mm/fs: add hooks to support cleancache
[linux-3.10.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 #include <linux/cleancache.h>
45
46 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
47
48 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
49
50 inline void
51 init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
52 {
53         bh->b_end_io = handler;
54         bh->b_private = private;
55 }
56 EXPORT_SYMBOL(init_buffer);
57
58 static int sleep_on_buffer(void *word)
59 {
60         io_schedule();
61         return 0;
62 }
63
64 void __lock_buffer(struct buffer_head *bh)
65 {
66         wait_on_bit_lock(&bh->b_state, BH_Lock, sleep_on_buffer,
67                                                         TASK_UNINTERRUPTIBLE);
68 }
69 EXPORT_SYMBOL(__lock_buffer);
70
71 void unlock_buffer(struct buffer_head *bh)
72 {
73         clear_bit_unlock(BH_Lock, &bh->b_state);
74         smp_mb__after_clear_bit();
75         wake_up_bit(&bh->b_state, BH_Lock);
76 }
77 EXPORT_SYMBOL(unlock_buffer);
78
79 /*
80  * Block until a buffer comes unlocked.  This doesn't stop it
81  * from becoming locked again - you have to lock it yourself
82  * if you want to preserve its state.
83  */
84 void __wait_on_buffer(struct buffer_head * bh)
85 {
86         wait_on_bit(&bh->b_state, BH_Lock, sleep_on_buffer, TASK_UNINTERRUPTIBLE);
87 }
88 EXPORT_SYMBOL(__wait_on_buffer);
89
90 static void
91 __clear_page_buffers(struct page *page)
92 {
93         ClearPagePrivate(page);
94         set_page_private(page, 0);
95         page_cache_release(page);
96 }
97
98
99 static int quiet_error(struct buffer_head *bh)
100 {
101         if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
102                 return 0;
103         return 1;
104 }
105
106
107 static void buffer_io_error(struct buffer_head *bh)
108 {
109         char b[BDEVNAME_SIZE];
110         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
111                         bdevname(bh->b_bdev, b),
112                         (unsigned long long)bh->b_blocknr);
113 }
114
115 /*
116  * End-of-IO handler helper function which does not touch the bh after
117  * unlocking it.
118  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
119  * a race there is benign: unlock_buffer() only use the bh's address for
120  * hashing after unlocking the buffer, so it doesn't actually touch the bh
121  * itself.
122  */
123 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
124 {
125         if (uptodate) {
126                 set_buffer_uptodate(bh);
127         } else {
128                 /* This happens, due to failed READA attempts. */
129                 clear_buffer_uptodate(bh);
130         }
131         unlock_buffer(bh);
132 }
133
134 /*
135  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
136  * unlock the buffer. This is what ll_rw_block uses too.
137  */
138 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
139 {
140         __end_buffer_read_notouch(bh, uptodate);
141         put_bh(bh);
142 }
143 EXPORT_SYMBOL(end_buffer_read_sync);
144
145 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
146 {
147         char b[BDEVNAME_SIZE];
148
149         if (uptodate) {
150                 set_buffer_uptodate(bh);
151         } else {
152                 if (!quiet_error(bh)) {
153                         buffer_io_error(bh);
154                         printk(KERN_WARNING "lost page write due to "
155                                         "I/O error on %s\n",
156                                        bdevname(bh->b_bdev, b));
157                 }
158                 set_buffer_write_io_error(bh);
159                 clear_buffer_uptodate(bh);
160         }
161         unlock_buffer(bh);
162         put_bh(bh);
163 }
164 EXPORT_SYMBOL(end_buffer_write_sync);
165
166 /*
167  * Various filesystems appear to want __find_get_block to be non-blocking.
168  * But it's the page lock which protects the buffers.  To get around this,
169  * we get exclusion from try_to_free_buffers with the blockdev mapping's
170  * private_lock.
171  *
172  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
173  * may be quite high.  This code could TryLock the page, and if that
174  * succeeds, there is no need to take private_lock. (But if
175  * private_lock is contended then so is mapping->tree_lock).
