714  * again at any time.  We handle that by only looking at the buffer
1715  * state inside lock_buffer().
1716  *
1717  * If block_write_full_page() is called for regular writeback
1718  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1719  * locked buffer.   This only can happen if someone has written the buffer
1720  * directly, with submit_bh().  At the address_space level PageWriteback
1721  * prevents this contention from occurring.
1722  */
1723 static int __block_write_full_page(struct inode *inode, struct page *page,
1724                         get_block_t *get_block, struct writeback_control *wbc)
1725 {
1726         int err;
1727         sector_t block;
1728         sector_t last_block;
1729         struct buffer_head *bh, *head;
1730         int nr_underway = 0;
1731
1732         BUG_ON(!PageLocked(page));
1733
1734         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1735
1736         if (!page_has_buffers(page)) {
1737                 create_empty_buffers(page, 1 << inode->i_blkbits,
1738                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1739         }
1740
1741         /*
1742          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1743          * here, and the (potentially unmapped) buffers may become dirty at
1744          * any time.  If a buffer becomes dirty here after we've inspected it
1745          * then we just miss that fact, and the page stays dirty.
1746          *
1747          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1748          * handle that here by just cleaning them.
1749          */
1750
1751         block = page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1752         head = page_buffers(page);
1753         bh = head;
1754
1755         /*
1756          * Get all the dirty buffers mapped to disk addresses and
1757          * handle any aliases from the underlying blockdev's mapping.
1758          */
1759         do {
1760                 if (block > last_block) {
1761                         /*
1762                          * mapped buffers outside i_size will occur, because
1763                          * this page can be outside i_size when there is a
1764                          * truncate in progress.
1765                          */
1766                         /*
1767                          * The buffer was zeroed by block_write_full_page()
1768                          */
1769                         clear_buffer_dirty(bh);
1770                         set_buffer_uptodate(bh);
1771                 } else if (!buffer_mapped(bh) && buffer_dirty(bh)) {
1772                         err = get_block(inode, block, bh, 1);
1773                         if (err)
1774                                 goto recover;
1775                         if (buffer_new(bh)) {
1776                                 /* blockdev mappings never come here */
1777                                 clear_buffer_new(bh);
1778                                 unmap_underlying_metadata(bh->b_bdev,
1779                                                         bh->b_blocknr);
1780                         }
1781                 }
1782                 bh = bh->b_this_page;
1783                 block++;
1784         } while (bh != head);
1785
1786         do {
1787                 if (!buffer_mapped(bh))
1788                         continue;
1789                 /*
1790                  * If it's a fully non-blocking write attempt and we cannot
1791                  * lock the buffer then redirty the page.  Note that this can
1792                  * potentially cause a busy-wait loop from pdflush and kswapd
1793                  * activity, but those code paths have their own higher-level
1794                  * throttling.
1795                  */
1796                 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1797                         lock_buffer(bh);
1798                 } else if (test_set_buffer_locked(bh)) {
1799                         redirty_page_for_writepage(wbc, page);
1800                         continue;
1801                 }
1802                 if (test_clear_buffer_dirty(bh)) {
1803                         mark_buffer_async_write(bh);
1804                 } else {
1805                         unlock_buffer(bh);
1806                 }
1807         } while ((bh = bh->b_this_page) != head);
1808
1809         /*
1810          * The page and its buffers are protected by PageWriteback(), so we can
1811          * drop the bh refcounts early.
1812          */
1813         BUG_ON(PageWriteback(page));
1814         set_page_writeback(page);
1815
1816         do {
1817                 struct buffer_head *next = bh->b_this_page;
1818                 if (buffer_async_write(bh)) {
1819                         submit_bh(WRITE, bh);
1820                         nr_underway++;
1821                 }
1822                 bh = next;
1823         } while (bh != head);
1824         unlock_page(page);
1825
1826         err = 0;
1827 done:
1828         if (nr_underway == 0) {
1829                 /*
1830                  * The page was marked dirty, but the buffers were
1831                  * clean.  Someone wrote them back by hand with
1832                  * ll_rw_block/submit_bh.  A rare case.
1833                  */
1834                 int uptodate = 1;
1835                 do {
1836                         if (!buffer_uptodate(bh)) {
1837                                 uptodate = 0;
1838                                 break;
1839                         }
1840                         bh = bh->b_this_page;
1841                 } while (bh != head);
1842                 if (uptodate)
1843                         SetPageUptodate(page);
1844                 end_page_writeback(page);
1845                 /*
1846                  * The page and buffer_heads can be released at any time from
1847                  * here on.
1848                  */
1849                 wbc->pages_skipped++;   /* We didn't write this page */
1850         }
1851         return err;
1852
1853 recover:
1854         /*
1855          * ENOSPC, or some other error.  We may already have added some
1856          * blocks to the file, so we need to write these out to avoid
1857          * exposing stale data.
1858          * The page is currently locked and not marked for writeback
1859          */
1860         bh = head;
1861         /* Recovery: lock and submit the mapped buffers */
1862         do {
1863                 if (buffer_mapped(bh) && buffer_dirty(bh)) {
1864                         lock_buffer(bh);
1865                         mark_buffer_async_write(bh);
1866                 } else {
1867                         /*
1868                          * The buffer may have been set dirty during
1869                          * attachment to a dirty page.
