Merge branch 'ext3-latency-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git...
[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;
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, 1, &bh);
766                                 brelse(bh);
767                                 spin_lock(lock);
768                         }
769                 }
770         }
771
772         while (!list_empty(&tmp)) {
773                 bh = BH_ENTRY(tmp.prev);
774                 get_bh(bh);
775                 mapping = bh->b_assoc_map;
776                 __remove_assoc_queue(bh);
777                 /* Avoid race with mark_buffer_dirty_inode() which does
778                  * a lockless check and we rely on seeing the dirty bit */
779                 smp_mb();
780                 if (buffer_dirty(bh)) {
781                         list_add(&bh->b_assoc_buffers,
782                                  &mapping->private_list);
783                         bh->b_assoc_map = mapping;
784                 }
785                 spin_unlock(lock);
786                 wait_on_buffer(bh);
787                 if (!buffer_uptodate(bh))
788                         err = -EIO;
789                 brelse(bh);
790                 spin_lock(lock);
791         }
792         
793         spin_unlock(lock);
794         err2 = osync_buffers_list(lock, list);
795         if (err)
796                 return err;
797         else
798                 return err2;
799 }
800
801 /*
802  * Invalidate any and all dirty buffers on a given inode.  We are
803  * probably unmounting the fs, but that doesn't mean we have already
804  * done a sync().  Just drop the buffers from the inode list.
805  *
806  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
807  * assumes that all the buffers are against the blockdev.  Not true
808  * for reiserfs.
809  */
810 void invalidate_inode_buffers(struct inode *inode)
811 {
812         if (inode_has_buffers(inode)) {
813                 struct address_space *mapping = &inode->i_data;
814                 struct list_head *list = &mapping->private_list;
815                 struct address_space *buffer_mapping = mapping->assoc_mapping;
816
817                 spin_lock(&buffer_mapping->private_lock);
818                 while (!list_empty(list))
819                         __remove_assoc_queue(BH_ENTRY(list->next));
820                 spin_unlock(&buffer_mapping->private_lock);
821         }
822 }
823 EXPORT_SYMBOL(invalidate_inode_buffers);
824
825 /*
826  * Remove any clean buffers from the inode's buffer list.  This is called
827  * when we're trying to free the inode itself.  Those buffers can pin it.
828  *
829  * Returns true if all buffers were removed.
830  */
831 int remove_inode_buffers(struct inode *inode)
832 {
833         int ret = 1;
834
835         if (inode_has_buffers(inode)) {
836                 struct address_space *mapping = &inode->i_data;
837                 struct list_head *list = &mapping->private_list;
838                 struct address_space *buffer_mapping = mapping->assoc_mapping;
839
840                 spin_lock(&buffer_mapping->private_lock);
841                 while (!list_empty(list)) {
842                         struct buffer_head *bh = BH_ENTRY(list->next);
843                         if (buffer_dirty(bh)) {
844                                 ret = 0;
845                                 break;
846                         }
847                         __remove_assoc_queue(bh);
848                 }
849                 spin_unlock(&buffer_mapping->private_lock);
850         }
851         return ret;
852 }
853
854 /*
855  * Create the appropriate buffers when given a page for data area and
856  * the size of each buffer.. Use the bh->b_this_page linked list to
857  * follow the buffers created.  Return NULL if unable to create more
858  * buffers.
859  *
860  * The retry flag is used to differentiate async IO (paging, swapping)
861  * which may not fail from ordinary buffer allocations.
862  */
863 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
864                 int retry)
865 {
866         struct buffer_head *bh, *head;
867         long offset;
868
869 try_again:
870         head = NULL;
871         offset = PAGE_SIZE;
872         while ((offset -= size) >= 0) {
873                 bh = alloc_buffer_head(GFP_NOFS);
874                 if (!bh)
875                         goto no_grow;
876
877                 bh->b_bdev = NULL;
878                 bh->b_this_page = head;
879                 bh->b_blocknr = -1;
880                 head = bh;
881
882                 bh->b_state = 0;
883                 atomic_set(&bh->b_count, 0);
884                 bh->b_private = NULL;
885                 bh->b_size = size;
886
887                 /* Link the buffer to its page */
888                 set_bh_page(bh, page, offset);
889
890                 init_buffer(bh, NULL, NULL);
891         }
892         return head;
893 /*
894  * In case anything failed, we just free everything we got.
895  */
896 no_grow:
897         if (head) {
898                 do {
899                         bh = head;
900                         head = head->b_this_page;
901                         free_buffer_head(bh);
902                 } while (head);
903         }
904
905         /*
906          * Return failure for non-async IO requests.  Async IO requests
907          * are not allowed to fail, so we have to wait until buffer heads
908          * become available.  But we don't want tasks sleeping with 
909          * partially complete buffers, so all were released above.
910          */
911         if (!retry)
912                 return NULL;
913
914         /* We're _really_ low on memory. Now we just
915          * wait for old buffer heads to become free due to
916          * finishing IO.  Since this is an async request and
917          * the reserve list is empty, we're sure there are 
918          * async buffer heads in use.
919          */
920         free_more_memory();
921         goto try_again;
922 }
923 EXPORT_SYMBOL_GPL(alloc_page_buffers);
924
925 static inline void
926 link_dev_buffers(struct page *page, struct buffer_head *head)
927 {
928         struct buffer_head *bh, *tail;
929
930         bh = head;
931         do {
932                 tail = bh;
933                 bh = bh->b_this_page;
934         } while (bh);
935         tail->b_this_page = head;
936         attach_page_buffers(page, head);
937 }
938
939 /*
940  * Initialise the state of a blockdev page's buffers.
941  */ 
942 static void
943 init_page_buffers(struct page *page, struct block_device *bdev,
944                         sector_t block, int size)
945 {
946         struct buffer_head *head = page_buffers(page);
947         struct buffer_head *bh = head;
948         int uptodate = PageUptodate(page);
949
950         do {
951                 if (!buffer_mapped(bh)) {
952                         init_buffer(bh, NULL, NULL);
953                         bh->b_bdev = bdev;
954                         bh->b_blocknr = block;
955                         if (uptodate)
956                                 set_buffer_uptodate(bh);
957                         set_buffer_mapped(bh);
958                 }
959                 block++;
960                 bh = bh->b_this_page;
961         } while (bh != head);
962 }
963
964 /*
965  * Create the page-cache page that contains the requested block.
966  *
967  * This is user purely for blockdev mappings.
968  */
969 static struct page *
970 grow_dev_page(struct block_device *bdev, sector_t block,
971                 pgoff_t index, int size)
972 {
973         struct inode *inode = bdev->bd_inode;
974         struct page *page;
975         struct buffer_head *bh;
976
977         page = find_or_create_page(inode->i_mapping, index,
978                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
979         if (!page)
980                 return NULL;
981
982         BUG_ON(!PageLocked(page));
983
984         if (page_has_buffers(page)) {
985                 bh = page_buffers(page);
986                 if (bh->b_size == size) {
987                         init_page_buffers(page, bdev, block, size);
988                         return page;
989                 }
990                 if (!try_to_free_buffers(page))
991                         goto failed;
992         }
993
994         /*
995          * Allocate some buffers for this page
996          */
997         bh = alloc_page_buffers(page, size, 0);
998         if (!bh)
999                 goto failed;
1000
1001         /*
1002          * Link the page to the buffers and initialise them.  Take the
1003          * lock to be atomic wrt __find_get_block(), which does not
1004          * run under the page lock.
1005          */
1006         spin_lock(&inode->i_mapping->private_lock);
1007         link_dev_buffers(page, bh);
1008         init_page_buffers(page, bdev, block, size);
1009         spin_unlock(&inode->i_mapping->private_lock);
1010         return page;
1011
1012 failed:
1013         BUG();
1014         unlock_page(page);
1015         page_cache_release(page);
1016         return NULL;
1017 }
1018
1019 /*
1020  * Create buffers for the specified block device block's page.  If
1021  * that page was dirty, the buffers are set dirty also.
1022  */
1023 static int
1024 grow_buffers(struct block_device *bdev, sector_t block, int size)
1025 {
1026         struct page *page;
1027         pgoff_t index;
1028         int sizebits;
1029
1030         sizebits = -1;
1031         do {
1032                 sizebits++;
1033         } while ((size << sizebits) < PAGE_SIZE);
1034
1035         index = block >> sizebits;
1036
1037         /*
1038          * Check for a block which wants to lie outside our maximum possible
1039          * pagecache index.  (this comparison is done using sector_t types).
1040          */
1041         if (unlikely(index != block >> sizebits)) {
1042                 char b[BDEVNAME_SIZE];
1043
1044                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1045                         "device %s\n",
1046                         __func__, (unsigned long long)block,
1047                         bdevname(bdev, b));
1048                 return -EIO;
1049         }
1050         block = index << sizebits;
1051         /* Create a page with the proper size buffers.. */
1052         page = grow_dev_page(bdev, block, index, size);
1053         if (!page)
1054                 return 0;
1055         unlock_page(page);
1056         page_cache_release(page);
1057         return 1;
1058 }
1059
1060 static struct buffer_head *
1061 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1062 {
1063         /* Size must be multiple of hard sectorsize */
1064         if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1065                         (size < 512 || size > PAGE_SIZE))) {
1066                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1067                                         size);
1068                 printk(KERN_ERR "hardsect size: %d\n",
1069                                         bdev_hardsect_size(bdev));
1070
1071                 dump_stack();
1072                 return NULL;
1073         }
1074
1075         for (;;) {
1076                 struct buffer_head * bh;
1077                 int ret;
1078
1079                 bh = __find_get_block(bdev, block, size);
1080                 if (bh)
1081                         return bh;
1082
1083                 ret = grow_buffers(bdev, block, size);
1084                 if (ret < 0)
1085                         return NULL;
1086                 if (ret == 0)
1087                         free_more_memory();
1088         }
1089 }
1090
1091 /*
1092  * The relationship between dirty buffers and dirty pages:
1093  *
1094  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1095  * the page is tagged dirty in its radix tree.
1096  *
1097  * At all times, the dirtiness of the buffers represents the dirtiness of
1098  * subsections of the page.  If the page has buffers, the page dirty bit is
1099  * merely a hint about the true dirty state.
1100  *
1101  * When a page is set dirty in its entirety, all its buffers are marked dirty
1102  * (if the page has buffers).
1103  *
1104  * When a buffer is marked dirty, its page is dirtied, but the page's other
1105  * buffers are not.