176  */
177 static struct buffer_head *
178 __find_get_block_slow(struct block_device *bdev, sector_t block)
179 {
180         struct inode *bd_inode = bdev->bd_inode;
181         struct address_space *bd_mapping = bd_inode->i_mapping;
182         struct buffer_head *ret = NULL;
183         pgoff_t index;
184         struct buffer_head *bh;
185         struct buffer_head *head;
186         struct page *page;
187         int all_mapped = 1;
188
189         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
190         page = find_get_page(bd_mapping, index);
191         if (!page)
192                 goto out;
193
194         spin_lock(&bd_mapping->private_lock);
195         if (!page_has_buffers(page))
196                 goto out_unlock;
197         head = page_buffers(page);
198         bh = head;
199         do {
200                 if (!buffer_mapped(bh))
201                         all_mapped = 0;
202                 else if (bh->b_blocknr == block) {
203                         ret = bh;
204                         get_bh(bh);
205                         goto out_unlock;
206                 }
207                 bh = bh->b_this_page;
208         } while (bh != head);
209
210         /* we might be here because some of the buffers on this page are
211          * not mapped.  This is due to various races between
212          * file io on the block device and getblk.  It gets dealt with
213          * elsewhere, don't buffer_error if we had some unmapped buffers
214          */
215         if (all_mapped) {
216                 printk("__find_get_block_slow() failed. "
217                         "block=%llu, b_blocknr=%llu\n",
218                         (unsigned long long)block,
219                         (unsigned long long)bh->b_blocknr);
220                 printk("b_state=0x%08lx, b_size=%zu\n",
221                         bh->b_state, bh->b_size);
222                 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
223         }
224 out_unlock:
225         spin_unlock(&bd_mapping->private_lock);
226         page_cache_release(page);
227 out:
228         return ret;
229 }
230
231 /* If invalidate_buffers() will trash dirty buffers, it means some kind
232    of fs corruption is going on. Trashing dirty data always imply losing
233    information that was supposed to be just stored on the physical layer
234    by the user.
235
236    Thus invalidate_buffers in general usage is not allwowed to trash
237    dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
238    be preserved.  These buffers are simply skipped.
239   
240    We also skip buffers which are still in use.  For example this can
241    happen if a userspace program is reading the block device.
242
243    NOTE: In the case where the user removed a removable-media-disk even if
244    there's still dirty data not synced on disk (due a bug in the device driver
245    or due an error of the user), by not destroying the dirty buffers we could
246    generate corruption also on the next media inserted, thus a parameter is
247    necessary to handle this case in the most safe way possible (trying
248    to not corrupt also the new disk inserted with the data belonging to
249    the old now corrupted disk). Also for the ramdisk the natural thing
250    to do in order to release the ramdisk memory is to destroy dirty buffers.
251
252    These are two special cases. Normal usage imply the device driver
253    to issue a sync on the device (without waiting I/O completion) and
254    then an invalidate_buffers call that doesn't trash dirty buffers.
255
256    For handling cache coherency with the blkdev pagecache the 'update' case
257    is been introduced. It is needed to re-read from disk any pinned
258    buffer. NOTE: re-reading from disk is destructive so we can do it only
259    when we assume nobody is changing the buffercache under our I/O and when
260    we think the disk contains more recent information than the buffercache.
261    The update == 1 pass marks the buffers we need to update, the update == 2
262    pass does the actual I/O. */
263 void invalidate_bdev(struct block_device *bdev)
264 {
265         struct address_space *mapping = bdev->bd_inode->i_mapping;
266
267         if (mapping->nrpages == 0)
268                 return;
269
270         invalidate_bh_lrus();
271         lru_add_drain_all();    /* make sure all lru add caches are flushed */
272         invalidate_mapping_pages(mapping, 0, -1);
273         /* 99% of the time, we don't need to flush the cleancache on the bdev.
274          * But, for the strange corners, lets be cautious
275          */
276         cleancache_flush_inode(mapping);
277 }
278 EXPORT_SYMBOL(invalidate_bdev);
279
280 /*
281  * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
282  */
283 static void free_more_memory(void)
284 {
285         struct zone *zone;
286         int nid;
287
288         wakeup_flusher_threads(1024);
289         yield();
290
291         for_each_online_node(nid) {
292                 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
293                                                 gfp_zone(GFP_NOFS), NULL,
294                                                 &zone);
295                 if (zone)
296                         try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
297                                                 GFP_NOFS, NULL);
298         }
299 }
300
301 /*
302  * I/O completion handler for block_read_full_page() - pages
303  * which come unlocked at the end of I/O.
304  */
305 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
306 {
307         unsigned long flags;
308         struct buffer_head *first;
309         struct buffer_head *tmp;
310         struct page *page;
311         int page_uptodate = 1;
312
313         BUG_ON(!buffer_async_read(bh));
314
315         page = bh->b_page;
316         if (uptodate) {
317                 set_buffer_uptodate(bh);
318         } else {
319                 clear_buffer_uptodate(bh);
320                 if (!quiet_error(bh))
321                         buffer_io_error(bh);
322                 SetPageError(page);
323         }
324
325         /*
326          * Be _very_ careful from here on. Bad things can happen if
327          * two buffer heads end IO at almost the same time and both
328          * decide that the page is now completely done.
