[PATCH] Make address_space_operations->invalidatepage return void
[linux-2.6.git] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/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/capability.h>
30 #include <linux/blkdev.h>
31 #include <linux/file.h>
32 #include <linux/quotaops.h>
33 #include <linux/highmem.h>
34 #include <linux/module.h>
35 #include <linux/writeback.h>
36 #include <linux/hash.h>
37 #include <linux/suspend.h>
38 #include <linux/buffer_head.h>
39 #include <linux/bio.h>
40 #include <linux/notifier.h>
41 #include <linux/cpu.h>
42 #include <linux/bitops.h>
43 #include <linux/mpage.h>
44 #include <linux/bit_spinlock.h>
45
46 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
47 static void invalidate_bh_lrus(void);
48
49 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
50
51 inline void
52 init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
53 {
54         bh->b_end_io = handler;
55         bh->b_private = private;
56 }
57
58 static int sync_buffer(void *word)
59 {
60         struct block_device *bd;
61         struct buffer_head *bh
62                 = container_of(word, struct buffer_head, b_state);
63
64         smp_mb();
65         bd = bh->b_bdev;
66         if (bd)
67                 blk_run_address_space(bd->bd_inode->i_mapping);
68         io_schedule();
69         return 0;
70 }
71
72 void fastcall __lock_buffer(struct buffer_head *bh)
73 {
74         wait_on_bit_lock(&bh->b_state, BH_Lock, sync_buffer,
75                                                         TASK_UNINTERRUPTIBLE);
76 }
77 EXPORT_SYMBOL(__lock_buffer);
78
79 void fastcall unlock_buffer(struct buffer_head *bh)
80 {
81         clear_buffer_locked(bh);
82         smp_mb__after_clear_bit();
83         wake_up_bit(&bh->b_state, BH_Lock);
84 }
85
86 /*
87  * Block until a buffer comes unlocked.  This doesn't stop it
88  * from becoming locked again - you have to lock it yourself
89  * if you want to preserve its state.
90  */
91 void __wait_on_buffer(struct buffer_head * bh)
92 {
93         wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
94 }
95
96 static void
97 __clear_page_buffers(struct page *page)
98 {
99         ClearPagePrivate(page);
100         set_page_private(page, 0);
101         page_cache_release(page);
102 }
103
104 static void buffer_io_error(struct buffer_head *bh)
105 {
106         char b[BDEVNAME_SIZE];
107
108         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
109                         bdevname(bh->b_bdev, b),
110                         (unsigned long long)bh->b_blocknr);
111 }
112
113 /*
114  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
115  * unlock the buffer. This is what ll_rw_block uses too.
116  */
117 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
118 {
119         if (uptodate) {
120                 set_buffer_uptodate(bh);
121         } else {
122                 /* This happens, due to failed READA attempts. */
123                 clear_buffer_uptodate(bh);
124         }
125         unlock_buffer(bh);
126         put_bh(bh);
127 }
128
129 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
130 {
131         char b[BDEVNAME_SIZE];
132
133         if (uptodate) {
134                 set_buffer_uptodate(bh);
135         } else {
136                 if (!buffer_eopnotsupp(bh) && printk_ratelimit()) {
137                         buffer_io_error(bh);
138                         printk(KERN_WARNING "lost page write due to "
139                                         "I/O error on %s\n",
140                                        bdevname(bh->b_bdev, b));
141                 }
142                 set_buffer_write_io_error(bh);
143                 clear_buffer_uptodate(bh);
144         }
145         unlock_buffer(bh);
146         put_bh(bh);
147 }
148
149 /*
150  * Write out and wait upon all the dirty data associated with a block
151  * device via its mapping.  Does not take the superblock lock.
152  */
153 int sync_blockdev(struct block_device *bdev)
154 {
155         int ret = 0;
156
157         if (bdev)
158                 ret = filemap_write_and_wait(bdev->bd_inode->i_mapping);
159         return ret;
160 }
161 EXPORT_SYMBOL(sync_blockdev);
162
163 static void __fsync_super(struct super_block *sb)
164 {
165         sync_inodes_sb(sb, 0);
166         DQUOT_SYNC(sb);
167         lock_super(sb);
168         if (sb->s_dirt && sb->s_op->write_super)
169                 sb->s_op->write_super(sb);
170         unlock_super(sb);
171         if (sb->s_op->sync_fs)
172                 sb->s_op->sync_fs(sb, 1);
173         sync_blockdev(sb->s_bdev);
174         sync_inodes_sb(sb, 1);
175 }
176
177 /*
178  * Write out and wait upon all dirty data associated with this
179  * superblock.  Filesystem data as well as the underlying block
180  * device.  Takes the superblock lock.
181  */
182 int fsync_super(struct super_block *sb)
183 {
184         __fsync_super(sb);
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_mutex 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         mutex_lock(&bdev->bd_mount_mutex);
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                 __fsync_super(sb);
224
225                 sb->s_frozen = SB_FREEZE_TRANS;
226                 smp_wmb();
227
228                 sync_blockdev(sb->s_bdev);
229
230                 if (sb->s_op->write_super_lockfs)
231                         sb->s_op->write_super_lockfs(sb);
232         }
233
234         sync_blockdev(bdev);
235         return sb;      /* thaw_bdev releases s->s_umount and bd_mount_sem */
236 }
237 EXPORT_SYMBOL(freeze_bdev);
238
239 /**
240  * thaw_bdev  -- unlock filesystem
241  * @bdev:       blockdevice to unlock
242  * @sb:         associated superblock
243  *
244  * Unlocks the filesystem and marks it writeable again after freeze_bdev().
245  */
246 void thaw_bdev(struct block_device *bdev, struct super_block *sb)
247 {
248         if (sb) {
249                 BUG_ON(sb->s_bdev != bdev);
250
251                 if (sb->s_op->unlockfs)
252                         sb->s_op->unlockfs(sb);
253                 sb->s_frozen = SB_UNFROZEN;
254                 smp_wmb();
255                 wake_up(&sb->s_wait_unfrozen);
256                 drop_super(sb);
257         }
258
259         mutex_unlock(&bdev->bd_mount_mutex);
260 }
261 EXPORT_SYMBOL(thaw_bdev);
262
263 /*
264  * sync everything.  Start out by waking pdflush, because that writes back
265  * all queues in parallel.
266  */
267 static void do_sync(unsigned long wait)
268 {
269         wakeup_pdflush(0);
270         sync_inodes(0);         /* All mappings, inodes and their blockdevs */
271         DQUOT_SYNC(NULL);
272         sync_supers();          /* Write the superblocks */
273         sync_filesystems(0);    /* Start syncing the filesystems */
274         sync_filesystems(wait); /* Waitingly sync the filesystems */
275         sync_inodes(wait);      /* Mappings, inodes and blockdevs, again. */
276         if (!wait)
277                 printk("Emergency Sync complete\n");
278         if (unlikely(laptop_mode))
279                 laptop_sync_completion();
280 }
281
282 asmlinkage long sys_sync(void)
283 {
284         do_sync(1);
285         return 0;
286 }
287
288 void emergency_sync(void)
289 {
290         pdflush_operation(do_sync, 0);
291 }
292
293 /*
294  * Generic function to fsync a file.
295  *
296  * filp may be NULL if called via the msync of a vma.
297  */
298  
299 int file_fsync(struct file *filp, struct dentry *dentry, int datasync)
300 {
301         struct inode * inode = dentry->d_inode;
302         struct super_block * sb;
303         int ret, err;
304
305         /* sync the inode to buffers */
306         ret = write_inode_now(inode, 0);
307
308         /* sync the superblock to buffers */
309         sb = inode->i_sb;
310         lock_super(sb);
311         if (sb->s_op->write_super)
312                 sb->s_op->write_super(sb);
313         unlock_super(sb);
314
315         /* .. finally sync the buffers to disk */
316         err = sync_blockdev(sb->s_bdev);
317         if (!ret)
318                 ret = err;
319         return ret;
320 }
321
322 long do_fsync(struct file *file, int datasync)
323 {
324         int ret;
325         int err;
326         struct address_space *mapping = file->f_mapping;
327
328         if (!file->f_op || !file->f_op->fsync) {
329                 /* Why?  We can still call filemap_fdatawrite */
330                 ret = -EINVAL;
331                 goto out;
332         }
333
334         current->flags |= PF_SYNCWRITE;
335         ret = filemap_fdatawrite(mapping);
336
337         /*
338          * We need to protect against concurrent writers, which could cause
339          * livelocks in fsync_buffers_list().
340          */
341         mutex_lock(&mapping->host->i_mutex);
342         err = file->f_op->fsync(file, file->f_dentry, datasync);
343         if (!ret)
344                 ret = err;
345         mutex_unlock(&mapping->host->i_mutex);
346         err = filemap_fdatawait(mapping);
347         if (!ret)
348                 ret = err;
349         current->flags &= ~PF_SYNCWRITE;
350 out:
351         return ret;
352 }
353
354 static long __do_fsync(unsigned int fd, int datasync)
355 {
356         struct file *file;
357         int ret = -EBADF;
358
359         file = fget(fd);
360         if (file) {
361                 ret = do_fsync(file, datasync);
362                 fput(file);
363         }
364         return ret;
365 }
366
367 asmlinkage long sys_fsync(unsigned int fd)
368 {
369         return __do_fsync(fd, 0);
370 }
371
372 asmlinkage long sys_fdatasync(unsigned int fd)
373 {
374         return __do_fsync(fd, 1);
375 }
376
377 /*
378  * Various filesystems appear to want __find_get_block to be non-blocking.
379  * But it's the page lock which protects the buffers.  To get around this,
380  * we get exclusion from try_to_free_buffers with the blockdev mapping's
381  * private_lock.
382  *
383  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
384  * may be quite high.  This code could TryLock the page, and if that
385  * succeeds, there is no need to take private_lock. (But if
386  * private_lock is contended then so is mapping->tree_lock).