1870                          */
1871                         clear_buffer_dirty(bh);
1872                 }
1873         } while ((bh = bh->b_this_page) != head);
1874         SetPageError(page);
1875         BUG_ON(PageWriteback(page));
1876         set_page_writeback(page);
1877         unlock_page(page);
1878         do {
1879                 struct buffer_head *next = bh->b_this_page;
1880                 if (buffer_async_write(bh)) {
1881                         clear_buffer_dirty(bh);
1882                         submit_bh(WRITE, bh);
1883                         nr_underway++;
1884                 }
1885                 bh = next;
1886         } while (bh != head);
1887         goto done;
1888 }
1889
1890 static int __block_prepare_write(struct inode *inode, struct page *page,
1891                 unsigned from, unsigned to, get_block_t *get_block)
1892 {
1893         unsigned block_start, block_end;
1894         sector_t block;
1895         int err = 0;
1896         unsigned blocksize, bbits;
1897         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1898
1899         BUG_ON(!PageLocked(page));
1900         BUG_ON(from > PAGE_CACHE_SIZE);
1901         BUG_ON(to > PAGE_CACHE_SIZE);
1902         BUG_ON(from > to);
1903
1904         blocksize = 1 << inode->i_blkbits;
1905         if (!page_has_buffers(page))
1906                 create_empty_buffers(page, blocksize, 0);
1907         head = page_buffers(page);
1908
1909         bbits = inode->i_blkbits;
1910         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1911
1912         for(bh = head, block_start = 0; bh != head || !block_start;
1913             block++, block_start=block_end, bh = bh->b_this_page) {
1914                 block_end = block_start + blocksize;
1915                 if (block_end <= from || block_start >= to) {
1916                         if (PageUptodate(page)) {
1917                                 if (!buffer_uptodate(bh))
1918                                         set_buffer_uptodate(bh);
1919                         }
1920                         continue;
1921                 }
1922                 if (buffer_new(bh))
1923                         clear_buffer_new(bh);
1924                 if (!buffer_mapped(bh)) {
1925                         err = get_block(inode, block, bh, 1);
1926                         if (err)
1927                                 break;
1928                         if (buffer_new(bh)) {
1929                                 unmap_underlying_metadata(bh->b_bdev,
1930                                                         bh->b_blocknr);
1931                                 if (PageUptodate(page)) {
1932                                         set_buffer_uptodate(bh);
1933                                         continue;
1934                                 }
1935                                 if (block_end > to || block_start < from) {
1936                                         void *kaddr;
1937
1938                                         kaddr = kmap_atomic(page, KM_USER0);
1939                                         if (block_end > to)
1940                                                 memset(kaddr+to, 0,
1941                                                         block_end-to);
1942                                         if (block_start < from)
1943                                                 memset(kaddr+block_start,
1944                                                         0, from-block_start);
1945                                         flush_dcache_page(page);
1946                                         kunmap_atomic(kaddr, KM_USER0);
1947                                 }
1948                                 continue;
1949                         }
1950                 }
1951                 if (PageUptodate(page)) {
1952                         if (!buffer_uptodate(bh))
1953                                 set_buffer_uptodate(bh);
1954                         continue; 
1955                 }
1956                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1957                      (block_start < from || block_end > to)) {
1958                         ll_rw_block(READ, 1, &bh);
1959                         *wait_bh++=bh;
1960                 }
1961         }
1962         /*
1963          * If we issued read requests - let them complete.
1964          */
1965         while(wait_bh > wait) {
1966                 wait_on_buffer(*--wait_bh);
1967                 if (!buffer_uptodate(*wait_bh))
1968                         err = -EIO;
1969         }
1970         if (!err) {
1971                 bh = head;
1972                 do {
1973                         if (buffer_new(bh))
1974                                 clear_buffer_new(bh);
1975                 } while ((bh = bh->b_this_page) != head);
1976                 return 0;
1977         }
1978         /* Error case: */
1979         /*
1980          * Zero out any newly allocated blocks to avoid exposing stale
1981          * data.  If BH_New is set, we know that the block was newly
1982          * allocated in the above loop.
1983          */
1984         bh = head;
1985         block_start = 0;
1986         do {
1987                 block_end = block_start+blocksize;
1988                 if (block_end <= from)
1989                         goto next_bh;
1990                 if (block_start >= to)
1991                         break;
1992                 if (buffer_new(bh)) {
1993                         void *kaddr;
1994
1995                         clear_buffer_new(bh);
1996                         kaddr = kmap_atomic(page, KM_USER0);
1997                         memset(kaddr+block_start, 0, bh->b_size);
1998                         kunmap_atomic(kaddr, KM_USER0);
1999                         set_buffer_uptodate(bh);
2000                         mark_buffer_dirty(bh);
2001                 }
2002 next_bh:
2003                 block_start = block_end;
2004                 bh = bh->b_this_page;
2005         } while (bh != head);
2006         return err;
2007 }
2008
2009 static int __block_commit_write(struct inode *inode, struct page *page,
2010                 unsigned from, unsigned to)
2011 {
2012         unsigned block_start, block_end;
2013         int partial = 0;
2014         unsigned blocksize;
2015         struct buffer_head *bh, *head;
2016
2017         blocksize = 1 << inode->i_blkbits;
2018
2019         for(bh = head = page_buffers(page), block_start = 0;
2020             bh != head || !block_start;
2021             block_start=block_end, bh = bh->b_this_page) {
2022                 block_end = block_start + blocksize;
2023                 if (block_end <= from || block_start >= to) {
2024                         if (!buffer_uptodate(bh))
2025                                 partial = 1;
2026                 } else {
2027                         set_buffer_uptodate(bh);
2028                         mark_buffer_dirty(bh);
2029                 }
2030         }
2031
2032         /*
2033          * If this is a partial write which happened to make all buffers
2034          * uptodate then we can optimize away a bogus readpage() for
2035          * the next read(). Here we 'discover' whether the page went
2036          * uptodate as a result of this (potentially partial) write.
2037          */
2038         if (!partial)
2039                 SetPageUptodate(page);
2040         return 0;
2041 }
2042
2043 /*
2044  * Generic "read page" function for block devices that have the normal
2045  * get_block functionality. This is most of the block device filesystems.
2046  * Reads the page asynchronously --- the unlock_buffer() and
2047  * set/clear_buffer_uptodate() functions propagate buffer state into the
2048  * page struct once IO has completed.