329          */
330         first = page_buffers(page);
331         local_irq_save(flags);
332         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
333         clear_buffer_async_read(bh);
334         unlock_buffer(bh);
335         tmp = bh;
336         do {
337                 if (!buffer_uptodate(tmp))
338                         page_uptodate = 0;
339                 if (buffer_async_read(tmp)) {
340                         BUG_ON(!buffer_locked(tmp));
341                         goto still_busy;
342                 }
343                 tmp = tmp->b_this_page;
344         } while (tmp != bh);
345         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
346         local_irq_restore(flags);
347
348         /*
349          * If none of the buffers had errors and they are all
350          * uptodate then we can set the page uptodate.
351          */
352         if (page_uptodate && !PageError(page))
353                 SetPageUptodate(page);
354         unlock_page(page);
355         return;
356
357 still_busy:
358         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
359         local_irq_restore(flags);
360         return;
361 }
362
363 /*
364  * Completion handler for block_write_full_page() - pages which are unlocked
365  * during I/O, and which have PageWriteback cleared upon I/O completion.
366  */
367 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
368 {
369         char b[BDEVNAME_SIZE];
370         unsigned long flags;
371         struct buffer_head *first;
372         struct buffer_head *tmp;
373         struct page *page;
374
375         BUG_ON(!buffer_async_write(bh));
376
377         page = bh->b_page;
378         if (uptodate) {
379                 set_buffer_uptodate(bh);
380         } else {
381                 if (!quiet_error(bh)) {
382                         buffer_io_error(bh);
383                         printk(KERN_WARNING "lost page write due to "
384                                         "I/O error on %s\n",
385                                bdevname(bh->b_bdev, b));
386                 }
387                 set_bit(AS_EIO, &page->mapping->flags);
388                 set_buffer_write_io_error(bh);
389                 clear_buffer_uptodate(bh);
390                 SetPageError(page);
391         }
392
393         first = page_buffers(page);
394         local_irq_save(flags);
395         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
396
397         clear_buffer_async_write(bh);
398         unlock_buffer(bh);
399         tmp = bh->b_this_page;
400         while (tmp != bh) {
401                 if (buffer_async_write(tmp)) {
402                         BUG_ON(!buffer_locked(tmp));
403                         goto still_busy;
404                 }
405                 tmp = tmp->b_this_page;
406         }
407         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
408         local_irq_restore(flags);
409         end_page_writeback(page);
410         return;
411
412 still_busy:
413         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
414         local_irq_restore(flags);
415         return;
416 }
417 EXPORT_SYMBOL(end_buffer_async_write);
418
419 /*
420  * If a page's buffers are under async readin (end_buffer_async_read
421  * completion) then there is a possibility that another thread of
422  * control could lock one of the buffers after it has completed
423  * but while some of the other buffers have not completed.  This
424  * locked buffer would confuse end_buffer_async_read() into not unlocking
425  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
426  * that this buffer is not under async I/O.
427  *
428  * The page comes unlocked when it has no locked buffer_async buffers
429  * left.
430  *
431  * PageLocked prevents anyone starting new async I/O reads any of
432  * the buffers.
433  *
434  * PageWriteback is used to prevent simultaneous writeout of the same
435  * page.
436  *
437  * PageLocked prevents anyone from starting writeback of a page which is
438  * under read I/O (PageWriteback is only ever set against a locked page).
439  */
440 static void mark_buffer_async_read(struct buffer_head *bh)
441 {
442         bh->b_end_io = end_buffer_async_read;
443         set_buffer_async_read(bh);
444 }
445
446 static void mark_buffer_async_write_endio(struct buffer_head *bh,
447                                           bh_end_io_t *handler)
448 {
449         bh->b_end_io = handler;
450         set_buffer_async_write(bh);
451 }
452
453 void mark_buffer_async_write(struct buffer_head *bh)
454 {
455         mark_buffer_async_write_endio(bh, end_buffer_async_write);
456 }
457 EXPORT_SYMBOL(mark_buffer_async_write);
458
459
460 /*
461  * fs/buffer.c contains helper functions for buffer-backed address space's
462  * fsync functions.  A common requirement for buffer-based filesystems is
463  * that certain data from the backing blockdev needs to be written out for
464  * a successful fsync().  For example, ext2 indirect blocks need to be
465  * written back and waited upon before fsync() returns.
466  *
467  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
468  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
469  * management of a list of dependent buffers at ->i_mapping->private_list.
470  *
471  * Locking is a little subtle: try_to_free_buffers() will remove buffers
472  * from their controlling inode's queue when they are being freed.  But
473  * try_to_free_buffers() will be operating against the *blockdev* mapping
474  * at the time, not against the S_ISREG file which depends on those buffers.