387  */
388 static struct buffer_head *
389 __find_get_block_slow(struct block_device *bdev, sector_t block)
390 {
391         struct inode *bd_inode = bdev->bd_inode;
392         struct address_space *bd_mapping = bd_inode->i_mapping;
393         struct buffer_head *ret = NULL;
394         pgoff_t index;
395         struct buffer_head *bh;
396         struct buffer_head *head;
397         struct page *page;
398         int all_mapped = 1;
399
400         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
401         page = find_get_page(bd_mapping, index);
402         if (!page)
403                 goto out;
404
405         spin_lock(&bd_mapping->private_lock);
406         if (!page_has_buffers(page))
407                 goto out_unlock;
408         head = page_buffers(page);
409         bh = head;
410         do {
411                 if (bh->b_blocknr == block) {
412                         ret = bh;
413                         get_bh(bh);
414                         goto out_unlock;
415                 }
416                 if (!buffer_mapped(bh))
417                         all_mapped = 0;
418                 bh = bh->b_this_page;
419         } while (bh != head);
420
421         /* we might be here because some of the buffers on this page are
422          * not mapped.  This is due to various races between
423          * file io on the block device and getblk.  It gets dealt with
424          * elsewhere, don't buffer_error if we had some unmapped buffers
425          */
426         if (all_mapped) {
427                 printk("__find_get_block_slow() failed. "
428                         "block=%llu, b_blocknr=%llu\n",
429                         (unsigned long long)block, (unsigned long long)bh->b_blocknr);
430                 printk("b_state=0x%08lx, b_size=%u\n", bh->b_state, bh->b_size);
431                 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
432         }
433 out_unlock:
434         spin_unlock(&bd_mapping->private_lock);
435         page_cache_release(page);
436 out:
437         return ret;
438 }
439
440 /* If invalidate_buffers() will trash dirty buffers, it means some kind
441    of fs corruption is going on. Trashing dirty data always imply losing
442    information that was supposed to be just stored on the physical layer
443    by the user.
444
445    Thus invalidate_buffers in general usage is not allwowed to trash
446    dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
447    be preserved.  These buffers are simply skipped.
448   
449    We also skip buffers which are still in use.  For example this can
450    happen if a userspace program is reading the block device.
451
452    NOTE: In the case where the user removed a removable-media-disk even if
453    there's still dirty data not synced on disk (due a bug in the device driver
454    or due an error of the user), by not destroying the dirty buffers we could
455    generate corruption also on the next media inserted, thus a parameter is
456    necessary to handle this case in the most safe way possible (trying
457    to not corrupt also the new disk inserted with the data belonging to
458    the old now corrupted disk). Also for the ramdisk the natural thing
459    to do in order to release the ramdisk memory is to destroy dirty buffers.
460
461    These are two special cases. Normal usage imply the device driver
462    to issue a sync on the device (without waiting I/O completion) and
463    then an invalidate_buffers call that doesn't trash dirty buffers.
464
465    For handling cache coherency with the blkdev pagecache the 'update' case
466    is been introduced. It is needed to re-read from disk any pinned
467    buffer. NOTE: re-reading from disk is destructive so we can do it only
468    when we assume nobody is changing the buffercache under our I/O and when
469    we think the disk contains more recent information than the buffercache.
470    The update == 1 pass marks the buffers we need to update, the update == 2
471    pass does the actual I/O. */
472 void invalidate_bdev(struct block_device *bdev, int destroy_dirty_buffers)
473 {
474         invalidate_bh_lrus();
475         /*
476          * FIXME: what about destroy_dirty_buffers?
477          * We really want to use invalidate_inode_pages2() for
478          * that, but not until that's cleaned up.
479          */
480         invalidate_inode_pages(bdev->bd_inode->i_mapping);
481 }
482
483 /*
484  * Kick pdflush then try to free up some ZONE_NORMAL memory.
485  */
486 static void free_more_memory(void)
487 {
488         struct zone **zones;
489         pg_data_t *pgdat;
490
491         wakeup_pdflush(1024);
492         yield();
493
494         for_each_pgdat(pgdat) {
495                 zones = pgdat->node_zonelists[gfp_zone(GFP_NOFS)].zones;
496                 if (*zones)
497                         try_to_free_pages(zones, GFP_NOFS);
498         }
499 }
500
501 /*
502  * I/O completion handler for block_read_full_page() - pages
503  * which come unlocked at the end of I/O.
504  */
505 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
506 {
507         unsigned long flags;
508         struct buffer_head *first;
509         struct buffer_head *tmp;
510         struct page *page;
511         int page_uptodate = 1;
512
513         BUG_ON(!buffer_async_read(bh));
514
515         page = bh->b_page;
516         if (uptodate) {
517                 set_buffer_uptodate(bh);
518         } else {
519                 clear_buffer_uptodate(bh);
520                 if (printk_ratelimit())
521                         buffer_io_error(bh);
522                 SetPageError(page);
523         }
524
525         /*
526          * Be _very_ careful from here on. Bad things can happen if
527          * two buffer heads end IO at almost the same time and both
528          * decide that the page is now completely done.
529          */
530         first = page_buffers(page);
531         local_irq_save(flags);
532         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
533         clear_buffer_async_read(bh);
534         unlock_buffer(bh);
535         tmp = bh;
536         do {
537                 if (!buffer_uptodate(tmp))
538                         page_uptodate = 0;
539                 if (buffer_async_read(tmp)) {
540                         BUG_ON(!buffer_locked(tmp));
541                         goto still_busy;
542                 }
543                 tmp = tmp->b_this_page;
544         } while (tmp != bh);
545         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
546         local_irq_restore(flags);
547
548         /*
549          * If none of the buffers had errors and they are all
550          * uptodate then we can set the page uptodate.
551          */
552         if (page_uptodate && !PageError(page))
553                 SetPageUptodate(page);
554         unlock_page(page);
555         return;
556
557 still_busy:
558         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
559         local_irq_restore(flags);
560         return;
561 }
562
563 /*
564  * Completion handler for block_write_full_page() - pages which are unlocked
565  * during I/O, and which have PageWriteback cleared upon I/O completion.
566  */
567 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
568 {
569         char b[BDEVNAME_SIZE];
570         unsigned long flags;
571         struct buffer_head *first;
572         struct buffer_head *tmp;
573         struct page *page;
574
575         BUG_ON(!buffer_async_write(bh));
576
577         page = bh->b_page;
578         if (uptodate) {
579                 set_buffer_uptodate(bh);
580         } else {
581                 if (printk_ratelimit()) {
582                         buffer_io_error(bh);
583                         printk(KERN_WARNING "lost page write due to "
584                                         "I/O error on %s\n",
585                                bdevname(bh->b_bdev, b));
586                 }
587                 set_bit(AS_EIO, &page->mapping->flags);
588                 clear_buffer_uptodate(bh);
589                 SetPageError(page);
590         }
591
592         first = page_buffers(page);
593         local_irq_save(flags);
594         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
595
596         clear_buffer_async_write(bh);
597         unlock_buffer(bh);
598         tmp = bh->b_this_page;
599         while (tmp != bh) {
600                 if (buffer_async_write(tmp)) {
601                         BUG_ON(!buffer_locked(tmp));
602                         goto still_busy;
603                 }
604                 tmp = tmp->b_this_page;
605         }
606         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
607         local_irq_restore(flags);
608         end_page_writeback(page);
609         return;
610
611 still_busy:
612         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
613         local_irq_restore(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                 return 1;
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                                 ll_rw_block(SWRITE, 1, &bh);
912                                 brelse(bh);
913                                 spin_lock(lock);
914                         }
915                 }
916         }
917
918         while (!list_empty(&tmp)) {
919                 bh = BH_ENTRY(tmp.prev);
920                 __remove_assoc_queue(bh);
921                 get_bh(bh);
922                 spin_unlock(lock);
923                 wait_on_buffer(bh);
924                 if (!buffer_uptodate(bh))
925                         err = -EIO;
926                 brelse(bh);
927                 spin_lock(lock);
928         }
929         
930         spin_unlock(lock);
931         err2 = osync_buffers_list(lock, list);
932         if (err)
933                 return err;
934         else
935                 return err2;
936 }
937
938 /*
939  * Invalidate any and all dirty buffers on a given inode.  We are
940  * probably unmounting the fs, but that doesn't mean we have already
941  * done a sync().  Just drop the buffers from the inode list.
942  *
943  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
944  * assumes that all the buffers are against the blockdev.  Not true
945  * for reiserfs.
946  */
947 void invalidate_inode_buffers(struct inode *inode)
948 {
949         if (inode_has_buffers(inode)) {
950                 struct address_space *mapping = &inode->i_data;
951                 struct list_head *list = &mapping->private_list;
952                 struct address_space *buffer_mapping = mapping->assoc_mapping;
953
954                 spin_lock(&buffer_mapping->private_lock);
955                 while (!list_empty(list))
956                         __remove_assoc_queue(BH_ENTRY(list->next));
957                 spin_unlock(&buffer_mapping->private_lock);
958         }
959 }
960
961 /*
962  * Remove any clean buffers from the inode's buffer list.  This is called
963  * when we're trying to free the inode itself.  Those buffers can pin it.
964  *
965  * Returns true if all buffers were removed.
966  */
967 int remove_inode_buffers(struct inode *inode)
968 {
969         int ret = 1;
970
971         if (inode_has_buffers(inode)) {
972                 struct address_space *mapping = &inode->i_data;
973                 struct list_head *list = &mapping->private_list;
974                 struct address_space *buffer_mapping = mapping->assoc_mapping;
975
976                 spin_lock(&buffer_mapping->private_lock);
977                 while (!list_empty(list)) {
978                         struct buffer_head *bh = BH_ENTRY(list->next);
979                         if (buffer_dirty(bh)) {
980                                 ret = 0;
981                                 break;
982                         }
983                         __remove_assoc_queue(bh);
984                 }
985                 spin_unlock(&buffer_mapping->private_lock);
986         }
987         return ret;
988 }
989
990 /*
991  * Create the appropriate buffers when given a page for data area and
992  * the size of each buffer.. Use the bh->b_this_page linked list to
993  * follow the buffers created.  Return NULL if unable to create more
994  * buffers.