1106  *
1107  * Also.  When blockdev buffers are explicitly read with bread(), they
1108  * individually become uptodate.  But their backing page remains not
1109  * uptodate - even if all of its buffers are uptodate.  A subsequent
1110  * block_read_full_page() against that page will discover all the uptodate
1111  * buffers, will set the page uptodate and will perform no I/O.
1112  */
1113
1114 /**
1115  * mark_buffer_dirty - mark a buffer_head as needing writeout
1116  * @bh: the buffer_head to mark dirty
1117  *
1118  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1119  * backing page dirty, then tag the page as dirty in its address_space's radix
1120  * tree and then attach the address_space's inode to its superblock's dirty
1121  * inode list.
1122  *
1123  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1124  * mapping->tree_lock and the global inode_lock.
1125  */
1126 void mark_buffer_dirty(struct buffer_head *bh)
1127 {
1128         WARN_ON_ONCE(!buffer_uptodate(bh));
1129
1130         /*
1131          * Very *carefully* optimize the it-is-already-dirty case.
1132          *
1133          * Don't let the final "is it dirty" escape to before we
1134          * perhaps modified the buffer.
1135          */
1136         if (buffer_dirty(bh)) {
1137                 smp_mb();
1138                 if (buffer_dirty(bh))
1139                         return;
1140         }
1141
1142         if (!test_set_buffer_dirty(bh)) {
1143                 struct page *page = bh->b_page;
1144                 if (!TestSetPageDirty(page))
1145                         __set_page_dirty(page, page_mapping(page), 0);
1146         }
1147 }
1148
1149 /*
1150  * Decrement a buffer_head's reference count.  If all buffers against a page
1151  * have zero reference count, are clean and unlocked, and if the page is clean
1152  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1153  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1154  * a page but it ends up not being freed, and buffers may later be reattached).
1155  */
1156 void __brelse(struct buffer_head * buf)
1157 {
1158         if (atomic_read(&buf->b_count)) {
1159                 put_bh(buf);
1160                 return;
1161         }
1162         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1163 }
1164
1165 /*
1166  * bforget() is like brelse(), except it discards any
1167  * potentially dirty data.
1168  */
1169 void __bforget(struct buffer_head *bh)
1170 {
1171         clear_buffer_dirty(bh);
1172         if (bh->b_assoc_map) {
1173                 struct address_space *buffer_mapping = bh->b_page->mapping;
1174
1175                 spin_lock(&buffer_mapping->private_lock);
1176                 list_del_init(&bh->b_assoc_buffers);
1177                 bh->b_assoc_map = NULL;
1178                 spin_unlock(&buffer_mapping->private_lock);
1179         }
1180         __brelse(bh);
1181 }
1182
1183 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1184 {
1185         lock_buffer(bh);
1186         if (buffer_uptodate(bh)) {
1187                 unlock_buffer(bh);
1188                 return bh;
1189         } else {
1190                 get_bh(bh);
1191                 bh->b_end_io = end_buffer_read_sync;
1192                 submit_bh(READ, bh);
1193                 wait_on_buffer(bh);
1194                 if (buffer_uptodate(bh))
1195                         return bh;
1196         }
1197         brelse(bh);
1198         return NULL;
1199 }
1200
1201 /*
1202  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1203  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1204  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1205  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1206  * CPU's LRUs at the same time.
1207  *
1208  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1209  * sb_find_get_block().
1210  *
1211  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1212  * a local interrupt disable for that.
1213  */
1214
1215 #define BH_LRU_SIZE     8
1216
1217 struct bh_lru {
1218         struct buffer_head *bhs[BH_LRU_SIZE];
1219 };
1220
1221 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1222
1223 #ifdef CONFIG_SMP
1224 #define bh_lru_lock()   local_irq_disable()
1225 #define bh_lru_unlock() local_irq_enable()
1226 #else
1227 #define bh_lru_lock()   preempt_disable()
1228 #define bh_lru_unlock() preempt_enable()
1229 #endif
1230
1231 static inline void check_irqs_on(void)
1232 {
1233 #ifdef irqs_disabled
1234         BUG_ON(irqs_disabled());
1235 #endif
1236 }
1237
1238 /*
1239  * The LRU management algorithm is dopey-but-simple.  Sorry.
1240  */
1241 static void bh_lru_install(struct buffer_head *bh)
1242 {
1243         struct buffer_head *evictee = NULL;
1244         struct bh_lru *lru;
1245
1246         check_irqs_on();
1247         bh_lru_lock();
1248         lru = &__get_cpu_var(bh_lrus);
1249         if (lru->bhs[0] != bh) {
1250                 struct buffer_head *bhs[BH_LRU_SIZE];
1251                 int in;
1252                 int out = 0;
1253
1254                 get_bh(bh);
1255                 bhs[out++] = bh;
1256                 for (in = 0; in < BH_LRU_SIZE; in++) {
1257                         struct buffer_head *bh2 = lru->bhs[in];
1258
1259                         if (bh2 == bh) {
1260                                 __brelse(bh2);
1261                         } else {
1262                                 if (out >= BH_LRU_SIZE) {
1263                                         BUG_ON(evictee != NULL);
1264                                         evictee = bh2;
1265                                 } else {
1266                                         bhs[out++] = bh2;
1267                                 }
1268                         }
1269                 }
1270                 while (out < BH_LRU_SIZE)
1271                         bhs[out++] = NULL;
1272                 memcpy(lru->bhs, bhs, sizeof(bhs));
1273         }
1274         bh_lru_unlock();
1275
1276         if (evictee)
1277                 __brelse(evictee);
1278 }
1279
1280 /*
1281  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1282  */
1283 static struct buffer_head *
1284 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1285 {
1286         struct buffer_head *ret = NULL;
1287         struct bh_lru *lru;
1288         unsigned int i;
1289
1290         check_irqs_on();
1291         bh_lru_lock();
1292         lru = &__get_cpu_var(bh_lrus);
1293         for (i = 0; i < BH_LRU_SIZE; i++) {
1294                 struct buffer_head *bh = lru->bhs[i];
1295
1296                 if (bh && bh->b_bdev == bdev &&
1297                                 bh->b_blocknr == block && bh->b_size == size) {
1298                         if (i) {
1299                                 while (i) {
1300                                         lru->bhs[i] = lru->bhs[i - 1];
1301                                         i--;
1302                                 }
1303                                 lru->bhs[0] = bh;
1304                         }
1305                         get_bh(bh);
1306                         ret = bh;
1307                         break;
1308                 }
1309         }
1310         bh_lru_unlock();
1311         return ret;
1312 }
1313
1314 /*
1315  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1316  * it in the LRU and mark it as accessed.  If it is not present then return
1317  * NULL
1318  */
1319 struct buffer_head *
1320 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1321 {
1322         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1323
1324         if (bh == NULL) {
1325                 bh = __find_get_block_slow(bdev, block);
1326                 if (bh)
1327                         bh_lru_install(bh);
1328         }
1329         if (bh)
1330                 touch_buffer(bh);
1331         return bh;
1332 }
1333 EXPORT_SYMBOL(__find_get_block);
1334
1335 /*
1336  * __getblk will locate (and, if necessary, create) the buffer_head
1337  * which corresponds to the passed block_device, block and size. The
1338  * returned buffer has its reference count incremented.
1339  *
1340  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1341  * illegal block number, __getblk() will happily return a buffer_head
1342  * which represents the non-existent block.  Very weird.
1343  *
1344  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1345  * attempt is failing.  FIXME, perhaps?
1346  */
1347 struct buffer_head *
1348 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1349 {
1350         struct buffer_head *bh = __find_get_block(bdev, block, size);
1351
1352         might_sleep();
1353         if (bh == NULL)
1354                 bh = __getblk_slow(bdev, block, size);
1355         return bh;
1356 }
1357 EXPORT_SYMBOL(__getblk);
1358
1359 /*
1360  * Do async read-ahead on a buffer..
1361  */
1362 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1363 {
1364         struct buffer_head *bh = __getblk(bdev, block, size);
1365         if (likely(bh)) {
1366                 ll_rw_block(READA, 1, &bh);
1367                 brelse(bh);
1368         }
1369 }
1370 EXPORT_SYMBOL(__breadahead);
1371
1372 /**
1373  *  __bread() - reads a specified block and returns the bh
1374  *  @bdev: the block_device to read from
1375  *  @block: number of block
1376  *  @size: size (in bytes) to read
1377  * 
1378  *  Reads a specified block, and returns buffer head that contains it.
1379  *  It returns NULL if the block was unreadable.
1380  */
1381 struct buffer_head *
1382 __bread(struct block_device *bdev, sector_t block, unsigned size)
1383 {
1384         struct buffer_head *bh = __getblk(bdev, block, size);
1385
1386         if (likely(bh) && !buffer_uptodate(bh))
1387                 bh = __bread_slow(bh);
1388         return bh;
1389 }
1390 EXPORT_SYMBOL(__bread);
1391
1392 /*
1393  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1394  * This doesn't race because it runs in each cpu either in irq
1395  * or with preempt disabled.
1396  */
1397 static void invalidate_bh_lru(void *arg)
1398 {
1399         struct bh_lru *b = &get_cpu_var(bh_lrus);
1400         int i;
1401
1402         for (i = 0; i < BH_LRU_SIZE; i++) {
1403                 brelse(b->bhs[i]);
1404                 b->bhs[i] = NULL;
1405         }
1406         put_cpu_var(bh_lrus);
1407 }
1408         
1409 void invalidate_bh_lrus(void)
1410 {
1411         on_each_cpu(invalidate_bh_lru, NULL, 1);
1412 }
1413 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1414
1415 void set_bh_page(struct buffer_head *bh,
1416                 struct page *page, unsigned long offset)
1417 {
1418         bh->b_page = page;
1419         BUG_ON(offset >= PAGE_SIZE);
1420         if (PageHighMem(page))
1421                 /*
1422                  * This catches illegal uses and preserves the offset:
1423                  */
1424                 bh->b_data = (char *)(0 + offset);
1425         else
1426                 bh->b_data = page_address(page) + offset;
1427 }
1428 EXPORT_SYMBOL(set_bh_page);
1429
1430 /*
1431  * Called when truncating a buffer on a page completely.