475  * So the locking for private_list is via the private_lock in the address_space
476  * which backs the buffers.  Which is different from the address_space 
477  * against which the buffers are listed.  So for a particular address_space,
478  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
479  * mapping->private_list will always be protected by the backing blockdev's
480  * ->private_lock.
481  *
482  * Which introduces a requirement: all buffers on an address_space's
483  * ->private_list must be from the same address_space: the blockdev's.
484  *
485  * address_spaces which do not place buffers at ->private_list via these
486  * utility functions are free to use private_lock and private_list for
487  * whatever they want.  The only requirement is that list_empty(private_list)
488  * be true at clear_inode() time.
489  *
490  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
491  * filesystems should do that.  invalidate_inode_buffers() should just go
492  * BUG_ON(!list_empty).
493  *
494  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
495  * take an address_space, not an inode.  And it should be called
496  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
497  * queued up.
498  *
499  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
500  * list if it is already on a list.  Because if the buffer is on a list,
501  * it *must* already be on the right one.  If not, the filesystem is being
502  * silly.  This will save a ton of locking.  But first we have to ensure
503  * that buffers are taken *off* the old inode's list when they are freed
504  * (presumably in truncate).  That requires careful auditing of all
505  * filesystems (do it inside bforget()).  It could also be done by bringing
506  * b_inode back.
507  */
508
509 /*
510  * The buffer's backing address_space's private_lock must be held
511  */
512 static void __remove_assoc_queue(struct buffer_head *bh)
513 {
514         list_del_init(&bh->b_assoc_buffers);
515         WARN_ON(!bh->b_assoc_map);
516         if (buffer_write_io_error(bh))
517                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
518         bh->b_assoc_map = NULL;
519 }
520
521 int inode_has_buffers(struct inode *inode)
522 {
523         return !list_empty(&inode->i_data.private_list);
524 }
525
526 /*
527  * osync is designed to support O_SYNC io.  It waits synchronously for
528  * all already-submitted IO to complete, but does not queue any new
529  * writes to the disk.
530  *
531  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
532  * you dirty the buffers, and then use osync_inode_buffers to wait for
533  * completion.  Any other dirty buffers which are not yet queued for
534  * write will not be flushed to disk by the osync.
535  */
536 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
537 {
538         struct buffer_head *bh;
539         struct list_head *p;
540         int err = 0;
541
542         spin_lock(lock);
543 repeat:
544         list_for_each_prev(p, list) {
545                 bh = BH_ENTRY(p);
546                 if (buffer_locked(bh)) {
547                         get_bh(bh);
548                         spin_unlock(lock);
549                         wait_on_buffer(bh);
550                         if (!buffer_uptodate(bh))
551                                 err = -EIO;
552                         brelse(bh);
553                         spin_lock(lock);
554                         goto repeat;
555                 }
556         }
557         spin_unlock(lock);
558         return err;
559 }
560
561 static void do_thaw_one(struct super_block *sb, void *unused)
562 {
563         char b[BDEVNAME_SIZE];
564         while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
565                 printk(KERN_WARNING "Emergency Thaw on %s\n",
566                        bdevname(sb->s_bdev, b));
567 }
568
569 static void do_thaw_all(struct work_struct *work)
570 {
571         iterate_supers(do_thaw_one, NULL);
572         kfree(work);
573         printk(KERN_WARNING "Emergency Thaw complete\n");
574 }
575
576 /**
577  * emergency_thaw_all -- forcibly thaw every frozen filesystem
578  *
579  * Used for emergency unfreeze of all filesystems via SysRq
580  */
581 void emergency_thaw_all(void)
582 {
583         struct work_struct *work;
584
585         work = kmalloc(sizeof(*work), GFP_ATOMIC);
586         if (work) {
587                 INIT_WORK(work, do_thaw_all);
588                 schedule_work(work);
589         }
590 }
591
592 /**
593  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
594  * @mapping: the mapping which wants those buffers written
595  *
596  * Starts I/O against the buffers at mapping->private_list, and waits upon
597  * that I/O.
598  *
599  * Basically, this is a convenience function for fsync().
600  * @mapping is a file or directory which needs those buffers to be written for
601  * a successful fsync().