995  *
996  * The retry flag is used to differentiate async IO (paging, swapping)
997  * which may not fail from ordinary buffer allocations.
998  */
999 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
1000                 int retry)
1001 {
1002         struct buffer_head *bh, *head;
1003         long offset;
1004
1005 try_again:
1006         head = NULL;
1007         offset = PAGE_SIZE;
1008         while ((offset -= size) >= 0) {
1009                 bh = alloc_buffer_head(GFP_NOFS);
1010                 if (!bh)
1011                         goto no_grow;
1012
1013                 bh->b_bdev = NULL;
1014                 bh->b_this_page = head;
1015                 bh->b_blocknr = -1;
1016                 head = bh;
1017
1018                 bh->b_state = 0;
1019                 atomic_set(&bh->b_count, 0);
1020                 bh->b_private = NULL;
1021                 bh->b_size = size;
1022
1023                 /* Link the buffer to its page */
1024                 set_bh_page(bh, page, offset);
1025
1026                 init_buffer(bh, NULL, 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 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 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);
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         if (likely(bh)) {
1473                 ll_rw_block(READA, 1, &bh);
1474                 brelse(bh);
1475         }
1476 }
1477 EXPORT_SYMBOL(__breadahead);
1478
1479 /**
1480  *  __bread() - reads a specified block and returns the bh
1481  *  @bdev: the block_device to read from
1482  *  @block: number of block
1483  *  @size: size (in bytes) to read
1484  * 
1485  *  Reads a specified block, and returns buffer head that contains it.
1486  *  It returns NULL if the block was unreadable.
1487  */
1488 struct buffer_head *
1489 __bread(struct block_device *bdev, sector_t block, int size)
1490 {
1491         struct buffer_head *bh = __getblk(bdev, block, size);
1492
1493         if (likely(bh) && !buffer_uptodate(bh))
1494                 bh = __bread_slow(bh);
1495         return bh;
1496 }
1497 EXPORT_SYMBOL(__bread);
1498
1499 /*
1500  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1501  * This doesn't race because it runs in each cpu either in irq
1502  * or with preempt disabled.
1503  */
1504 static void invalidate_bh_lru(void *arg)
1505 {
1506         struct bh_lru *b = &get_cpu_var(bh_lrus);
1507         int i;
1508
1509         for (i = 0; i < BH_LRU_SIZE; i++) {
1510                 brelse(b->bhs[i]);
1511                 b->bhs[i] = NULL;
1512         }
1513         put_cpu_var(bh_lrus);
1514 }
1515         
1516 static void invalidate_bh_lrus(void)
1517 {
1518         on_each_cpu(invalidate_bh_lru, NULL, 1, 1);
1519 }
1520
1521 void set_bh_page(struct buffer_head *bh,
1522                 struct page *page, unsigned long offset)
1523 {
1524         bh->b_page = page;
1525         if (offset >= PAGE_SIZE)
1526                 BUG();
1527         if (PageHighMem(page))
1528                 /*
1529                  * This catches illegal uses and preserves the offset:
1530                  */
1531                 bh->b_data = (char *)(0 + offset);
1532         else
1533                 bh->b_data = page_address(page) + offset;
1534 }
1535 EXPORT_SYMBOL(set_bh_page);
1536
1537 /*
1538  * Called when truncating a buffer on a page completely.
1539  */
1540 static void discard_buffer(struct buffer_head * bh)
1541 {
1542         lock_buffer(bh);
1543         clear_buffer_dirty(bh);
1544         bh->b_bdev = NULL;
1545         clear_buffer_mapped(bh);
1546         clear_buffer_req(bh);
1547         clear_buffer_new(bh);
1548         clear_buffer_delay(bh);
1549         unlock_buffer(bh);
1550 }
1551
1552 /**
1553  * try_to_release_page() - release old fs-specific metadata on a page
1554  *
1555  * @page: the page which the kernel is trying to free
1556  * @gfp_mask: memory allocation flags (and I/O mode)
1557  *
1558  * The address_space is to try to release any data against the page
1559  * (presumably at page->private).  If the release was successful, return `1'.
1560  * Otherwise return zero.
1561  *
1562  * The @gfp_mask argument specifies whether I/O may be performed to release
1563  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
1564  *
1565  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
1566  */
1567 int try_to_release_page(struct page *page, gfp_t gfp_mask)
1568 {
1569         struct address_space * const mapping = page->mapping;
1570
1571         BUG_ON(!PageLocked(page));
1572         if (PageWriteback(page))
1573                 return 0;
1574         
1575         if (mapping && mapping->a_ops->releasepage)
1576                 return mapping->a_ops->releasepage(page, gfp_mask);
1577         return try_to_free_buffers(page);
1578 }
1579 EXPORT_SYMBOL(try_to_release_page);
1580
1581 /**
1582  * block_invalidatepage - invalidate part of all of a buffer-backed page
1583  *
1584  * @page: the page which is affected
1585  * @offset: the index of the truncation point
1586  *
1587  * block_invalidatepage() is called when all or part of the page has become
1588  * invalidatedby a truncate operation.
1589  *
1590  * block_invalidatepage() does not have to release all buffers, but it must
1591  * ensure that no dirty buffer is left outside @offset and that no I/O
1592  * is underway against any of the blocks which are outside the truncation
1593  * point.  Because the caller is about to free (and possibly reuse) those
1594  * blocks on-disk.
1595  */
1596 void block_invalidatepage(struct page *page, unsigned long offset)
1597 {
1598         struct buffer_head *head, *bh, *next;
1599         unsigned int curr_off = 0;
1600
1601         BUG_ON(!PageLocked(page));
1602         if (!page_has_buffers(page))
1603                 goto out;
1604
1605         head = page_buffers(page);
1606         bh = head;
1607         do {
1608                 unsigned int next_off = curr_off + bh->b_size;
1609                 next = bh->b_this_page;
1610
1611                 /*
1612                  * is this block fully invalidated?
1613                  */
1614                 if (offset <= curr_off)
1615                         discard_buffer(bh);
1616                 curr_off = next_off;
1617                 bh = next;
1618         } while (bh != head);
1619
1620         /*
1621          * We release buffers only if the entire page is being invalidated.
1622          * The get_block cached value has been unconditionally invalidated,
1623          * so real IO is not possible anymore.
1624          */
1625         if (offset == 0)
1626                 try_to_release_page(page, 0);
1627 out:
1628         return;
1629 }
1630 EXPORT_SYMBOL(block_invalidatepage);
1631
1632 void do_invalidatepage(struct page *page, unsigned long offset)
1633 {
1634         void (*invalidatepage)(struct page *, unsigned long);
1635         invalidatepage = page->mapping->a_ops->invalidatepage ? :
1636                 block_invalidatepage;
1637         (*invalidatepage)(page, offset);
1638 }
1639
1640 /*
1641  * We attach and possibly dirty the buffers atomically wrt
1642  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1643  * is already excluded via the page lock.
1644  */
1645 void create_empty_buffers(struct page *page,
1646                         unsigned long blocksize, unsigned long b_state)
1647 {
1648         struct buffer_head *bh, *head, *tail;
1649
1650         head = alloc_page_buffers(page, blocksize, 1);
1651         bh = head;
1652         do {
1653                 bh->b_state |= b_state;
1654                 tail = bh;
1655                 bh = bh->b_this_page;
1656         } while (bh);
1657         tail->b_this_page = head;
1658
1659         spin_lock(&page->mapping->private_lock);
1660         if (PageUptodate(page) || PageDirty(page)) {
1661                 bh = head;
1662                 do {
1663                         if (PageDirty(page))
1664                                 set_buffer_dirty(bh);
1665                         if (PageUptodate(page))
1666                                 set_buffer_uptodate(bh);
1667                         bh = bh->b_this_page;
1668                 } while (bh != head);
1669         }
1670         attach_page_buffers(page, head);
1671         spin_unlock(&page->mapping->private_lock);
1672 }
1673 EXPORT_SYMBOL(create_empty_buffers);
1674
1675 /*
1676  * We are taking a block for data and we don't want any output from any
1677  * buffer-cache aliases starting from return from that function and
1678  * until the moment when something will explicitly mark the buffer
1679  * dirty (hopefully that will not happen until we will free that block ;-)
1680  * We don't even need to mark it not-uptodate - nobody can expect
1681  * anything from a newly allocated buffer anyway. We used to used
1682  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1683  * don't want to mark the alias unmapped, for example - it would confuse
1684  * anyone who might pick it with bread() afterwards...
1685  *
1686  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1687  * be writeout I/O going on against recently-freed buffers.  We don't
1688  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1689  * only if we really need to.  That happens here.
1690  */
1691 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1692 {
1693         struct buffer_head *old_bh;
1694
1695         might_sleep();
1696
1697         old_bh = __find_get_block_slow(bdev, block);
1698         if (old_bh) {
1699                 clear_buffer_dirty(old_bh);
1700                 wait_on_buffer(old_bh);
1701                 clear_buffer_req(old_bh);
1702                 __brelse(old_bh);
1703         }
1704 }
1705 EXPORT_SYMBOL(unmap_underlying_metadata);
1706
1707 /*
1708  * NOTE! All mapped/uptodate combinations are valid:
1709  *
1710  *      Mapped  Uptodate        Meaning
1711  *
1712  *      No      No              "unknown" - must do get_block()
1713  *      No      Yes             "hole" - zero-filled
1714  *      Yes     No              "allocated" - allocated on disk, not read in
1715  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1716  *
1717  * "Dirty" is valid only with the last case (mapped+uptodate).
1718  */
1719
1720 /*
1721  * While block_write_full_page is writing back the dirty buffers under
1722  * the page lock, whoever dirtied the buffers may decide to clean them
1723  * again at any time.  We handle that by only looking at the buffer
1724  * state inside lock_buffer().
1725  *
1726  * If block_write_full_page() is called for regular writeback
1727  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1728  * locked buffer.   This only can happen if someone has written the buffer
1729  * directly, with submit_bh().  At the address_space level PageWriteback
1730  * prevents this contention from occurring.