1432  */
1433 static void discard_buffer(struct buffer_head * bh)
1434 {
1435         lock_buffer(bh);
1436         clear_buffer_dirty(bh);
1437         bh->b_bdev = NULL;
1438         clear_buffer_mapped(bh);
1439         clear_buffer_req(bh);
1440         clear_buffer_new(bh);
1441         clear_buffer_delay(bh);
1442         clear_buffer_unwritten(bh);
1443         unlock_buffer(bh);
1444 }
1445
1446 /**
1447  * block_invalidatepage - invalidate part of all of a buffer-backed page
1448  *
1449  * @page: the page which is affected
1450  * @offset: the index of the truncation point
1451  *
1452  * block_invalidatepage() is called when all or part of the page has become
1453  * invalidatedby a truncate operation.
1454  *
1455  * block_invalidatepage() does not have to release all buffers, but it must
1456  * ensure that no dirty buffer is left outside @offset and that no I/O
1457  * is underway against any of the blocks which are outside the truncation
1458  * point.  Because the caller is about to free (and possibly reuse) those
1459  * blocks on-disk.
1460  */
1461 void block_invalidatepage(struct page *page, unsigned long offset)
1462 {
1463         struct buffer_head *head, *bh, *next;
1464         unsigned int curr_off = 0;
1465
1466         BUG_ON(!PageLocked(page));
1467         if (!page_has_buffers(page))
1468                 goto out;
1469
1470         head = page_buffers(page);
1471         bh = head;
1472         do {
1473                 unsigned int next_off = curr_off + bh->b_size;
1474                 next = bh->b_this_page;
1475
1476                 /*
1477                  * is this block fully invalidated?
1478                  */
1479                 if (offset <= curr_off)
1480                         discard_buffer(bh);
1481                 curr_off = next_off;
1482                 bh = next;
1483         } while (bh != head);
1484
1485         /*
1486          * We release buffers only if the entire page is being invalidated.
1487          * The get_block cached value has been unconditionally invalidated,
1488          * so real IO is not possible anymore.
1489          */
1490         if (offset == 0)
1491                 try_to_release_page(page, 0);
1492 out:
1493         return;
1494 }
1495 EXPORT_SYMBOL(block_invalidatepage);
1496
1497 /*
1498  * We attach and possibly dirty the buffers atomically wrt
1499  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1500  * is already excluded via the page lock.
1501  */
1502 void create_empty_buffers(struct page *page,
1503                         unsigned long blocksize, unsigned long b_state)
1504 {
1505         struct buffer_head *bh, *head, *tail;
1506
1507         head = alloc_page_buffers(page, blocksize, 1);
1508         bh = head;
1509         do {
1510                 bh->b_state |= b_state;
1511                 tail = bh;
1512                 bh = bh->b_this_page;
1513         } while (bh);
1514         tail->b_this_page = head;
1515
1516         spin_lock(&page->mapping->private_lock);
1517         if (PageUptodate(page) || PageDirty(page)) {
1518                 bh = head;
1519                 do {
1520                         if (PageDirty(page))
1521                                 set_buffer_dirty(bh);
1522                         if (PageUptodate(page))
1523                                 set_buffer_uptodate(bh);
1524                         bh = bh->b_this_page;
1525                 } while (bh != head);
1526         }
1527         attach_page_buffers(page, head);
1528         spin_unlock(&page->mapping->private_lock);
1529 }
1530 EXPORT_SYMBOL(create_empty_buffers);
1531
1532 /*
1533  * We are taking a block for data and we don't want any output from any
1534  * buffer-cache aliases starting from return from that function and
1535  * until the moment when something will explicitly mark the buffer
1536  * dirty (hopefully that will not happen until we will free that block ;-)
1537  * We don't even need to mark it not-uptodate - nobody can expect
1538  * anything from a newly allocated buffer anyway. We used to used
1539  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1540  * don't want to mark the alias unmapped, for example - it would confuse
1541  * anyone who might pick it with bread() afterwards...
1542  *
1543  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1544  * be writeout I/O going on against recently-freed buffers.  We don't
1545  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1546  * only if we really need to.  That happens here.
1547  */
1548 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1549 {
1550         struct buffer_head *old_bh;
1551
1552         might_sleep();
1553
1554         old_bh = __find_get_block_slow(bdev, block);
1555         if (old_bh) {
1556                 clear_buffer_dirty(old_bh);
1557                 wait_on_buffer(old_bh);
1558                 clear_buffer_req(old_bh);
1559                 __brelse(old_bh);
1560         }
1561 }
1562 EXPORT_SYMBOL(unmap_underlying_metadata);
1563
1564 /*
1565  * NOTE! All mapped/uptodate combinations are valid:
1566  *
1567  *      Mapped  Uptodate        Meaning
1568  *
1569  *      No      No              "unknown" - must do get_block()
1570  *      No      Yes             "hole" - zero-filled
1571  *      Yes     No              "allocated" - allocated on disk, not read in
1572  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1573  *
1574  * "Dirty" is valid only with the last case (mapped+uptodate).
1575  */
1576
1577 /*
1578  * While block_write_full_page is writing back the dirty buffers under
1579  * the page lock, whoever dirtied the buffers may decide to clean them
1580  * again at any time.  We handle that by only looking at the buffer
1581  * state inside lock_buffer().
1582  *
1583  * If block_write_full_page() is called for regular writeback
1584  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1585  * locked buffer.   This only can happen if someone has written the buffer
1586  * directly, with submit_bh().  At the address_space level PageWriteback
1587  * prevents this contention from occurring.
1588  */
1589 static int __block_write_full_page(struct inode *inode, struct page *page,
1590                         get_block_t *get_block, struct writeback_control *wbc)
1591 {
1592         int err;
1593         sector_t block;
1594         sector_t last_block;
1595         struct buffer_head *bh, *head;
1596         const unsigned blocksize = 1 << inode->i_blkbits;
1597         int nr_underway = 0;
1598         int write_op = (wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE);
1599
1600         BUG_ON(!PageLocked(page));
1601
1602         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1603
1604         if (!page_has_buffers(page)) {
1605                 create_empty_buffers(page, blocksize,
1606                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1607         }
1608
1609         /*
1610          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1611          * here, and the (potentially unmapped) buffers may become dirty at
1612          * any time.  If a buffer becomes dirty here after we've inspected it
1613          * then we just miss that fact, and the page stays dirty.
1614          *
1615          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1616          * handle that here by just cleaning them.
1617          */
1618
1619         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1620         head = page_buffers(page);
1621         bh = head;
1622
1623         /*
1624          * Get all the dirty buffers mapped to disk addresses and
1625          * handle any aliases from the underlying blockdev's mapping.
1626          */
1627         do {
1628                 if (block > last_block) {
1629                         /*
1630                          * mapped buffers outside i_size will occur, because
1631                          * this page can be outside i_size when there is a
1632                          * truncate in progress.
1633                          */
1634                         /*
1635                          * The buffer was zeroed by block_write_full_page()
1636                          */
1637                         clear_buffer_dirty(bh);
1638                         set_buffer_uptodate(bh);
1639                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1640                            buffer_dirty(bh)) {
1641                         WARN_ON(bh->b_size != blocksize);
1642                         err = get_block(inode, block, bh, 1);
1643                         if (err)
1644                                 goto recover;
1645                         clear_buffer_delay(bh);
1646                         if (buffer_new(bh)) {
1647                                 /* blockdev mappings never come here */
1648                                 clear_buffer_new(bh);
1649                                 unmap_underlying_metadata(bh->b_bdev,
1650                                                         bh->b_blocknr);
1651                         }
1652                 }
1653                 bh = bh->b_this_page;
1654                 block++;
1655         } while (bh != head);
1656
1657         do {
1658                 if (!buffer_mapped(bh))
1659                         continue;
1660                 /*
1661                  * If it's a fully non-blocking write attempt and we cannot
1662                  * lock the buffer then redirty the page.  Note that this can
1663                  * potentially cause a busy-wait loop from pdflush and kswapd
1664                  * activity, but those code paths have their own higher-level
1665                  * throttling.
1666                  */
1667                 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1668                         lock_buffer(bh);
1669                 } else if (!trylock_buffer(bh)) {
1670                         redirty_page_for_writepage(wbc, page);
1671                         continue;
1672                 }
1673                 if (test_clear_buffer_dirty(bh)) {
1674                         mark_buffer_async_write(bh);
1675                 } else {
1676                         unlock_buffer(bh);
1677                 }
1678         } while ((bh = bh->b_this_page) != head);
1679
1680         /*
1681          * The page and its buffers are protected by PageWriteback(), so we can
1682          * drop the bh refcounts early.
1683          */
1684         BUG_ON(PageWriteback(page));
1685         set_page_writeback(page);
1686
1687         do {
1688                 struct buffer_head *next = bh->b_this_page;
1689                 if (buffer_async_write(bh)) {
1690                         submit_bh(write_op, bh);
1691                         nr_underway++;
1692                 }
1693                 bh = next;
1694         } while (bh != head);
1695         unlock_page(page);
1696
1697         err = 0;
1698 done:
1699         if (nr_underway == 0) {
1700                 /*
1701                  * The page was marked dirty, but the buffers were
1702                  * clean.  Someone wrote them back by hand with
1703                  * ll_rw_block/submit_bh.  A rare case.
1704                  */
1705                 end_page_writeback(page);
1706
1707                 /*
1708                  * The page and buffer_heads can be released at any time from
1709                  * here on.
1710                  */
1711         }
1712         return err;
1713
1714 recover:
1715         /*
1716          * ENOSPC, or some other error.  We may already have added some
1717          * blocks to the file, so we need to write these out to avoid
1718          * exposing stale data.
1719          * The page is currently locked and not marked for writeback
1720          */
1721         bh = head;
1722         /* Recovery: lock and submit the mapped buffers */
1723         do {
1724                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1725                     !buffer_delay(bh)) {
1726                         lock_buffer(bh);
1727                         mark_buffer_async_write(bh);
1728                 } else {
1729                         /*
1730                          * The buffer may have been set dirty during
1731                          * attachment to a dirty page.
1732                          */
1733                         clear_buffer_dirty(bh);
1734                 }
1735         } while ((bh = bh->b_this_page) != head);
1736         SetPageError(page);
1737         BUG_ON(PageWriteback(page));
1738         mapping_set_error(page->mapping, err);
1739         set_page_writeback(page);
1740         do {
1741                 struct buffer_head *next = bh->b_this_page;
1742                 if (buffer_async_write(bh)) {
1743                         clear_buffer_dirty(bh);
1744                         submit_bh(write_op, bh);
1745                         nr_underway++;
1746                 }
1747                 bh = next;
1748         } while (bh != head);
1749         unlock_page(page);
1750         goto done;
1751 }
1752
1753 /*
1754  * If a page has any new buffers, zero them out here, and mark them uptodate
1755  * and dirty so they'll be written out (in order to prevent uninitialised
1756  * block data from leaking). And clear the new bit.