602  */
603 int sync_mapping_buffers(struct address_space *mapping)
604 {
605         struct address_space *buffer_mapping = mapping->assoc_mapping;
606
607         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
608                 return 0;
609
610         return fsync_buffers_list(&buffer_mapping->private_lock,
611                                         &mapping->private_list);
612 }
613 EXPORT_SYMBOL(sync_mapping_buffers);
614
615 /*
616  * Called when we've recently written block `bblock', and it is known that
617  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
618  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
619  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
620  */
621 void write_boundary_block(struct block_device *bdev,
622                         sector_t bblock, unsigned blocksize)
623 {
624         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
625         if (bh) {
626                 if (buffer_dirty(bh))
627                         ll_rw_block(WRITE, 1, &bh);
628                 put_bh(bh);
629         }
630 }
631
632 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
633 {
634         struct address_space *mapping = inode->i_mapping;
635         struct address_space *buffer_mapping = bh->b_page->mapping;
636
637         mark_buffer_dirty(bh);
638         if (!mapping->assoc_mapping) {
639                 mapping->assoc_mapping = buffer_mapping;
640         } else {
641                 BUG_ON(mapping->assoc_mapping != buffer_mapping);
642         }
643         if (!bh->b_assoc_map) {
644                 spin_lock(&buffer_mapping->private_lock);
645                 list_move_tail(&bh->b_assoc_buffers,
646                                 &mapping->private_list);
647                 bh->b_assoc_map = mapping;
648                 spin_unlock(&buffer_mapping->private_lock);
649         }
650 }
651 EXPORT_SYMBOL(mark_buffer_dirty_inode);
652
653 /*
654  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
655  * dirty.
656  *
657  * If warn is true, then emit a warning if the page is not uptodate and has
658  * not been truncated.
659  */
660 static void __set_page_dirty(struct page *page,
661                 struct address_space *mapping, int warn)
662 {
663         spin_lock_irq(&mapping->tree_lock);
664         if (page->mapping) {    /* Race with truncate? */
665                 WARN_ON_ONCE(warn && !PageUptodate(page));
666                 account_page_dirtied(page, mapping);
667                 radix_tree_tag_set(&mapping->page_tree,
668                                 page_index(page), PAGECACHE_TAG_DIRTY);
669         }
670         spin_unlock_irq(&mapping->tree_lock);
671         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
672 }
673
674 /*
675  * Add a page to the dirty page list.
676  *
677  * It is a sad fact of life that this function is called from several places
678  * deeply under spinlocking.  It may not sleep.
679  *
680  * If the page has buffers, the uptodate buffers are set dirty, to preserve
681  * dirty-state coherency between the page and the buffers.  It the page does
682  * not have buffers then when they are later attached they will all be set
683  * dirty.
684  *
685  * The buffers are dirtied before the page is dirtied.  There's a small race
686  * window in which a writepage caller may see the page cleanness but not the
687  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
688  * before the buffers, a concurrent writepage caller could clear the page dirty
689  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
690  * page on the dirty page list.
691  *
692  * We use private_lock to lock against try_to_free_buffers while using the
693  * page's buffer list.  Also use this to protect against clean buffers being
694  * added to the page after it was set dirty.
695  *
696  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
697  * address_space though.
698  */
699 int __set_page_dirty_buffers(struct page *page)
700 {
701         int newly_dirty;
702         struct address_space *mapping = page_mapping(page);
703
704         if (unlikely(!mapping))
705                 return !TestSetPageDirty(page);
706
707         spin_lock(&mapping->private_lock);
708         if (page_has_buffers(page)) {
709                 struct buffer_head *head = page_buffers(page);
710                 struct buffer_head *bh = head;
711
712                 do {
713                         set_buffer_dirty(bh);
714                         bh = bh->b_this_page;
715                 } while (bh != head);
716         }
717         newly_dirty = !TestSetPageDirty(page);
718         spin_unlock(&mapping->private_lock);
719
720         if (newly_dirty)
721                 __set_page_dirty(page, mapping, 1);
722         return newly_dirty;
723 }
724 EXPORT_SYMBOL(__set_page_dirty_buffers);
725
726 /*
727  * Write out and wait upon a list of buffers.
728  *
729  * We have conflicting pressures: we want to make sure that all
730  * initially dirty buffers get waited on, but that any subsequently
731  * dirtied buffers don't.  After all, we don't want fsync to last
732  * forever if somebody is actively writing to the file.
733  *
734  * Do this in two main stages: first we copy dirty buffers to a
735  * temporary inode list, queueing the writes as we go.  Then we clean
736  * up, waiting for those writes to complete.
737  * 
738  * During this second stage, any subsequent updates to the file may end
739  * up refiling the buffer on the original inode's dirty list again, so
740  * there is a chance we will end up with a buffer queued for write but
741  * not yet completed on that list.  So, as a final cleanup we go through
742  * the osync code to catch these locked, dirty buffers without requeuing
743  * any newly dirty buffers for write.
