6e35762b6169be2a30a69462764b71967369a1d8
[linux-2.6.git] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/capability.h>
28 #include <linux/blkdev.h>
29 #include <linux/file.h>
30 #include <linux/quotaops.h>
31 #include <linux/highmem.h>
32 #include <linux/module.h>
33 #include <linux/writeback.h>
34 #include <linux/hash.h>
35 #include <linux/suspend.h>
36 #include <linux/buffer_head.h>
37 #include <linux/task_io_accounting_ops.h>
38 #include <linux/bio.h>
39 #include <linux/notifier.h>
40 #include <linux/cpu.h>
41 #include <linux/bitops.h>
42 #include <linux/mpage.h>
43 #include <linux/bit_spinlock.h>
44
45 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
46
47 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
48
49 inline void
50 init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
51 {
52         bh->b_end_io = handler;
53         bh->b_private = private;
54 }
55
56 static int sync_buffer(void *word)
57 {
58         struct block_device *bd;
59         struct buffer_head *bh
60                 = container_of(word, struct buffer_head, b_state);
61
62         smp_mb();
63         bd = bh->b_bdev;
64         if (bd)
65                 blk_run_address_space(bd->bd_inode->i_mapping);
66         io_schedule();
67         return 0;
68 }
69
70 void __lock_buffer(struct buffer_head *bh)
71 {
72         wait_on_bit_lock(&bh->b_state, BH_Lock, sync_buffer,
73                                                         TASK_UNINTERRUPTIBLE);
74 }
75 EXPORT_SYMBOL(__lock_buffer);
76
77 void unlock_buffer(struct buffer_head *bh)
78 {
79         clear_bit_unlock(BH_Lock, &bh->b_state);
80         smp_mb__after_clear_bit();
81         wake_up_bit(&bh->b_state, BH_Lock);
82 }
83
84 /*
85  * Block until a buffer comes unlocked.  This doesn't stop it
86  * from becoming locked again - you have to lock it yourself
87  * if you want to preserve its state.
88  */
89 void __wait_on_buffer(struct buffer_head * bh)
90 {
91         wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
92 }
93
94 static void
95 __clear_page_buffers(struct page *page)
96 {
97         ClearPagePrivate(page);
98         set_page_private(page, 0);
99         page_cache_release(page);
100 }
101
102
103 static int quiet_error(struct buffer_head *bh)
104 {
105         if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
106                 return 0;
107         return 1;
108 }
109
110
111 static void buffer_io_error(struct buffer_head *bh)
112 {
113         char b[BDEVNAME_SIZE];
114         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
115                         bdevname(bh->b_bdev, b),
116                         (unsigned long long)bh->b_blocknr);
117 }
118
119 /*
120  * End-of-IO handler helper function which does not touch the bh after
121  * unlocking it.
122  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
123  * a race there is benign: unlock_buffer() only use the bh's address for
124  * hashing after unlocking the buffer, so it doesn't actually touch the bh
125  * itself.
126  */
127 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
128 {
129         if (uptodate) {
130                 set_buffer_uptodate(bh);
131         } else {
132                 /* This happens, due to failed READA attempts. */
133                 clear_buffer_uptodate(bh);
134         }
135         unlock_buffer(bh);
136 }
137
138 /*
139  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
140  * unlock the buffer. This is what ll_rw_block uses too.
141  */
142 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
143 {
144         __end_buffer_read_notouch(bh, uptodate);
145         put_bh(bh);
146 }
147
148 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
149 {
150         char b[BDEVNAME_SIZE];
151
152         if (uptodate) {
153                 set_buffer_uptodate(bh);
154         } else {
155                 if (!buffer_eopnotsupp(bh) && !quiet_error(bh)) {
156                         buffer_io_error(bh);
157                         printk(KERN_WARNING "lost page write due to "
158                                         "I/O error on %s\n",
159                                        bdevname(bh->b_bdev, b));
160                 }
161                 set_buffer_write_io_error(bh);
162                 clear_buffer_uptodate(bh);
163         }
164         unlock_buffer(bh);
165         put_bh(bh);
166 }
167
168 /*
169  * Various filesystems appear to want __find_get_block to be non-blocking.
170  * But it's the page lock which protects the buffers.  To get around this,
171  * we get exclusion from try_to_free_buffers with the blockdev mapping's
172  * private_lock.
173  *
174  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
175  * may be quite high.  This code could TryLock the page, and if that
176  * succeeds, there is no need to take private_lock. (But if
177  * private_lock is contended then so is mapping->tree_lock).
178  */
179 static struct buffer_head *
180 __find_get_block_slow(struct block_device *bdev, sector_t block)
181 {
182         struct inode *bd_inode = bdev->bd_inode;
183         struct address_space *bd_mapping = bd_inode->i_mapping;
184         struct buffer_head *ret = NULL;
185         pgoff_t index;
186         struct buffer_head *bh;
187         struct buffer_head *head;
188         struct page *page;
189         int all_mapped = 1;
190
191         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
192         page = find_get_page(bd_mapping, index);
193         if (!page)
194                 goto out;
195
196         spin_lock(&bd_mapping->private_lock);
197         if (!page_has_buffers(page))
198                 goto out_unlock;
199         head = page_buffers(page);
200         bh = head;
201         do {
202                 if (!buffer_mapped(bh))
203                         all_mapped = 0;
204                 else if (bh->b_blocknr == block) {
205                         ret = bh;
206                         get_bh(bh);
207                         goto out_unlock;
208                 }
209                 bh = bh->b_this_page;
210         } while (bh != head);
211
212         /* we might be here because some of the buffers on this page are
213          * not mapped.  This is due to various races between
214          * file io on the block device and getblk.  It gets dealt with
215          * elsewhere, don't buffer_error if we had some unmapped buffers
216          */
217         if (all_mapped) {
218                 printk("__find_get_block_slow() failed. "
219                         "block=%llu, b_blocknr=%llu\n",
220                         (unsigned long long)block,
221                         (unsigned long long)bh->b_blocknr);
222                 printk("b_state=0x%08lx, b_size=%zu\n",
223                         bh->b_state, bh->b_size);
224                 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
225         }
226 out_unlock:
227         spin_unlock(&bd_mapping->private_lock);
228         page_cache_release(page);
229 out:
230         return ret;
231 }
232
233 /* If invalidate_buffers() will trash dirty buffers, it means some kind
234    of fs corruption is going on. Trashing dirty data always imply losing
235    information that was supposed to be just stored on the physical layer
236    by the user.
237
238    Thus invalidate_buffers in general usage is not allwowed to trash
239    dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
240    be preserved.  These buffers are simply skipped.
241   
242    We also skip buffers which are still in use.  For example this can
243    happen if a userspace program is reading the block device.
244
245    NOTE: In the case where the user removed a removable-media-disk even if
246    there's still dirty data not synced on disk (due a bug in the device driver
247    or due an error of the user), by not destroying the dirty buffers we could
248    generate corruption also on the next media inserted, thus a parameter is
249    necessary to handle this case in the most safe way possible (trying
250    to not corrupt also the new disk inserted with the data belonging to
251    the old now corrupted disk). Also for the ramdisk the natural thing
252    to do in order to release the ramdisk memory is to destroy dirty buffers.
253
254    These are two special cases. Normal usage imply the device driver
255    to issue a sync on the device (without waiting I/O completion) and
256    then an invalidate_buffers call that doesn't trash dirty buffers.
257
258    For handling cache coherency with the blkdev pagecache the 'update' case
259    is been introduced. It is needed to re-read from disk any pinned
260    buffer. NOTE: re-reading from disk is destructive so we can do it only
261    when we assume nobody is changing the buffercache under our I/O and when
262    we think the disk contains more recent information than the buffercache.
263    The update == 1 pass marks the buffers we need to update, the update == 2
264    pass does the actual I/O. */
265 void invalidate_bdev(struct block_device *bdev)
266 {
267         struct address_space *mapping = bdev->bd_inode->i_mapping;
268
269         if (mapping->nrpages == 0)
270                 return;
271
272         invalidate_bh_lrus();
273         invalidate_mapping_pages(mapping, 0, -1);
274 }
275
276 /*
277  * Kick pdflush then try to free up some ZONE_NORMAL memory.
278  */
279 static void free_more_memory(void)
280 {
281         struct zone *zone;
282         int nid;
283
284         wakeup_pdflush(1024);
285         yield();
286
287         for_each_online_node(nid) {
288                 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
289                                                 gfp_zone(GFP_NOFS), NULL,
290                                                 &zone);
291                 if (zone)
292                         try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
293                                                 GFP_NOFS, NULL);
294         }
295 }
296
297 /*
298  * I/O completion handler for block_read_full_page() - pages
299  * which come unlocked at the end of I/O.
300  */
301 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
302 {
303         unsigned long flags;
304         struct buffer_head *first;
305         struct buffer_head *tmp;
306         struct page *page;
307         int page_uptodate = 1;
308
309         BUG_ON(!buffer_async_read(bh));
310
311         page = bh->b_page;
312         if (uptodate) {
313                 set_buffer_uptodate(bh);
314         } else {
315                 clear_buffer_uptodate(bh);
316                 if (!quiet_error(bh))
317                         buffer_io_error(bh);
318                 SetPageError(page);
319         }
320
321         /*
322          * Be _very_ careful from here on. Bad things can happen if
323          * two buffer heads end IO at almost the same time and both
324          * decide that the page is now completely done.
325          */
326         first = page_buffers(page);
327         local_irq_save(flags);
328         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
329         clear_buffer_async_read(bh);
330         unlock_buffer(bh);
331         tmp = bh;
332         do {
333                 if (!buffer_uptodate(tmp))
334                         page_uptodate = 0;
335                 if (buffer_async_read(tmp)) {
336                         BUG_ON(!buffer_locked(tmp));
337                         goto still_busy;
338                 }
339                 tmp = tmp->b_this_page;
340         } while (tmp != bh);
341         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
342         local_irq_restore(flags);
343
344         /*
345          * If none of the buffers had errors and they are all
346          * uptodate then we can set the page uptodate.
347          */
348         if (page_uptodate && !PageError(page))
349                 SetPageUptodate(page);
350         unlock_page(page);
351         return;
352
353 still_busy:
354         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
355         local_irq_restore(flags);
356         return;
357 }
358
359 /*
360  * Completion handler for block_write_full_page() - pages which are unlocked
361  * during I/O, and which have PageWriteback cleared upon I/O completion.
362  */
363 static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
364 {
365         char b[BDEVNAME_SIZE];
366         unsigned long flags;
367         struct buffer_head *first;
368         struct buffer_head *tmp;
369         struct page *page;
370
371         BUG_ON(!buffer_async_write(bh));
372
373         page = bh->b_page;
374         if (uptodate) {
375                 set_buffer_uptodate(bh);
376         } else {
377                 if (!quiet_error(bh)) {
378                         buffer_io_error(bh);
379                         printk(KERN_WARNING "lost page write due to "
380                                         "I/O error on %s\n",
381                                bdevname(bh->b_bdev, b));
382                 }
383                 set_bit(AS_EIO, &page->mapping->flags);
384                 set_buffer_write_io_error(bh);
385                 clear_buffer_uptodate(bh);
386                 SetPageError(page);
387         }
388
389         first = page_buffers(page);
390         local_irq_save(flags);
391         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
392
393         clear_buffer_async_write(bh);
394         unlock_buffer(bh);
395         tmp = bh->b_this_page;
396         while (tmp != bh) {
397                 if (buffer_async_write(tmp)) {
398                         BUG_ON(!buffer_locked(tmp));
399                         goto still_busy;
400                 }
401                 tmp = tmp->b_this_page;
402         }
403         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
404         local_irq_restore(flags);
405         end_page_writeback(page);
406         return;
407
408 still_busy:
409         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
410         local_irq_restore(flags);
411         return;
412 }
413
414 /*
415  * If a page's buffers are under async readin (end_buffer_async_read
416  * completion) then there is a possibility that another thread of
417  * control could lock one of the buffers after it has completed
418  * but while some of the other buffers have not completed.  This
419  * locked buffer would confuse end_buffer_async_read() into not unlocking
420  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
421  * that this buffer is not under async I/O.
422  *
423  * The page comes unlocked when it has no locked buffer_async buffers
424  * left.
425  *
426  * PageLocked prevents anyone starting new async I/O reads any of
427  * the buffers.
428  *
429  * PageWriteback is used to prevent simultaneous writeout of the same
430  * page.
431  *
432  * PageLocked prevents anyone from starting writeback of a page which is
433  * under read I/O (PageWriteback is only ever set against a locked page).
434  */
435 static void mark_buffer_async_read(struct buffer_head *bh)
436 {
437         bh->b_end_io = end_buffer_async_read;
438         set_buffer_async_read(bh);
439 }
440
441 void mark_buffer_async_write(struct buffer_head *bh)
442 {
443         bh->b_end_io = end_buffer_async_write;
444         set_buffer_async_write(bh);
445 }
446 EXPORT_SYMBOL(mark_buffer_async_write);
447
448
449 /*
450  * fs/buffer.c contains helper functions for buffer-backed address space's
451  * fsync functions.  A common requirement for buffer-based filesystems is
452  * that certain data from the backing blockdev needs to be written out for
453  * a successful fsync().  For example, ext2 indirect blocks need to be
454  * written back and waited upon before fsync() returns.
455  *
456  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
457  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
458  * management of a list of dependent buffers at ->i_mapping->private_list.