1731  */
1732 static int __block_write_full_page(struct inode *inode, struct page *page,
1733                         get_block_t *get_block, struct writeback_control *wbc)
1734 {
1735         int err;
1736         sector_t block;
1737         sector_t last_block;
1738         struct buffer_head *bh, *head;
1739         int nr_underway = 0;
1740
1741         BUG_ON(!PageLocked(page));
1742
1743         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1744
1745         if (!page_has_buffers(page)) {
1746                 create_empty_buffers(page, 1 << inode->i_blkbits,
1747                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1748         }
1749
1750         /*
1751          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1752          * here, and the (potentially unmapped) buffers may become dirty at
1753          * any time.  If a buffer becomes dirty here after we've inspected it
1754          * then we just miss that fact, and the page stays dirty.
1755          *
1756          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1757          * handle that here by just cleaning them.
1758          */
1759
1760         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1761         head = page_buffers(page);
1762         bh = head;
1763
1764         /*
1765          * Get all the dirty buffers mapped to disk addresses and
1766          * handle any aliases from the underlying blockdev's mapping.
1767          */
1768         do {
1769                 if (block > last_block) {
1770                         /*
1771                          * mapped buffers outside i_size will occur, because
1772                          * this page can be outside i_size when there is a
1773                          * truncate in progress.
1774                          */
1775                         /*
1776                          * The buffer was zeroed by block_write_full_page()
1777                          */
1778                         clear_buffer_dirty(bh);
1779                         set_buffer_uptodate(bh);
1780                 } else if (!buffer_mapped(bh) && buffer_dirty(bh)) {
1781                         err = get_block(inode, block, bh, 1);
1782                         if (err)
1783                                 goto recover;
1784                         if (buffer_new(bh)) {
1785                                 /* blockdev mappings never come here */
1786                                 clear_buffer_new(bh);
1787                                 unmap_underlying_metadata(bh->b_bdev,
1788                                                         bh->b_blocknr);
1789                         }
1790                 }
1791                 bh = bh->b_this_page;
1792                 block++;
1793         } while (bh != head);
1794
1795         do {
1796                 if (!buffer_mapped(bh))
1797                         continue;
1798                 /*
1799                  * If it's a fully non-blocking write attempt and we cannot
1800                  * lock the buffer then redirty the page.  Note that this can
1801                  * potentially cause a busy-wait loop from pdflush and kswapd
1802                  * activity, but those code paths have their own higher-level
1803                  * throttling.
1804                  */
1805                 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1806                         lock_buffer(bh);
1807                 } else if (test_set_buffer_locked(bh)) {
1808                         redirty_page_for_writepage(wbc, page);
1809                         continue;
1810                 }
1811                 if (test_clear_buffer_dirty(bh)) {
1812                         mark_buffer_async_write(bh);
1813                 } else {
1814                         unlock_buffer(bh);
1815                 }
1816         } while ((bh = bh->b_this_page) != head);
1817
1818         /*
1819          * The page and its buffers are protected by PageWriteback(), so we can
1820          * drop the bh refcounts early.
1821          */
1822         BUG_ON(PageWriteback(page));
1823         set_page_writeback(page);
1824
1825         do {
1826                 struct buffer_head *next = bh->b_this_page;
1827                 if (buffer_async_write(bh)) {
1828                         submit_bh(WRITE, bh);
1829                         nr_underway++;
1830                 }
1831                 bh = next;
1832         } while (bh != head);
1833         unlock_page(page);
1834
1835         err = 0;
1836 done:
1837         if (nr_underway == 0) {
1838                 /*
1839                  * The page was marked dirty, but the buffers were
1840                  * clean.  Someone wrote them back by hand with
1841                  * ll_rw_block/submit_bh.  A rare case.
1842                  */
1843                 int uptodate = 1;
1844                 do {
1845                         if (!buffer_uptodate(bh)) {
1846                                 uptodate = 0;
1847                                 break;
1848                         }
1849                         bh = bh->b_this_page;
1850                 } while (bh != head);
1851                 if (uptodate)
1852                         SetPageUptodate(page);
1853                 end_page_writeback(page);
1854                 /*
1855                  * The page and buffer_heads can be released at any time from
1856                  * here on.
1857                  */
1858                 wbc->pages_skipped++;   /* We didn't write this page */
1859         }
1860         return err;
1861
1862 recover:
1863         /*
1864          * ENOSPC, or some other error.  We may already have added some
1865          * blocks to the file, so we need to write these out to avoid
1866          * exposing stale data.
1867          * The page is currently locked and not marked for writeback
1868          */
1869         bh = head;
1870         /* Recovery: lock and submit the mapped buffers */
1871         do {
1872                 if (buffer_mapped(bh) && buffer_dirty(bh)) {
1873                         lock_buffer(bh);
1874                         mark_buffer_async_write(bh);
1875                 } else {
1876                         /*
1877                          * The buffer may have been set dirty during
1878                          * attachment to a dirty page.
1879                          */
1880                         clear_buffer_dirty(bh);
1881                 }
1882         } while ((bh = bh->b_this_page) != head);
1883         SetPageError(page);
1884         BUG_ON(PageWriteback(page));
1885         set_page_writeback(page);
1886         unlock_page(page);
1887         do {
1888                 struct buffer_head *next = bh->b_this_page;
1889                 if (buffer_async_write(bh)) {
1890                         clear_buffer_dirty(bh);
1891                         submit_bh(WRITE, bh);
1892                         nr_underway++;
1893                 }
1894                 bh = next;
1895         } while (bh != head);
1896         goto done;
1897 }
1898
1899 static int __block_prepare_write(struct inode *inode, struct page *page,
1900                 unsigned from, unsigned to, get_block_t *get_block)
1901 {
1902         unsigned block_start, block_end;
1903         sector_t block;
1904         int err = 0;
1905         unsigned blocksize, bbits;
1906         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1907
1908         BUG_ON(!PageLocked(page));
1909         BUG_ON(from > PAGE_CACHE_SIZE);
1910         BUG_ON(to > PAGE_CACHE_SIZE);
1911         BUG_ON(from > to);
1912
1913         blocksize = 1 << inode->i_blkbits;
1914         if (!page_has_buffers(page))
1915                 create_empty_buffers(page, blocksize, 0);
1916         head = page_buffers(page);
1917
1918         bbits = inode->i_blkbits;
1919         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1920
1921         for(bh = head, block_start = 0; bh != head || !block_start;
1922             block++, block_start=block_end, bh = bh->b_this_page) {
1923                 block_end = block_start + blocksize;
1924                 if (block_end <= from || block_start >= to) {
1925                         if (PageUptodate(page)) {
1926                                 if (!buffer_uptodate(bh))
1927                                         set_buffer_uptodate(bh);
1928                         }
1929                         continue;
1930                 }
1931                 if (buffer_new(bh))
1932                         clear_buffer_new(bh);
1933                 if (!buffer_mapped(bh)) {
1934                         err = get_block(inode, block, bh, 1);
1935                         if (err)
1936                                 break;
1937                         if (buffer_new(bh)) {
1938                                 unmap_underlying_metadata(bh->b_bdev,
1939                                                         bh->b_blocknr);
1940                                 if (PageUptodate(page)) {
1941                                         set_buffer_uptodate(bh);
1942                                         continue;
1943                                 }
1944                                 if (block_end > to || block_start < from) {
1945                                         void *kaddr;
1946
1947                                         kaddr = kmap_atomic(page, KM_USER0);
1948                                         if (block_end > to)
1949                                                 memset(kaddr+to, 0,
1950                                                         block_end-to);
1951                                         if (block_start < from)
1952                                                 memset(kaddr+block_start,
1953                                                         0, from-block_start);
1954                                         flush_dcache_page(page);
1955                                         kunmap_atomic(kaddr, KM_USER0);
1956                                 }
1957                                 continue;
1958                         }
1959                 }
1960                 if (PageUptodate(page)) {
1961                         if (!buffer_uptodate(bh))
1962                                 set_buffer_uptodate(bh);
1963                         continue; 
1964                 }
1965                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1966                      (block_start < from || block_end > to)) {
1967                         ll_rw_block(READ, 1, &bh);
1968                         *wait_bh++=bh;
1969                 }
1970         }
1971         /*
1972          * If we issued read requests - let them complete.
1973          */
1974         while(wait_bh > wait) {
1975                 wait_on_buffer(*--wait_bh);
1976                 if (!buffer_uptodate(*wait_bh))
1977                         err = -EIO;
1978         }
1979         if (!err) {
1980                 bh = head;
1981                 do {
1982                         if (buffer_new(bh))
1983                                 clear_buffer_new(bh);
1984                 } while ((bh = bh->b_this_page) != head);
1985                 return 0;
1986         }
1987         /* Error case: */
1988         /*
1989          * Zero out any newly allocated blocks to avoid exposing stale
1990          * data.  If BH_New is set, we know that the block was newly
1991          * allocated in the above loop.
1992          */
1993         bh = head;
1994         block_start = 0;
1995         do {
1996                 block_end = block_start+blocksize;
1997                 if (block_end <= from)
1998                         goto next_bh;
1999                 if (block_start >= to)
2000                         break;
2001                 if (buffer_new(bh)) {
2002                         void *kaddr;
2003
2004                         clear_buffer_new(bh);
2005                         kaddr = kmap_atomic(page, KM_USER0);
2006                         memset(kaddr+block_start, 0, bh->b_size);
2007                         kunmap_atomic(kaddr, KM_USER0);
2008                         set_buffer_uptodate(bh);
2009                         mark_buffer_dirty(bh);
2010                 }
2011 next_bh:
2012                 block_start = block_end;
2013                 bh = bh->b_this_page;
2014         } while (bh != head);
2015         return err;
2016 }
2017
2018 static int __block_commit_write(struct inode *inode, struct page *page,
2019                 unsigned from, unsigned to)
2020 {
2021         unsigned block_start, block_end;
2022         int partial = 0;
2023         unsigned blocksize;
2024         struct buffer_head *bh, *head;
2025
2026         blocksize = 1 << inode->i_blkbits;
2027
2028         for(bh = head = page_buffers(page), block_start = 0;
2029             bh != head || !block_start;
2030             block_start=block_end, bh = bh->b_this_page) {
2031                 block_end = block_start + blocksize;
2032                 if (block_end <= from || block_start >= to) {
2033                         if (!buffer_uptodate(bh))
2034                                 partial = 1;
2035                 } else {
2036                         set_buffer_uptodate(bh);
2037                         mark_buffer_dirty(bh);
2038                 }
2039         }
2040
2041         /*
2042          * If this is a partial write which happened to make all buffers
2043          * uptodate then we can optimize away a bogus readpage() for
2044          * the next read(). Here we 'discover' whether the page went
2045          * uptodate as a result of this (potentially partial) write.