744  */
745 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
746 {
747         struct buffer_head *bh;
748         struct list_head tmp;
749         struct address_space *mapping;
750         int err = 0, err2;
751         struct blk_plug plug;
752
753         INIT_LIST_HEAD(&tmp);
754         blk_start_plug(&plug);
755
756         spin_lock(lock);
757         while (!list_empty(list)) {
758                 bh = BH_ENTRY(list->next);
759                 mapping = bh->b_assoc_map;
760                 __remove_assoc_queue(bh);
761                 /* Avoid race with mark_buffer_dirty_inode() which does
762                  * a lockless check and we rely on seeing the dirty bit */
763                 smp_mb();
764                 if (buffer_dirty(bh) || buffer_locked(bh)) {
765                         list_add(&bh->b_assoc_buffers, &tmp);
766                         bh->b_assoc_map = mapping;
767                         if (buffer_dirty(bh)) {
768                                 get_bh(bh);
769                                 spin_unlock(lock);
770                                 /*
771                                  * Ensure any pending I/O completes so that
772                                  * write_dirty_buffer() actually writes the
773                                  * current contents - it is a noop if I/O is
774                                  * still in flight on potentially older
775                                  * contents.
776                                  */
777                                 write_dirty_buffer(bh, WRITE_SYNC);
778
779                                 /*
780                                  * Kick off IO for the previous mapping. Note
781                                  * that we will not run the very last mapping,
782                                  * wait_on_buffer() will do that for us
783                                  * through sync_buffer().
784                                  */
785                                 brelse(bh);
786                                 spin_lock(lock);
787                         }
788                 }
789         }
790
791         spin_unlock(lock);
792         blk_finish_plug(&plug);
793         spin_lock(lock);
794
795         while (!list_empty(&tmp)) {
796                 bh = BH_ENTRY(tmp.prev);
797                 get_bh(bh);
798                 mapping = bh->b_assoc_map;
799                 __remove_assoc_queue(bh);
800                 /* Avoid race with mark_buffer_dirty_inode() which does
801                  * a lockless check and we rely on seeing the dirty bit */
802                 smp_mb();
803                 if (buffer_dirty(bh)) {
804                         list_add(&bh->b_assoc_buffers,
805                                  &mapping->private_list);
806                         bh->b_assoc_map = mapping;
807                 }
808                 spin_unlock(lock);
809                 wait_on_buffer(bh);
810                 if (!buffer_uptodate(bh))
811                         err = -EIO;
812                 brelse(bh);
813                 spin_lock(lock);
814         }
815         
816         spin_unlock(lock);
817         err2 = osync_buffers_list(lock, list);
818         if (err)
819                 return err;
820         else
821                 return err2;
822 }
823
824 /*
825  * Invalidate any and all dirty buffers on a given inode.  We are
826  * probably unmounting the fs, but that doesn't mean we have already
827  * done a sync().  Just drop the buffers from the inode list.
828  *
829  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
830  * assumes that all the buffers are against the blockdev.  Not true
831  * for reiserfs.
832  */
833 void invalidate_inode_buffers(struct inode *inode)
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                         __remove_assoc_queue(BH_ENTRY(list->next));
843                 spin_unlock(&buffer_mapping->private_lock);
844         }
845 }
846 EXPORT_SYMBOL(invalidate_inode_buffers);
847
848 /*
849  * Remove any clean buffers from the inode's buffer list.  This is called
850  * when we're trying to free the inode itself.  Those buffers can pin it.
851  *
852  * Returns true if all buffers were removed.
853  */
854 int remove_inode_buffers(struct inode *inode)
855 {
856         int ret = 1;
857
858         if (inode_has_buffers(inode)) {
859                 struct address_space *mapping = &inode->i_data;
860                 struct list_head *list = &mapping->private_list;
861                 struct address_space *buffer_mapping = mapping->assoc_mapping;
862
863                 spin_lock(&buffer_mapping->private_lock);
864                 while (!list_empty(list)) {
865                         struct buffer_head *bh = BH_ENTRY(list->next);
866                         if (buffer_dirty(bh)) {
867                                 ret = 0;
868                                 break;
869                         }
870                         __remove_assoc_queue(bh);
871                 }
872                 spin_unlock(&buffer_mapping->private_lock);
873         }
874         return ret;
875 }
876
877 /*
878  * Create the appropriate buffers when given a page for data area and
879  * the size of each buffer.. Use the bh->b_this_page linked list to
880  * follow the buffers created.  Return NULL if unable to create more
881  * buffers.
882  *
883  * The retry flag is used to differentiate async IO (paging, swapping)
884  * which may not fail from ordinary buffer allocations.