459  *
460  * Locking is a little subtle: try_to_free_buffers() will remove buffers
461  * from their controlling inode's queue when they are being freed.  But
462  * try_to_free_buffers() will be operating against the *blockdev* mapping
463  * at the time, not against the S_ISREG file which depends on those buffers.
464  * So the locking for private_list is via the private_lock in the address_space
465  * which backs the buffers.  Which is different from the address_space 
466  * against which the buffers are listed.  So for a particular address_space,
467  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
468  * mapping->private_list will always be protected by the backing blockdev's
469  * ->private_lock.
470  *
471  * Which introduces a requirement: all buffers on an address_space's
472  * ->private_list must be from the same address_space: the blockdev's.
473  *
474  * address_spaces which do not place buffers at ->private_list via these
475  * utility functions are free to use private_lock and private_list for
476  * whatever they want.  The only requirement is that list_empty(private_list)
477  * be true at clear_inode() time.
478  *
479  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
480  * filesystems should do that.  invalidate_inode_buffers() should just go
481  * BUG_ON(!list_empty).
482  *
483  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
484  * take an address_space, not an inode.  And it should be called
485  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
486  * queued up.
487  *
488  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
489  * list if it is already on a list.  Because if the buffer is on a list,
490  * it *must* already be on the right one.  If not, the filesystem is being
491  * silly.  This will save a ton of locking.  But first we have to ensure
492  * that buffers are taken *off* the old inode's list when they are freed
493  * (presumably in truncate).  That requires careful auditing of all
494  * filesystems (do it inside bforget()).  It could also be done by bringing
495  * b_inode back.
496  */
497
498 /*
499  * The buffer's backing address_space's private_lock must be held
500  */
501 static void __remove_assoc_queue(struct buffer_head *bh)
502 {
503         list_del_init(&bh->b_assoc_buffers);
504         WARN_ON(!bh->b_assoc_map);
505         if (buffer_write_io_error(bh))
506                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
507         bh->b_assoc_map = NULL;
508 }
509
510 int inode_has_buffers(struct inode *inode)
511 {
512         return !list_empty(&inode->i_data.private_list);
513 }
514
515 /*
516  * osync is designed to support O_SYNC io.  It waits synchronously for
517  * all already-submitted IO to complete, but does not queue any new
518  * writes to the disk.
519  *
520  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
521  * you dirty the buffers, and then use osync_inode_buffers to wait for
522  * completion.  Any other dirty buffers which are not yet queued for
523  * write will not be flushed to disk by the osync.
524  */
525 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
526 {
527         struct buffer_head *bh;
528         struct list_head *p;
529         int err = 0;
530
531         spin_lock(lock);
532 repeat:
533         list_for_each_prev(p, list) {
534                 bh = BH_ENTRY(p);
535                 if (buffer_locked(bh)) {
536                         get_bh(bh);
537                         spin_unlock(lock);
538                         wait_on_buffer(bh);
539                         if (!buffer_uptodate(bh))
540                                 err = -EIO;
541                         brelse(bh);
542                         spin_lock(lock);
543                         goto repeat;
544                 }
545         }
546         spin_unlock(lock);
547         return err;
548 }
549
550 void do_thaw_all(unsigned long unused)
551 {
552         struct super_block *sb;
553         char b[BDEVNAME_SIZE];
554
555         spin_lock(&sb_lock);
556 restart:
557         list_for_each_entry(sb, &super_blocks, s_list) {
558                 sb->s_count++;
559                 spin_unlock(&sb_lock);
560                 down_read(&sb->s_umount);
561                 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
562                         printk(KERN_WARNING "Emergency Thaw on %s\n",
563                                bdevname(sb->s_bdev, b));
564                 up_read(&sb->s_umount);
565                 spin_lock(&sb_lock);
566                 if (__put_super_and_need_restart(sb))
567                         goto restart;
568         }
569         spin_unlock(&sb_lock);
570         printk(KERN_WARNING "Emergency Thaw complete\n");
571 }
572
573 /**
574  * emergency_thaw_all -- forcibly thaw every frozen filesystem
575  *
576  * Used for emergency unfreeze of all filesystems via SysRq
577  */
578 void emergency_thaw_all(void)
579 {
580         pdflush_operation(do_thaw_all, 0);
581 }
582
583 /**
584  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
585  * @mapping: the mapping which wants those buffers written
586  *
587  * Starts I/O against the buffers at mapping->private_list, and waits upon
588  * that I/O.
589  *
590  * Basically, this is a convenience function for fsync().
591  * @mapping is a file or directory which needs those buffers to be written for
592  * a successful fsync().
593  */
594 int sync_mapping_buffers(struct address_space *mapping)
595 {
596         struct address_space *buffer_mapping = mapping->assoc_mapping;
597
598         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
599                 return 0;
600
601         return fsync_buffers_list(&buffer_mapping->private_lock,
602                                         &mapping->private_list);
603 }
604 EXPORT_SYMBOL(sync_mapping_buffers);
605
606 /*
607  * Called when we've recently written block `bblock', and it is known that
608  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
609  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
610  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
611  */
612 void write_boundary_block(struct block_device *bdev,
613                         sector_t bblock, unsigned blocksize)
614 {
615         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
616         if (bh) {
617                 if (buffer_dirty(bh))
618                         ll_rw_block(WRITE, 1, &bh);
619                 put_bh(bh);
620         }
621 }
622
623 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
624 {
625         struct address_space *mapping = inode->i_mapping;
626         struct address_space *buffer_mapping = bh->b_page->mapping;
627
628         mark_buffer_dirty(bh);
629         if (!mapping->assoc_mapping) {
630                 mapping->assoc_mapping = buffer_mapping;
631         } else {
632                 BUG_ON(mapping->assoc_mapping != buffer_mapping);
633         }
634         if (!bh->b_assoc_map) {
635                 spin_lock(&buffer_mapping->private_lock);
636                 list_move_tail(&bh->b_assoc_buffers,
637                                 &mapping->private_list);
638                 bh->b_assoc_map = mapping;
639                 spin_unlock(&buffer_mapping->private_lock);
640         }
641 }
642 EXPORT_SYMBOL(mark_buffer_dirty_inode);
643
644 /*
645  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
646  * dirty.
647  *
648  * If warn is true, then emit a warning if the page is not uptodate and has
649  * not been truncated.
650  */
651 static void __set_page_dirty(struct page *page,
652                 struct address_space *mapping, int warn)
653 {
654         spin_lock_irq(&mapping->tree_lock);
655         if (page->mapping) {    /* Race with truncate? */
656                 WARN_ON_ONCE(warn && !PageUptodate(page));
657                 account_page_dirtied(page, mapping);
658                 radix_tree_tag_set(&mapping->page_tree,
659                                 page_index(page), PAGECACHE_TAG_DIRTY);
660         }
661         spin_unlock_irq(&mapping->tree_lock);
662         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
663 }
664
665 /*
666  * Add a page to the dirty page list.
667  *
668  * It is a sad fact of life that this function is called from several places
669  * deeply under spinlocking.  It may not sleep.
670  *
671  * If the page has buffers, the uptodate buffers are set dirty, to preserve
672  * dirty-state coherency between the page and the buffers.  It the page does
673  * not have buffers then when they are later attached they will all be set
674  * dirty.
675  *
676  * The buffers are dirtied before the page is dirtied.  There's a small race
677  * window in which a writepage caller may see the page cleanness but not the
678  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
679  * before the buffers, a concurrent writepage caller could clear the page dirty
680  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
681  * page on the dirty page list.
682  *
683  * We use private_lock to lock against try_to_free_buffers while using the
684  * page's buffer list.  Also use this to protect against clean buffers being
685  * added to the page after it was set dirty.
686  *
687  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
688  * address_space though.
689  */
690 int __set_page_dirty_buffers(struct page *page)
691 {
692         int newly_dirty;
693         struct address_space *mapping = page_mapping(page);
694
695         if (unlikely(!mapping))
696                 return !TestSetPageDirty(page);
697
698         spin_lock(&mapping->private_lock);
699         if (page_has_buffers(page)) {
700                 struct buffer_head *head = page_buffers(page);
701                 struct buffer_head *bh = head;
702
703                 do {
704                         set_buffer_dirty(bh);
705                         bh = bh->b_this_page;
706                 } while (bh != head);
707         }
708         newly_dirty = !TestSetPageDirty(page);
709         spin_unlock(&mapping->private_lock);
710
711         if (newly_dirty)
712                 __set_page_dirty(page, mapping, 1);
713         return newly_dirty;
714 }
715 EXPORT_SYMBOL(__set_page_dirty_buffers);
716
717 /*
718  * Write out and wait upon a list of buffers.
719  *
720  * We have conflicting pressures: we want to make sure that all
721  * initially dirty buffers get waited on, but that any subsequently
722  * dirtied buffers don't.  After all, we don't want fsync to last
723  * forever if somebody is actively writing to the file.
724  *
725  * Do this in two main stages: first we copy dirty buffers to a
726  * temporary inode list, queueing the writes as we go.  Then we clean
727  * up, waiting for those writes to complete.
728  * 
729  * During this second stage, any subsequent updates to the file may end
730  * up refiling the buffer on the original inode's dirty list again, so
731  * there is a chance we will end up with a buffer queued for write but
732  * not yet completed on that list.  So, as a final cleanup we go through
733  * the osync code to catch these locked, dirty buffers without requeuing
734  * any newly dirty buffers for write.
735  */
736 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
737 {
738         struct buffer_head *bh;
739         struct list_head tmp;
740         struct address_space *mapping, *prev_mapping = NULL;
741         int err = 0, err2;
742
743         INIT_LIST_HEAD(&tmp);
744
745         spin_lock(lock);
746         while (!list_empty(list)) {
747                 bh = BH_ENTRY(list->next);
748                 mapping = bh->b_assoc_map;
749                 __remove_assoc_queue(bh);
750                 /* Avoid race with mark_buffer_dirty_inode() which does
751                  * a lockless check and we rely on seeing the dirty bit */
752                 smp_mb();
753                 if (buffer_dirty(bh) || buffer_locked(bh)) {
754                         list_add(&bh->b_assoc_buffers, &tmp);
755                         bh->b_assoc_map = mapping;
756                         if (buffer_dirty(bh)) {
757                                 get_bh(bh);
758                                 spin_unlock(lock);
759                                 /*
760                                  * Ensure any pending I/O completes so that
761                                  * ll_rw_block() actually writes the current
762                                  * contents - it is a noop if I/O is still in
763                                  * flight on potentially older contents.
764                                  */
765                                 ll_rw_block(SWRITE_SYNC_PLUG, 1, &bh);
766
767                                 /*
768                                  * Kick off IO for the previous mapping. Note
769                                  * that we will not run the very last mapping,
770                                  * wait_on_buffer() will do that for us
771                                  * through sync_buffer().
772                                  */
773                                 if (prev_mapping && prev_mapping != mapping)
774                                         blk_run_address_space(prev_mapping);
775                                 prev_mapping = mapping;
776
777                                 brelse(bh);
778                                 spin_lock(lock);
779                         }
780                 }
781         }
782
783         while (!list_empty(&tmp)) {
784                 bh = BH_ENTRY(tmp.prev);
785                 get_bh(bh);
786                 mapping = bh->b_assoc_map;
787                 __remove_assoc_queue(bh);
788                 /* Avoid race with mark_buffer_dirty_inode() which does
789                  * a lockless check and we rely on seeing the dirty bit */
790                 smp_mb();
791                 if (buffer_dirty(bh)) {
792                         list_add(&bh->b_assoc_buffers,
793                                  &mapping->private_list);
794                         bh->b_assoc_map = mapping;
795                 }
796                 spin_unlock(lock);
797                 wait_on_buffer(bh);
798                 if (!buffer_uptodate(bh))
799                         err = -EIO;
800                 brelse(bh);
801                 spin_lock(lock);
802         }
803         
804         spin_unlock(lock);
805         err2 = osync_buffers_list(lock, list);
806         if (err)
807                 return err;
808         else
809                 return err2;
810 }
811
812 /*
813  * Invalidate any and all dirty buffers on a given inode.  We are
814  * probably unmounting the fs, but that doesn't mean we have already
815  * done a sync().  Just drop the buffers from the inode list.
816  *
817  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
818  * assumes that all the buffers are against the blockdev.  Not true
819  * for reiserfs.
820  */
821 void invalidate_inode_buffers(struct inode *inode)
822 {
823         if (inode_has_buffers(inode)) {
824                 struct address_space *mapping = &inode->i_data;
825                 struct list_head *list = &mapping->private_list;
826                 struct address_space *buffer_mapping = mapping->assoc_mapping;
827
828                 spin_lock(&buffer_mapping->private_lock);
829                 while (!list_empty(list))
830                         __remove_assoc_queue(BH_ENTRY(list->next));
831                 spin_unlock(&buffer_mapping->private_lock);
832         }
833 }
834 EXPORT_SYMBOL(invalidate_inode_buffers);
835
836 /*
837  * Remove any clean buffers from the inode's buffer list.  This is called
838  * when we're trying to free the inode itself.  Those buffers can pin it.
839  *
840  * Returns true if all buffers were removed.