1757  */
1758 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1759 {
1760         unsigned int block_start, block_end;
1761         struct buffer_head *head, *bh;
1762
1763         BUG_ON(!PageLocked(page));
1764         if (!page_has_buffers(page))
1765                 return;
1766
1767         bh = head = page_buffers(page);
1768         block_start = 0;
1769         do {
1770                 block_end = block_start + bh->b_size;
1771
1772                 if (buffer_new(bh)) {
1773                         if (block_end > from && block_start < to) {
1774                                 if (!PageUptodate(page)) {
1775                                         unsigned start, size;
1776
1777                                         start = max(from, block_start);
1778                                         size = min(to, block_end) - start;
1779
1780                                         zero_user(page, start, size);
1781                                         set_buffer_uptodate(bh);
1782                                 }
1783
1784                                 clear_buffer_new(bh);
1785                                 mark_buffer_dirty(bh);
1786                         }
1787                 }
1788
1789                 block_start = block_end;
1790                 bh = bh->b_this_page;
1791         } while (bh != head);
1792 }
1793 EXPORT_SYMBOL(page_zero_new_buffers);
1794
1795 static int __block_prepare_write(struct inode *inode, struct page *page,
1796                 unsigned from, unsigned to, get_block_t *get_block)
1797 {
1798         unsigned block_start, block_end;
1799         sector_t block;
1800         int err = 0;
1801         unsigned blocksize, bbits;
1802         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1803
1804         BUG_ON(!PageLocked(page));
1805         BUG_ON(from > PAGE_CACHE_SIZE);
1806         BUG_ON(to > PAGE_CACHE_SIZE);
1807         BUG_ON(from > to);
1808
1809         blocksize = 1 << inode->i_blkbits;
1810         if (!page_has_buffers(page))
1811                 create_empty_buffers(page, blocksize, 0);
1812         head = page_buffers(page);
1813
1814         bbits = inode->i_blkbits;
1815         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1816
1817         for(bh = head, block_start = 0; bh != head || !block_start;
1818             block++, block_start=block_end, bh = bh->b_this_page) {
1819                 block_end = block_start + blocksize;
1820                 if (block_end <= from || block_start >= to) {
1821                         if (PageUptodate(page)) {
1822                                 if (!buffer_uptodate(bh))
1823                                         set_buffer_uptodate(bh);
1824                         }
1825                         continue;
1826                 }
1827                 if (buffer_new(bh))
1828                         clear_buffer_new(bh);
1829                 if (!buffer_mapped(bh)) {
1830                         WARN_ON(bh->b_size != blocksize);
1831                         err = get_block(inode, block, bh, 1);
1832                         if (err)
1833                                 break;
1834                         if (buffer_new(bh)) {
1835                                 unmap_underlying_metadata(bh->b_bdev,
1836                                                         bh->b_blocknr);
1837                                 if (PageUptodate(page)) {
1838                                         clear_buffer_new(bh);
1839                                         set_buffer_uptodate(bh);
1840                                         mark_buffer_dirty(bh);
1841                                         continue;
1842                                 }
1843                                 if (block_end > to || block_start < from)
1844                                         zero_user_segments(page,
1845                                                 to, block_end,
1846                                                 block_start, from);
1847                                 continue;
1848                         }
1849                 }
1850                 if (PageUptodate(page)) {
1851                         if (!buffer_uptodate(bh))
1852                                 set_buffer_uptodate(bh);
1853                         continue; 
1854                 }
1855                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1856                     !buffer_unwritten(bh) &&
1857                      (block_start < from || block_end > to)) {
1858                         ll_rw_block(READ, 1, &bh);
1859                         *wait_bh++=bh;
1860                 }
1861         }
1862         /*
1863          * If we issued read requests - let them complete.
1864          */
1865         while(wait_bh > wait) {
1866                 wait_on_buffer(*--wait_bh);
1867                 if (!buffer_uptodate(*wait_bh))
1868                         err = -EIO;
1869         }
1870         if (unlikely(err))
1871                 page_zero_new_buffers(page, from, to);
1872         return err;
1873 }
1874
1875 static int __block_commit_write(struct inode *inode, struct page *page,
1876                 unsigned from, unsigned to)
1877 {
1878         unsigned block_start, block_end;
1879         int partial = 0;
1880         unsigned blocksize;
1881         struct buffer_head *bh, *head;
1882
1883         blocksize = 1 << inode->i_blkbits;
1884
1885         for(bh = head = page_buffers(page), block_start = 0;
1886             bh != head || !block_start;
1887             block_start=block_end, bh = bh->b_this_page) {
1888                 block_end = block_start + blocksize;
1889                 if (block_end <= from || block_start >= to) {
1890                         if (!buffer_uptodate(bh))
1891                                 partial = 1;
1892                 } else {
1893                         set_buffer_uptodate(bh);
1894                         mark_buffer_dirty(bh);
1895                 }
1896                 clear_buffer_new(bh);
1897         }
1898
1899         /*
1900          * If this is a partial write which happened to make all buffers
1901          * uptodate then we can optimize away a bogus readpage() for
1902          * the next read(). Here we 'discover' whether the page went
1903          * uptodate as a result of this (potentially partial) write.
1904          */
1905         if (!partial)
1906                 SetPageUptodate(page);
1907         return 0;
1908 }
1909
1910 /*
1911  * block_write_begin takes care of the basic task of block allocation and
1912  * bringing partial write blocks uptodate first.
1913  *
1914  * If *pagep is not NULL, then block_write_begin uses the locked page
1915  * at *pagep rather than allocating its own. In this case, the page will
1916  * not be unlocked or deallocated on failure.
1917  */
1918 int block_write_begin(struct file *file, struct address_space *mapping,
1919                         loff_t pos, unsigned len, unsigned flags,
1920                         struct page **pagep, void **fsdata,
1921                         get_block_t *get_block)
1922 {
1923         struct inode *inode = mapping->host;
1924         int status = 0;
1925         struct page *page;
1926         pgoff_t index;
1927         unsigned start, end;
1928         int ownpage = 0;
1929
1930         index = pos >> PAGE_CACHE_SHIFT;
1931         start = pos & (PAGE_CACHE_SIZE - 1);
1932         end = start + len;
1933
1934         page = *pagep;
1935         if (page == NULL) {
1936                 ownpage = 1;
1937                 page = grab_cache_page_write_begin(mapping, index, flags);
1938                 if (!page) {
1939                         status = -ENOMEM;
1940                         goto out;
1941                 }
1942                 *pagep = page;
1943         } else
1944                 BUG_ON(!PageLocked(page));
1945
1946         status = __block_prepare_write(inode, page, start, end, get_block);
1947         if (unlikely(status)) {
1948                 ClearPageUptodate(page);
1949
1950                 if (ownpage) {
1951                         unlock_page(page);
1952                         page_cache_release(page);
1953                         *pagep = NULL;
1954
1955                         /*
1956                          * prepare_write() may have instantiated a few blocks
1957                          * outside i_size.  Trim these off again. Don't need
1958                          * i_size_read because we hold i_mutex.
1959                          */
1960                         if (pos + len > inode->i_size)
1961                                 vmtruncate(inode, inode->i_size);
1962                 }
1963         }
1964
1965 out:
1966         return status;
1967 }
1968 EXPORT_SYMBOL(block_write_begin);
1969
1970 int block_write_end(struct file *file, struct address_space *mapping,
1971                         loff_t pos, unsigned len, unsigned copied,
1972                         struct page *page, void *fsdata)
1973 {
1974         struct inode *inode = mapping->host;
1975         unsigned start;
1976
1977         start = pos & (PAGE_CACHE_SIZE - 1);
1978
1979         if (unlikely(copied < len)) {
1980                 /*
1981                  * The buffers that were written will now be uptodate, so we
1982                  * don't have to worry about a readpage reading them and
1983                  * overwriting a partial write. However if we have encountered
1984                  * a short write and only partially written into a buffer, it
1985                  * will not be marked uptodate, so a readpage might come in and
1986                  * destroy our partial write.
1987                  *
1988                  * Do the simplest thing, and just treat any short write to a
1989                  * non uptodate page as a zero-length write, and force the
1990                  * caller to redo the whole thing.
1991                  */
1992                 if (!PageUptodate(page))
1993                         copied = 0;
1994
1995                 page_zero_new_buffers(page, start+copied, start+len);
1996         }
1997         flush_dcache_page(page);
1998
1999         /* This could be a short (even 0-length) commit */
2000         __block_commit_write(inode, page, start, start+copied);
2001
2002         return copied;
2003 }
2004 EXPORT_SYMBOL(block_write_end);
2005
2006 int generic_write_end(struct file *file, struct address_space *mapping,
2007                         loff_t pos, unsigned len, unsigned copied,
2008                         struct page *page, void *fsdata)
2009 {
2010         struct inode *inode = mapping->host;
2011         int i_size_changed = 0;
2012
2013         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2014
2015         /*
2016          * No need to use i_size_read() here, the i_size
2017          * cannot change under us because we hold i_mutex.
2018          *
2019          * But it's important to update i_size while still holding page lock:
2020          * page writeout could otherwise come in and zero beyond i_size.
2021          */
2022         if (pos+copied > inode->i_size) {
2023                 i_size_write(inode, pos+copied);
2024                 i_size_changed = 1;
2025         }
2026
2027         unlock_page(page);
2028         page_cache_release(page);
2029
2030         /*
2031          * Don't mark the inode dirty under page lock. First, it unnecessarily
2032          * makes the holding time of page lock longer. Second, it forces lock
2033          * ordering of page lock and transaction start for journaling
2034          * filesystems.
2035          */
2036         if (i_size_changed)
2037                 mark_inode_dirty(inode);
2038
2039         return copied;
2040 }
2041 EXPORT_SYMBOL(generic_write_end);
2042
2043 /*
2044  * block_is_partially_uptodate checks whether buffers within a page are
2045  * uptodate or not.
2046  *
2047  * Returns true if all buffers which correspond to a file portion
2048  * we want to read are uptodate.