2046          */
2047         if (!partial)
2048                 SetPageUptodate(page);
2049         return 0;
2050 }
2051
2052 /*
2053  * Generic "read page" function for block devices that have the normal
2054  * get_block functionality. This is most of the block device filesystems.
2055  * Reads the page asynchronously --- the unlock_buffer() and
2056  * set/clear_buffer_uptodate() functions propagate buffer state into the
2057  * page struct once IO has completed.
2058  */
2059 int block_read_full_page(struct page *page, get_block_t *get_block)
2060 {
2061         struct inode *inode = page->mapping->host;
2062         sector_t iblock, lblock;
2063         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2064         unsigned int blocksize;
2065         int nr, i;
2066         int fully_mapped = 1;
2067
2068         BUG_ON(!PageLocked(page));
2069         blocksize = 1 << inode->i_blkbits;
2070         if (!page_has_buffers(page))
2071                 create_empty_buffers(page, blocksize, 0);
2072         head = page_buffers(page);
2073
2074         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2075         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2076         bh = head;
2077         nr = 0;
2078         i = 0;
2079
2080         do {
2081                 if (buffer_uptodate(bh))
2082                         continue;
2083
2084                 if (!buffer_mapped(bh)) {
2085                         int err = 0;
2086
2087                         fully_mapped = 0;
2088                         if (iblock < lblock) {
2089                                 err = get_block(inode, iblock, bh, 0);
2090                                 if (err)
2091                                         SetPageError(page);
2092                         }
2093                         if (!buffer_mapped(bh)) {
2094                                 void *kaddr = kmap_atomic(page, KM_USER0);
2095                                 memset(kaddr + i * blocksize, 0, blocksize);
2096                                 flush_dcache_page(page);
2097                                 kunmap_atomic(kaddr, KM_USER0);
2098                                 if (!err)
2099                                         set_buffer_uptodate(bh);
2100                                 continue;
2101                         }
2102                         /*
2103                          * get_block() might have updated the buffer
2104                          * synchronously
2105                          */
2106                         if (buffer_uptodate(bh))
2107                                 continue;
2108                 }
2109                 arr[nr++] = bh;
2110         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2111
2112         if (fully_mapped)
2113                 SetPageMappedToDisk(page);
2114
2115         if (!nr) {
2116                 /*
2117                  * All buffers are uptodate - we can set the page uptodate
2118                  * as well. But not if get_block() returned an error.
2119                  */
2120                 if (!PageError(page))
2121                         SetPageUptodate(page);
2122                 unlock_page(page);
2123                 return 0;
2124         }
2125
2126         /* Stage two: lock the buffers */
2127         for (i = 0; i < nr; i++) {
2128                 bh = arr[i];
2129                 lock_buffer(bh);
2130                 mark_buffer_async_read(bh);
2131         }
2132
2133         /*
2134          * Stage 3: start the IO.  Check for uptodateness
2135          * inside the buffer lock in case another process reading
2136          * the underlying blockdev brought it uptodate (the sct fix).
2137          */
2138         for (i = 0; i < nr; i++) {
2139                 bh = arr[i];
2140                 if (buffer_uptodate(bh))
2141                         end_buffer_async_read(bh, 1);
2142                 else
2143                         submit_bh(READ, bh);
2144         }
2145         return 0;
2146 }
2147
2148 /* utility function for filesystems that need to do work on expanding
2149  * truncates.  Uses prepare/commit_write to allow the filesystem to
2150  * deal with the hole.  
2151  */
2152 static int __generic_cont_expand(struct inode *inode, loff_t size,
2153                                  pgoff_t index, unsigned int offset)
2154 {
2155         struct address_space *mapping = inode->i_mapping;
2156         struct page *page;
2157         unsigned long limit;
2158         int err;
2159
2160         err = -EFBIG;
2161         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2162         if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2163                 send_sig(SIGXFSZ, current, 0);
2164                 goto out;
2165         }
2166         if (size > inode->i_sb->s_maxbytes)
2167                 goto out;
2168
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                 /*
2176                  * ->prepare_write() may have instantiated a few blocks
2177                  * outside i_size.  Trim these off again.
2178                  */
2179                 unlock_page(page);
2180                 page_cache_release(page);
2181                 vmtruncate(inode, inode->i_size);
2182                 goto out;
2183         }
2184
2185         err = mapping->a_ops->commit_write(NULL, page, offset, offset);
2186
2187         unlock_page(page);
2188         page_cache_release(page);
2189         if (err > 0)
2190                 err = 0;
2191 out:
2192         return err;
2193 }
2194
2195 int generic_cont_expand(struct inode *inode, loff_t size)
2196 {
2197         pgoff_t index;
2198         unsigned int offset;
2199
2200         offset = (size & (PAGE_CACHE_SIZE - 1)); /* Within page */
2201
2202         /* ugh.  in prepare/commit_write, if from==to==start of block, we
2203         ** skip the prepare.  make sure we never send an offset for the start
2204         ** of a block
2205         */
2206         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
2207                 /* caller must handle this extra byte. */
2208                 offset++;
2209         }
2210         index = size >> PAGE_CACHE_SHIFT;
2211
2212         return __generic_cont_expand(inode, size, index, offset);
2213 }
2214
2215 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2216 {
2217         loff_t pos = size - 1;
2218         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
2219         unsigned int offset = (pos & (PAGE_CACHE_SIZE - 1)) + 1;
2220
2221         /* prepare/commit_write can handle even if from==to==start of block. */
2222         return __generic_cont_expand(inode, size, index, offset);
2223 }
2224
2225 /*
2226  * For moronic filesystems that do not allow holes in file.
2227  * We may have to extend the file.
2228  */
2229
2230 int cont_prepare_write(struct page *page, unsigned offset,
2231                 unsigned to, get_block_t *get_block, loff_t *bytes)
2232 {
2233         struct address_space *mapping = page->mapping;
2234         struct inode *inode = mapping->host;
2235         struct page *new_page;
2236         pgoff_t pgpos;
2237         long status;
2238         unsigned zerofrom;
2239         unsigned blocksize = 1 << inode->i_blkbits;
2240         void *kaddr;
2241
2242         while(page->index > (pgpos = *bytes>>PAGE_CACHE_SHIFT)) {
2243                 status = -ENOMEM;
2244                 new_page = grab_cache_page(mapping, pgpos);
2245                 if (!new_page)
2246                         goto out;
2247                 /* we might sleep */
2248                 if (*bytes>>PAGE_CACHE_SHIFT != pgpos) {
2249                         unlock_page(new_page);
2250                         page_cache_release(new_page);
2251                         continue;
2252                 }
2253                 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2254                 if (zerofrom & (blocksize-1)) {
2255                         *bytes |= (blocksize-1);
2256                         (*bytes)++;
2257                 }
2258                 status = __block_prepare_write(inode, new_page, zerofrom,
2259                                                 PAGE_CACHE_SIZE, get_block);
2260                 if (status)
2261                         goto out_unmap;
2262                 kaddr = kmap_atomic(new_page, KM_USER0);
2263                 memset(kaddr+zerofrom, 0, PAGE_CACHE_SIZE-zerofrom);
2264                 flush_dcache_page(new_page);
2265                 kunmap_atomic(kaddr, KM_USER0);
2266                 generic_commit_write(NULL, new_page, zerofrom, PAGE_CACHE_SIZE);
2267                 unlock_page(new_page);
2268                 page_cache_release(new_page);
2269         }
2270
2271         if (page->index < pgpos) {
2272                 /* completely inside the area */
2273                 zerofrom = offset;
2274         } else {
2275                 /* page covers the boundary, find the boundary offset */
2276                 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2277
2278                 /* if we will expand the thing last block will be filled */
2279                 if (to > zerofrom && (zerofrom & (blocksize-1))) {
2280                         *bytes |= (blocksize-1);
2281                         (*bytes)++;
2282                 }
2283
2284                 /* starting below the boundary? Nothing to zero out */
2285                 if (offset <= zerofrom)
2286                         zerofrom = offset;
2287         }
2288         status = __block_prepare_write(inode, page, zerofrom, to, get_block);
2289         if (status)
2290                 goto out1;
2291         if (zerofrom < offset) {
2292                 kaddr = kmap_atomic(page, KM_USER0);
2293                 memset(kaddr+zerofrom, 0, offset-zerofrom);
2294                 flush_dcache_page(page);
2295                 kunmap_atomic(kaddr, KM_USER0);
2296                 __block_commit_write(inode, page, zerofrom, offset);
2297         }
2298         return 0;
2299 out1:
2300         ClearPageUptodate(page);
2301         return status;
2302
2303 out_unmap:
2304         ClearPageUptodate(new_page);
2305         unlock_page(new_page);
2306         page_cache_release(new_page);
2307 out:
2308         return status;
2309 }
2310
2311 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2312                         get_block_t *get_block)
2313 {
2314         struct inode *inode = page->mapping->host;
2315         int err = __block_prepare_write(inode, page, from, to, get_block);
2316         if (err)
2317                 ClearPageUptodate(page);
2318         return err;
2319 }
2320
2321 int block_commit_write(struct page *page, unsigned from, unsigned to)
2322 {
2323         struct inode *inode = page->mapping->host;
2324         __block_commit_write(inode,page,from,to);
2325         return 0;
2326 }
2327
2328 int generic_commit_write(struct file *file, struct page *page,
2329                 unsigned from, unsigned to)
2330 {
2331         struct inode *inode = page->mapping->host;
2332         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2333         __block_commit_write(inode,page,from,to);
2334         /*
2335          * No need to use i_size_read() here, the i_size
2336          * cannot change under us because we hold i_mutex.