841  */
842 int remove_inode_buffers(struct inode *inode)
843 {
844         int ret = 1;
845
846         if (inode_has_buffers(inode)) {
847                 struct address_space *mapping = &inode->i_data;
848                 struct list_head *list = &mapping->private_list;
849                 struct address_space *buffer_mapping = mapping->assoc_mapping;
850
851                 spin_lock(&buffer_mapping->private_lock);
852                 while (!list_empty(list)) {
853                         struct buffer_head *bh = BH_ENTRY(list->next);
854                         if (buffer_dirty(bh)) {
855                                 ret = 0;
856                                 break;
857                         }
858                         __remove_assoc_queue(bh);
859                 }
860                 spin_unlock(&buffer_mapping->private_lock);
861         }
862         return ret;
863 }
864
865 /*
866  * Create the appropriate buffers when given a page for data area and
867  * the size of each buffer.. Use the bh->b_this_page linked list to
868  * follow the buffers created.  Return NULL if unable to create more
869  * buffers.
870  *
871  * The retry flag is used to differentiate async IO (paging, swapping)
872  * which may not fail from ordinary buffer allocations.
873  */
874 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
875                 int retry)
876 {
877         struct buffer_head *bh, *head;
878         long offset;
879
880 try_again:
881         head = NULL;
882         offset = PAGE_SIZE;
883         while ((offset -= size) >= 0) {
884                 bh = alloc_buffer_head(GFP_NOFS);
885                 if (!bh)
886                         goto no_grow;
887
888                 bh->b_bdev = NULL;
889                 bh->b_this_page = head;
890                 bh->b_blocknr = -1;
891                 head = bh;
892
893                 bh->b_state = 0;
894                 atomic_set(&bh->b_count, 0);
895                 bh->b_private = NULL;
896                 bh->b_size = size;
897
898                 /* Link the buffer to its page */
899                 set_bh_page(bh, page, offset);
900
901                 init_buffer(bh, NULL, NULL);
902         }
903         return head;
904 /*
905  * In case anything failed, we just free everything we got.
906  */
907 no_grow:
908         if (head) {
909                 do {
910                         bh = head;
911                         head = head->b_this_page;
912                         free_buffer_head(bh);
913                 } while (head);
914         }
915
916         /*
917          * Return failure for non-async IO requests.  Async IO requests
918          * are not allowed to fail, so we have to wait until buffer heads
919          * become available.  But we don't want tasks sleeping with 
920          * partially complete buffers, so all were released above.
921          */
922         if (!retry)
923                 return NULL;
924
925         /* We're _really_ low on memory. Now we just
926          * wait for old buffer heads to become free due to
927          * finishing IO.  Since this is an async request and
928          * the reserve list is empty, we're sure there are 
929          * async buffer heads in use.
930          */
931         free_more_memory();
932         goto try_again;
933 }
934 EXPORT_SYMBOL_GPL(alloc_page_buffers);
935
936 static inline void
937 link_dev_buffers(struct page *page, struct buffer_head *head)
938 {
939         struct buffer_head *bh, *tail;
940
941         bh = head;
942         do {
943                 tail = bh;
944                 bh = bh->b_this_page;
945         } while (bh);
946         tail->b_this_page = head;
947         attach_page_buffers(page, head);
948 }
949
950 /*
951  * Initialise the state of a blockdev page's buffers.
952  */ 
953 static void
954 init_page_buffers(struct page *page, struct block_device *bdev,
955                         sector_t block, int size)
956 {
957         struct buffer_head *head = page_buffers(page);
958         struct buffer_head *bh = head;
959         int uptodate = PageUptodate(page);
960
961         do {
962                 if (!buffer_mapped(bh)) {
963                         init_buffer(bh, NULL, NULL);
964                         bh->b_bdev = bdev;
965                         bh->b_blocknr = block;
966                         if (uptodate)
967                                 set_buffer_uptodate(bh);
968                         set_buffer_mapped(bh);
969                 }
970                 block++;
971                 bh = bh->b_this_page;
972         } while (bh != head);
973 }
974
975 /*
976  * Create the page-cache page that contains the requested block.
977  *
978  * This is user purely for blockdev mappings.
979  */
980 static struct page *
981 grow_dev_page(struct block_device *bdev, sector_t block,
982                 pgoff_t index, int size)
983 {
984         struct inode *inode = bdev->bd_inode;
985         struct page *page;
986         struct buffer_head *bh;
987
988         page = find_or_create_page(inode->i_mapping, index,
989                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
990         if (!page)
991                 return NULL;
992
993         BUG_ON(!PageLocked(page));
994
995         if (page_has_buffers(page)) {
996                 bh = page_buffers(page);
997                 if (bh->b_size == size) {
998                         init_page_buffers(page, bdev, block, size);
999                         return page;
1000                 }
1001                 if (!try_to_free_buffers(page))
1002                         goto failed;
1003         }
1004
1005         /*
1006          * Allocate some buffers for this page
1007          */
1008         bh = alloc_page_buffers(page, size, 0);
1009         if (!bh)
1010                 goto failed;
1011
1012         /*
1013          * Link the page to the buffers and initialise them.  Take the
1014          * lock to be atomic wrt __find_get_block(), which does not
1015          * run under the page lock.
1016          */
1017         spin_lock(&inode->i_mapping->private_lock);
1018         link_dev_buffers(page, bh);
1019         init_page_buffers(page, bdev, block, size);
1020         spin_unlock(&inode->i_mapping->private_lock);
1021         return page;
1022
1023 failed:
1024         BUG();
1025         unlock_page(page);
1026         page_cache_release(page);
1027         return NULL;
1028 }
1029
1030 /*
1031  * Create buffers for the specified block device block's page.  If
1032  * that page was dirty, the buffers are set dirty also.
1033  */
1034 static int
1035 grow_buffers(struct block_device *bdev, sector_t block, int size)
1036 {
1037         struct page *page;
1038         pgoff_t index;
1039         int sizebits;
1040
1041         sizebits = -1;
1042         do {
1043                 sizebits++;
1044         } while ((size << sizebits) < PAGE_SIZE);
1045
1046         index = block >> sizebits;
1047
1048         /*
1049          * Check for a block which wants to lie outside our maximum possible
1050          * pagecache index.  (this comparison is done using sector_t types).
1051          */
1052         if (unlikely(index != block >> sizebits)) {
1053                 char b[BDEVNAME_SIZE];
1054
1055                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1056                         "device %s\n",
1057                         __func__, (unsigned long long)block,
1058                         bdevname(bdev, b));
1059                 return -EIO;
1060         }
1061         block = index << sizebits;
1062         /* Create a page with the proper size buffers.. */
1063         page = grow_dev_page(bdev, block, index, size);
1064         if (!page)
1065                 return 0;
1066         unlock_page(page);
1067         page_cache_release(page);
1068         return 1;
1069 }
1070
1071 static struct buffer_head *
1072 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1073 {
1074         /* Size must be multiple of hard sectorsize */
1075         if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1076                         (size < 512 || size > PAGE_SIZE))) {
1077                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1078                                         size);
1079                 printk(KERN_ERR "hardsect size: %d\n",
1080                                         bdev_hardsect_size(bdev));
1081
1082                 dump_stack();
1083                 return NULL;
1084         }
1085
1086         for (;;) {
1087                 struct buffer_head * bh;
1088                 int ret;
1089
1090                 bh = __find_get_block(bdev, block, size);
1091                 if (bh)
1092                         return bh;
1093
1094                 ret = grow_buffers(bdev, block, size);
1095                 if (ret < 0)
1096                         return NULL;
1097                 if (ret == 0)
1098                         free_more_memory();
1099         }
1100 }
1101
1102 /*
1103  * The relationship between dirty buffers and dirty pages:
1104  *
1105  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1106  * the page is tagged dirty in its radix tree.
1107  *
1108  * At all times, the dirtiness of the buffers represents the dirtiness of
1109  * subsections of the page.  If the page has buffers, the page dirty bit is
1110  * merely a hint about the true dirty state.
1111  *
1112  * When a page is set dirty in its entirety, all its buffers are marked dirty
1113  * (if the page has buffers).
1114  *
1115  * When a buffer is marked dirty, its page is dirtied, but the page's other
1116  * buffers are not.
1117  *
1118  * Also.  When blockdev buffers are explicitly read with bread(), they
1119  * individually become uptodate.  But their backing page remains not
1120  * uptodate - even if all of its buffers are uptodate.  A subsequent
1121  * block_read_full_page() against that page will discover all the uptodate
1122  * buffers, will set the page uptodate and will perform no I/O.
1123  */
1124
1125 /**
1126  * mark_buffer_dirty - mark a buffer_head as needing writeout
1127  * @bh: the buffer_head to mark dirty
1128  *
1129  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1130  * backing page dirty, then tag the page as dirty in its address_space's radix
1131  * tree and then attach the address_space's inode to its superblock's dirty
1132  * inode list.
1133  *
1134  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1135  * mapping->tree_lock and the global inode_lock.
1136  */
1137 void mark_buffer_dirty(struct buffer_head *bh)
1138 {
1139         WARN_ON_ONCE(!buffer_uptodate(bh));
1140
1141         /*
1142          * Very *carefully* optimize the it-is-already-dirty case.
1143          *
1144          * Don't let the final "is it dirty" escape to before we
1145          * perhaps modified the buffer.
1146          */
1147         if (buffer_dirty(bh)) {
1148                 smp_mb();
1149                 if (buffer_dirty(bh))
1150                         return;
1151         }
1152
1153         if (!test_set_buffer_dirty(bh)) {
1154                 struct page *page = bh->b_page;
1155                 if (!TestSetPageDirty(page))
1156                         __set_page_dirty(page, page_mapping(page), 0);
1157         }
1158 }
1159
1160 /*
1161  * Decrement a buffer_head's reference count.  If all buffers against a page
1162  * have zero reference count, are clean and unlocked, and if the page is clean
1163  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1164  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1165  * a page but it ends up not being freed, and buffers may later be reattached).
1166  */
1167 void __brelse(struct buffer_head * buf)
1168 {
1169         if (atomic_read(&buf->b_count)) {
1170                 put_bh(buf);
1171                 return;
1172         }
1173         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1174 }
1175
1176 /*
1177  * bforget() is like brelse(), except it discards any
1178  * potentially dirty data.
1179  */
1180 void __bforget(struct buffer_head *bh)
1181 {
1182         clear_buffer_dirty(bh);
1183         if (bh->b_assoc_map) {
1184                 struct address_space *buffer_mapping = bh->b_page->mapping;
1185
1186                 spin_lock(&buffer_mapping->private_lock);
1187                 list_del_init(&bh->b_assoc_buffers);
1188                 bh->b_assoc_map = NULL;
1189                 spin_unlock(&buffer_mapping->private_lock);
1190         }
1191         __brelse(bh);
1192 }
1193
1194 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1195 {
1196         lock_buffer(bh);
1197         if (buffer_uptodate(bh)) {
1198                 unlock_buffer(bh);
1199                 return bh;
1200         } else {
1201                 get_bh(bh);
1202                 bh->b_end_io = end_buffer_read_sync;
1203                 submit_bh(READ, bh);
1204                 wait_on_buffer(bh);
1205                 if (buffer_uptodate(bh))
1206                         return bh;
1207         }
1208         brelse(bh);
1209         return NULL;
1210 }
1211
1212 /*
1213  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1214  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1215  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1216  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1217  * CPU's LRUs at the same time.
1218  *
1219  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1220  * sb_find_get_block().
1221  *
1222  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1223  * a local interrupt disable for that.
1224  */
1225
1226 #define BH_LRU_SIZE     8
1227
1228 struct bh_lru {
1229         struct buffer_head *bhs[BH_LRU_SIZE];
1230 };
1231
1232 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1233
1234 #ifdef CONFIG_SMP
1235 #define bh_lru_lock()   local_irq_disable()
1236 #define bh_lru_unlock() local_irq_enable()
1237 #else
1238 #define bh_lru_lock()   preempt_disable()
1239 #define bh_lru_unlock() preempt_enable()
1240 #endif
1241
1242 static inline void check_irqs_on(void)
1243 {
1244 #ifdef irqs_disabled
1245         BUG_ON(irqs_disabled());
1246 #endif
1247 }
1248
1249 /*
1250  * The LRU management algorithm is dopey-but-simple.  Sorry.