2049  */
2050 int block_read_full_page(struct page *page, get_block_t *get_block)
2051 {
2052         struct inode *inode = page->mapping->host;
2053         sector_t iblock, lblock;
2054         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2055         unsigned int blocksize;
2056         int nr, i;
2057         int fully_mapped = 1;
2058
2059         BUG_ON(!PageLocked(page));
2060         blocksize = 1 << inode->i_blkbits;
2061         if (!page_has_buffers(page))
2062                 create_empty_buffers(page, blocksize, 0);
2063         head = page_buffers(page);
2064
2065         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2066         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2067         bh = head;
2068         nr = 0;
2069         i = 0;
2070
2071         do {
2072                 if (buffer_uptodate(bh))
2073                         continue;
2074
2075                 if (!buffer_mapped(bh)) {
2076                         int err = 0;
2077
2078                         fully_mapped = 0;
2079                         if (iblock < lblock) {
2080                                 err = get_block(inode, iblock, bh, 0);
2081                                 if (err)
2082                                         SetPageError(page);
2083                         }
2084                         if (!buffer_mapped(bh)) {
2085                                 void *kaddr = kmap_atomic(page, KM_USER0);
2086                                 memset(kaddr + i * blocksize, 0, blocksize);
2087                                 flush_dcache_page(page);
2088                                 kunmap_atomic(kaddr, KM_USER0);
2089                                 if (!err)
2090                                         set_buffer_uptodate(bh);
2091                                 continue;
2092                         }
2093                         /*
2094                          * get_block() might have updated the buffer
2095                          * synchronously
2096                          */
2097                         if (buffer_uptodate(bh))
2098                                 continue;
2099                 }
2100                 arr[nr++] = bh;
2101         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2102
2103         if (fully_mapped)
2104                 SetPageMappedToDisk(page);
2105
2106         if (!nr) {
2107                 /*
2108                  * All buffers are uptodate - we can set the page uptodate
2109                  * as well. But not if get_block() returned an error.
2110                  */
2111                 if (!PageError(page))
2112                         SetPageUptodate(page);
2113                 unlock_page(page);
2114                 return 0;
2115         }
2116
2117         /* Stage two: lock the buffers */
2118         for (i = 0; i < nr; i++) {
2119                 bh = arr[i];
2120                 lock_buffer(bh);
2121                 mark_buffer_async_read(bh);
2122         }
2123
2124         /*
2125          * Stage 3: start the IO.  Check for uptodateness
2126          * inside the buffer lock in case another process reading
2127          * the underlying blockdev brought it uptodate (the sct fix).
2128          */
2129         for (i = 0; i < nr; i++) {
2130                 bh = arr[i];
2131                 if (buffer_uptodate(bh))
2132                         end_buffer_async_read(bh, 1);
2133                 else
2134                         submit_bh(READ, bh);
2135         }
2136         return 0;
2137 }
2138
2139 /* utility function for filesystems that need to do work on expanding
2140  * truncates.  Uses prepare/commit_write to allow the filesystem to
2141  * deal with the hole.  
2142  */
2143 int generic_cont_expand(struct inode *inode, loff_t size)
2144 {
2145         struct address_space *mapping = inode->i_mapping;
2146         struct page *page;
2147         unsigned long index, offset, limit;
2148         int err;
2149
2150         err = -EFBIG;
2151         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2152         if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2153                 send_sig(SIGXFSZ, current, 0);
2154                 goto out;
2155         }
2156         if (size > inode->i_sb->s_maxbytes)
2157                 goto out;
2158
2159         offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
2160
2161         /* ugh.  in prepare/commit_write, if from==to==start of block, we 
2162         ** skip the prepare.  make sure we never send an offset for the start
2163         ** of a block
2164         */
2165         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
2166                 offset++;
2167         }
2168         index = size >> PAGE_CACHE_SHIFT;
2169         err = -ENOMEM;
2170         page = grab_cache_page(mapping, index);
2171         if (!page)
2172                 goto out;
2173         err = mapping->a_ops->prepare_write(NULL, page, offset, offset);
2174         if (!err) {
2175                 err = mapping->a_ops->commit_write(NULL, page, offset, offset);
2176         }
2177         unlock_page(page);
2178         page_cache_release(page);
2179         if (err > 0)
2180                 err = 0;
2181 out:
2182         return err;
2183 }
2184
2185 /*
2186  * For moronic filesystems that do not allow holes in file.
2187  * We may have to extend the file.
2188  */
2189
2190 int cont_prepare_write(struct page *page, unsigned offset,
2191                 unsigned to, get_block_t *get_block, loff_t *bytes)
2192 {
2193         struct address_space *mapping = page->mapping;
2194         struct inode *inode = mapping->host;
2195         struct page *new_page;
2196         pgoff_t pgpos;
2197         long status;
2198         unsigned zerofrom;
2199         unsigned blocksize = 1 << inode->i_blkbits;
2200         void *kaddr;
2201
2202         while(page->index > (pgpos = *bytes>>PAGE_CACHE_SHIFT)) {
2203                 status = -ENOMEM;
2204                 new_page = grab_cache_page(mapping, pgpos);
2205                 if (!new_page)
2206                         goto out;
2207                 /* we might sleep */
2208                 if (*bytes>>PAGE_CACHE_SHIFT != pgpos) {
2209                         unlock_page(new_page);
2210                         page_cache_release(new_page);
2211                         continue;
2212                 }
2213                 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2214                 if (zerofrom & (blocksize-1)) {
2215                         *bytes |= (blocksize-1);
2216                         (*bytes)++;
2217                 }
2218                 status = __block_prepare_write(inode, new_page, zerofrom,
2219                                                 PAGE_CACHE_SIZE, get_block);
2220                 if (status)
2221                         goto out_unmap;
2222                 kaddr = kmap_atomic(new_page, KM_USER0);
2223                 memset(kaddr+zerofrom, 0, PAGE_CACHE_SIZE-zerofrom);
2224                 flush_dcache_page(new_page);
2225                 kunmap_atomic(kaddr, KM_USER0);
2226                 generic_commit_write(NULL, new_page, zerofrom, PAGE_CACHE_SIZE);
2227                 unlock_page(new_page);
2228                 page_cache_release(new_page);
2229         }
2230
2231         if (page->index < pgpos) {
2232                 /* completely inside the area */
2233                 zerofrom = offset;
2234         } else {
2235                 /* page covers the boundary, find the boundary offset */
2236                 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2237
2238                 /* if we will expand the thing last block will be filled */
2239                 if (to > zerofrom && (zerofrom & (blocksize-1))) {
2240                         *bytes |= (blocksize-1);
2241                         (*bytes)++;
2242                 }
2243
2244                 /* starting below the boundary? Nothing to zero out */
2245                 if (offset <= zerofrom)
2246                         zerofrom = offset;
2247         }
2248         status = __block_prepare_write(inode, page, zerofrom, to, get_block);
2249         if (status)
2250                 goto out1;
2251         if (zerofrom < offset) {
2252                 kaddr = kmap_atomic(page, KM_USER0);
2253                 memset(kaddr+zerofrom, 0, offset-zerofrom);
2254                 flush_dcache_page(page);
2255                 kunmap_atomic(kaddr, KM_USER0);
2256                 __block_commit_write(inode, page, zerofrom, offset);
2257         }
2258         return 0;
2259 out1:
2260         ClearPageUptodate(page);
2261         return status;
2262
2263 out_unmap:
2264         ClearPageUptodate(new_page);
2265         unlock_page(new_page);
2266         page_cache_release(new_page);
2267 out:
2268         return status;
2269 }
2270
2271 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2272                         get_block_t *get_block)
2273 {
2274         struct inode *inode = page->mapping->host;
2275         int err = __block_prepare_write(inode, page, from, to, get_block);
2276         if (err)
2277                 ClearPageUptodate(page);
2278         return err;
2279 }
2280
2281 int block_commit_write(struct page *page, unsigned from, unsigned to)
2282 {
2283         struct inode *inode = page->mapping->host;
2284         __block_commit_write(inode,page,from,to);
2285         return 0;
2286 }
2287
2288 int generic_commit_write(struct file *file, struct page *page,
2289                 unsigned from, unsigned to)
2290 {
2291         struct inode *inode = page->mapping->host;
2292         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2293         __block_commit_write(inode,page,from,to);
2294         /*
2295          * No need to use i_size_read() here, the i_size
2296          * cannot change under us because we hold i_sem.