885  */
886 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
887                 int retry)
888 {
889         struct buffer_head *bh, *head;
890         long offset;
891
892 try_again:
893         head = NULL;
894         offset = PAGE_SIZE;
895         while ((offset -= size) >= 0) {
896                 bh = alloc_buffer_head(GFP_NOFS);
897                 if (!bh)
898                         goto no_grow;
899
900                 bh->b_bdev = NULL;
901                 bh->b_this_page = head;
902                 bh->b_blocknr = -1;
903                 head = bh;
904
905                 bh->b_state = 0;
906                 atomic_set(&bh->b_count, 0);
907                 bh->b_size = size;
908
909                 /* Link the buffer to its page */
910                 set_bh_page(bh, page, offset);
911
912                 init_buffer(bh, NULL, NULL);
913         }
914         return head;
915 /*
916  * In case anything failed, we just free everything we got.
917  */
918 no_grow:
919         if (head) {
920                 do {
921                         bh = head;
922                         head = head->b_this_page;
923                         free_buffer_head(bh);
924                 } while (head);
925         }
926
927         /*
928          * Return failure for non-async IO requests.  Async IO requests
929          * are not allowed to fail, so we have to wait until buffer heads
930          * become available.  But we don't want tasks sleeping with 
931          * partially complete buffers, so all were released above.
932          */
933         if (!retry)
934                 return NULL;
935
936         /* We're _really_ low on memory. Now we just
937          * wait for old buffer heads to become free due to
938          * finishing IO.  Since this is an async request and
939          * the reserve list is empty, we're sure there are 
940          * async buffer heads in use.
941          */
942         free_more_memory();
943         goto try_again;
944 }
945 EXPORT_SYMBOL_GPL(alloc_page_buffers);
946
947 static inline void
948 link_dev_buffers(struct page *page, struct buffer_head *head)
949 {
950         struct buffer_head *bh, *tail;
951
952         bh = head;
953         do {
954                 tail = bh;
955                 bh = bh->b_this_page;
956         } while (bh);
957         tail->b_this_page = head;
958         attach_page_buffers(page, head);
959 }
960
961 /*
962  * Initialise the state of a blockdev page's buffers.
963  */ 
964 static void
965 init_page_buffers(struct page *page, struct block_device *bdev,
966                         sector_t block, int size)
967 {
968         struct buffer_head *head = page_buffers(page);
969         struct buffer_head *bh = head;
970         int uptodate = PageUptodate(page);
971
972         do {
973                 if (!buffer_mapped(bh)) {
974                         init_buffer(bh, NULL, NULL);
975                         bh->b_bdev = bdev;
976                         bh->b_blocknr = block;
977                         if (uptodate)
978                                 set_buffer_uptodate(bh);
979                         set_buffer_mapped(bh);
980                 }
981                 block++;
982                 bh = bh->b_this_page;
983         } while (bh != head);
984 }
985
986 /*
987  * Create the page-cache page that contains the requested block.
988  *
989  * This is user purely for blockdev mappings.
990  */
991 static struct page *
992 grow_dev_page(struct block_device *bdev, sector_t block,
993                 pgoff_t index, int size)
994 {
995         struct inode *inode = bdev->bd_inode;
996         struct page *page;
997         struct buffer_head *bh;
998
999         page = find_or_create_page(inode->i_mapping, index,
1000                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
1001         if (!page)
1002                 return NULL;
1003
1004         BUG_ON(!PageLocked(page));
1005
1006         if (page_has_buffers(page)) {
1007                 bh = page_buffers(page);
1008                 if (bh->b_size == size) {
1009                         init_page_buffers(page, bdev, block, size);
1010                         return page;
1011                 }
1012                 if (!try_to_free_buffers(page))
1013                         goto failed;
1014         }
1015
1016         /*
1017          * Allocate some buffers for this page
1018          */
1019         bh = alloc_page_buffers(page, size, 0);
1020         if (!bh)
1021                 goto failed;
1022
1023         /*
1024          * Link the page to the buffers and initialise them.  Take the
1025          * lock to be atomic wrt __find_get_block(), which does not
1026          * run under the page lock.
1027          */
1028         spin_lock(&inode->i_mapping->private_lock);
1029         link_dev_buffers(page, bh);
1030         init_page_buffers(page, bdev, block, size);
1031         spin_unlock(&inode->i_mapping->private_lock);
1032         return page;
1033
1034 failed:
1035         BUG();
1036         unlock_page(page);
1037         page_cache_release(page);
1038         return NULL;
1039 }
1040
1041 /*
1042  * Create buffers for the specified block device block's page.  If
1043  * that page was dirty, the buffers are set dirty also.
1044  */
1045 static int
1046 grow_buffers(struct block_device *bdev, sector_t block, int size)
1047 {
1048         struct page *page;
1049         pgoff_t index;
1050         int sizebits;
1051
1052         sizebits = -1;
1053         do {
1054                 sizebits++;
1055         } while ((size << sizebits) < PAGE_SIZE);
1056
1057         index = block >> sizebits;
1058
1059         /*
1060          * Check for a block which wants to lie outside our maximum possible
1061          * pagecache index.  (this comparison is done using sector_t types).