2337          */
2338         if (pos > inode->i_size) {
2339                 i_size_write(inode, pos);
2340                 mark_inode_dirty(inode);
2341         }
2342         return 0;
2343 }
2344
2345
2346 /*
2347  * nobh_prepare_write()'s prereads are special: the buffer_heads are freed
2348  * immediately, while under the page lock.  So it needs a special end_io
2349  * handler which does not touch the bh after unlocking it.
2350  *
2351  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
2352  * a race there is benign: unlock_buffer() only use the bh's address for
2353  * hashing after unlocking the buffer, so it doesn't actually touch the bh
2354  * itself.
2355  */
2356 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2357 {
2358         if (uptodate) {
2359                 set_buffer_uptodate(bh);
2360         } else {
2361                 /* This happens, due to failed READA attempts. */
2362                 clear_buffer_uptodate(bh);
2363         }
2364         unlock_buffer(bh);
2365 }
2366
2367 /*
2368  * On entry, the page is fully not uptodate.
2369  * On exit the page is fully uptodate in the areas outside (from,to)
2370  */
2371 int nobh_prepare_write(struct page *page, unsigned from, unsigned to,
2372                         get_block_t *get_block)
2373 {
2374         struct inode *inode = page->mapping->host;
2375         const unsigned blkbits = inode->i_blkbits;
2376         const unsigned blocksize = 1 << blkbits;
2377         struct buffer_head map_bh;
2378         struct buffer_head *read_bh[MAX_BUF_PER_PAGE];
2379         unsigned block_in_page;
2380         unsigned block_start;
2381         sector_t block_in_file;
2382         char *kaddr;
2383         int nr_reads = 0;
2384         int i;
2385         int ret = 0;
2386         int is_mapped_to_disk = 1;
2387         int dirtied_it = 0;
2388
2389         if (PageMappedToDisk(page))
2390                 return 0;
2391
2392         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2393         map_bh.b_page = page;
2394
2395         /*
2396          * We loop across all blocks in the page, whether or not they are
2397          * part of the affected region.  This is so we can discover if the
2398          * page is fully mapped-to-disk.
2399          */
2400         for (block_start = 0, block_in_page = 0;
2401                   block_start < PAGE_CACHE_SIZE;
2402                   block_in_page++, block_start += blocksize) {
2403                 unsigned block_end = block_start + blocksize;
2404                 int create;
2405
2406                 map_bh.b_state = 0;
2407                 create = 1;
2408                 if (block_start >= to)
2409                         create = 0;
2410                 ret = get_block(inode, block_in_file + block_in_page,
2411                                         &map_bh, create);
2412                 if (ret)
2413                         goto failed;
2414                 if (!buffer_mapped(&map_bh))
2415                         is_mapped_to_disk = 0;
2416                 if (buffer_new(&map_bh))
2417                         unmap_underlying_metadata(map_bh.b_bdev,
2418                                                         map_bh.b_blocknr);
2419                 if (PageUptodate(page))
2420                         continue;
2421                 if (buffer_new(&map_bh) || !buffer_mapped(&map_bh)) {
2422                         kaddr = kmap_atomic(page, KM_USER0);
2423                         if (block_start < from) {
2424                                 memset(kaddr+block_start, 0, from-block_start);
2425                                 dirtied_it = 1;
2426                         }
2427                         if (block_end > to) {
2428                                 memset(kaddr + to, 0, block_end - to);
2429                                 dirtied_it = 1;
2430                         }
2431                         flush_dcache_page(page);
2432                         kunmap_atomic(kaddr, KM_USER0);
2433                         continue;
2434                 }
2435                 if (buffer_uptodate(&map_bh))
2436                         continue;       /* reiserfs does this */
2437                 if (block_start < from || block_end > to) {
2438                         struct buffer_head *bh = alloc_buffer_head(GFP_NOFS);
2439
2440                         if (!bh) {
2441                                 ret = -ENOMEM;
2442                                 goto failed;
2443                         }
2444                         bh->b_state = map_bh.b_state;
2445                         atomic_set(&bh->b_count, 0);
2446                         bh->b_this_page = NULL;
2447                         bh->b_page = page;
2448                         bh->b_blocknr = map_bh.b_blocknr;
2449                         bh->b_size = blocksize;
2450                         bh->b_data = (char *)(long)block_start;
2451                         bh->b_bdev = map_bh.b_bdev;
2452                         bh->b_private = NULL;
2453                         read_bh[nr_reads++] = bh;
2454                 }
2455         }
2456
2457         if (nr_reads) {
2458                 struct buffer_head *bh;
2459
2460                 /*
2461                  * The page is locked, so these buffers are protected from
2462                  * any VM or truncate activity.  Hence we don't need to care
2463                  * for the buffer_head refcounts.
2464                  */
2465                 for (i = 0; i < nr_reads; i++) {
2466                         bh = read_bh[i];
2467                         lock_buffer(bh);
2468                         bh->b_end_io = end_buffer_read_nobh;
2469                         submit_bh(READ, bh);
2470                 }
2471                 for (i = 0; i < nr_reads; i++) {
2472                         bh = read_bh[i];
2473                         wait_on_buffer(bh);
2474                         if (!buffer_uptodate(bh))
2475                                 ret = -EIO;
2476                         free_buffer_head(bh);
2477                         read_bh[i] = NULL;
2478                 }
2479                 if (ret)
2480                         goto failed;
2481         }
2482
2483         if (is_mapped_to_disk)
2484                 SetPageMappedToDisk(page);
2485         SetPageUptodate(page);
2486
2487         /*
2488          * Setting the page dirty here isn't necessary for the prepare_write
2489          * function - commit_write will do that.  But if/when this function is
2490          * used within the pagefault handler to ensure that all mmapped pages
2491          * have backing space in the filesystem, we will need to dirty the page
2492          * if its contents were altered.
2493          */
2494         if (dirtied_it)
2495                 set_page_dirty(page);
2496
2497         return 0;
2498
2499 failed:
2500         for (i = 0; i < nr_reads; i++) {
2501                 if (read_bh[i])
2502                         free_buffer_head(read_bh[i]);
2503         }
2504
2505         /*
2506          * Error recovery is pretty slack.  Clear the page and mark it dirty
2507          * so we'll later zero out any blocks which _were_ allocated.
2508          */
2509         kaddr = kmap_atomic(page, KM_USER0);
2510         memset(kaddr, 0, PAGE_CACHE_SIZE);
2511         kunmap_atomic(kaddr, KM_USER0);
2512         SetPageUptodate(page);
2513         set_page_dirty(page);
2514         return ret;
2515 }
2516 EXPORT_SYMBOL(nobh_prepare_write);
2517
2518 int nobh_commit_write(struct file *file, struct page *page,
2519                 unsigned from, unsigned to)
2520 {
2521         struct inode *inode = page->mapping->host;
2522         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2523
2524         set_page_dirty(page);
2525         if (pos > inode->i_size) {
2526                 i_size_write(inode, pos);
2527                 mark_inode_dirty(inode);
2528         }
2529         return 0;
2530 }
2531 EXPORT_SYMBOL(nobh_commit_write);
2532
2533 /*
2534  * nobh_writepage() - based on block_full_write_page() except
2535  * that it tries to operate without attaching bufferheads to
2536  * the page.
2537  */
2538 int nobh_writepage(struct page *page, get_block_t *get_block,
2539                         struct writeback_control *wbc)
2540 {
2541         struct inode * const inode = page->mapping->host;
2542         loff_t i_size = i_size_read(inode);
2543         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2544         unsigned offset;
2545         void *kaddr;
2546         int ret;
2547
2548         /* Is the page fully inside i_size? */
2549         if (page->index < end_index)
2550                 goto out;
2551
2552         /* Is the page fully outside i_size? (truncate in progress) */
2553         offset = i_size & (PAGE_CACHE_SIZE-1);
2554         if (page->index >= end_index+1 || !offset) {
2555                 /*
2556                  * The page may have dirty, unmapped buffers.  For example,
2557                  * they may have been added in ext3_writepage().  Make them
2558                  * freeable here, so the page does not leak.
2559                  */
2560 #if 0
2561                 /* Not really sure about this  - do we need this ? */
2562                 if (page->mapping->a_ops->invalidatepage)
2563                         page->mapping->a_ops->invalidatepage(page, offset);
2564 #endif
2565                 unlock_page(page);
2566                 return 0; /* don't care */
2567         }
2568
2569         /*
2570          * The page straddles i_size.  It must be zeroed out on each and every
2571          * writepage invocation because it may be mmapped.  "A file is mapped
2572          * in multiples of the page size.  For a file that is not a multiple of
2573          * the  page size, the remaining memory is zeroed when mapped, and
2574          * writes to that region are not written out to the file."
2575          */
2576         kaddr = kmap_atomic(page, KM_USER0);
2577         memset(kaddr + offset, 0, PAGE_CACHE_SIZE - offset);
2578         flush_dcache_page(page);
2579         kunmap_atomic(kaddr, KM_USER0);
2580 out:
2581         ret = mpage_writepage(page, get_block, wbc);
2582         if (ret == -EAGAIN)
2583                 ret = __block_write_full_page(inode, page, get_block, wbc);
2584         return ret;
2585 }
2586 EXPORT_SYMBOL(nobh_writepage);
2587
2588 /*
2589  * This function assumes that ->prepare_write() uses nobh_prepare_write().