2297          */
2298         if (pos > inode->i_size) {
2299                 i_size_write(inode, pos);
2300                 mark_inode_dirty(inode);
2301         }
2302         return 0;
2303 }
2304
2305
2306 /*
2307  * nobh_prepare_write()'s prereads are special: the buffer_heads are freed
2308  * immediately, while under the page lock.  So it needs a special end_io
2309  * handler which does not touch the bh after unlocking it.
2310  *
2311  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
2312  * a race there is benign: unlock_buffer() only use the bh's address for
2313  * hashing after unlocking the buffer, so it doesn't actually touch the bh
2314  * itself.
2315  */
2316 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2317 {
2318         if (uptodate) {
2319                 set_buffer_uptodate(bh);
2320         } else {
2321                 /* This happens, due to failed READA attempts. */
2322                 clear_buffer_uptodate(bh);
2323         }
2324         unlock_buffer(bh);
2325 }
2326
2327 /*
2328  * On entry, the page is fully not uptodate.
2329  * On exit the page is fully uptodate in the areas outside (from,to)
2330  */
2331 int nobh_prepare_write(struct page *page, unsigned from, unsigned to,
2332                         get_block_t *get_block)
2333 {
2334         struct inode *inode = page->mapping->host;
2335         const unsigned blkbits = inode->i_blkbits;
2336         const unsigned blocksize = 1 << blkbits;
2337         struct buffer_head map_bh;
2338         struct buffer_head *read_bh[MAX_BUF_PER_PAGE];
2339         unsigned block_in_page;
2340         unsigned block_start;
2341         sector_t block_in_file;
2342         char *kaddr;
2343         int nr_reads = 0;
2344         int i;
2345         int ret = 0;
2346         int is_mapped_to_disk = 1;
2347         int dirtied_it = 0;
2348
2349         if (PageMappedToDisk(page))
2350                 return 0;
2351
2352         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2353         map_bh.b_page = page;
2354
2355         /*
2356          * We loop across all blocks in the page, whether or not they are
2357          * part of the affected region.  This is so we can discover if the
2358          * page is fully mapped-to-disk.
2359          */
2360         for (block_start = 0, block_in_page = 0;
2361                   block_start < PAGE_CACHE_SIZE;
2362                   block_in_page++, block_start += blocksize) {
2363                 unsigned block_end = block_start + blocksize;
2364                 int create;
2365
2366                 map_bh.b_state = 0;
2367                 create = 1;
2368                 if (block_start >= to)
2369                         create = 0;
2370                 ret = get_block(inode, block_in_file + block_in_page,
2371                                         &map_bh, create);
2372                 if (ret)
2373                         goto failed;
2374                 if (!buffer_mapped(&map_bh))
2375                         is_mapped_to_disk = 0;
2376                 if (buffer_new(&map_bh))
2377                         unmap_underlying_metadata(map_bh.b_bdev,
2378                                                         map_bh.b_blocknr);
2379                 if (PageUptodate(page))
2380                         continue;
2381                 if (buffer_new(&map_bh) || !buffer_mapped(&map_bh)) {
2382                         kaddr = kmap_atomic(page, KM_USER0);
2383                         if (block_start < from) {
2384                                 memset(kaddr+block_start, 0, from-block_start);
2385                                 dirtied_it = 1;
2386                         }
2387                         if (block_end > to) {
2388                                 memset(kaddr + to, 0, block_end - to);
2389                                 dirtied_it = 1;
2390                         }
2391                         flush_dcache_page(page);
2392                         kunmap_atomic(kaddr, KM_USER0);
2393                         continue;
2394                 }
2395                 if (buffer_uptodate(&map_bh))
2396                         continue;       /* reiserfs does this */
2397                 if (block_start < from || block_end > to) {
2398                         struct buffer_head *bh = alloc_buffer_head(GFP_NOFS);
2399
2400                         if (!bh) {
2401                                 ret = -ENOMEM;
2402                                 goto failed;
2403                         }
2404                         bh->b_state = map_bh.b_state;
2405                         atomic_set(&bh->b_count, 0);
2406                         bh->b_this_page = NULL;
2407                         bh->b_page = page;
2408                         bh->b_blocknr = map_bh.b_blocknr;
2409                         bh->b_size = blocksize;
2410                         bh->b_data = (char *)(long)block_start;
2411                         bh->b_bdev = map_bh.b_bdev;
2412                         bh->b_private = NULL;
2413                         read_bh[nr_reads++] = bh;
2414                 }
2415         }
2416
2417         if (nr_reads) {
2418                 struct buffer_head *bh;
2419
2420                 /*
2421                  * The page is locked, so these buffers are protected from
2422                  * any VM or truncate activity.  Hence we don't need to care
2423                  * for the buffer_head refcounts.