1251  */
1252 static void bh_lru_install(struct buffer_head *bh)
1253 {
1254         struct buffer_head *evictee = NULL;
1255         struct bh_lru *lru;
1256
1257         check_irqs_on();
1258         bh_lru_lock();
1259         lru = &__get_cpu_var(bh_lrus);
1260         if (lru->bhs[0] != bh) {
1261                 struct buffer_head *bhs[BH_LRU_SIZE];
1262                 int in;
1263                 int out = 0;
1264
1265                 get_bh(bh);
1266                 bhs[out++] = bh;
1267                 for (in = 0; in < BH_LRU_SIZE; in++) {
1268                         struct buffer_head *bh2 = lru->bhs[in];
1269
1270                         if (bh2 == bh) {
1271                                 __brelse(bh2);
1272                         } else {
1273                                 if (out >= BH_LRU_SIZE) {
1274                                         BUG_ON(evictee != NULL);
1275                                         evictee = bh2;
1276                                 } else {
1277                                         bhs[out++] = bh2;
1278                                 }
1279                         }
1280                 }
1281                 while (out < BH_LRU_SIZE)
1282                         bhs[out++] = NULL;
1283                 memcpy(lru->bhs, bhs, sizeof(bhs));
1284         }
1285         bh_lru_unlock();
1286
1287         if (evictee)
1288                 __brelse(evictee);
1289 }
1290
1291 /*
1292  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1293  */
1294 static struct buffer_head *
1295 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1296 {
1297         struct buffer_head *ret = NULL;
1298         struct bh_lru *lru;
1299         unsigned int i;
1300
1301         check_irqs_on();
1302         bh_lru_lock();
1303         lru = &__get_cpu_var(bh_lrus);
1304         for (i = 0; i < BH_LRU_SIZE; i++) {
1305                 struct buffer_head *bh = lru->bhs[i];
1306
1307                 if (bh && bh->b_bdev == bdev &&
1308                                 bh->b_blocknr == block && bh->b_size == size) {
1309                         if (i) {
1310                                 while (i) {
1311                                         lru->bhs[i] = lru->bhs[i - 1];
1312                                         i--;
1313                                 }
1314                                 lru->bhs[0] = bh;
1315                         }
1316                         get_bh(bh);
1317                         ret = bh;
1318                         break;
1319                 }
1320         }
1321         bh_lru_unlock();
1322         return ret;
1323 }
1324
1325 /*
1326  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1327  * it in the LRU and mark it as accessed.  If it is not present then return
1328  * NULL
1329  */
1330 struct buffer_head *
1331 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1332 {
1333         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1334
1335         if (bh == NULL) {
1336                 bh = __find_get_block_slow(bdev, block);
1337                 if (bh)
1338                         bh_lru_install(bh);
1339         }
1340         if (bh)
1341                 touch_buffer(bh);
1342         return bh;
1343 }
1344 EXPORT_SYMBOL(__find_get_block);
1345
1346 /*
1347  * __getblk will locate (and, if necessary, create) the buffer_head
1348  * which corresponds to the passed block_device, block and size. The
1349  * returned buffer has its reference count incremented.
1350  *
1351  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1352  * illegal block number, __getblk() will happily return a buffer_head
1353  * which represents the non-existent block.  Very weird.
1354  *
1355  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1356  * attempt is failing.  FIXME, perhaps?
1357  */
1358 struct buffer_head *
1359 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1360 {
1361         struct buffer_head *bh = __find_get_block(bdev, block, size);
1362
1363         might_sleep();
1364         if (bh == NULL)
1365                 bh = __getblk_slow(bdev, block, size);
1366         return bh;
1367 }
1368 EXPORT_SYMBOL(__getblk);
1369
1370 /*
1371  * Do async read-ahead on a buffer..
1372  */
1373 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1374 {
1375         struct buffer_head *bh = __getblk(bdev, block, size);
1376         if (likely(bh)) {
1377                 ll_rw_block(READA, 1, &bh);
1378                 brelse(bh);
1379         }
1380 }
1381 EXPORT_SYMBOL(__breadahead);
1382
1383 /**
1384  *  __bread() - reads a specified block and returns the bh
1385  *  @bdev: the block_device to read from
1386  *  @block: number of block
1387  *  @size: size (in bytes) to read
1388  * 
1389  *  Reads a specified block, and returns buffer head that contains it.
1390  *  It returns NULL if the block was unreadable.
1391  */
1392 struct buffer_head *
1393 __bread(struct block_device *bdev, sector_t block, unsigned size)
1394 {
1395         struct buffer_head *bh = __getblk(bdev, block, size);
1396
1397         if (likely(bh) && !buffer_uptodate(bh))
1398                 bh = __bread_slow(bh);
1399         return bh;
1400 }
1401 EXPORT_SYMBOL(__bread);
1402
1403 /*
1404  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1405  * This doesn't race because it runs in each cpu either in irq
1406  * or with preempt disabled.
1407  */
1408 static void invalidate_bh_lru(void *arg)
1409 {
1410         struct bh_lru *b = &get_cpu_var(bh_lrus);
1411         int i;
1412
1413         for (i = 0; i < BH_LRU_SIZE; i++) {
1414                 brelse(b->bhs[i]);
1415                 b->bhs[i] = NULL;
1416         }
1417         put_cpu_var(bh_lrus);
1418 }
1419         
1420 void invalidate_bh_lrus(void)
1421 {
1422         on_each_cpu(invalidate_bh_lru, NULL, 1);
1423 }
1424 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1425
1426 void set_bh_page(struct buffer_head *bh,
1427                 struct page *page, unsigned long offset)
1428 {
1429         bh->b_page = page;
1430         BUG_ON(offset >= PAGE_SIZE);
1431         if (PageHighMem(page))
1432                 /*
1433                  * This catches illegal uses and preserves the offset:
1434                  */
1435                 bh->b_data = (char *)(0 + offset);
1436         else
1437                 bh->b_data = page_address(page) + offset;
1438 }
1439 EXPORT_SYMBOL(set_bh_page);
1440
1441 /*
1442  * Called when truncating a buffer on a page completely.
1443  */
1444 static void discard_buffer(struct buffer_head * bh)
1445 {
1446         lock_buffer(bh);
1447         clear_buffer_dirty(bh);
1448         bh->b_bdev = NULL;
1449         clear_buffer_mapped(bh);
1450         clear_buffer_req(bh);
1451         clear_buffer_new(bh);
1452         clear_buffer_delay(bh);
1453         clear_buffer_unwritten(bh);
1454         unlock_buffer(bh);
1455 }
1456
1457 /**
1458  * block_invalidatepage - invalidate part of all of a buffer-backed page
1459  *
1460  * @page: the page which is affected
1461  * @offset: the index of the truncation point
1462  *
1463  * block_invalidatepage() is called when all or part of the page has become
1464  * invalidatedby a truncate operation.
1465  *
1466  * block_invalidatepage() does not have to release all buffers, but it must
1467  * ensure that no dirty buffer is left outside @offset and that no I/O
1468  * is underway against any of the blocks which are outside the truncation
1469  * point.  Because the caller is about to free (and possibly reuse) those
1470  * blocks on-disk.
1471  */
1472 void block_invalidatepage(struct page *page, unsigned long offset)
1473 {
1474         struct buffer_head *head, *bh, *next;
1475         unsigned int curr_off = 0;
1476
1477         BUG_ON(!PageLocked(page));
1478         if (!page_has_buffers(page))
1479                 goto out;
1480
1481         head = page_buffers(page);
1482         bh = head;
1483         do {
1484                 unsigned int next_off = curr_off + bh->b_size;
1485                 next = bh->b_this_page;
1486
1487                 /*
1488                  * is this block fully invalidated?
1489                  */
1490                 if (offset <= curr_off)
1491                         discard_buffer(bh);
1492                 curr_off = next_off;
1493                 bh = next;
1494         } while (bh != head);
1495
1496         /*
1497          * We release buffers only if the entire page is being invalidated.
1498          * The get_block cached value has been unconditionally invalidated,
1499          * so real IO is not possible anymore.
1500          */
1501         if (offset == 0)
1502                 try_to_release_page(page, 0);
1503 out:
1504         return;
1505 }
1506 EXPORT_SYMBOL(block_invalidatepage);
1507
1508 /*
1509  * We attach and possibly dirty the buffers atomically wrt
1510  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1511  * is already excluded via the page lock.
1512  */
1513 void create_empty_buffers(struct page *page,
1514                         unsigned long blocksize, unsigned long b_state)
1515 {
1516         struct buffer_head *bh, *head, *tail;
1517
1518         head = alloc_page_buffers(page, blocksize, 1);
1519         bh = head;
1520         do {
1521                 bh->b_state |= b_state;
1522                 tail = bh;
1523                 bh = bh->b_this_page;
1524         } while (bh);
1525         tail->b_this_page = head;
1526
1527         spin_lock(&page->mapping->private_lock);
1528         if (PageUptodate(page) || PageDirty(page)) {
1529                 bh = head;
1530                 do {
1531                         if (PageDirty(page))
1532                                 set_buffer_dirty(bh);
1533                         if (PageUptodate(page))
1534                                 set_buffer_uptodate(bh);
1535                         bh = bh->b_this_page;
1536                 } while (bh != head);
1537         }
1538         attach_page_buffers(page, head);
1539         spin_unlock(&page->mapping->private_lock);
1540 }
1541 EXPORT_SYMBOL(create_empty_buffers);
1542
1543 /*
1544  * We are taking a block for data and we don't want any output from any
1545  * buffer-cache aliases starting from return from that function and
1546  * until the moment when something will explicitly mark the buffer
1547  * dirty (hopefully that will not happen until we will free that block ;-)
1548  * We don't even need to mark it not-uptodate - nobody can expect
1549  * anything from a newly allocated buffer anyway. We used to used
1550  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1551  * don't want to mark the alias unmapped, for example - it would confuse
1552  * anyone who might pick it with bread() afterwards...
1553  *
1554  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1555  * be writeout I/O going on against recently-freed buffers.  We don't
1556  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1557  * only if we really need to.  That happens here.
1558  */
1559 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1560 {
1561         struct buffer_head *old_bh;
1562
1563         might_sleep();
1564
1565         old_bh = __find_get_block_slow(bdev, block);
1566         if (old_bh) {
1567                 clear_buffer_dirty(old_bh);
1568                 wait_on_buffer(old_bh);
1569                 clear_buffer_req(old_bh);
1570                 __brelse(old_bh);
1571         }
1572 }
1573 EXPORT_SYMBOL(unmap_underlying_metadata);
1574
1575 /*
1576  * NOTE! All mapped/uptodate combinations are valid:
1577  *
1578  *      Mapped  Uptodate        Meaning
1579  *
1580  *      No      No              "unknown" - must do get_block()
1581  *      No      Yes             "hole" - zero-filled
1582  *      Yes     No              "allocated" - allocated on disk, not read in
1583  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1584  *
1585  * "Dirty" is valid only with the last case (mapped+uptodate).
1586  */
1587
1588 /*
1589  * While block_write_full_page is writing back the dirty buffers under
1590  * the page lock, whoever dirtied the buffers may decide to clean them
1591  * again at any time.  We handle that by only looking at the buffer
1592  * state inside lock_buffer().
1593  *
1594  * If block_write_full_page() is called for regular writeback
1595  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1596  * locked buffer.   This only can happen if someone has written the buffer
1597  * directly, with submit_bh().  At the address_space level PageWriteback
1598  * prevents this contention from occurring.
1599  */
1600 static int __block_write_full_page(struct inode *inode, struct page *page,
1601                         get_block_t *get_block, struct writeback_control *wbc)
1602 {
1603         int err;
1604         sector_t block;
1605         sector_t last_block;
1606         struct buffer_head *bh, *head;
1607         const unsigned blocksize = 1 << inode->i_blkbits;
1608         int nr_underway = 0;
1609         int write_op = (wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE);
1610
1611         BUG_ON(!PageLocked(page));
1612
1613         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1614
1615         if (!page_has_buffers(page)) {
1616                 create_empty_buffers(page, blocksize,
1617                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1618         }
1619
1620         /*
1621          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1622          * here, and the (potentially unmapped) buffers may become dirty at
1623          * any time.  If a buffer becomes dirty here after we've inspected it
1624          * then we just miss that fact, and the page stays dirty.
1625          *
1626          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1627          * handle that here by just cleaning them.
1628          */
1629
1630         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1631         head = page_buffers(page);
1632         bh = head;
1633
1634         /*
1635          * Get all the dirty buffers mapped to disk addresses and
1636          * handle any aliases from the underlying blockdev's mapping.
1637          */
1638         do {
1639                 if (block > last_block) {
1640                         /*
1641                          * mapped buffers outside i_size will occur, because
1642                          * this page can be outside i_size when there is a
1643                          * truncate in progress.
1644                          */
1645                         /*
1646                          * The buffer was zeroed by block_write_full_page()
1647                          */
1648                         clear_buffer_dirty(bh);
1649                         set_buffer_uptodate(bh);
1650                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1651                            buffer_dirty(bh)) {
1652                         WARN_ON(bh->b_size != blocksize);
1653                         err = get_block(inode, block, bh, 1);
1654                         if (err)
1655                                 goto recover;
1656                         clear_buffer_delay(bh);
1657                         if (buffer_new(bh)) {
1658                                 /* blockdev mappings never come here */
1659                                 clear_buffer_new(bh);
1660                                 unmap_underlying_metadata(bh->b_bdev,
1661                                                         bh->b_blocknr);
1662                         }
1663                 }
1664                 bh = bh->b_this_page;
1665                 block++;
1666         } while (bh != head);
1667
1668         do {
1669                 if (!buffer_mapped(bh))
1670                         continue;
1671                 /*
1672                  * If it's a fully non-blocking write attempt and we cannot
1673                  * lock the buffer then redirty the page.  Note that this can
1674                  * potentially cause a busy-wait loop from pdflush and kswapd
1675                  * activity, but those code paths have their own higher-level
1676                  * throttling.
1677                  */
1678                 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1679                         lock_buffer(bh);
1680                 } else if (!trylock_buffer(bh)) {
1681                         redirty_page_for_writepage(wbc, page);
1682                         continue;
1683                 }
1684                 if (test_clear_buffer_dirty(bh)) {
1685                         mark_buffer_async_write(bh);
1686                 } else {
1687                         unlock_buffer(bh);
1688                 }
1689         } while ((bh = bh->b_this_page) != head);
1690
1691         /*
1692          * The page and its buffers are protected by PageWriteback(), so we can
1693          * drop the bh refcounts early.
1694          */
1695         BUG_ON(PageWriteback(page));
1696         set_page_writeback(page);
1697
1698         do {
1699                 struct buffer_head *next = bh->b_this_page;
1700                 if (buffer_async_write(bh)) {
1701                         submit_bh(write_op, bh);
1702                         nr_underway++;
1703                 }
1704                 bh = next;
1705         } while (bh != head);
1706         unlock_page(page);
1707
1708         err = 0;
1709 done:
1710         if (nr_underway == 0) {
1711                 /*
1712                  * The page was marked dirty, but the buffers were
1713                  * clean.  Someone wrote them back by hand with
1714                  * ll_rw_block/submit_bh.  A rare case.