2590  */
2591 int nobh_truncate_page(struct address_space *mapping, loff_t from)
2592 {
2593         struct inode *inode = mapping->host;
2594         unsigned blocksize = 1 << inode->i_blkbits;
2595         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2596         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2597         unsigned to;
2598         struct page *page;
2599         struct address_space_operations *a_ops = mapping->a_ops;
2600         char *kaddr;
2601         int ret = 0;
2602
2603         if ((offset & (blocksize - 1)) == 0)
2604                 goto out;
2605
2606         ret = -ENOMEM;
2607         page = grab_cache_page(mapping, index);
2608         if (!page)
2609                 goto out;
2610
2611         to = (offset + blocksize) & ~(blocksize - 1);
2612         ret = a_ops->prepare_write(NULL, page, offset, to);
2613         if (ret == 0) {
2614                 kaddr = kmap_atomic(page, KM_USER0);
2615                 memset(kaddr + offset, 0, PAGE_CACHE_SIZE - offset);
2616                 flush_dcache_page(page);
2617                 kunmap_atomic(kaddr, KM_USER0);
2618                 set_page_dirty(page);
2619         }
2620         unlock_page(page);
2621         page_cache_release(page);
2622 out:
2623         return ret;
2624 }
2625 EXPORT_SYMBOL(nobh_truncate_page);
2626
2627 int block_truncate_page(struct address_space *mapping,
2628                         loff_t from, get_block_t *get_block)
2629 {
2630         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2631         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2632         unsigned blocksize;
2633         sector_t iblock;
2634         unsigned length, pos;
2635         struct inode *inode = mapping->host;
2636         struct page *page;
2637         struct buffer_head *bh;
2638         void *kaddr;
2639         int err;
2640
2641         blocksize = 1 << inode->i_blkbits;
2642         length = offset & (blocksize - 1);
2643
2644         /* Block boundary? Nothing to do */
2645         if (!length)
2646                 return 0;
2647
2648         length = blocksize - length;
2649         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2650         
2651         page = grab_cache_page(mapping, index);
2652         err = -ENOMEM;
2653         if (!page)
2654                 goto out;
2655
2656         if (!page_has_buffers(page))
2657                 create_empty_buffers(page, blocksize, 0);
2658
2659         /* Find the buffer that contains "offset" */
2660         bh = page_buffers(page);
2661         pos = blocksize;
2662         while (offset >= pos) {
2663                 bh = bh->b_this_page;
2664                 iblock++;
2665                 pos += blocksize;
2666         }
2667
2668         err = 0;
2669         if (!buffer_mapped(bh)) {
2670                 err = get_block(inode, iblock, bh, 0);
2671                 if (err)
2672                         goto unlock;
2673                 /* unmapped? It's a hole - nothing to do */
2674                 if (!buffer_mapped(bh))
2675                         goto unlock;
2676         }
2677
2678         /* Ok, it's mapped. Make sure it's up-to-date */
2679         if (PageUptodate(page))
2680                 set_buffer_uptodate(bh);
2681
2682         if (!buffer_uptodate(bh) && !buffer_delay(bh)) {
2683                 err = -EIO;
2684                 ll_rw_block(READ, 1, &bh);
2685                 wait_on_buffer(bh);
2686                 /* Uhhuh. Read error. Complain and punt. */
2687                 if (!buffer_uptodate(bh))
2688                         goto unlock;
2689         }
2690
2691         kaddr = kmap_atomic(page, KM_USER0);
2692         memset(kaddr + offset, 0, length);
2693         flush_dcache_page(page);
2694         kunmap_atomic(kaddr, KM_USER0);
2695
2696         mark_buffer_dirty(bh);
2697         err = 0;
2698
2699 unlock:
2700         unlock_page(page);
2701         page_cache_release(page);
2702 out:
2703         return err;
2704 }
2705
2706 /*
2707  * The generic ->writepage function for buffer-backed address_spaces
2708  */
2709 int block_write_full_page(struct page *page, get_block_t *get_block,
2710                         struct writeback_control *wbc)
2711 {
2712         struct inode * const inode = page->mapping->host;
2713         loff_t i_size = i_size_read(inode);
2714         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2715         unsigned offset;
2716         void *kaddr;
2717
2718         /* Is the page fully inside i_size? */
2719         if (page->index < end_index)
2720                 return __block_write_full_page(inode, page, get_block, wbc);
2721
2722         /* Is the page fully outside i_size? (truncate in progress) */
2723         offset = i_size & (PAGE_CACHE_SIZE-1);
2724         if (page->index >= end_index+1 || !offset) {
2725                 /*
2726                  * The page may have dirty, unmapped buffers.  For example,
2727                  * they may have been added in ext3_writepage().  Make them
2728                  * freeable here, so the page does not leak.
2729                  */
2730                 do_invalidatepage(page, 0);
2731                 unlock_page(page);
2732                 return 0; /* don't care */
2733         }
2734
2735         /*
2736          * The page straddles i_size.  It must be zeroed out on each and every
2737          * writepage invokation because it may be mmapped.  "A file is mapped
2738          * in multiples of the page size.  For a file that is not a multiple of
2739          * the  page size, the remaining memory is zeroed when mapped, and
2740          * writes to that region are not written out to the file."
2741          */
2742         kaddr = kmap_atomic(page, KM_USER0);
2743         memset(kaddr + offset, 0, PAGE_CACHE_SIZE - offset);
2744         flush_dcache_page(page);
2745         kunmap_atomic(kaddr, KM_USER0);
2746         return __block_write_full_page(inode, page, get_block, wbc);
2747 }
2748
2749 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2750                             get_block_t *get_block)
2751 {
2752         struct buffer_head tmp;
2753         struct inode *inode = mapping->host;
2754         tmp.b_state = 0;
2755         tmp.b_blocknr = 0;
2756         get_block(inode, block, &tmp, 0);
2757         return tmp.b_blocknr;
2758 }
2759
2760 static int end_bio_bh_io_sync(struct bio *bio, unsigned int bytes_done, int err)
2761 {
2762         struct buffer_head *bh = bio->bi_private;
2763
2764         if (bio->bi_size)
2765                 return 1;
2766
2767         if (err == -EOPNOTSUPP) {
2768                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2769                 set_bit(BH_Eopnotsupp, &bh->b_state);
2770         }
2771
2772         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2773         bio_put(bio);
2774         return 0;
2775 }
2776
2777 int submit_bh(int rw, struct buffer_head * bh)
2778 {
2779         struct bio *bio;
2780         int ret = 0;
2781
2782         BUG_ON(!buffer_locked(bh));
2783         BUG_ON(!buffer_mapped(bh));
2784         BUG_ON(!bh->b_end_io);
2785
2786         if (buffer_ordered(bh) && (rw == WRITE))
2787                 rw = WRITE_BARRIER;
2788
2789         /*
2790          * Only clear out a write error when rewriting, should this
2791          * include WRITE_SYNC as well?
2792          */
2793         if (test_set_buffer_req(bh) && (rw == WRITE || rw == WRITE_BARRIER))
2794                 clear_buffer_write_io_error(bh);
2795
2796         /*
2797          * from here on down, it's all bio -- do the initial mapping,
2798          * submit_bio -> generic_make_request may further map this bio around
2799          */
2800         bio = bio_alloc(GFP_NOIO, 1);
2801
2802         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2803         bio->bi_bdev = bh->b_bdev;
2804         bio->bi_io_vec[0].bv_page = bh->b_page;
2805         bio->bi_io_vec[0].bv_len = bh->b_size;
2806         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2807
2808         bio->bi_vcnt = 1;
2809         bio->bi_idx = 0;
2810         bio->bi_size = bh->b_size;
2811
2812         bio->bi_end_io = end_bio_bh_io_sync;
2813         bio->bi_private = bh;
2814
2815         bio_get(bio);
2816         submit_bio(rw, bio);
2817
2818         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2819                 ret = -EOPNOTSUPP;
2820
2821         bio_put(bio);
2822         return ret;
2823 }
2824
2825 /**
2826  * ll_rw_block: low-level access to block devices (DEPRECATED)
2827  * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
2828  * @nr: number of &struct buffer_heads in the array
2829  * @bhs: array of pointers to &struct buffer_head
2830  *
2831  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2832  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2833  * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2834  * are sent to disk. The fourth %READA option is described in the documentation
2835  * for generic_make_request() which ll_rw_block() calls.
2836  *
2837  * This function drops any buffer that it cannot get a lock on (with the
2838  * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
2839  * clean when doing a write request, and any buffer that appears to be
2840  * up-to-date when doing read request.  Further it marks as clean buffers that
2841  * are processed for writing (the buffer cache won't assume that they are
2842  * actually clean until the buffer gets unlocked).
2843  *
2844  * ll_rw_block sets b_end_io to simple completion handler that marks
2845  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2846  * any waiters. 
2847  *
2848  * All of the buffers must be for the same device, and must also be a
2849  * multiple of the current approved size for the device.
2850  */
2851 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2852 {
2853         int i;
2854
2855         for (i = 0; i < nr; i++) {
2856                 struct buffer_head *bh = bhs[i];
2857
2858                 if (rw == SWRITE)
2859                         lock_buffer(bh);
2860                 else if (test_set_buffer_locked(bh))
2861                         continue;
2862
2863                 if (rw == WRITE || rw == SWRITE) {
2864                         if (test_clear_buffer_dirty(bh)) {
2865                                 bh->b_end_io = end_buffer_write_sync;
2866                                 get_bh(bh);
2867                                 submit_bh(WRITE, bh);
2868                                 continue;
2869                         }
2870                 } else {
2871                         if (!buffer_uptodate(bh)) {
2872                                 bh->b_end_io = end_buffer_read_sync;
2873                                 get_bh(bh);
2874                                 submit_bh(rw, bh);
2875                                 continue;
2876                         }
2877                 }
2878                 unlock_buffer(bh);
2879         }
2880 }
2881
2882 /*
2883  * For a data-integrity writeout, we need to wait upon any in-progress I/O
2884  * and then start new I/O and then wait upon it.  The caller must have a ref on
2885  * the buffer_head.
2886  */
2887 int sync_dirty_buffer(struct buffer_head *bh)
2888 {
2889         int ret = 0;
2890
2891         WARN_ON(atomic_read(&bh->b_count) < 1);
2892         lock_buffer(bh);
2893         if (test_clear_buffer_dirty(bh)) {
2894                 get_bh(bh);
2895                 bh->b_end_io = end_buffer_write_sync;
2896                 ret = submit_bh(WRITE, bh);
2897                 wait_on_buffer(bh);
2898                 if (buffer_eopnotsupp(bh)) {
2899                         clear_buffer_eopnotsupp(bh);
2900                         ret = -EOPNOTSUPP;
2901                 }
2902                 if (!ret && !buffer_uptodate(bh))
2903                         ret = -EIO;
2904         } else {
2905                 unlock_buffer(bh);
2906         }
2907         return ret;
2908 }
2909
2910 /*
2911  * try_to_free_buffers() checks if all the buffers on this particular page
2912  * are unused, and releases them if so.