2424                  */
2425                 for (i = 0; i < nr_reads; i++) {
2426                         bh = read_bh[i];
2427                         lock_buffer(bh);
2428                         bh->b_end_io = end_buffer_read_nobh;
2429                         submit_bh(READ, bh);
2430                 }
2431                 for (i = 0; i < nr_reads; i++) {
2432                         bh = read_bh[i];
2433                         wait_on_buffer(bh);
2434                         if (!buffer_uptodate(bh))
2435                                 ret = -EIO;
2436                         free_buffer_head(bh);
2437                         read_bh[i] = NULL;
2438                 }
2439                 if (ret)
2440                         goto failed;
2441         }
2442
2443         if (is_mapped_to_disk)
2444                 SetPageMappedToDisk(page);
2445         SetPageUptodate(page);
2446
2447         /*
2448          * Setting the page dirty here isn't necessary for the prepare_write
2449          * function - commit_write will do that.  But if/when this function is
2450          * used within the pagefault handler to ensure that all mmapped pages
2451          * have backing space in the filesystem, we will need to dirty the page
2452          * if its contents were altered.
2453          */
2454         if (dirtied_it)
2455                 set_page_dirty(page);
2456
2457         return 0;
2458
2459 failed:
2460         for (i = 0; i < nr_reads; i++) {
2461                 if (read_bh[i])
2462                         free_buffer_head(read_bh[i]);
2463         }
2464
2465         /*
2466          * Error recovery is pretty slack.  Clear the page and mark it dirty
2467          * so we'll later zero out any blocks which _were_ allocated.
2468          */
2469         kaddr = kmap_atomic(page, KM_USER0);
2470         memset(kaddr, 0, PAGE_CACHE_SIZE);
2471         kunmap_atomic(kaddr, KM_USER0);
2472         SetPageUptodate(page);
2473         set_page_dirty(page);
2474         return ret;
2475 }
2476 EXPORT_SYMBOL(nobh_prepare_write);
2477
2478 int nobh_commit_write(struct file *file, struct page *page,
2479                 unsigned from, unsigned to)
2480 {
2481         struct inode *inode = page->mapping->host;
2482         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2483
2484         set_page_dirty(page);
2485         if (pos > inode->i_size) {
2486                 i_size_write(inode, pos);
2487                 mark_inode_dirty(inode);
2488         }
2489         return 0;
2490 }
2491 EXPORT_SYMBOL(nobh_commit_write);
2492
2493 /*
2494  * nobh_writepage() - based on block_full_write_page() except
2495  * that it tries to operate without attaching bufferheads to
2496  * the page.
2497  */
2498 int nobh_writepage(struct page *page, get_block_t *get_block,
2499                         struct writeback_control *wbc)
2500 {
2501         struct inode * const inode = page->mapping->host;
2502         loff_t i_size = i_size_read(inode);
2503         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2504         unsigned offset;
2505         void *kaddr;
2506         int ret;
2507
2508         /* Is the page fully inside i_size? */
2509         if (page->index < end_index)
2510                 goto out;
2511
2512         /* Is the page fully outside i_size? (truncate in progress) */
2513         offset = i_size & (PAGE_CACHE_SIZE-1);
2514         if (page->index >= end_index+1 || !offset) {
2515                 /*
2516                  * The page may have dirty, unmapped buffers.  For example,
2517                  * they may have been added in ext3_writepage().  Make them
2518                  * freeable here, so the page does not leak.
2519                  */
2520 #if 0
2521                 /* Not really sure about this  - do we need this ? */
2522                 if (page->mapping->a_ops->invalidatepage)
2523                         page->mapping->a_ops->invalidatepage(page, offset);
2524 #endif
2525                 unlock_page(page);
2526                 return 0; /* don't care */
2527         }
2528
2529         /*
2530          * The page straddles i_size.  It must be zeroed out on each and every
2531          * writepage invocation because it may be mmapped.  "A file is mapped
2532          * in multiples of the page size.  For a file that is not a multiple of
2533          * the  page size, the remaining memory is zeroed when mapped, and
2534          * writes to that region are not written out to the file."
2535          */
2536         kaddr = kmap_atomic(page, KM_USER0);
2537         memset(kaddr + offset, 0, PAGE_CACHE_SIZE - offset);
2538         flush_dcache_page(page);
2539         kunmap_atomic(kaddr, KM_USER0);
2540 out:
2541         ret = mpage_writepage(page, get_block, wbc);
2542         if (ret == -EAGAIN)
2543                 ret = __block_write_full_page(inode, page, get_block, wbc);
2544         return ret;
2545 }
2546 EXPORT_SYMBOL(nobh_writepage);
2547
2548 /*
2549  * This function assumes that ->prepare_write() uses nobh_prepare_write().
2550  */
2551 int nobh_truncate_page(struct address_space *mapping, loff_t from)
2552 {
2553         struct inode *inode = mapping->host;
2554         unsigned blocksize = 1 << inode->i_blkbits;
2555         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2556         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2557         unsigned to;
2558         struct page *page;
2559         struct address_space_operations *a_ops = mapping->a_ops;
2560         char *kaddr;
2561         int ret = 0;
2562
2563         if ((offset & (blocksize - 1)) == 0)
2564                 goto out;
2565
2566         ret = -ENOMEM;
2567         page = grab_cache_page(mapping, index);
2568         if (!page)
2569                 goto out;
2570
2571         to = (offset + blocksize) & ~(blocksize - 1);
2572         ret = a_ops->prepare_write(NULL, page, offset, to);
2573         if (ret == 0) {
2574                 kaddr = kmap_atomic(page, KM_USER0);
2575                 memset(kaddr + offset, 0, PAGE_CACHE_SIZE - offset);
2576                 flush_dcache_page(page);
2577                 kunmap_atomic(kaddr, KM_USER0);
2578                 set_page_dirty(page);
2579         }
2580         unlock_page(page);
2581         page_cache_release(page);
2582 out:
2583         return ret;
2584 }
2585 EXPORT_SYMBOL(nobh_truncate_page);
2586
2587 int block_truncate_page(struct address_space *mapping,
2588                         loff_t from, get_block_t *get_block)
2589 {
2590         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2591         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2592         unsigned blocksize;
2593         pgoff_t iblock;
2594         unsigned length, pos;
2595         struct inode *inode = mapping->host;
2596         struct page *page;
2597         struct buffer_head *bh;
2598         void *kaddr;
2599         int err;
2600
2601         blocksize = 1 << inode->i_blkbits;