2049  */
2050 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2051                                         unsigned long from)
2052 {
2053         struct inode *inode = page->mapping->host;
2054         unsigned block_start, block_end, blocksize;
2055         unsigned to;
2056         struct buffer_head *bh, *head;
2057         int ret = 1;
2058
2059         if (!page_has_buffers(page))
2060                 return 0;
2061
2062         blocksize = 1 << inode->i_blkbits;
2063         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2064         to = from + to;
2065         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2066                 return 0;
2067
2068         head = page_buffers(page);
2069         bh = head;
2070         block_start = 0;
2071         do {
2072                 block_end = block_start + blocksize;
2073                 if (block_end > from && block_start < to) {
2074                         if (!buffer_uptodate(bh)) {
2075                                 ret = 0;
2076                                 break;
2077                         }
2078                         if (block_end >= to)
2079                                 break;
2080                 }
2081                 block_start = block_end;
2082                 bh = bh->b_this_page;
2083         } while (bh != head);
2084
2085         return ret;
2086 }
2087 EXPORT_SYMBOL(block_is_partially_uptodate);
2088
2089 /*
2090  * Generic "read page" function for block devices that have the normal
2091  * get_block functionality. This is most of the block device filesystems.
2092  * Reads the page asynchronously --- the unlock_buffer() and
2093  * set/clear_buffer_uptodate() functions propagate buffer state into the
2094  * page struct once IO has completed.
2095  */
2096 int block_read_full_page(struct page *page, get_block_t *get_block)
2097 {
2098         struct inode *inode = page->mapping->host;
2099         sector_t iblock, lblock;
2100         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2101         unsigned int blocksize;
2102         int nr, i;
2103         int fully_mapped = 1;
2104
2105         BUG_ON(!PageLocked(page));
2106         blocksize = 1 << inode->i_blkbits;
2107         if (!page_has_buffers(page))
2108                 create_empty_buffers(page, blocksize, 0);
2109         head = page_buffers(page);
2110
2111         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2112         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2113         bh = head;
2114         nr = 0;
2115         i = 0;
2116
2117         do {
2118                 if (buffer_uptodate(bh))
2119                         continue;
2120
2121                 if (!buffer_mapped(bh)) {
2122                         int err = 0;
2123
2124                         fully_mapped = 0;
2125                         if (iblock < lblock) {
2126                                 WARN_ON(bh->b_size != blocksize);
2127                                 err = get_block(inode, iblock, bh, 0);
2128                                 if (err)
2129                                         SetPageError(page);
2130                         }
2131                         if (!buffer_mapped(bh)) {
2132                                 zero_user(page, i * blocksize, blocksize);
2133                                 if (!err)
2134                                         set_buffer_uptodate(bh);
2135                                 continue;
2136                         }
2137                         /*
2138                          * get_block() might have updated the buffer
2139                          * synchronously
2140                          */
2141                         if (buffer_uptodate(bh))
2142                                 continue;
2143                 }
2144                 arr[nr++] = bh;
2145         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2146
2147         if (fully_mapped)
2148                 SetPageMappedToDisk(page);
2149
2150         if (!nr) {
2151                 /*
2152                  * All buffers are uptodate - we can set the page uptodate
2153                  * as well. But not if get_block() returned an error.
2154                  */
2155                 if (!PageError(page))
2156                         SetPageUptodate(page);
2157                 unlock_page(page);
2158                 return 0;
2159         }
2160
2161         /* Stage two: lock the buffers */
2162         for (i = 0; i < nr; i++) {
2163                 bh = arr[i];
2164                 lock_buffer(bh);
2165                 mark_buffer_async_read(bh);
2166         }
2167
2168         /*
2169          * Stage 3: start the IO.  Check for uptodateness
2170          * inside the buffer lock in case another process reading
2171          * the underlying blockdev brought it uptodate (the sct fix).
2172          */
2173         for (i = 0; i < nr; i++) {
2174                 bh = arr[i];
2175                 if (buffer_uptodate(bh))
2176                         end_buffer_async_read(bh, 1);
2177                 else
2178                         submit_bh(READ, bh);
2179         }
2180         return 0;
2181 }
2182
2183 /* utility function for filesystems that need to do work on expanding
2184  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2185  * deal with the hole.  
2186  */
2187 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2188 {
2189         struct address_space *mapping = inode->i_mapping;
2190         struct page *page;
2191         void *fsdata;
2192         unsigned long limit;
2193         int err;
2194
2195         err = -EFBIG;
2196         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2197         if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2198                 send_sig(SIGXFSZ, current, 0);
2199                 goto out;
2200         }
2201         if (size > inode->i_sb->s_maxbytes)
2202                 goto out;
2203
2204         err = pagecache_write_begin(NULL, mapping, size, 0,
2205                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2206                                 &page, &fsdata);
2207         if (err)
2208                 goto out;
2209
2210         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2211         BUG_ON(err > 0);
2212
2213 out:
2214         return err;
2215 }
2216
2217 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2218                             loff_t pos, loff_t *bytes)
2219 {
2220         struct inode *inode = mapping->host;
2221         unsigned blocksize = 1 << inode->i_blkbits;
2222         struct page *page;
2223         void *fsdata;
2224         pgoff_t index, curidx;
2225         loff_t curpos;
2226         unsigned zerofrom, offset, len;
2227         int err = 0;
2228
2229         index = pos >> PAGE_CACHE_SHIFT;
2230         offset = pos & ~PAGE_CACHE_MASK;
2231
2232         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2233                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2234                 if (zerofrom & (blocksize-1)) {
2235                         *bytes |= (blocksize-1);
2236                         (*bytes)++;
2237                 }
2238                 len = PAGE_CACHE_SIZE - zerofrom;
2239
2240                 err = pagecache_write_begin(file, mapping, curpos, len,
2241                                                 AOP_FLAG_UNINTERRUPTIBLE,
2242                                                 &page, &fsdata);
2243                 if (err)
2244                         goto out;
2245                 zero_user(page, zerofrom, len);
2246                 err = pagecache_write_end(file, mapping, curpos, len, len,
2247                                                 page, fsdata);
2248                 if (err < 0)
2249                         goto out;
2250                 BUG_ON(err != len);
2251                 err = 0;
2252
2253                 balance_dirty_pages_ratelimited(mapping);
2254         }
2255
2256         /* page covers the boundary, find the boundary offset */
2257         if (index == curidx) {
2258                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2259                 /* if we will expand the thing last block will be filled */
2260                 if (offset <= zerofrom) {
2261                         goto out;
2262                 }
2263                 if (zerofrom & (blocksize-1)) {
2264                         *bytes |= (blocksize-1);
2265                         (*bytes)++;
2266                 }
2267                 len = offset - zerofrom;
2268
2269                 err = pagecache_write_begin(file, mapping, curpos, len,
2270                                                 AOP_FLAG_UNINTERRUPTIBLE,
2271                                                 &page, &fsdata);
2272                 if (err)
2273                         goto out;
2274                 zero_user(page, zerofrom, len);
2275                 err = pagecache_write_end(file, mapping, curpos, len, len,
2276                                                 page, fsdata);
2277                 if (err < 0)
2278                         goto out;
2279                 BUG_ON(err != len);
2280                 err = 0;
2281         }
2282 out:
2283         return err;
2284 }
2285
2286 /*
2287  * For moronic filesystems that do not allow holes in file.
2288  * We may have to extend the file.
2289  */
2290 int cont_write_begin(struct file *file, struct address_space *mapping,
2291                         loff_t pos, unsigned len, unsigned flags,
2292                         struct page **pagep, void **fsdata,
2293                         get_block_t *get_block, loff_t *bytes)
2294 {
2295         struct inode *inode = mapping->host;
2296         unsigned blocksize = 1 << inode->i_blkbits;
2297         unsigned zerofrom;
2298         int err;
2299
2300         err = cont_expand_zero(file, mapping, pos, bytes);
2301         if (err)
2302                 goto out;
2303
2304         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2305         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2306                 *bytes |= (blocksize-1);
2307                 (*bytes)++;
2308         }
2309
2310         *pagep = NULL;
2311         err = block_write_begin(file, mapping, pos, len,
2312                                 flags, pagep, fsdata, get_block);
2313 out:
2314         return err;
2315 }
2316
2317 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2318                         get_block_t *get_block)
2319 {
2320         struct inode *inode = page->mapping->host;
2321         int err = __block_prepare_write(inode, page, from, to, get_block);
2322         if (err)
2323                 ClearPageUptodate(page);
2324         return err;
2325 }
2326
2327 int block_commit_write(struct page *page, unsigned from, unsigned to)
2328 {
2329         struct inode *inode = page->mapping->host;
2330         __block_commit_write(inode,page,from,to);
2331         return 0;
2332 }
2333
2334 /*
2335  * block_page_mkwrite() is not allowed to change the file size as it gets
2336  * called from a page fault handler when a page is first dirtied. Hence we must
2337  * be careful to check for EOF conditions here. We set the page up correctly
2338  * for a written page which means we get ENOSPC checking when writing into
2339  * holes and correct delalloc and unwritten extent mapping on filesystems that
2340  * support these features.
2341  *
2342  * We are not allowed to take the i_mutex here so we have to play games to
2343  * protect against truncate races as the page could now be beyond EOF.  Because
2344  * vmtruncate() writes the inode size before removing pages, once we have the
2345  * page lock we can determine safely if the page is beyond EOF. If it is not
2346  * beyond EOF, then the page is guaranteed safe against truncation until we
2347  * unlock the page.
2348  */
2349 int
2350 block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2351                    get_block_t get_block)
2352 {
2353         struct page *page = vmf->page;
2354         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2355         unsigned long end;
2356         loff_t size;
2357         int ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
2358
2359         lock_page(page);
2360         size = i_size_read(inode);
2361         if ((page->mapping != inode->i_mapping) ||
2362             (page_offset(page) > size)) {
2363                 /* page got truncated out from underneath us */
2364                 goto out_unlock;
2365         }
2366
2367         /* page is wholly or partially inside EOF */
2368         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2369                 end = size & ~PAGE_CACHE_MASK;
2370         else
2371                 end = PAGE_CACHE_SIZE;
2372
2373         ret = block_prepare_write(page, 0, end, get_block);
2374         if (!ret)
2375                 ret = block_commit_write(page, 0, end);
2376
2377         if (unlikely(ret)) {
2378                 if (ret == -ENOMEM)
2379                         ret = VM_FAULT_OOM;
2380                 else /* -ENOSPC, -EIO, etc */
2381                         ret = VM_FAULT_SIGBUS;
2382         }
2383
2384 out_unlock:
2385         unlock_page(page);
2386         return ret;
2387 }
2388
2389 /*
2390  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2391  * immediately, while under the page lock.  So it needs a special end_io
2392  * handler which does not touch the bh after unlocking it.