2913  *
2914  * Exclusion against try_to_free_buffers may be obtained by either
2915  * locking the page or by holding its mapping's private_lock.
2916  *
2917  * If the page is dirty but all the buffers are clean then we need to
2918  * be sure to mark the page clean as well.  This is because the page
2919  * may be against a block device, and a later reattachment of buffers
2920  * to a dirty page will set *all* buffers dirty.  Which would corrupt
2921  * filesystem data on the same device.
2922  *
2923  * The same applies to regular filesystem pages: if all the buffers are
2924  * clean then we set the page clean and proceed.  To do that, we require
2925  * total exclusion from __set_page_dirty_buffers().  That is obtained with
2926  * private_lock.
2927  *
2928  * try_to_free_buffers() is non-blocking.
2929  */
2930 static inline int buffer_busy(struct buffer_head *bh)
2931 {
2932         return atomic_read(&bh->b_count) |
2933                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
2934 }
2935
2936 static int
2937 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
2938 {
2939         struct buffer_head *head = page_buffers(page);
2940         struct buffer_head *bh;
2941
2942         bh = head;
2943         do {
2944                 if (buffer_write_io_error(bh) && page->mapping)
2945                         set_bit(AS_EIO, &page->mapping->flags);
2946                 if (buffer_busy(bh))
2947                         goto failed;
2948                 bh = bh->b_this_page;
2949         } while (bh != head);
2950
2951         do {
2952                 struct buffer_head *next = bh->b_this_page;
2953
2954                 if (!list_empty(&bh->b_assoc_buffers))
2955                         __remove_assoc_queue(bh);
2956                 bh = next;
2957         } while (bh != head);
2958         *buffers_to_free = head;
2959         __clear_page_buffers(page);
2960         return 1;
2961 failed:
2962         return 0;
2963 }
2964
2965 int try_to_free_buffers(struct page *page)
2966 {
2967         struct address_space * const mapping = page->mapping;
2968         struct buffer_head *buffers_to_free = NULL;
2969         int ret = 0;
2970
2971         BUG_ON(!PageLocked(page));
2972         if (PageWriteback(page))
2973                 return 0;
2974
2975         if (mapping == NULL) {          /* can this still happen? */
2976                 ret = drop_buffers(page, &buffers_to_free);
2977                 goto out;
2978         }
2979
2980         spin_lock(&mapping->private_lock);
2981         ret = drop_buffers(page, &buffers_to_free);
2982         if (ret) {
2983                 /*
2984                  * If the filesystem writes its buffers by hand (eg ext3)
2985                  * then we can have clean buffers against a dirty page.  We
2986                  * clean the page here; otherwise later reattachment of buffers
2987                  * could encounter a non-uptodate page, which is unresolvable.
2988                  * This only applies in the rare case where try_to_free_buffers
2989                  * succeeds but the page is not freed.
2990                  */
2991                 clear_page_dirty(page);
2992         }
2993         spin_unlock(&mapping->private_lock);
2994 out:
2995         if (buffers_to_free) {
2996                 struct buffer_head *bh = buffers_to_free;
2997
2998                 do {
2999                         struct buffer_head *next = bh->b_this_page;
3000                         free_buffer_head(bh);
3001                         bh = next;
3002                 } while (bh != buffers_to_free);
3003         }
3004         return ret;
3005 }
3006 EXPORT_SYMBOL(try_to_free_buffers);
3007
3008 void block_sync_page(struct page *page)
3009 {
3010         struct address_space *mapping;
3011
3012         smp_mb();
3013         mapping = page_mapping(page);
3014         if (mapping)
3015                 blk_run_backing_dev(mapping->backing_dev_info, page);
3016 }
3017
3018 /*
3019  * There are no bdflush tunables left.  But distributions are
3020  * still running obsolete flush daemons, so we terminate them here.
3021  *
3022  * Use of bdflush() is deprecated and will be removed in a future kernel.
3023  * The `pdflush' kernel threads fully replace bdflush daemons and this call.
3024  */
3025 asmlinkage long sys_bdflush(int func, long data)
3026 {
3027         static int msg_count;
3028
3029         if (!capable(CAP_SYS_ADMIN))
3030                 return -EPERM;
3031
3032         if (msg_count < 5) {
3033                 msg_count++;
3034                 printk(KERN_INFO
3035                         "warning: process `%s' used the obsolete bdflush"
3036                         " system call\n", current->comm);
3037                 printk(KERN_INFO "Fix your initscripts?\n");
3038         }
3039
3040         if (func == 1)
3041                 do_exit(0);
3042         return 0;
3043 }
3044
3045 /*
3046  * Buffer-head allocation
3047  */
3048 static kmem_cache_t *bh_cachep;
3049
3050 /*
3051  * Once the number of bh's in the machine exceeds this level, we start
3052  * stripping them in writeback.
3053  */
3054 static int max_buffer_heads;
3055
3056 int buffer_heads_over_limit;
3057
3058 struct bh_accounting {
3059         int nr;                 /* Number of live bh's */
3060         int ratelimit;          /* Limit cacheline bouncing */
3061 };
3062
3063 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3064
3065 static void recalc_bh_state(void)
3066 {
3067         int i;
3068         int tot = 0;
3069
3070         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3071                 return;
3072         __get_cpu_var(bh_accounting).ratelimit = 0;
3073         for_each_online_cpu(i)
3074                 tot += per_cpu(bh_accounting, i).nr;
3075         buffer_heads_over_limit = (tot > max_buffer_heads);
3076 }
3077         
3078 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3079 {
3080         struct buffer_head *ret = kmem_cache_alloc(bh_cachep, gfp_flags);
3081         if (ret) {
3082                 get_cpu_var(bh_accounting).nr++;
3083                 recalc_bh_state();
3084                 put_cpu_var(bh_accounting);
3085         }
3086         return ret;
3087 }
3088 EXPORT_SYMBOL(alloc_buffer_head);
3089
3090 void free_buffer_head(struct buffer_head *bh)
3091 {
3092         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3093         kmem_cache_free(bh_cachep, bh);
3094         get_cpu_var(bh_accounting).nr--;
3095         recalc_bh_state();
3096         put_cpu_var(bh_accounting);
3097 }
3098 EXPORT_SYMBOL(free_buffer_head);
3099
3100 static void
3101 init_buffer_head(void *data, kmem_cache_t *cachep, unsigned long flags)
3102 {
3103         if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) ==
3104                             SLAB_CTOR_CONSTRUCTOR) {
3105                 struct buffer_head * bh = (struct buffer_head *)data;
3106
3107                 memset(bh, 0, sizeof(*bh));
3108                 INIT_LIST_HEAD(&bh->b_assoc_buffers);
3109         }
3110 }
3111
3112 #ifdef CONFIG_HOTPLUG_CPU
3113 static void buffer_exit_cpu(int cpu)
3114 {
3115         int i;
3116         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3117
3118         for (i = 0; i < BH_LRU_SIZE; i++) {
3119                 brelse(b->bhs[i]);
3120                 b->bhs[i] = NULL;
3121         }
3122         get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3123         per_cpu(bh_accounting, cpu).nr = 0;
3124         put_cpu_var(bh_accounting);
3125 }
3126
3127 static int buffer_cpu_notify(struct notifier_block *self,
3128                               unsigned long action, void *hcpu)
3129 {
3130         if (action == CPU_DEAD)
3131                 buffer_exit_cpu((unsigned long)hcpu);
3132         return NOTIFY_OK;
3133 }
3134 #endif /* CONFIG_HOTPLUG_CPU */
3135
3136 void __init buffer_init(void)
3137 {
3138         int nrpages;
3139
3140         bh_cachep = kmem_cache_create("buffer_head",
3141                                         sizeof(struct buffer_head), 0,
3142                                         (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3143                                         SLAB_MEM_SPREAD),
3144                                         init_buffer_head,
3145                                         NULL);
3146
3147         /*
3148          * Limit the bh occupancy to 10% of ZONE_NORMAL
3149          */
3150         nrpages = (nr_free_buffer_pages() * 10) / 100;
3151         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3152         hotcpu_notifier(buffer_cpu_notify, 0);
3153 }
3154
3155 EXPORT_SYMBOL(__bforget);
3156 EXPORT_SYMBOL(__brelse);
3157 EXPORT_SYMBOL(__wait_on_buffer);
3158 EXPORT_SYMBOL(block_commit_write);
3159 EXPORT_SYMBOL(block_prepare_write);
3160 EXPORT_SYMBOL(block_read_full_page);
3161 EXPORT_SYMBOL(block_sync_page);
3162 EXPORT_SYMBOL(block_truncate_page);
3163 EXPORT_SYMBOL(block_write_full_page);
3164 EXPORT_SYMBOL(cont_prepare_write);
3165 EXPORT_SYMBOL(end_buffer_async_write);
3166 EXPORT_SYMBOL(end_buffer_read_sync);
3167 EXPORT_SYMBOL(end_buffer_write_sync);
3168 EXPORT_SYMBOL(file_fsync);
3169 EXPORT_SYMBOL(fsync_bdev);
3170 EXPORT_SYMBOL(generic_block_bmap);
3171 EXPORT_SYMBOL(generic_commit_write);
3172 EXPORT_SYMBOL(generic_cont_expand);
3173 EXPORT_SYMBOL(generic_cont_expand_simple);
3174 EXPORT_SYMBOL(init_buffer);
3175 EXPORT_SYMBOL(invalidate_bdev);
3176 EXPORT_SYMBOL(ll_rw_block);
3177 EXPORT_SYMBOL(mark_buffer_dirty);
3178 EXPORT_SYMBOL(submit_bh);
3179 EXPORT_SYMBOL(sync_dirty_buffer);
3180 EXPORT_SYMBOL(unlock_buffer);