2602         length = offset & (blocksize - 1);
2603
2604         /* Block boundary? Nothing to do */
2605         if (!length)
2606                 return 0;
2607
2608         length = blocksize - length;
2609         iblock = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2610         
2611         page = grab_cache_page(mapping, index);
2612         err = -ENOMEM;
2613         if (!page)
2614                 goto out;
2615
2616         if (!page_has_buffers(page))
2617                 create_empty_buffers(page, blocksize, 0);
2618
2619         /* Find the buffer that contains "offset" */
2620         bh = page_buffers(page);
2621         pos = blocksize;
2622         while (offset >= pos) {
2623                 bh = bh->b_this_page;
2624                 iblock++;
2625                 pos += blocksize;
2626         }
2627
2628         err = 0;
2629         if (!buffer_mapped(bh)) {
2630                 err = get_block(inode, iblock, bh, 0);
2631                 if (err)
2632                         goto unlock;
2633                 /* unmapped? It's a hole - nothing to do */
2634                 if (!buffer_mapped(bh))
2635                         goto unlock;
2636         }
2637
2638         /* Ok, it's mapped. Make sure it's up-to-date */
2639         if (PageUptodate(page))
2640                 set_buffer_uptodate(bh);
2641
2642         if (!buffer_uptodate(bh) && !buffer_delay(bh)) {
2643                 err = -EIO;
2644                 ll_rw_block(READ, 1, &bh);
2645                 wait_on_buffer(bh);
2646                 /* Uhhuh. Read error. Complain and punt. */
2647                 if (!buffer_uptodate(bh))
2648                         goto unlock;
2649         }
2650
2651         kaddr = kmap_atomic(page, KM_USER0);
2652         memset(kaddr + offset, 0, length);
2653         flush_dcache_page(page);
2654         kunmap_atomic(kaddr, KM_USER0);
2655
2656         mark_buffer_dirty(bh);
2657         err = 0;
2658
2659 unlock:
2660         unlock_page(page);
2661         page_cache_release(page);
2662 out:
2663         return err;
2664 }
2665
2666 /*
2667  * The generic ->writepage function for buffer-backed address_spaces
2668  */
2669 int block_write_full_page(struct page *page, get_block_t *get_block,
2670                         struct writeback_control *wbc)
2671 {
2672         struct inode * const inode = page->mapping->host;
2673         loff_t i_size = i_size_read(inode);
2674         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2675         unsigned offset;
2676         void *kaddr;
2677
2678         /* Is the page fully inside i_size? */
2679         if (page->index < end_index)
2680                 return __block_write_full_page(inode, page, get_block, wbc);
2681
2682         /* Is the page fully outside i_size? (truncate in progress) */
2683         offset = i_size & (PAGE_CACHE_SIZE-1);
2684         if (page->index >= end_index+1 || !offset) {
2685                 /*
2686                  * The page may have dirty, unmapped buffers.  For example,
2687                  * they may have been added in ext3_writepage().  Make them
2688                  * freeable here, so the page does not leak.
2689                  */
2690                 block_invalidatepage(page, 0);
2691                 unlock_page(page);
2692                 return 0; /* don't care */
2693         }
2694
2695         /*
2696          * The page straddles i_size.  It must be zeroed out on each and every
2697          * writepage invokation because it may be mmapped.  "A file is mapped
2698          * in multiples of the page size.  For a file that is not a multiple of
2699          * the  page size, the remaining memory is zeroed when mapped, and
2700          * writes to that region are not written out to the file."
2701          */
2702         kaddr = kmap_atomic(page, KM_USER0);
2703         memset(kaddr + offset, 0, PAGE_CACHE_SIZE - offset);
2704         flush_dcache_page(page);
2705         kunmap_atomic(kaddr, KM_USER0);
2706         return __block_write_full_page(inode, page, get_block, wbc);
2707 }
2708
2709 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2710                             get_block_t *get_block)
2711 {
2712         struct buffer_head tmp;
2713         struct inode *inode = mapping->host;
2714         tmp.b_state = 0;
2715         tmp.b_blocknr = 0;
2716         get_block(inode, block, &tmp, 0);
2717         return tmp.b_blocknr;
2718 }
2719
2720 static int end_bio_bh_io_sync(struct bio *bio, unsigned int bytes_done, int err)
2721 {
2722         struct buffer_head *bh = bio->bi_private;
2723
2724         if (bio->bi_size)
2725                 return 1;
2726
2727         if (err == -EOPNOTSUPP) {
2728                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2729                 set_bit(BH_Eopnotsupp, &bh->b_state);
2730         }
2731
2732         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2733         bio_put(bio);
2734         return 0;
2735 }
2736
2737 int submit_bh(int rw, struct buffer_head * bh)
2738 {
2739         struct bio *bio;
2740         int ret = 0;
2741
2742         BUG_ON(!buffer_locked(bh));
2743         BUG_ON(!buffer_mapped(bh));
2744         BUG_ON(!bh->b_end_io);
2745
2746         if (buffer_ordered(bh) && (rw == WRITE))
2747                 rw = WRITE_BARRIER;
2748
2749         /*
2750          * Only clear out a write error when rewriting, should this
2751          * include WRITE_SYNC as well?
2752          */
2753         if (test_set_buffer_req(bh) && (rw == WRITE || rw == WRITE_BARRIER))
2754                 clear_buffer_write_io_error(bh);
2755
2756         /*
2757          * from here on down, it's all bio -- do the initial mapping,
2758          * submit_bio -> generic_make_request may further map this bio around
2759          */
2760         bio = bio_alloc(GFP_NOIO, 1);
2761
2762         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2763         bio->bi_bdev = bh->b_bdev;
2764         bio->bi_io_vec[0].bv_page = bh->b_page;
2765         bio->bi_io_vec[0].bv_len = bh->b_size;
2766         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2767
2768         bio->bi_vcnt = 1;
2769         bio->bi_idx = 0;
2770         bio->bi_size = bh->b_size;
2771
2772         bio->bi_end_io = end_bio_bh_io_sync;
2773         bio->bi_private = bh;
2774
2775         bio_get(bio);
2776         submit_bio(rw, bio);
2777
2778         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2779                 ret = -EOPNOTSUPP;
2780
2781         bio_put(bio);
2782         return ret;
2783 }
2784
2785 /**
2786  * ll_rw_block: low-level access to block devices (DEPRECATED)
2787  * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
2788  * @nr: number of &struct buffer_heads in the array
2789  * @bhs: array of pointers to &struct buffer_head
2790  *
2791  * ll_rw_block() takes an array of pointers to &struct buffer_heads,
2792  * and requests an I/O operation on them, either a %READ or a %WRITE.