1715                  */
1716                 end_page_writeback(page);
1717
1718                 /*
1719                  * The page and buffer_heads can be released at any time from
1720                  * here on.
1721                  */
1722         }
1723         return err;
1724
1725 recover:
1726         /*
1727          * ENOSPC, or some other error.  We may already have added some
1728          * blocks to the file, so we need to write these out to avoid
1729          * exposing stale data.
1730          * The page is currently locked and not marked for writeback
1731          */
1732         bh = head;
1733         /* Recovery: lock and submit the mapped buffers */
1734         do {
1735                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1736                     !buffer_delay(bh)) {
1737                         lock_buffer(bh);
1738                         mark_buffer_async_write(bh);
1739                 } else {
1740                         /*
1741                          * The buffer may have been set dirty during
1742                          * attachment to a dirty page.
1743                          */
1744                         clear_buffer_dirty(bh);
1745                 }
1746         } while ((bh = bh->b_this_page) != head);
1747         SetPageError(page);
1748         BUG_ON(PageWriteback(page));
1749         mapping_set_error(page->mapping, err);
1750         set_page_writeback(page);
1751         do {
1752                 struct buffer_head *next = bh->b_this_page;
1753                 if (buffer_async_write(bh)) {
1754                         clear_buffer_dirty(bh);
1755                         submit_bh(write_op, bh);
1756                         nr_underway++;
1757                 }
1758                 bh = next;
1759         } while (bh != head);
1760         unlock_page(page);
1761         goto done;
1762 }
1763
1764 /*
1765  * If a page has any new buffers, zero them out here, and mark them uptodate
1766  * and dirty so they'll be written out (in order to prevent uninitialised
1767  * block data from leaking). And clear the new bit.
1768  */
1769 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1770 {
1771         unsigned int block_start, block_end;
1772         struct buffer_head *head, *bh;
1773
1774         BUG_ON(!PageLocked(page));
1775         if (!page_has_buffers(page))
1776                 return;
1777
1778         bh = head = page_buffers(page);
1779         block_start = 0;
1780         do {
1781                 block_end = block_start + bh->b_size;
1782
1783                 if (buffer_new(bh)) {
1784                         if (block_end > from && block_start < to) {
1785                                 if (!PageUptodate(page)) {
1786                                         unsigned start, size;
1787
1788                                         start = max(from, block_start);
1789                                         size = min(to, block_end) - start;
1790
1791                                         zero_user(page, start, size);
1792                                         set_buffer_uptodate(bh);
1793                                 }
1794
1795                                 clear_buffer_new(bh);
1796                                 mark_buffer_dirty(bh);
1797                         }
1798                 }
1799
1800                 block_start = block_end;
1801                 bh = bh->b_this_page;
1802         } while (bh != head);
1803 }
1804 EXPORT_SYMBOL(page_zero_new_buffers);
1805
1806 static int __block_prepare_write(struct inode *inode, struct page *page,
1807                 unsigned from, unsigned to, get_block_t *get_block)
1808 {
1809         unsigned block_start, block_end;
1810         sector_t block;
1811         int err = 0;
1812         unsigned blocksize, bbits;
1813         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1814
1815         BUG_ON(!PageLocked(page));
1816         BUG_ON(from > PAGE_CACHE_SIZE);
1817         BUG_ON(to > PAGE_CACHE_SIZE);
1818         BUG_ON(from > to);
1819
1820         blocksize = 1 << inode->i_blkbits;
1821         if (!page_has_buffers(page))
1822                 create_empty_buffers(page, blocksize, 0);
1823         head = page_buffers(page);
1824
1825         bbits = inode->i_blkbits;
1826         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1827
1828         for(bh = head, block_start = 0; bh != head || !block_start;
1829             block++, block_start=block_end, bh = bh->b_this_page) {
1830                 block_end = block_start + blocksize;
1831                 if (block_end <= from || block_start >= to) {
1832                         if (PageUptodate(page)) {
1833                                 if (!buffer_uptodate(bh))
1834                                         set_buffer_uptodate(bh);
1835                         }
1836                         continue;
1837                 }
1838                 if (buffer_new(bh))
1839                         clear_buffer_new(bh);
1840                 if (!buffer_mapped(bh)) {
1841                         WARN_ON(bh->b_size != blocksize);
1842                         err = get_block(inode, block, bh, 1);
1843                         if (err)
1844                                 break;
1845                         if (buffer_new(bh)) {
1846                                 unmap_underlying_metadata(bh->b_bdev,
1847                                                         bh->b_blocknr);
1848                                 if (PageUptodate(page)) {
1849                                         clear_buffer_new(bh);
1850                                         set_buffer_uptodate(bh);
1851                                         mark_buffer_dirty(bh);
1852                                         continue;
1853                                 }
1854                                 if (block_end > to || block_start < from)
1855                                         zero_user_segments(page,
1856                                                 to, block_end,
1857                                                 block_start, from);
1858                                 continue;
1859                         }
1860                 }
1861                 if (PageUptodate(page)) {
1862                         if (!buffer_uptodate(bh))
1863                                 set_buffer_uptodate(bh);
1864                         continue; 
1865                 }
1866                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1867                     !buffer_unwritten(bh) &&
1868                      (block_start < from || block_end > to)) {
1869                         ll_rw_block(READ, 1, &bh);
1870                         *wait_bh++=bh;
1871                 }
1872         }
1873         /*
1874          * If we issued read requests - let them complete.
1875          */
1876         while(wait_bh > wait) {
1877                 wait_on_buffer(*--wait_bh);
1878                 if (!buffer_uptodate(*wait_bh))
1879                         err = -EIO;
1880         }
1881         if (unlikely(err))
1882                 page_zero_new_buffers(page, from, to);
1883         return err;
1884 }
1885
1886 static int __block_commit_write(struct inode *inode, struct page *page,
1887                 unsigned from, unsigned to)
1888 {
1889         unsigned block_start, block_end;
1890         int partial = 0;
1891         unsigned blocksize;
1892         struct buffer_head *bh, *head;
1893
1894         blocksize = 1 << inode->i_blkbits;
1895
1896         for(bh = head = page_buffers(page), block_start = 0;
1897             bh != head || !block_start;
1898             block_start=block_end, bh = bh->b_this_page) {
1899                 block_end = block_start + blocksize;
1900                 if (block_end <= from || block_start >= to) {
1901                         if (!buffer_uptodate(bh))
1902                                 partial = 1;
1903                 } else {
1904                         set_buffer_uptodate(bh);
1905                         mark_buffer_dirty(bh);
1906                 }
1907                 clear_buffer_new(bh);
1908         }
1909
1910         /*
1911          * If this is a partial write which happened to make all buffers
1912          * uptodate then we can optimize away a bogus readpage() for
1913          * the next read(). Here we 'discover' whether the page went
1914          * uptodate as a result of this (potentially partial) write.
1915          */
1916         if (!partial)
1917                 SetPageUptodate(page);
1918         return 0;
1919 }
1920
1921 /*
1922  * block_write_begin takes care of the basic task of block allocation and
1923  * bringing partial write blocks uptodate first.
1924  *
1925  * If *pagep is not NULL, then block_write_begin uses the locked page
1926  * at *pagep rather than allocating its own. In this case, the page will
1927  * not be unlocked or deallocated on failure.
1928  */
1929 int block_write_begin(struct file *file, struct address_space *mapping,
1930                         loff_t pos, unsigned len, unsigned flags,
1931                         struct page **pagep, void **fsdata,
1932                         get_block_t *get_block)
1933 {
1934         struct inode *inode = mapping->host;
1935         int status = 0;
1936         struct page *page;
1937         pgoff_t index;
1938         unsigned start, end;
1939         int ownpage = 0;
1940
1941         index = pos >> PAGE_CACHE_SHIFT;
1942         start = pos & (PAGE_CACHE_SIZE - 1);
1943         end = start + len;
1944
1945         page = *pagep;
1946         if (page == NULL) {
1947                 ownpage = 1;
1948                 page = grab_cache_page_write_begin(mapping, index, flags);
1949                 if (!page) {
1950                         status = -ENOMEM;
1951                         goto out;
1952                 }
1953                 *pagep = page;
1954         } else
1955                 BUG_ON(!PageLocked(page));
1956
1957         status = __block_prepare_write(inode, page, start, end, get_block);
1958         if (unlikely(status)) {
1959                 ClearPageUptodate(page);
1960
1961                 if (ownpage) {
1962                         unlock_page(page);
1963                         page_cache_release(page);
1964                         *pagep = NULL;
1965
1966                         /*
1967                          * prepare_write() may have instantiated a few blocks
1968                          * outside i_size.  Trim these off again. Don't need
1969                          * i_size_read because we hold i_mutex.
1970                          */
1971                         if (pos + len > inode->i_size)
1972                                 vmtruncate(inode, inode->i_size);
1973                 }
1974         }
1975
1976 out:
1977         return status;
1978 }
1979 EXPORT_SYMBOL(block_write_begin);
1980
1981 int block_write_end(struct file *file, struct address_space *mapping,
1982                         loff_t pos, unsigned len, unsigned copied,
1983                         struct page *page, void *fsdata)
1984 {
1985         struct inode *inode = mapping->host;
1986         unsigned start;
1987
1988         start = pos & (PAGE_CACHE_SIZE - 1);
1989
1990         if (unlikely(copied < len)) {
1991                 /*
1992                  * The buffers that were written will now be uptodate, so we
1993                  * don't have to worry about a readpage reading them and
1994                  * overwriting a partial write. However if we have encountered
1995                  * a short write and only partially written into a buffer, it
1996                  * will not be marked uptodate, so a readpage might come in and
1997                  * destroy our partial write.
1998                  *
1999                  * Do the simplest thing, and just treat any short write to a
2000                  * non uptodate page as a zero-length write, and force the
2001                  * caller to redo the whole thing.
2002                  */
2003                 if (!PageUptodate(page))
2004                         copied = 0;
2005
2006                 page_zero_new_buffers(page, start+copied, start+len);
2007         }
2008         flush_dcache_page(page);
2009
2010         /* This could be a short (even 0-length) commit */
2011         __block_commit_write(inode, page, start, start+copied);
2012
2013         return copied;
2014 }
2015 EXPORT_SYMBOL(block_write_end);
2016
2017 int generic_write_end(struct file *file, struct address_space *mapping,
2018                         loff_t pos, unsigned len, unsigned copied,
2019                         struct page *page, void *fsdata)
2020 {
2021         struct inode *inode = mapping->host;
2022         int i_size_changed = 0;
2023
2024         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2025
2026         /*
2027          * No need to use i_size_read() here, the i_size
2028          * cannot change under us because we hold i_mutex.
2029          *
2030          * But it's important to update i_size while still holding page lock:
2031          * page writeout could otherwise come in and zero beyond i_size.
2032          */
2033         if (pos+copied > inode->i_size) {
2034                 i_size_write(inode, pos+copied);
2035                 i_size_changed = 1;
2036         }
2037
2038         unlock_page(page);
2039         page_cache_release(page);
2040
2041         /*
2042          * Don't mark the inode dirty under page lock. First, it unnecessarily
2043          * makes the holding time of page lock longer. Second, it forces lock
2044          * ordering of page lock and transaction start for journaling
2045          * filesystems.
2046          */
2047         if (i_size_changed)
2048                 mark_inode_dirty(inode);
2049
2050         return copied;
2051 }
2052 EXPORT_SYMBOL(generic_write_end);
2053
2054 /*
2055  * block_is_partially_uptodate checks whether buffers within a page are
2056  * uptodate or not.
2057  *
2058  * Returns true if all buffers which correspond to a file portion
2059  * we want to read are uptodate.
2060  */
2061 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2062                                         unsigned long from)
2063 {
2064         struct inode *inode = page->mapping->host;
2065         unsigned block_start, block_end, blocksize;
2066         unsigned to;
2067         struct buffer_head *bh, *head;
2068         int ret = 1;
2069
2070         if (!page_has_buffers(page))
2071                 return 0;
2072
2073         blocksize = 1 << inode->i_blkbits;
2074         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2075         to = from + to;
2076         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2077                 return 0;
2078
2079         head = page_buffers(page);
2080         bh = head;
2081         block_start = 0;
2082         do {
2083                 block_end = block_start + blocksize;
2084                 if (block_end > from && block_start < to) {
2085                         if (!buffer_uptodate(bh)) {
2086                                 ret = 0;
2087                                 break;
2088                         }
2089                         if (block_end >= to)
2090                                 break;
2091                 }
2092                 block_start = block_end;
2093                 bh = bh->b_this_page;
2094         } while (bh != head);
2095
2096         return ret;
2097 }
2098 EXPORT_SYMBOL(block_is_partially_uptodate);
2099
2100 /*
2101  * Generic "read page" function for block devices that have the normal
2102  * get_block functionality. This is most of the block device filesystems.
2103  * Reads the page asynchronously --- the unlock_buffer() and
2104  * set/clear_buffer_uptodate() functions propagate buffer state into the
2105  * page struct once IO has completed.