2393  */
2394 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2395 {
2396         __end_buffer_read_notouch(bh, uptodate);
2397 }
2398
2399 /*
2400  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2401  * the page (converting it to circular linked list and taking care of page
2402  * dirty races).
2403  */
2404 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2405 {
2406         struct buffer_head *bh;
2407
2408         BUG_ON(!PageLocked(page));
2409
2410         spin_lock(&page->mapping->private_lock);
2411         bh = head;
2412         do {
2413                 if (PageDirty(page))
2414                         set_buffer_dirty(bh);
2415                 if (!bh->b_this_page)
2416                         bh->b_this_page = head;
2417                 bh = bh->b_this_page;
2418         } while (bh != head);
2419         attach_page_buffers(page, head);
2420         spin_unlock(&page->mapping->private_lock);
2421 }
2422
2423 /*
2424  * On entry, the page is fully not uptodate.
2425  * On exit the page is fully uptodate in the areas outside (from,to)
2426  */
2427 int nobh_write_begin(struct file *file, struct address_space *mapping,
2428                         loff_t pos, unsigned len, unsigned flags,
2429                         struct page **pagep, void **fsdata,
2430                         get_block_t *get_block)
2431 {
2432         struct inode *inode = mapping->host;
2433         const unsigned blkbits = inode->i_blkbits;
2434         const unsigned blocksize = 1 << blkbits;
2435         struct buffer_head *head, *bh;
2436         struct page *page;
2437         pgoff_t index;
2438         unsigned from, to;
2439         unsigned block_in_page;
2440         unsigned block_start, block_end;
2441         sector_t block_in_file;
2442         int nr_reads = 0;
2443         int ret = 0;
2444         int is_mapped_to_disk = 1;
2445
2446         index = pos >> PAGE_CACHE_SHIFT;
2447         from = pos & (PAGE_CACHE_SIZE - 1);
2448         to = from + len;
2449
2450         page = grab_cache_page_write_begin(mapping, index, flags);
2451         if (!page)
2452                 return -ENOMEM;
2453         *pagep = page;
2454         *fsdata = NULL;
2455
2456         if (page_has_buffers(page)) {
2457                 unlock_page(page);
2458                 page_cache_release(page);
2459                 *pagep = NULL;
2460                 return block_write_begin(file, mapping, pos, len, flags, pagep,
2461                                         fsdata, get_block);
2462         }
2463
2464         if (PageMappedToDisk(page))
2465                 return 0;
2466
2467         /*
2468          * Allocate buffers so that we can keep track of state, and potentially
2469          * attach them to the page if an error occurs. In the common case of
2470          * no error, they will just be freed again without ever being attached
2471          * to the page (which is all OK, because we're under the page lock).
2472          *
2473          * Be careful: the buffer linked list is a NULL terminated one, rather
2474          * than the circular one we're used to.
2475          */
2476         head = alloc_page_buffers(page, blocksize, 0);
2477         if (!head) {
2478                 ret = -ENOMEM;
2479                 goto out_release;
2480         }
2481
2482         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2483
2484         /*
2485          * We loop across all blocks in the page, whether or not they are
2486          * part of the affected region.  This is so we can discover if the
2487          * page is fully mapped-to-disk.
2488          */
2489         for (block_start = 0, block_in_page = 0, bh = head;
2490                   block_start < PAGE_CACHE_SIZE;
2491                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2492                 int create;
2493
2494                 block_end = block_start + blocksize;
2495                 bh->b_state = 0;
2496                 create = 1;
2497                 if (block_start >= to)
2498                         create = 0;
2499                 ret = get_block(inode, block_in_file + block_in_page,
2500                                         bh, create);
2501                 if (ret)
2502                         goto failed;
2503                 if (!buffer_mapped(bh))
2504                         is_mapped_to_disk = 0;
2505                 if (buffer_new(bh))
2506                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2507                 if (PageUptodate(page)) {
2508                         set_buffer_uptodate(bh);
2509                         continue;
2510                 }
2511                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2512                         zero_user_segments(page, block_start, from,
2513                                                         to, block_end);
2514                         continue;
2515                 }
2516                 if (buffer_uptodate(bh))
2517                         continue;       /* reiserfs does this */
2518                 if (block_start < from || block_end > to) {
2519                         lock_buffer(bh);
2520                         bh->b_end_io = end_buffer_read_nobh;
2521                         submit_bh(READ, bh);
2522                         nr_reads++;
2523                 }
2524         }
2525
2526         if (nr_reads) {
2527                 /*
2528                  * The page is locked, so these buffers are protected from
2529                  * any VM or truncate activity.  Hence we don't need to care
2530                  * for the buffer_head refcounts.
2531                  */
2532                 for (bh = head; bh; bh = bh->b_this_page) {
2533                         wait_on_buffer(bh);
2534                         if (!buffer_uptodate(bh))
2535                                 ret = -EIO;
2536                 }
2537                 if (ret)
2538                         goto failed;
2539         }
2540
2541         if (is_mapped_to_disk)
2542                 SetPageMappedToDisk(page);
2543
2544         *fsdata = head; /* to be released by nobh_write_end */
2545
2546         return 0;
2547
2548 failed:
2549         BUG_ON(!ret);
2550         /*
2551          * Error recovery is a bit difficult. We need to zero out blocks that
2552          * were newly allocated, and dirty them to ensure they get written out.
2553          * Buffers need to be attached to the page at this point, otherwise
2554          * the handling of potential IO errors during writeout would be hard
2555          * (could try doing synchronous writeout, but what if that fails too?)
2556          */
2557         attach_nobh_buffers(page, head);
2558         page_zero_new_buffers(page, from, to);
2559
2560 out_release:
2561         unlock_page(page);
2562         page_cache_release(page);
2563         *pagep = NULL;
2564
2565         if (pos + len > inode->i_size)
2566                 vmtruncate(inode, inode->i_size);
2567
2568         return ret;
2569 }
2570 EXPORT_SYMBOL(nobh_write_begin);
2571
2572 int nobh_write_end(struct file *file, struct address_space *mapping,
2573                         loff_t pos, unsigned len, unsigned copied,
2574                         struct page *page, void *fsdata)
2575 {
2576         struct inode *inode = page->mapping->host;
2577         struct buffer_head *head = fsdata;
2578         struct buffer_head *bh;
2579         BUG_ON(fsdata != NULL && page_has_buffers(page));
2580
2581         if (unlikely(copied < len) && head)
2582                 attach_nobh_buffers(page, head);
2583         if (page_has_buffers(page))
2584                 return generic_write_end(file, mapping, pos, len,
2585                                         copied, page, fsdata);
2586
2587         SetPageUptodate(page);
2588         set_page_dirty(page);
2589         if (pos+copied > inode->i_size) {
2590                 i_size_write(inode, pos+copied);
2591                 mark_inode_dirty(inode);
2592         }
2593
2594         unlock_page(page);
2595         page_cache_release(page);
2596
2597         while (head) {
2598                 bh = head;
2599                 head = head->b_this_page;
2600                 free_buffer_head(bh);
2601         }
2602
2603         return copied;
2604 }
2605 EXPORT_SYMBOL(nobh_write_end);
2606
2607 /*
2608  * nobh_writepage() - based on block_full_write_page() except
2609  * that it tries to operate without attaching bufferheads to
2610  * the page.
2611  */
2612 int nobh_writepage(struct page *page, get_block_t *get_block,
2613                         struct writeback_control *wbc)
2614 {
2615         struct inode * const inode = page->mapping->host;
2616         loff_t i_size = i_size_read(inode);
2617         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2618         unsigned offset;
2619         int ret;
2620
2621         /* Is the page fully inside i_size? */
2622         if (page->index < end_index)
2623                 goto out;
2624
2625         /* Is the page fully outside i_size? (truncate in progress) */
2626         offset = i_size & (PAGE_CACHE_SIZE-1);
2627         if (page->index >= end_index+1 || !offset) {
2628                 /*
2629                  * The page may have dirty, unmapped buffers.  For example,
2630                  * they may have been added in ext3_writepage().  Make them
2631                  * freeable here, so the page does not leak.
2632                  */
2633 #if 0
2634                 /* Not really sure about this  - do we need this ? */
2635                 if (page->mapping->a_ops->invalidatepage)
2636                         page->mapping->a_ops->invalidatepage(page, offset);
2637 #endif
2638                 unlock_page(page);
2639                 return 0; /* don't care */
2640         }
2641
2642         /*
2643          * The page straddles i_size.  It must be zeroed out on each and every
2644          * writepage invocation because it may be mmapped.  "A file is mapped
2645          * in multiples of the page size.  For a file that is not a multiple of
2646          * the  page size, the remaining memory is zeroed when mapped, and
2647          * writes to that region are not written out to the file."