2793  * The third %READA option is described in the documentation for
2794  * generic_make_request() which ll_rw_block() calls.
2795  *
2796  * This function drops any buffer that it cannot get a lock on (with the
2797  * BH_Lock state bit), any buffer that appears to be clean when doing a
2798  * write request, and any buffer that appears to be up-to-date when doing
2799  * read request.  Further it marks as clean buffers that are processed for
2800  * writing (the buffer cache won't assume that they are actually clean until
2801  * the buffer gets unlocked).
2802  *
2803  * ll_rw_block sets b_end_io to simple completion handler that marks
2804  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2805  * any waiters. 
2806  *
2807  * All of the buffers must be for the same device, and must also be a
2808  * multiple of the current approved size for the device.
2809  */
2810 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2811 {
2812         int i;
2813
2814         for (i = 0; i < nr; i++) {
2815                 struct buffer_head *bh = bhs[i];
2816
2817                 if (test_set_buffer_locked(bh))
2818                         continue;
2819
2820                 get_bh(bh);
2821                 if (rw == WRITE) {
2822                         if (test_clear_buffer_dirty(bh)) {
2823                                 bh->b_end_io = end_buffer_write_sync;
2824                                 submit_bh(WRITE, bh);
2825                                 continue;
2826                         }
2827                 } else {
2828                         if (!buffer_uptodate(bh)) {
2829                                 bh->b_end_io = end_buffer_read_sync;
2830                                 submit_bh(rw, bh);
2831                                 continue;
2832                         }
2833                 }
2834                 unlock_buffer(bh);
2835                 put_bh(bh);
2836         }
2837 }
2838
2839 /*
2840  * For a data-integrity writeout, we need to wait upon any in-progress I/O
2841  * and then start new I/O and then wait upon it.  The caller must have a ref on
2842  * the buffer_head.
2843  */
2844 int sync_dirty_buffer(struct buffer_head *bh)
2845 {
2846         int ret = 0;
2847
2848         WARN_ON(atomic_read(&bh->b_count) < 1);
2849         lock_buffer(bh);
2850         if (test_clear_buffer_dirty(bh)) {
2851                 get_bh(bh);
2852                 bh->b_end_io = end_buffer_write_sync;
2853                 ret = submit_bh(WRITE, bh);
2854                 wait_on_buffer(bh);
2855                 if (buffer_eopnotsupp(bh)) {
2856                         clear_buffer_eopnotsupp(bh);
2857                         ret = -EOPNOTSUPP;
2858                 }
2859                 if (!ret && !buffer_uptodate(bh))
2860                         ret = -EIO;
2861         } else {
2862                 unlock_buffer(bh);
2863         }
2864         return ret;
2865 }
2866
2867 /*
2868  * try_to_free_buffers() checks if all the buffers on this particular page
2869  * are unused, and releases them if so.
2870  *
2871  * Exclusion against try_to_free_buffers may be obtained by either
2872  * locking the page or by holding its mapping's private_lock.
2873  *
2874  * If the page is dirty but all the buffers are clean then we need to
2875  * be sure to mark the page clean as well.  This is because the page
2876  * may be against a block device, and a later reattachment of buffers
2877  * to a dirty page will set *all* buffers dirty.  Which would corrupt
2878  * filesystem data on the same device.
2879  *
2880  * The same applies to regular filesystem pages: if all the buffers are
2881  * clean then we set the page clean and proceed.  To do that, we require
2882  * total exclusion from __set_page_dirty_buffers().  That is obtained with
2883  * private_lock.
2884  *
2885  * try_to_free_buffers() is non-blocking.
2886  */
2887 static inline int buffer_busy(struct buffer_head *bh)
2888 {
2889         return atomic_read(&bh->b_count) |
2890                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
2891 }
2892
2893 static int
2894 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
2895 {
2896         struct buffer_head *head = page_buffers(page);
2897         struct buffer_head *bh;
2898
2899         bh = head;
2900         do {
2901                 if (buffer_write_io_error(bh) && page->mapping)
2902                         set_bit(AS_EIO, &page->mapping->flags);
2903                 if (buffer_busy(bh))
2904                         goto failed;
2905                 bh = bh->b_this_page;
2906         } while (bh != head);
2907
2908         do {
2909                 struct buffer_head *next = bh->b_this_page;
2910
2911                 if (!list_empty(&bh->b_assoc_buffers))
2912                         __remove_assoc_queue(bh);
2913                 bh = next;
2914         } while (bh != head);
2915         *buffers_to_free = head;
2916         __clear_page_buffers(page);
2917         return 1;
2918 failed:
2919         return 0;
2920 }
2921
2922 int try_to_free_buffers(struct page *page)
2923 {
2924         struct address_space * const mapping = page->mapping;
2925         struct buffer_head *buffers_to_free = NULL;
2926         int ret = 0;
2927
2928         BUG_ON(!PageLocked(page));
2929         if (PageWriteback(page))
2930                 return 0;
2931
2932         if (mapping == NULL) {          /* can this still happen? */
2933                 ret = drop_buffers(page, &buffers_to_free);
2934                 goto out;
2935         }
2936
2937         spin_lock(&mapping->private_lock);
2938         ret = drop_buffers(page, &buffers_to_free);
2939         if (ret) {
2940                 /*
2941                  * If the filesystem writes its buffers by hand (eg ext3)
2942                  * then we can have clean buffers against a dirty page.  We
2943                  * clean the page here; otherwise later reattachment of buffers
2944                  * could encounter a non-uptodate page, which is unresolvable.
2945                  * This only applies in the rare case where try_to_free_buffers
2946                  * succeeds but the page is not freed.
2947                  */
2948                 clear_page_dirty(page);
2949         }
2950         spin_unlock(&mapping->private_lock);
2951 out:
2952         if (buffers_to_free) {
2953                 struct buffer_head *bh = buffers_to_free;
2954
2955                 do {
2956                         struct buffer_head *next = bh->b_this_page;
2957                         free_buffer_head(bh);
2958                         bh = next;
2959                 } while (bh != buffers_to_free);
2960         }
2961         return ret;
2962 }
2963 EXPORT_SYMBOL(try_to_free_buffers);
2964
2965 int block_sync_page(struct page *page)
2966 {
2967         struct address_space *mapping;
2968
2969         smp_mb();
2970         mapping = page_mapping(page);
2971         if (mapping)
2972                 blk_run_backing_dev(mapping->backing_dev_info, page);
2973         return 0;
2974 }
2975
2976 /*
2977  * There are no bdflush tunables left.  But distributions are
2978  * still running obsolete flush daemons, so we terminate them here.