2106  */
2107 int block_read_full_page(struct page *page, get_block_t *get_block)
2108 {
2109         struct inode *inode = page->mapping->host;
2110         sector_t iblock, lblock;
2111         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2112         unsigned int blocksize;
2113         int nr, i;
2114         int fully_mapped = 1;
2115
2116         BUG_ON(!PageLocked(page));
2117         blocksize = 1 << inode->i_blkbits;
2118         if (!page_has_buffers(page))
2119                 create_empty_buffers(page, blocksize, 0);
2120         head = page_buffers(page);
2121
2122         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2123         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2124         bh = head;
2125         nr = 0;
2126         i = 0;
2127
2128         do {
2129                 if (buffer_uptodate(bh))
2130                         continue;
2131
2132                 if (!buffer_mapped(bh)) {
2133                         int err = 0;
2134
2135                         fully_mapped = 0;
2136                         if (iblock < lblock) {
2137                                 WARN_ON(bh->b_size != blocksize);
2138                                 err = get_block(inode, iblock, bh, 0);
2139                                 if (err)
2140                                         SetPageError(page);
2141                         }
2142                         if (!buffer_mapped(bh)) {
2143                                 zero_user(page, i * blocksize, blocksize);
2144                                 if (!err)
2145                                         set_buffer_uptodate(bh);
2146                                 continue;
2147                         }
2148                         /*
2149                          * get_block() might have updated the buffer
2150                          * synchronously
2151                          */
2152                         if (buffer_uptodate(bh))
2153                                 continue;
2154                 }
2155                 arr[nr++] = bh;
2156         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2157
2158         if (fully_mapped)
2159                 SetPageMappedToDisk(page);
2160
2161         if (!nr) {
2162                 /*
2163                  * All buffers are uptodate - we can set the page uptodate
2164                  * as well. But not if get_block() returned an error.
2165                  */
2166                 if (!PageError(page))
2167                         SetPageUptodate(page);
2168                 unlock_page(page);
2169                 return 0;
2170         }
2171
2172         /* Stage two: lock the buffers */
2173         for (i = 0; i < nr; i++) {
2174                 bh = arr[i];
2175                 lock_buffer(bh);
2176                 mark_buffer_async_read(bh);
2177         }
2178
2179         /*
2180          * Stage 3: start the IO.  Check for uptodateness
2181          * inside the buffer lock in case another process reading
2182          * the underlying blockdev brought it uptodate (the sct fix).
2183          */
2184         for (i = 0; i < nr; i++) {
2185                 bh = arr[i];
2186                 if (buffer_uptodate(bh))
2187                         end_buffer_async_read(bh, 1);
2188                 else
2189                         submit_bh(READ, bh);
2190         }
2191         return 0;
2192 }
2193
2194 /* utility function for filesystems that need to do work on expanding
2195  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2196  * deal with the hole.  
2197  */
2198 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2199 {
2200         struct address_space *mapping = inode->i_mapping;
2201         struct page *page;
2202         void *fsdata;
2203         unsigned long limit;
2204         int err;
2205
2206         err = -EFBIG;
2207         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2208         if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2209                 send_sig(SIGXFSZ, current, 0);
2210                 goto out;
2211         }
2212         if (size > inode->i_sb->s_maxbytes)
2213                 goto out;
2214
2215         err = pagecache_write_begin(NULL, mapping, size, 0,
2216                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2217                                 &page, &fsdata);
2218         if (err)
2219                 goto out;
2220
2221         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2222         BUG_ON(err > 0);
2223
2224 out:
2225         return err;
2226 }
2227
2228 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2229                             loff_t pos, loff_t *bytes)
2230 {
2231         struct inode *inode = mapping->host;
2232         unsigned blocksize = 1 << inode->i_blkbits;
2233         struct page *page;
2234         void *fsdata;
2235         pgoff_t index, curidx;
2236         loff_t curpos;
2237         unsigned zerofrom, offset, len;
2238         int err = 0;
2239
2240         index = pos >> PAGE_CACHE_SHIFT;
2241         offset = pos & ~PAGE_CACHE_MASK;
2242
2243         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2244                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2245                 if (zerofrom & (blocksize-1)) {
2246                         *bytes |= (blocksize-1);
2247                         (*bytes)++;
2248                 }
2249                 len = PAGE_CACHE_SIZE - zerofrom;
2250
2251                 err = pagecache_write_begin(file, mapping, curpos, len,
2252                                                 AOP_FLAG_UNINTERRUPTIBLE,
2253                                                 &page, &fsdata);
2254                 if (err)
2255                         goto out;
2256                 zero_user(page, zerofrom, len);
2257                 err = pagecache_write_end(file, mapping, curpos, len, len,
2258                                                 page, fsdata);
2259                 if (err < 0)
2260                         goto out;
2261                 BUG_ON(err != len);
2262                 err = 0;
2263
2264                 balance_dirty_pages_ratelimited(mapping);
2265         }
2266
2267         /* page covers the boundary, find the boundary offset */
2268         if (index == curidx) {
2269                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2270                 /* if we will expand the thing last block will be filled */
2271                 if (offset <= zerofrom) {
2272                         goto out;
2273                 }
2274                 if (zerofrom & (blocksize-1)) {
2275                         *bytes |= (blocksize-1);
2276                         (*bytes)++;
2277                 }
2278                 len = offset - zerofrom;
2279
2280                 err = pagecache_write_begin(file, mapping, curpos, len,
2281                                                 AOP_FLAG_UNINTERRUPTIBLE,
2282                                                 &page, &fsdata);
2283                 if (err)
2284                         goto out;
2285                 zero_user(page, zerofrom, len);
2286                 err = pagecache_write_end(file, mapping, curpos, len, len,
2287                                                 page, fsdata);
2288                 if (err < 0)
2289                         goto out;
2290                 BUG_ON(err != len);
2291                 err = 0;
2292         }
2293 out:
2294         return err;
2295 }
2296
2297 /*
2298  * For moronic filesystems that do not allow holes in file.
2299  * We may have to extend the file.
2300  */
2301 int cont_write_begin(struct file *file, struct address_space *mapping,
2302                         loff_t pos, unsigned len, unsigned flags,
2303                         struct page **pagep, void **fsdata,
2304                         get_block_t *get_block, loff_t *bytes)
2305 {
2306         struct inode *inode = mapping->host;
2307         unsigned blocksize = 1 << inode->i_blkbits;
2308         unsigned zerofrom;
2309         int err;
2310
2311         err = cont_expand_zero(file, mapping, pos, bytes);
2312         if (err)
2313                 goto out;
2314
2315         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2316         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2317                 *bytes |= (blocksize-1);
2318                 (*bytes)++;
2319         }
2320
2321         *pagep = NULL;
2322         err = block_write_begin(file, mapping, pos, len,
2323                                 flags, pagep, fsdata, get_block);
2324 out:
2325         return err;
2326 }
2327
2328 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2329                         get_block_t *get_block)
2330 {
2331         struct inode *inode = page->mapping->host;
2332         int err = __block_prepare_write(inode, page, from, to, get_block);
2333         if (err)
2334                 ClearPageUptodate(page);
2335         return err;
2336 }
2337
2338 int block_commit_write(struct page *page, unsigned from, unsigned to)
2339 {
2340         struct inode *inode = page->mapping->host;
2341         __block_commit_write(inode,page,from,to);
2342         return 0;
2343 }
2344
2345 /*
2346  * block_page_mkwrite() is not allowed to change the file size as it gets
2347  * called from a page fault handler when a page is first dirtied. Hence we must
2348  * be careful to check for EOF conditions here. We set the page up correctly
2349  * for a written page which means we get ENOSPC checking when writing into
2350  * holes and correct delalloc and unwritten extent mapping on filesystems that
2351  * support these features.
2352  *
2353  * We are not allowed to take the i_mutex here so we have to play games to
2354  * protect against truncate races as the page could now be beyond EOF.  Because
2355  * vmtruncate() writes the inode size before removing pages, once we have the
2356  * page lock we can determine safely if the page is beyond EOF. If it is not
2357  * beyond EOF, then the page is guaranteed safe against truncation until we
2358  * unlock the page.
2359  */
2360 int
2361 block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2362                    get_block_t get_block)
2363 {
2364         struct page *page = vmf->page;
2365         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2366         unsigned long end;
2367         loff_t size;
2368         int ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
2369
2370         lock_page(page);
2371         size = i_size_read(inode);
2372         if ((page->mapping != inode->i_mapping) ||
2373             (page_offset(page) > size)) {
2374                 /* page got truncated out from underneath us */
2375                 goto out_unlock;
2376         }
2377
2378         /* page is wholly or partially inside EOF */
2379         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2380                 end = size & ~PAGE_CACHE_MASK;
2381         else
2382                 end = PAGE_CACHE_SIZE;
2383
2384         ret = block_prepare_write(page, 0, end, get_block);
2385         if (!ret)
2386                 ret = block_commit_write(page, 0, end);
2387
2388         if (unlikely(ret)) {
2389                 if (ret == -ENOMEM)
2390                         ret = VM_FAULT_OOM;
2391                 else /* -ENOSPC, -EIO, etc */
2392                         ret = VM_FAULT_SIGBUS;
2393         }
2394
2395 out_unlock:
2396         unlock_page(page);
2397         return ret;
2398 }
2399
2400 /*
2401  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2402  * immediately, while under the page lock.  So it needs a special end_io
2403  * handler which does not touch the bh after unlocking it.
2404  */
2405 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2406 {
2407         __end_buffer_read_notouch(bh, uptodate);
2408 }
2409
2410 /*
2411  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2412  * the page (converting it to circular linked list and taking care of page
2413  * dirty races).
2414  */
2415 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2416 {
2417         struct buffer_head *bh;
2418
2419         BUG_ON(!PageLocked(page));
2420
2421         spin_lock(&page->mapping->private_lock);
2422         bh = head;
2423         do {
2424                 if (PageDirty(page))
2425                         set_buffer_dirty(bh);
2426                 if (!bh->b_this_page)
2427                         bh->b_this_page = head;
2428                 bh = bh->b_this_page;
2429         } while (bh != head);
2430         attach_page_buffers(page, head);
2431         spin_unlock(&page->mapping->private_lock);
2432 }
2433
2434 /*
2435  * On entry, the page is fully not uptodate.
2436  * On exit the page is fully uptodate in the areas outside (from,to)
2437  */
2438 int nobh_write_begin(struct file *file, struct address_space *mapping,
2439                         loff_t pos, unsigned len, unsigned flags,
2440                         struct page **pagep, void **fsdata,
2441                         get_block_t *get_block)
2442 {
2443         struct inode *inode = mapping->host;
2444         const unsigned blkbits = inode->i_blkbits;
2445         const unsigned blocksize = 1 << blkbits;
2446         struct buffer_head *head, *bh;
2447         struct page *page;
2448         pgoff_t index;
2449         unsigned from, to;
2450         unsigned block_in_page;
2451         unsigned block_start, block_end;
2452         sector_t block_in_file;
2453         int nr_reads = 0;
2454         int ret = 0;
2455         int is_mapped_to_disk = 1;
2456
2457         index = pos >> PAGE_CACHE_SHIFT;
2458         from = pos & (PAGE_CACHE_SIZE - 1);
2459         to = from + len;
2460
2461         page = grab_cache_page_write_begin(mapping, index, flags);
2462         if (!page)
2463                 return -ENOMEM;
2464         *pagep = page;
2465         *fsdata = NULL;
2466
2467         if (page_has_buffers(page)) {
2468                 unlock_page(page);
2469                 page_cache_release(page);
2470                 *pagep = NULL;
2471                 return block_write_begin(file, mapping, pos, len, flags, pagep,
2472                                         fsdata, get_block);
2473         }
2474
2475         if (PageMappedToDisk(page))
2476                 return 0;
2477
2478         /*
2479          * Allocate buffers so that we can keep track of state, and potentially
2480          * attach them to the page if an error occurs. In the common case of
2481          * no error, they will just be freed again without ever being attached
2482          * to the page (which is all OK, because we're under the page lock).
2483          *
2484          * Be careful: the buffer linked list is a NULL terminated one, rather
2485          * than the circular one we're used to.
2486          */
2487         head = alloc_page_buffers(page, blocksize, 0);
2488         if (!head) {
2489                 ret = -ENOMEM;
2490                 goto out_release;
2491         }
2492
2493         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2494
2495         /*
2496          * We loop across all blocks in the page, whether or not they are
2497          * part of the affected region.  This is so we can discover if the
2498          * page is fully mapped-to-disk.
2499          */
2500         for (block_start = 0, block_in_page = 0, bh = head;
2501                   block_start < PAGE_CACHE_SIZE;
2502                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2503                 int create;
2504
2505                 block_end = block_start + blocksize;
2506                 bh->b_state = 0;
2507                 create = 1;
2508                 if (block_start >= to)
2509                         create = 0;
2510                 ret = get_block(inode, block_in_file + block_in_page,
2511                                         bh, create);
2512                 if (ret)
2513                         goto failed;
2514                 if (!buffer_mapped(bh))
2515                         is_mapped_to_disk = 0;
2516                 if (buffer_new(bh))
2517                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2518                 if (PageUptodate(page)) {
2519                         set_buffer_uptodate(bh);
2520                         continue;
2521                 }
2522                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2523                         zero_user_segments(page, block_start, from,
2524                                                         to, block_end);
2525                         continue;
2526                 }
2527                 if (buffer_uptodate(bh))
2528                         continue;       /* reiserfs does this */
2529                 if (block_start < from || block_end > to) {
2530                         lock_buffer(bh);
2531                         bh->b_end_io = end_buffer_read_nobh;
2532                         submit_bh(READ, bh);
2533                         nr_reads++;
2534                 }
2535         }
2536
2537         if (nr_reads) {
2538                 /*
2539                  * The page is locked, so these buffers are protected from
2540                  * any VM or truncate activity.  Hence we don't need to care
2541                  * for the buffer_head refcounts.