2648          */
2649         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2650 out:
2651         ret = mpage_writepage(page, get_block, wbc);
2652         if (ret == -EAGAIN)
2653                 ret = __block_write_full_page(inode, page, get_block, wbc);
2654         return ret;
2655 }
2656 EXPORT_SYMBOL(nobh_writepage);
2657
2658 int nobh_truncate_page(struct address_space *mapping,
2659                         loff_t from, get_block_t *get_block)
2660 {
2661         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2662         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2663         unsigned blocksize;
2664         sector_t iblock;
2665         unsigned length, pos;
2666         struct inode *inode = mapping->host;
2667         struct page *page;
2668         struct buffer_head map_bh;
2669         int err;
2670
2671         blocksize = 1 << inode->i_blkbits;
2672         length = offset & (blocksize - 1);
2673
2674         /* Block boundary? Nothing to do */
2675         if (!length)
2676                 return 0;
2677
2678         length = blocksize - length;
2679         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2680
2681         page = grab_cache_page(mapping, index);
2682         err = -ENOMEM;
2683         if (!page)
2684                 goto out;
2685
2686         if (page_has_buffers(page)) {
2687 has_buffers:
2688                 unlock_page(page);
2689                 page_cache_release(page);
2690                 return block_truncate_page(mapping, from, get_block);
2691         }
2692
2693         /* Find the buffer that contains "offset" */
2694         pos = blocksize;
2695         while (offset >= pos) {
2696                 iblock++;
2697                 pos += blocksize;
2698         }
2699
2700         err = get_block(inode, iblock, &map_bh, 0);
2701         if (err)
2702                 goto unlock;
2703         /* unmapped? It's a hole - nothing to do */
2704         if (!buffer_mapped(&map_bh))
2705                 goto unlock;
2706
2707         /* Ok, it's mapped. Make sure it's up-to-date */
2708         if (!PageUptodate(page)) {
2709                 err = mapping->a_ops->readpage(NULL, page);
2710                 if (err) {
2711                         page_cache_release(page);
2712                         goto out;
2713                 }
2714                 lock_page(page);
2715                 if (!PageUptodate(page)) {
2716                         err = -EIO;
2717                         goto unlock;
2718                 }
2719                 if (page_has_buffers(page))
2720                         goto has_buffers;
2721         }
2722         zero_user(page, offset, length);
2723         set_page_dirty(page);
2724         err = 0;
2725
2726 unlock:
2727         unlock_page(page);
2728         page_cache_release(page);
2729 out:
2730         return err;
2731 }
2732 EXPORT_SYMBOL(nobh_truncate_page);
2733
2734 int block_truncate_page(struct address_space *mapping,
2735                         loff_t from, get_block_t *get_block)
2736 {
2737         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2738         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2739         unsigned blocksize;
2740         sector_t iblock;
2741         unsigned length, pos;
2742         struct inode *inode = mapping->host;
2743         struct page *page;
2744         struct buffer_head *bh;
2745         int err;
2746
2747         blocksize = 1 << inode->i_blkbits;
2748         length = offset & (blocksize - 1);
2749
2750         /* Block boundary? Nothing to do */
2751         if (!length)
2752                 return 0;
2753
2754         length = blocksize - length;
2755         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2756         
2757         page = grab_cache_page(mapping, index);
2758         err = -ENOMEM;
2759         if (!page)
2760                 goto out;
2761
2762         if (!page_has_buffers(page))
2763                 create_empty_buffers(page, blocksize, 0);
2764
2765         /* Find the buffer that contains "offset" */
2766         bh = page_buffers(page);
2767         pos = blocksize;
2768         while (offset >= pos) {
2769                 bh = bh->b_this_page;
2770                 iblock++;
2771                 pos += blocksize;
2772         }
2773
2774         err = 0;
2775         if (!buffer_mapped(bh)) {
2776                 WARN_ON(bh->b_size != blocksize);
2777                 err = get_block(inode, iblock, bh, 0);
2778                 if (err)
2779                         goto unlock;
2780                 /* unmapped? It's a hole - nothing to do */
2781                 if (!buffer_mapped(bh))
2782                         goto unlock;
2783         }
2784
2785         /* Ok, it's mapped. Make sure it's up-to-date */
2786         if (PageUptodate(page))
2787                 set_buffer_uptodate(bh);
2788
2789         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2790                 err = -EIO;
2791                 ll_rw_block(READ, 1, &bh);
2792                 wait_on_buffer(bh);
2793                 /* Uhhuh. Read error. Complain and punt. */
2794                 if (!buffer_uptodate(bh))
2795                         goto unlock;
2796         }
2797
2798         zero_user(page, offset, length);
2799         mark_buffer_dirty(bh);
2800         err = 0;
2801
2802 unlock:
2803         unlock_page(page);
2804         page_cache_release(page);
2805 out:
2806         return err;
2807 }
2808
2809 /*
2810  * The generic ->writepage function for buffer-backed address_spaces
2811  */
2812 int block_write_full_page(struct page *page, get_block_t *get_block,
2813                         struct writeback_control *wbc)
2814 {
2815         struct inode * const inode = page->mapping->host;
2816         loff_t i_size = i_size_read(inode);
2817         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2818         unsigned offset;
2819
2820         /* Is the page fully inside i_size? */
2821         if (page->index < end_index)
2822                 return __block_write_full_page(inode, page, get_block, wbc);
2823
2824         /* Is the page fully outside i_size? (truncate in progress) */
2825         offset = i_size & (PAGE_CACHE_SIZE-1);
2826         if (page->index >= end_index+1 || !offset) {
2827                 /*
2828                  * The page may have dirty, unmapped buffers.  For example,
2829                  * they may have been added in ext3_writepage().  Make them
2830                  * freeable here, so the page does not leak.
2831                  */
2832                 do_invalidatepage(page, 0);
2833                 unlock_page(page);
2834                 return 0; /* don't care */
2835         }
2836
2837         /*
2838          * The page straddles i_size.  It must be zeroed out on each and every
2839          * writepage invokation because it may be mmapped.  "A file is mapped
2840          * in multiples of the page size.  For a file that is not a multiple of
2841          * the  page size, the remaining memory is zeroed when mapped, and
2842          * writes to that region are not written out to the file."
2843          */
2844         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2845         return __block_write_full_page(inode, page, get_block, wbc);
2846 }
2847
2848 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2849                             get_block_t *get_block)
2850 {
2851         struct buffer_head tmp;
2852         struct inode *inode = mapping->host;
2853         tmp.b_state = 0;
2854         tmp.b_blocknr = 0;
2855         tmp.b_size = 1 << inode->i_blkbits;
2856         get_block(inode, block, &tmp, 0);
2857         return tmp.b_blocknr;
2858 }
2859
2860 static void end_bio_bh_io_sync(struct bio *bio, int err)
2861 {
2862         struct buffer_head *bh = bio->bi_private;
2863
2864         if (err == -EOPNOTSUPP) {
2865                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2866                 set_bit(BH_Eopnotsupp, &bh->b_state);
2867         }
2868
2869         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2870                 set_bit(BH_Quiet, &bh->b_state);
2871
2872         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2873         bio_put(bio);
2874 }
2875
2876 int submit_bh(int rw, struct buffer_head * bh)
2877 {
2878         struct bio *bio;
2879         int ret = 0;
2880
2881         BUG_ON(!buffer_locked(bh));
2882         BUG_ON(!buffer_mapped(bh));
2883         BUG_ON(!bh->b_end_io);
2884
2885         /*
2886          * Mask in barrier bit for a write (could be either a WRITE or a
2887          * WRITE_SYNC
2888          */
2889         if (buffer_ordered(bh) && (rw & WRITE))
2890                 rw |= WRITE_BARRIER;
2891
2892         /*
2893          * Only clear out a write error when rewriting
2894          */
2895         if (test_set_buffer_req(bh) && (rw & WRITE))
2896                 clear_buffer_write_io_error(bh);
2897
2898         /*
2899          * from here on down, it's all bio -- do the initial mapping,
2900          * submit_bio -> generic_make_request may further map this bio around
2901          */
2902         bio = bio_alloc(GFP_NOIO, 1);
2903
2904         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2905         bio->bi_bdev = bh->b_bdev;
2906         bio->bi_io_vec[0].bv_page = bh->b_page;
2907         bio->bi_io_vec[0].bv_len = bh->b_size;
2908         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2909
2910         bio->bi_vcnt = 1;
2911         bio->bi_idx = 0;
2912         bio->bi_size = bh->b_size;
2913
2914         bio->bi_end_io = end_bio_bh_io_sync;
2915         bio->bi_private = bh;
2916
2917         bio_get(bio);
2918         submit_bio(rw, bio);
2919
2920         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2921                 ret = -EOPNOTSUPP;
2922
2923         bio_put(bio);
2924         return ret;
2925 }
2926
2927 /**
2928  * ll_rw_block: low-level access to block devices (DEPRECATED)
2929  * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
2930  * @nr: number of &struct buffer_heads in the array
2931  * @bhs: array of pointers to &struct buffer_head
2932  *
2933  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2934  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2935  * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2936  * are sent to disk. The fourth %READA option is described in the documentation
2937  * for generic_make_request() which ll_rw_block() calls.
2938  *
2939  * This function drops any buffer that it cannot get a lock on (with the
2940  * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
2941  * clean when doing a write request, and any buffer that appears to be
2942  * up-to-date when doing read request.  Further it marks as clean buffers that
2943  * are processed for writing (the buffer cache won't assume that they are
2944  * actually clean until the buffer gets unlocked).
2945  *
2946  * ll_rw_block sets b_end_io to simple completion handler that marks
2947  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2948  * any waiters. 
2949  *
2950  * All of the buffers must be for the same device, and must also be a
2951  * multiple of the current approved size for the device.
2952  */
2953 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2954 {
2955         int i;
2956
2957         for (i = 0; i < nr; i++) {
2958                 struct buffer_head *bh = bhs[i];
2959
2960                 if (rw == SWRITE || rw == SWRITE_SYNC)
2961                         lock_buffer(bh);
2962                 else if (!trylock_buffer(bh))
2963                         continue;
2964
2965                 if (rw == WRITE || rw == SWRITE || rw == SWRITE_SYNC) {
2966                         if (test_clear_buffer_dirty(bh)) {
2967                                 bh->b_end_io = end_buffer_write_sync;
2968                                 get_bh(bh);
2969                                 if (rw == SWRITE_SYNC)
2970                                         submit_bh(WRITE_SYNC, bh);
2971                                 else
2972                                         submit_bh(WRITE, bh);
2973                                 continue;
2974                         }
2975                 } else {
2976                         if (!buffer_uptodate(bh)) {
2977                                 bh->b_end_io = end_buffer_read_sync;
2978                                 get_bh(bh);
2979                                 submit_bh(rw, bh);
2980                                 continue;
2981                         }
2982                 }
2983                 unlock_buffer(bh);
2984         }
2985 }
2986
2987 /*
2988  * For a data-integrity writeout, we need to wait upon any in-progress I/O
2989  * and then start new I/O and then wait upon it.  The caller must have a ref on
2990  * the buffer_head.
2991  */
2992 int sync_dirty_buffer(struct buffer_head *bh)
2993 {
2994         int ret = 0;
2995
2996         WARN_ON(atomic_read(&bh->b_count) < 1);
2997         lock_buffer(bh);
2998         if (test_clear_buffer_dirty(bh)) {
2999                 get_bh(bh);
3000                 bh->b_end_io = end_buffer_write_sync;
3001                 ret = submit_bh(WRITE, bh);
3002                 wait_on_buffer(bh);
3003                 if (buffer_eopnotsupp(bh)) {
3004                         clear_buffer_eopnotsupp(bh);
3005                         ret = -EOPNOTSUPP;
3006                 }
3007                 if (!ret && !buffer_uptodate(bh))
3008                         ret = -EIO;
3009         } else {
3010                 unlock_buffer(bh);
3011         }
3012         return ret;
3013 }
3014
3015 /*
3016  * try_to_free_buffers() checks if all the buffers on this particular page
3017  * are unused, and releases them if so.