2979  *
2980  * Use of bdflush() is deprecated and will be removed in a future kernel.
2981  * The `pdflush' kernel threads fully replace bdflush daemons and this call.
2982  */
2983 asmlinkage long sys_bdflush(int func, long data)
2984 {
2985         static int msg_count;
2986
2987         if (!capable(CAP_SYS_ADMIN))
2988                 return -EPERM;
2989
2990         if (msg_count < 5) {
2991                 msg_count++;
2992                 printk(KERN_INFO
2993                         "warning: process `%s' used the obsolete bdflush"
2994                         " system call\n", current->comm);
2995                 printk(KERN_INFO "Fix your initscripts?\n");
2996         }
2997
2998         if (func == 1)
2999                 do_exit(0);
3000         return 0;
3001 }
3002
3003 /*
3004  * Buffer-head allocation
3005  */
3006 static kmem_cache_t *bh_cachep;
3007
3008 /*
3009  * Once the number of bh's in the machine exceeds this level, we start
3010  * stripping them in writeback.
3011  */
3012 static int max_buffer_heads;
3013
3014 int buffer_heads_over_limit;
3015
3016 struct bh_accounting {
3017         int nr;                 /* Number of live bh's */
3018         int ratelimit;          /* Limit cacheline bouncing */
3019 };
3020
3021 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3022
3023 static void recalc_bh_state(void)
3024 {
3025         int i;
3026         int tot = 0;
3027
3028         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3029                 return;
3030         __get_cpu_var(bh_accounting).ratelimit = 0;
3031         for_each_cpu(i)
3032                 tot += per_cpu(bh_accounting, i).nr;
3033         buffer_heads_over_limit = (tot > max_buffer_heads);
3034 }
3035         
3036 struct buffer_head *alloc_buffer_head(unsigned int __nocast gfp_flags)
3037 {
3038         struct buffer_head *ret = kmem_cache_alloc(bh_cachep, gfp_flags);
3039         if (ret) {
3040                 preempt_disable();
3041                 __get_cpu_var(bh_accounting).nr++;
3042                 recalc_bh_state();
3043                 preempt_enable();
3044         }
3045         return ret;
3046 }
3047 EXPORT_SYMBOL(alloc_buffer_head);
3048
3049 void free_buffer_head(struct buffer_head *bh)
3050 {
3051         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3052         kmem_cache_free(bh_cachep, bh);
3053         preempt_disable();
3054         __get_cpu_var(bh_accounting).nr--;
3055         recalc_bh_state();
3056         preempt_enable();
3057 }
3058 EXPORT_SYMBOL(free_buffer_head);
3059
3060 static void
3061 init_buffer_head(void *data, kmem_cache_t *cachep, unsigned long flags)
3062 {
3063         if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) ==
3064                             SLAB_CTOR_CONSTRUCTOR) {
3065                 struct buffer_head * bh = (struct buffer_head *)data;
3066
3067                 memset(bh, 0, sizeof(*bh));
3068                 INIT_LIST_HEAD(&bh->b_assoc_buffers);
3069         }
3070 }
3071
3072 #ifdef CONFIG_HOTPLUG_CPU
3073 static void buffer_exit_cpu(int cpu)
3074 {
3075         int i;
3076         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3077
3078         for (i = 0; i < BH_LRU_SIZE; i++) {
3079                 brelse(b->bhs[i]);
3080                 b->bhs[i] = NULL;
3081         }
3082 }
3083
3084 static int buffer_cpu_notify(struct notifier_block *self,
3085                               unsigned long action, void *hcpu)
3086 {
3087         if (action == CPU_DEAD)
3088                 buffer_exit_cpu((unsigned long)hcpu);
3089         return NOTIFY_OK;
3090 }
3091 #endif /* CONFIG_HOTPLUG_CPU */
3092
3093 void __init buffer_init(void)
3094 {
3095         int nrpages;
3096
3097         bh_cachep = kmem_cache_create("buffer_head",
3098                         sizeof(struct buffer_head), 0,
3099                         SLAB_RECLAIM_ACCOUNT|SLAB_PANIC, init_buffer_head, NULL);
3100
3101         /*
3102          * Limit the bh occupancy to 10% of ZONE_NORMAL
3103          */
3104         nrpages = (nr_free_buffer_pages() * 10) / 100;
3105         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3106         hotcpu_notifier(buffer_cpu_notify, 0);
3107 }
3108
3109 EXPORT_SYMBOL(__bforget);
3110 EXPORT_SYMBOL(__brelse);
3111 EXPORT_SYMBOL(__wait_on_buffer);
3112 EXPORT_SYMBOL(block_commit_write);
3113 EXPORT_SYMBOL(block_prepare_write);
3114 EXPORT_SYMBOL(block_read_full_page);
3115 EXPORT_SYMBOL(block_sync_page);
3116 EXPORT_SYMBOL(block_truncate_page);
3117 EXPORT_SYMBOL(block_write_full_page);
3118 EXPORT_SYMBOL(cont_prepare_write);
3119 EXPORT_SYMBOL(end_buffer_async_write);
3120 EXPORT_SYMBOL(end_buffer_read_sync);
3121 EXPORT_SYMBOL(end_buffer_write_sync);
3122 EXPORT_SYMBOL(file_fsync);
3123 EXPORT_SYMBOL(fsync_bdev);
3124 EXPORT_SYMBOL(generic_block_bmap);
3125 EXPORT_SYMBOL(generic_commit_write);
3126 EXPORT_SYMBOL(generic_cont_expand);
3127 EXPORT_SYMBOL(init_buffer);
3128 EXPORT_SYMBOL(invalidate_bdev);
3129 EXPORT_SYMBOL(ll_rw_block);
3130 EXPORT_SYMBOL(mark_buffer_dirty);
3131 EXPORT_SYMBOL(submit_bh);
3132 EXPORT_SYMBOL(sync_dirty_buffer);
3133 EXPORT_SYMBOL(unlock_buffer);