2542                  */
2543                 for (bh = head; bh; bh = bh->b_this_page) {
2544                         wait_on_buffer(bh);
2545                         if (!buffer_uptodate(bh))
2546                                 ret = -EIO;
2547                 }
2548                 if (ret)
2549                         goto failed;
2550         }
2551
2552         if (is_mapped_to_disk)
2553                 SetPageMappedToDisk(page);
2554
2555         *fsdata = head; /* to be released by nobh_write_end */
2556
2557         return 0;
2558
2559 failed:
2560         BUG_ON(!ret);
2561         /*
2562          * Error recovery is a bit difficult. We need to zero out blocks that
2563          * were newly allocated, and dirty them to ensure they get written out.
2564          * Buffers need to be attached to the page at this point, otherwise
2565          * the handling of potential IO errors during writeout would be hard
2566          * (could try doing synchronous writeout, but what if that fails too?)
2567          */
2568         attach_nobh_buffers(page, head);
2569         page_zero_new_buffers(page, from, to);
2570
2571 out_release:
2572         unlock_page(page);
2573         page_cache_release(page);
2574         *pagep = NULL;
2575
2576         if (pos + len > inode->i_size)
2577                 vmtruncate(inode, inode->i_size);
2578
2579         return ret;
2580 }
2581 EXPORT_SYMBOL(nobh_write_begin);
2582
2583 int nobh_write_end(struct file *file, struct address_space *mapping,
2584                         loff_t pos, unsigned len, unsigned copied,
2585                         struct page *page, void *fsdata)
2586 {
2587         struct inode *inode = page->mapping->host;
2588         struct buffer_head *head = fsdata;
2589         struct buffer_head *bh;
2590         BUG_ON(fsdata != NULL && page_has_buffers(page));
2591
2592         if (unlikely(copied < len) && head)
2593                 attach_nobh_buffers(page, head);
2594         if (page_has_buffers(page))
2595                 return generic_write_end(file, mapping, pos, len,
2596                                         copied, page, fsdata);
2597
2598         SetPageUptodate(page);
2599         set_page_dirty(page);
2600         if (pos+copied > inode->i_size) {
2601                 i_size_write(inode, pos+copied);
2602                 mark_inode_dirty(inode);
2603         }
2604
2605         unlock_page(page);
2606         page_cache_release(page);
2607
2608         while (head) {
2609                 bh = head;
2610                 head = head->b_this_page;
2611                 free_buffer_head(bh);
2612         }
2613
2614         return copied;
2615 }
2616 EXPORT_SYMBOL(nobh_write_end);
2617
2618 /*
2619  * nobh_writepage() - based on block_full_write_page() except
2620  * that it tries to operate without attaching bufferheads to
2621  * the page.
2622  */
2623 int nobh_writepage(struct page *page, get_block_t *get_block,
2624                         struct writeback_control *wbc)
2625 {
2626         struct inode * const inode = page->mapping->host;
2627         loff_t i_size = i_size_read(inode);
2628         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2629         unsigned offset;
2630         int ret;
2631
2632         /* Is the page fully inside i_size? */
2633         if (page->index < end_index)
2634                 goto out;
2635
2636         /* Is the page fully outside i_size? (truncate in progress) */
2637         offset = i_size & (PAGE_CACHE_SIZE-1);
2638         if (page->index >= end_index+1 || !offset) {
2639                 /*
2640                  * The page may have dirty, unmapped buffers.  For example,
2641                  * they may have been added in ext3_writepage().  Make them
2642                  * freeable here, so the page does not leak.
2643                  */
2644 #if 0
2645                 /* Not really sure about this  - do we need this ? */
2646                 if (page->mapping->a_ops->invalidatepage)
2647                         page->mapping->a_ops->invalidatepage(page, offset);
2648 #endif
2649                 unlock_page(page);
2650                 return 0; /* don't care */
2651         }
2652
2653         /*
2654          * The page straddles i_size.  It must be zeroed out on each and every
2655          * writepage invocation because it may be mmapped.  "A file is mapped
2656          * in multiples of the page size.  For a file that is not a multiple of
2657          * the  page size, the remaining memory is zeroed when mapped, and
2658          * writes to that region are not written out to the file."
2659          */
2660         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2661 out:
2662         ret = mpage_writepage(page, get_block, wbc);
2663         if (ret == -EAGAIN)
2664                 ret = __block_write_full_page(inode, page, get_block, wbc);
2665         return ret;
2666 }
2667 EXPORT_SYMBOL(nobh_writepage);
2668
2669 int nobh_truncate_page(struct address_space *mapping,
2670                         loff_t from, get_block_t *get_block)
2671 {
2672         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2673         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2674         unsigned blocksize;
2675         sector_t iblock;
2676         unsigned length, pos;
2677         struct inode *inode = mapping->host;
2678         struct page *page;
2679         struct buffer_head map_bh;
2680         int err;
2681
2682         blocksize = 1 << inode->i_blkbits;
2683         length = offset & (blocksize - 1);
2684
2685         /* Block boundary? Nothing to do */
2686         if (!length)
2687                 return 0;
2688
2689         length = blocksize - length;
2690         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2691
2692         page = grab_cache_page(mapping, index);
2693         err = -ENOMEM;
2694         if (!page)
2695                 goto out;
2696
2697         if (page_has_buffers(page)) {
2698 has_buffers:
2699                 unlock_page(page);
2700                 page_cache_release(page);
2701                 return block_truncate_page(mapping, from, get_block);
2702         }
2703
2704         /* Find the buffer that contains "offset" */
2705         pos = blocksize;
2706         while (offset >= pos) {
2707                 iblock++;
2708                 pos += blocksize;
2709         }
2710
2711         err = get_block(inode, iblock, &map_bh, 0);
2712         if (err)
2713                 goto unlock;
2714         /* unmapped? It's a hole - nothing to do */
2715         if (!buffer_mapped(&map_bh))
2716                 goto unlock;
2717
2718         /* Ok, it's mapped. Make sure it's up-to-date */
2719         if (!PageUptodate(page)) {
2720                 err = mapping->a_ops->readpage(NULL, page);
2721                 if (err) {
2722                         page_cache_release(page);
2723                         goto out;
2724                 }
2725                 lock_page(page);
2726                 if (!PageUptodate(page)) {
2727                         err = -EIO;
2728                         goto unlock;
2729                 }
2730                 if (page_has_buffers(page))
2731                         goto has_buffers;
2732         }
2733         zero_user(page, offset, length);
2734         set_page_dirty(page);
2735         err = 0;
2736
2737 unlock:
2738         unlock_page(page);
2739         page_cache_release(page);
2740 out:
2741         return err;
2742 }
2743 EXPORT_SYMBOL(nobh_truncate_page);
2744
2745 int block_truncate_page(struct address_space *mapping,
2746                         loff_t from, get_block_t *get_block)
2747 {
2748         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2749         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2750         unsigned blocksize;
2751         sector_t iblock;
2752         unsigned length, pos;
2753         struct inode *inode = mapping->host;
2754         struct page *page;
2755         struct buffer_head *bh;
2756         int err;
2757
2758         blocksize = 1 << inode->i_blkbits;
2759         length = offset & (blocksize - 1);
2760
2761         /* Block boundary? Nothing to do */
2762         if (!length)
2763                 return 0;
2764
2765         length = blocksize - length;
2766         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2767         
2768         page = grab_cache_page(mapping, index);
2769         err = -ENOMEM;
2770         if (!page)
2771                 goto out;
2772
2773         if (!page_has_buffers(page))
2774                 create_empty_buffers(page, blocksize, 0);
2775
2776         /* Find the buffer that contains "offset" */
2777         bh = page_buffers(page);
2778         pos = blocksize;
2779         while (offset >= pos) {
2780                 bh = bh->b_this_page;
2781                 iblock++;
2782                 pos += blocksize;
2783         }
2784
2785         err = 0;
2786         if (!buffer_mapped(bh)) {
2787                 WARN_ON(bh->b_size != blocksize);
2788                 err = get_block(inode, iblock, bh, 0);
2789                 if (err)
2790                         goto unlock;
2791                 /* unmapped? It's a hole - nothing to do */
2792                 if (!buffer_mapped(bh))
2793                         goto unlock;
2794         }
2795
2796         /* Ok, it's mapped. Make sure it's up-to-date */
2797         if (PageUptodate(page))
2798                 set_buffer_uptodate(bh);
2799
2800         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2801                 err = -EIO;
2802                 ll_rw_block(READ, 1, &bh);
2803                 wait_on_buffer(bh);
2804                 /* Uhhuh. Read error. Complain and punt. */
2805                 if (!buffer_uptodate(bh))
2806                         goto unlock;
2807         }
2808
2809         zero_user(page, offset, length);
2810         mark_buffer_dirty(bh);
2811         err = 0;
2812
2813 unlock:
2814         unlock_page(page);
2815         page_cache_release(page);
2816 out:
2817         return err;
2818 }
2819
2820 /*
2821  * The generic ->writepage function for buffer-backed address_spaces
2822  */
2823 int block_write_full_page(struct page *page, get_block_t *get_block,
2824                         struct writeback_control *wbc)
2825 {
2826         struct inode * const inode = page->mapping->host;
2827         loff_t i_size = i_size_read(inode);
2828         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2829         unsigned offset;
2830
2831         /* Is the page fully inside i_size? */
2832         if (page->index < end_index)
2833                 return __block_write_full_page(inode, page, get_block, wbc);
2834
2835         /* Is the page fully outside i_size? (truncate in progress) */
2836         offset = i_size & (PAGE_CACHE_SIZE-1);
2837         if (page->index >= end_index+1 || !offset) {
2838                 /*
2839                  * The page may have dirty, unmapped buffers.  For example,
2840                  * they may have been added in ext3_writepage().  Make them
2841                  * freeable here, so the page does not leak.
2842                  */
2843                 do_invalidatepage(page, 0);
2844                 unlock_page(page);
2845                 return 0; /* don't care */
2846         }
2847
2848         /*
2849          * The page straddles i_size.  It must be zeroed out on each and every
2850          * writepage invokation because it may be mmapped.  "A file is mapped
2851          * in multiples of the page size.  For a file that is not a multiple of
2852          * the  page size, the remaining memory is zeroed when mapped, and
2853          * writes to that region are not written out to the file."
2854          */
2855         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2856         return __block_write_full_page(inode, page, get_block, wbc);
2857 }
2858
2859 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2860                             get_block_t *get_block)
2861 {
2862         struct buffer_head tmp;
2863         struct inode *inode = mapping->host;
2864         tmp.b_state = 0;
2865         tmp.b_blocknr = 0;
2866         tmp.b_size = 1 << inode->i_blkbits;
2867         get_block(inode, block, &tmp, 0);
2868         return tmp.b_blocknr;
2869 }
2870
2871 static void end_bio_bh_io_sync(struct bio *bio, int err)
2872 {
2873         struct buffer_head *bh = bio->bi_private;
2874
2875         if (err == -EOPNOTSUPP) {
2876                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2877                 set_bit(BH_Eopnotsupp, &bh->b_state);
2878         }
2879
2880         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2881                 set_bit(BH_Quiet, &bh->b_state);
2882
2883         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2884         bio_put(bio);
2885 }
2886
2887 int submit_bh(int rw, struct buffer_head * bh)
2888 {
2889         struct bio *bio;
2890         int ret = 0;
2891
2892         BUG_ON(!buffer_locked(bh));
2893         BUG_ON(!buffer_mapped(bh));
2894         BUG_ON(!bh->b_end_io);
2895
2896         /*
2897          * Mask in barrier bit for a write (could be either a WRITE or a
2898          * WRITE_SYNC
2899          */
2900         if (buffer_ordered(bh) && (rw & WRITE))
2901                 rw |= WRITE_BARRIER;
2902
2903         /*
2904          * Only clear out a write error when rewriting
2905          */
2906         if (test_set_buffer_req(bh) && (rw & WRITE))
2907                 clear_buffer_write_io_error(bh);
2908
2909         /*
2910          * from here on down, it's all bio -- do the initial mapping,
2911          * submit_bio -> generic_make_request may further map this bio around
2912          */
2913         bio = bio_alloc(GFP_NOIO, 1);
2914
2915         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2916         bio->bi_bdev = bh->b_bdev;
2917         bio->bi_io_vec[0].bv_page = bh->b_page;
2918         bio->bi_io_vec[0].bv_len = bh->b_size;
2919         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2920
2921         bio->bi_vcnt = 1;
2922         bio->bi_idx = 0;
2923         bio->bi_size = bh->b_size;
2924
2925         bio->bi_end_io = end_bio_bh_io_sync;
2926         bio->bi_private = bh;
2927
2928         bio_get(bio);
2929         submit_bio(rw, bio);
2930
2931         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2932                 ret = -EOPNOTSUPP;
2933
2934         bio_put(bio);
2935         return ret;
2936 }
2937
2938 /**
2939  * ll_rw_block: low-level access to block devices (DEPRECATED)
2940  * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
2941  * @nr: number of &struct buffer_heads in the array
2942  * @bhs: array of pointers to &struct buffer_head
2943  *
2944  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2945  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2946  * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2947  * are sent to disk. The fourth %READA option is described in the documentation
2948  * for generic_make_request() which ll_rw_block() calls.