3018  *
3019  * Exclusion against try_to_free_buffers may be obtained by either
3020  * locking the page or by holding its mapping's private_lock.
3021  *
3022  * If the page is dirty but all the buffers are clean then we need to
3023  * be sure to mark the page clean as well.  This is because the page
3024  * may be against a block device, and a later reattachment of buffers
3025  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3026  * filesystem data on the same device.
3027  *
3028  * The same applies to regular filesystem pages: if all the buffers are
3029  * clean then we set the page clean and proceed.  To do that, we require
3030  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3031  * private_lock.
3032  *
3033  * try_to_free_buffers() is non-blocking.
3034  */
3035 static inline int buffer_busy(struct buffer_head *bh)
3036 {
3037         return atomic_read(&bh->b_count) |
3038                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3039 }
3040
3041 static int
3042 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3043 {
3044         struct buffer_head *head = page_buffers(page);
3045         struct buffer_head *bh;
3046
3047         bh = head;
3048         do {
3049                 if (buffer_write_io_error(bh) && page->mapping)
3050                         set_bit(AS_EIO, &page->mapping->flags);
3051                 if (buffer_busy(bh))
3052                         goto failed;
3053                 bh = bh->b_this_page;
3054         } while (bh != head);
3055
3056         do {
3057                 struct buffer_head *next = bh->b_this_page;
3058
3059                 if (bh->b_assoc_map)
3060                         __remove_assoc_queue(bh);
3061                 bh = next;
3062         } while (bh != head);
3063         *buffers_to_free = head;
3064         __clear_page_buffers(page);
3065         return 1;
3066 failed:
3067         return 0;
3068 }
3069
3070 int try_to_free_buffers(struct page *page)
3071 {
3072         struct address_space * const mapping = page->mapping;
3073         struct buffer_head *buffers_to_free = NULL;
3074         int ret = 0;
3075
3076         BUG_ON(!PageLocked(page));
3077         if (PageWriteback(page))
3078                 return 0;
3079
3080         if (mapping == NULL) {          /* can this still happen? */
3081                 ret = drop_buffers(page, &buffers_to_free);
3082                 goto out;
3083         }
3084
3085         spin_lock(&mapping->private_lock);
3086         ret = drop_buffers(page, &buffers_to_free);
3087
3088         /*
3089          * If the filesystem writes its buffers by hand (eg ext3)
3090          * then we can have clean buffers against a dirty page.  We
3091          * clean the page here; otherwise the VM will never notice
3092          * that the filesystem did any IO at all.
3093          *
3094          * Also, during truncate, discard_buffer will have marked all
3095          * the page's buffers clean.  We discover that here and clean
3096          * the page also.
3097          *
3098          * private_lock must be held over this entire operation in order
3099          * to synchronise against __set_page_dirty_buffers and prevent the
3100          * dirty bit from being lost.
3101          */
3102         if (ret)
3103                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3104         spin_unlock(&mapping->private_lock);
3105 out:
3106         if (buffers_to_free) {
3107                 struct buffer_head *bh = buffers_to_free;
3108
3109                 do {
3110                         struct buffer_head *next = bh->b_this_page;
3111                         free_buffer_head(bh);
3112                         bh = next;
3113                 } while (bh != buffers_to_free);
3114         }
3115         return ret;
3116 }
3117 EXPORT_SYMBOL(try_to_free_buffers);
3118
3119 void block_sync_page(struct page *page)
3120 {
3121         struct address_space *mapping;
3122
3123         smp_mb();
3124         mapping = page_mapping(page);
3125         if (mapping)
3126                 blk_run_backing_dev(mapping->backing_dev_info, page);
3127 }
3128
3129 /*
3130  * There are no bdflush tunables left.  But distributions are
3131  * still running obsolete flush daemons, so we terminate them here.
3132  *
3133  * Use of bdflush() is deprecated and will be removed in a future kernel.
3134  * The `pdflush' kernel threads fully replace bdflush daemons and this call.
3135  */
3136 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3137 {
3138         static int msg_count;
3139
3140         if (!capable(CAP_SYS_ADMIN))
3141                 return -EPERM;
3142
3143         if (msg_count < 5) {
3144                 msg_count++;
3145                 printk(KERN_INFO
3146                         "warning: process `%s' used the obsolete bdflush"
3147                         " system call\n", current->comm);
3148                 printk(KERN_INFO "Fix your initscripts?\n");
3149         }
3150
3151         if (func == 1)
3152                 do_exit(0);
3153         return 0;
3154 }
3155
3156 /*
3157  * Buffer-head allocation
3158  */
3159 static struct kmem_cache *bh_cachep;
3160
3161 /*
3162  * Once the number of bh's in the machine exceeds this level, we start
3163  * stripping them in writeback.
3164  */
3165 static int max_buffer_heads;
3166
3167 int buffer_heads_over_limit;
3168
3169 struct bh_accounting {
3170         int nr;                 /* Number of live bh's */
3171         int ratelimit;          /* Limit cacheline bouncing */
3172 };
3173
3174 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3175
3176 static void recalc_bh_state(void)
3177 {
3178         int i;
3179         int tot = 0;
3180
3181         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3182                 return;
3183         __get_cpu_var(bh_accounting).ratelimit = 0;
3184         for_each_online_cpu(i)
3185                 tot += per_cpu(bh_accounting, i).nr;
3186         buffer_heads_over_limit = (tot > max_buffer_heads);
3187 }
3188         
3189 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3190 {
3191         struct buffer_head *ret = kmem_cache_alloc(bh_cachep, gfp_flags);
3192         if (ret) {
3193                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3194                 get_cpu_var(bh_accounting).nr++;
3195                 recalc_bh_state();
3196                 put_cpu_var(bh_accounting);
3197         }
3198         return ret;
3199 }
3200 EXPORT_SYMBOL(alloc_buffer_head);
3201
3202 void free_buffer_head(struct buffer_head *bh)
3203 {
3204         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3205         kmem_cache_free(bh_cachep, bh);
3206         get_cpu_var(bh_accounting).nr--;
3207         recalc_bh_state();
3208         put_cpu_var(bh_accounting);
3209 }
3210 EXPORT_SYMBOL(free_buffer_head);
3211
3212 static void buffer_exit_cpu(int cpu)
3213 {
3214         int i;
3215         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3216
3217         for (i = 0; i < BH_LRU_SIZE; i++) {
3218                 brelse(b->bhs[i]);
3219                 b->bhs[i] = NULL;
3220         }
3221         get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3222         per_cpu(bh_accounting, cpu).nr = 0;
3223         put_cpu_var(bh_accounting);
3224 }
3225
3226 static int buffer_cpu_notify(struct notifier_block *self,
3227                               unsigned long action, void *hcpu)
3228 {
3229         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3230                 buffer_exit_cpu((unsigned long)hcpu);
3231         return NOTIFY_OK;
3232 }
3233
3234 /**
3235  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3236  * @bh: struct buffer_head
3237  *
3238  * Return true if the buffer is up-to-date and false,
3239  * with the buffer locked, if not.
3240  */
3241 int bh_uptodate_or_lock(struct buffer_head *bh)
3242 {
3243         if (!buffer_uptodate(bh)) {
3244                 lock_buffer(bh);
3245                 if (!buffer_uptodate(bh))
3246                         return 0;
3247                 unlock_buffer(bh);
3248         }
3249         return 1;
3250 }
3251 EXPORT_SYMBOL(bh_uptodate_or_lock);
3252
3253 /**
3254  * bh_submit_read - Submit a locked buffer for reading
3255  * @bh: struct buffer_head
3256  *
3257  * Returns zero on success and -EIO on error.
3258  */
3259 int bh_submit_read(struct buffer_head *bh)
3260 {
3261         BUG_ON(!buffer_locked(bh));
3262
3263         if (buffer_uptodate(bh)) {
3264                 unlock_buffer(bh);
3265                 return 0;
3266         }
3267
3268         get_bh(bh);
3269         bh->b_end_io = end_buffer_read_sync;
3270         submit_bh(READ, bh);
3271         wait_on_buffer(bh);
3272         if (buffer_uptodate(bh))
3273                 return 0;
3274         return -EIO;
3275 }
3276 EXPORT_SYMBOL(bh_submit_read);
3277
3278 static void
3279 init_buffer_head(void *data)
3280 {
3281         struct buffer_head *bh = data;
3282
3283         memset(bh, 0, sizeof(*bh));
3284         INIT_LIST_HEAD(&bh->b_assoc_buffers);
3285 }
3286
3287 void __init buffer_init(void)
3288 {
3289         int nrpages;
3290
3291         bh_cachep = kmem_cache_create("buffer_head",
3292                         sizeof(struct buffer_head), 0,
3293                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3294                                 SLAB_MEM_SPREAD),
3295                                 init_buffer_head);
3296
3297         /*
3298          * Limit the bh occupancy to 10% of ZONE_NORMAL
3299          */
3300         nrpages = (nr_free_buffer_pages() * 10) / 100;
3301         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3302         hotcpu_notifier(buffer_cpu_notify, 0);
3303 }
3304
3305 EXPORT_SYMBOL(__bforget);
3306 EXPORT_SYMBOL(__brelse);
3307 EXPORT_SYMBOL(__wait_on_buffer);
3308 EXPORT_SYMBOL(block_commit_write);
3309 EXPORT_SYMBOL(block_prepare_write);
3310 EXPORT_SYMBOL(block_page_mkwrite);
3311 EXPORT_SYMBOL(block_read_full_page);
3312 EXPORT_SYMBOL(block_sync_page);
3313 EXPORT_SYMBOL(block_truncate_page);
3314 EXPORT_SYMBOL(block_write_full_page);
3315 EXPORT_SYMBOL(cont_write_begin);
3316 EXPORT_SYMBOL(end_buffer_read_sync);
3317 EXPORT_SYMBOL(end_buffer_write_sync);
3318 EXPORT_SYMBOL(file_fsync);
3319 EXPORT_SYMBOL(generic_block_bmap);
3320 EXPORT_SYMBOL(generic_cont_expand_simple);
3321 EXPORT_SYMBOL(init_buffer);
3322 EXPORT_SYMBOL(invalidate_bdev);
3323 EXPORT_SYMBOL(ll_rw_block);
3324 EXPORT_SYMBOL(mark_buffer_dirty);
3325 EXPORT_SYMBOL(submit_bh);
3326 EXPORT_SYMBOL(sync_dirty_buffer);
3327 EXPORT_SYMBOL(unlock_buffer);