2949  *
2950  * This function drops any buffer that it cannot get a lock on (with the
2951  * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
2952  * clean when doing a write request, and any buffer that appears to be
2953  * up-to-date when doing read request.  Further it marks as clean buffers that
2954  * are processed for writing (the buffer cache won't assume that they are
2955  * actually clean until the buffer gets unlocked).
2956  *
2957  * ll_rw_block sets b_end_io to simple completion handler that marks
2958  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2959  * any waiters. 
2960  *
2961  * All of the buffers must be for the same device, and must also be a
2962  * multiple of the current approved size for the device.
2963  */
2964 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2965 {
2966         int i;
2967
2968         for (i = 0; i < nr; i++) {
2969                 struct buffer_head *bh = bhs[i];
2970
2971                 if (rw == SWRITE || rw == SWRITE_SYNC || rw == SWRITE_SYNC_PLUG)
2972                         lock_buffer(bh);
2973                 else if (!trylock_buffer(bh))
2974                         continue;
2975
2976                 if (rw == WRITE || rw == SWRITE || rw == SWRITE_SYNC ||
2977                     rw == SWRITE_SYNC_PLUG) {
2978                         if (test_clear_buffer_dirty(bh)) {
2979                                 bh->b_end_io = end_buffer_write_sync;
2980                                 get_bh(bh);
2981                                 if (rw == SWRITE_SYNC)
2982                                         submit_bh(WRITE_SYNC, bh);
2983                                 else
2984                                         submit_bh(WRITE, bh);
2985                                 continue;
2986                         }
2987                 } else {
2988                         if (!buffer_uptodate(bh)) {
2989                                 bh->b_end_io = end_buffer_read_sync;
2990                                 get_bh(bh);
2991                                 submit_bh(rw, bh);
2992                                 continue;
2993                         }
2994                 }
2995                 unlock_buffer(bh);
2996         }
2997 }
2998
2999 /*
3000  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3001  * and then start new I/O and then wait upon it.  The caller must have a ref on
3002  * the buffer_head.
3003  */
3004 int sync_dirty_buffer(struct buffer_head *bh)
3005 {
3006         int ret = 0;
3007
3008         WARN_ON(atomic_read(&bh->b_count) < 1);
3009         lock_buffer(bh);
3010         if (test_clear_buffer_dirty(bh)) {
3011                 get_bh(bh);
3012                 bh->b_end_io = end_buffer_write_sync;
3013                 ret = submit_bh(WRITE_SYNC, bh);
3014                 wait_on_buffer(bh);
3015                 if (buffer_eopnotsupp(bh)) {
3016                         clear_buffer_eopnotsupp(bh);
3017                         ret = -EOPNOTSUPP;
3018                 }
3019                 if (!ret && !buffer_uptodate(bh))
3020                         ret = -EIO;
3021         } else {
3022                 unlock_buffer(bh);
3023         }
3024         return ret;
3025 }
3026
3027 /*
3028  * try_to_free_buffers() checks if all the buffers on this particular page
3029  * are unused, and releases them if so.
3030  *
3031  * Exclusion against try_to_free_buffers may be obtained by either
3032  * locking the page or by holding its mapping's private_lock.
3033  *
3034  * If the page is dirty but all the buffers are clean then we need to
3035  * be sure to mark the page clean as well.  This is because the page
3036  * may be against a block device, and a later reattachment of buffers
3037  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3038  * filesystem data on the same device.
3039  *
3040  * The same applies to regular filesystem pages: if all the buffers are
3041  * clean then we set the page clean and proceed.  To do that, we require
3042  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3043  * private_lock.
3044  *
3045  * try_to_free_buffers() is non-blocking.
3046  */
3047 static inline int buffer_busy(struct buffer_head *bh)
3048 {
3049         return atomic_read(&bh->b_count) |
3050                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3051 }
3052
3053 static int
3054 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3055 {
3056         struct buffer_head *head = page_buffers(page);
3057         struct buffer_head *bh;
3058
3059         bh = head;
3060         do {
3061                 if (buffer_write_io_error(bh) && page->mapping)
3062                         set_bit(AS_EIO, &page->mapping->flags);
3063                 if (buffer_busy(bh))
3064                         goto failed;
3065                 bh = bh->b_this_page;
3066         } while (bh != head);
3067
3068         do {
3069                 struct buffer_head *next = bh->b_this_page;
3070
3071                 if (bh->b_assoc_map)
3072                         __remove_assoc_queue(bh);
3073                 bh = next;
3074         } while (bh != head);
3075         *buffers_to_free = head;
3076         __clear_page_buffers(page);
3077         return 1;
3078 failed:
3079         return 0;
3080 }
3081
3082 int try_to_free_buffers(struct page *page)
3083 {
3084         struct address_space * const mapping = page->mapping;
3085         struct buffer_head *buffers_to_free = NULL;
3086         int ret = 0;
3087
3088         BUG_ON(!PageLocked(page));
3089         if (PageWriteback(page))
3090                 return 0;
3091
3092         if (mapping == NULL) {          /* can this still happen? */
3093                 ret = drop_buffers(page, &buffers_to_free);
3094                 goto out;
3095         }
3096
3097         spin_lock(&mapping->private_lock);
3098         ret = drop_buffers(page, &buffers_to_free);
3099
3100         /*
3101          * If the filesystem writes its buffers by hand (eg ext3)
3102          * then we can have clean buffers against a dirty page.  We
3103          * clean the page here; otherwise the VM will never notice
3104          * that the filesystem did any IO at all.
3105          *
3106          * Also, during truncate, discard_buffer will have marked all
3107          * the page's buffers clean.  We discover that here and clean
3108          * the page also.
3109          *
3110          * private_lock must be held over this entire operation in order
3111          * to synchronise against __set_page_dirty_buffers and prevent the
3112          * dirty bit from being lost.
3113          */
3114         if (ret)
3115                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3116         spin_unlock(&mapping->private_lock);
3117 out:
3118         if (buffers_to_free) {
3119                 struct buffer_head *bh = buffers_to_free;
3120
3121                 do {
3122                         struct buffer_head *next = bh->b_this_page;
3123                         free_buffer_head(bh);
3124                         bh = next;
3125                 } while (bh != buffers_to_free);
3126         }
3127         return ret;
3128 }
3129 EXPORT_SYMBOL(try_to_free_buffers);
3130
3131 void block_sync_page(struct page *page)
3132 {
3133         struct address_space *mapping;
3134
3135         smp_mb();
3136         mapping = page_mapping(page);
3137         if (mapping)
3138                 blk_run_backing_dev(mapping->backing_dev_info, page);
3139 }
3140
3141 /*
3142  * There are no bdflush tunables left.  But distributions are
3143  * still running obsolete flush daemons, so we terminate them here.
3144  *
3145  * Use of bdflush() is deprecated and will be removed in a future kernel.
3146  * The `pdflush' kernel threads fully replace bdflush daemons and this call.
3147  */
3148 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3149 {
3150         static int msg_count;
3151
3152         if (!capable(CAP_SYS_ADMIN))
3153                 return -EPERM;
3154
3155         if (msg_count < 5) {
3156                 msg_count++;
3157                 printk(KERN_INFO
3158                         "warning: process `%s' used the obsolete bdflush"
3159                         " system call\n", current->comm);
3160                 printk(KERN_INFO "Fix your initscripts?\n");
3161         }
3162
3163         if (func == 1)
3164                 do_exit(0);
3165         return 0;
3166 }
3167
3168 /*
3169  * Buffer-head allocation
3170  */
3171 static struct kmem_cache *bh_cachep;
3172
3173 /*
3174  * Once the number of bh's in the machine exceeds this level, we start
3175  * stripping them in writeback.
3176  */
3177 static int max_buffer_heads;
3178
3179 int buffer_heads_over_limit;
3180
3181 struct bh_accounting {
3182         int nr;                 /* Number of live bh's */
3183         int ratelimit;          /* Limit cacheline bouncing */
3184 };
3185
3186 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3187
3188 static void recalc_bh_state(void)
3189 {
3190         int i;
3191         int tot = 0;
3192
3193         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3194                 return;
3195         __get_cpu_var(bh_accounting).ratelimit = 0;
3196         for_each_online_cpu(i)
3197                 tot += per_cpu(bh_accounting, i).nr;
3198         buffer_heads_over_limit = (tot > max_buffer_heads);
3199 }
3200         
3201 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3202 {
3203         struct buffer_head *ret = kmem_cache_alloc(bh_cachep, gfp_flags);
3204         if (ret) {
3205                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3206                 get_cpu_var(bh_accounting).nr++;
3207                 recalc_bh_state();
3208                 put_cpu_var(bh_accounting);
3209         }
3210         return ret;
3211 }
3212 EXPORT_SYMBOL(alloc_buffer_head);
3213
3214 void free_buffer_head(struct buffer_head *bh)
3215 {
3216         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3217         kmem_cache_free(bh_cachep, bh);
3218         get_cpu_var(bh_accounting).nr--;
3219         recalc_bh_state();
3220         put_cpu_var(bh_accounting);
3221 }
3222 EXPORT_SYMBOL(free_buffer_head);
3223
3224 static void buffer_exit_cpu(int cpu)
3225 {
3226         int i;
3227         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3228
3229         for (i = 0; i < BH_LRU_SIZE; i++) {
3230                 brelse(b->bhs[i]);
3231                 b->bhs[i] = NULL;
3232         }
3233         get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3234         per_cpu(bh_accounting, cpu).nr = 0;
3235         put_cpu_var(bh_accounting);
3236 }
3237
3238 static int buffer_cpu_notify(struct notifier_block *self,
3239                               unsigned long action, void *hcpu)
3240 {
3241         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3242                 buffer_exit_cpu((unsigned long)hcpu);
3243         return NOTIFY_OK;
3244 }
3245
3246 /**
3247  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3248  * @bh: struct buffer_head
3249  *
3250  * Return true if the buffer is up-to-date and false,
3251  * with the buffer locked, if not.
3252  */
3253 int bh_uptodate_or_lock(struct buffer_head *bh)
3254 {
3255         if (!buffer_uptodate(bh)) {
3256                 lock_buffer(bh);
3257                 if (!buffer_uptodate(bh))
3258                         return 0;
3259                 unlock_buffer(bh);
3260         }
3261         return 1;
3262 }
3263 EXPORT_SYMBOL(bh_uptodate_or_lock);
3264
3265 /**
3266  * bh_submit_read - Submit a locked buffer for reading
3267  * @bh: struct buffer_head
3268  *
3269  * Returns zero on success and -EIO on error.
3270  */
3271 int bh_submit_read(struct buffer_head *bh)
3272 {
3273         BUG_ON(!buffer_locked(bh));
3274
3275         if (buffer_uptodate(bh)) {
3276                 unlock_buffer(bh);
3277                 return 0;
3278         }
3279
3280         get_bh(bh);
3281         bh->b_end_io = end_buffer_read_sync;
3282         submit_bh(READ, bh);
3283         wait_on_buffer(bh);
3284         if (buffer_uptodate(bh))
3285                 return 0;
3286         return -EIO;
3287 }
3288 EXPORT_SYMBOL(bh_submit_read);
3289
3290 static void
3291 init_buffer_head(void *data)
3292 {
3293         struct buffer_head *bh = data;
3294
3295         memset(bh, 0, sizeof(*bh));
3296         INIT_LIST_HEAD(&bh->b_assoc_buffers);
3297 }
3298
3299 void __init buffer_init(void)
3300 {
3301         int nrpages;
3302
3303         bh_cachep = kmem_cache_create("buffer_head",
3304                         sizeof(struct buffer_head), 0,
3305                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3306                                 SLAB_MEM_SPREAD),
3307                                 init_buffer_head);
3308
3309         /*
3310          * Limit the bh occupancy to 10% of ZONE_NORMAL
3311          */
3312         nrpages = (nr_free_buffer_pages() * 10) / 100;
3313         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3314         hotcpu_notifier(buffer_cpu_notify, 0);
3315 }
3316
3317 EXPORT_SYMBOL(__bforget);
3318 EXPORT_SYMBOL(__brelse);
3319 EXPORT_SYMBOL(__wait_on_buffer);
3320 EXPORT_SYMBOL(block_commit_write);
3321 EXPORT_SYMBOL(block_prepare_write);
3322 EXPORT_SYMBOL(block_page_mkwrite);
3323 EXPORT_SYMBOL(block_read_full_page);
3324 EXPORT_SYMBOL(block_sync_page);
3325 EXPORT_SYMBOL(block_truncate_page);
3326 EXPORT_SYMBOL(block_write_full_page);
3327 EXPORT_SYMBOL(cont_write_begin);
3328 EXPORT_SYMBOL(end_buffer_read_sync);
3329 EXPORT_SYMBOL(end_buffer_write_sync);
3330 EXPORT_SYMBOL(file_fsync);
3331 EXPORT_SYMBOL(generic_block_bmap);
3332 EXPORT_SYMBOL(generic_cont_expand_simple);
3333 EXPORT_SYMBOL(init_buffer);
3334 EXPORT_SYMBOL(invalidate_bdev);
3335 EXPORT_SYMBOL(ll_rw_block);
3336 EXPORT_SYMBOL(mark_buffer_dirty);
3337 EXPORT_SYMBOL(submit_bh);
3338 EXPORT_SYMBOL(sync_dirty_buffer);
3339 EXPORT_SYMBOL(unlock_buffer);