1062          */
1063         if (unlikely(index != block >> sizebits)) {
1064                 char b[BDEVNAME_SIZE];
1065
1066                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1067                         "device %s\n",
1068                         __func__, (unsigned long long)block,
1069                         bdevname(bdev, b));
1070                 return -EIO;
1071         }
1072         block = index << sizebits;
1073         /* Create a page with the proper size buffers.. */
1074         page = grow_dev_page(bdev, block, index, size);
1075         if (!page)
1076                 return 0;
1077         unlock_page(page);
1078         page_cache_release(page);
1079         return 1;
1080 }
1081
1082 static struct buffer_head *
1083 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1084 {
1085         /* Size must be multiple of hard sectorsize */
1086         if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1087                         (size < 512 || size > PAGE_SIZE))) {
1088                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1089                                         size);
1090                 printk(KERN_ERR "logical block size: %d\n",
1091                                         bdev_logical_block_size(bdev));
1092
1093                 dump_stack();
1094                 return NULL;
1095         }
1096
1097         for (;;) {
1098                 struct buffer_head * bh;
1099                 int ret;
1100
1101                 bh = __find_get_block(bdev, block, size);
1102                 if (bh)
1103                         return bh;
1104
1105                 ret = grow_buffers(bdev, block, size);
1106                 if (ret < 0)
1107                         return NULL;
1108                 if (ret == 0)
1109                         free_more_memory();
1110         }
1111 }
1112
1113 /*
1114  * The relationship between dirty buffers and dirty pages:
1115  *
1116  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1117  * the page is tagged dirty in its radix tree.
1118  *
1119  * At all times, the dirtiness of the buffers represents the dirtiness of
1120  * subsections of the page.  If the page has buffers, the page dirty bit is
1121  * merely a hint about the true dirty state.
1122  *
1123  * When a page is set dirty in its entirety, all its buffers are marked dirty
1124  * (if the page has buffers).
1125  *
1126  * When a buffer is marked dirty, its page is dirtied, but the page's other
1127  * buffers are not.
1128  *
1129  * Also.  When blockdev buffers are explicitly read with bread(), they
1130  * individually become uptodate.  But their backing page remains not
1131  * uptodate - even if all of its buffers are uptodate.  A subsequent
1132  * block_read_full_page() against that page will discover all the uptodate
1133  * buffers, will set the page uptodate and will perform no I/O.
1134  */
1135
1136 /**
1137  * mark_buffer_dirty - mark a buffer_head as needing writeout
1138  * @bh: the buffer_head to mark dirty
1139  *
1140  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1141  * backing page dirty, then tag the page as dirty in its address_space's radix
1142  * tree and then attach the address_space's inode to its superblock's dirty
1143  * inode list.
1144  *
1145  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1146  * mapping->tree_lock and mapping->host->i_lock.
1147  */
1148 void mark_buffer_dirty(struct buffer_head *bh)
1149 {
1150         WARN_ON_ONCE(!buffer_uptodate(bh));
1151
1152         /*
1153          * Very *carefully* optimize the it-is-already-dirty case.
1154          *
1155          * Don't let the final "is it dirty" escape to before we
1156          * perhaps modified the buffer.
1157          */
1158         if (buffer_dirty(bh)) {
1159                 smp_mb();
1160                 if (buffer_dirty(bh))
1161                         return;
1162         }
1163
1164         if (!test_set_buffer_dirty(bh)) {
1165                 struct page *page = bh->b_page;
1166                 if (!TestSetPageDirty(page)) {
1167                         struct address_space *mapping = page_mapping(page);
1168                         if (mapping)
1169                                 __set_page_dirty(page, mapping, 0);
1170                 }
1171         }
1172 }
1173 EXPORT_SYMBOL(mark_buffer_dirty);
1174
1175 /*
1176  * Decrement a buffer_head's reference count.  If all buffers against a page
1177  * have zero reference count, are clean and unlocked, and if the page is clean
1178  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1179  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1180  * a page but it ends up not being freed, and buffers may later be reattached).
1181  */
1182 void __brelse(struct buffer_head * buf)
1183 {
1184         if (atomic_read(&buf->b_count)) {
1185                 put_bh(buf);
1186                 return;
1187         }
1188         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1189 }
1190 EXPORT_SYMBOL(__brelse);
1191
1192 /*
1193  * bforget() is like brelse(), except it discards any
1194  * potentially dirty data.
1195  */
1196 void __bforget(struct buffer_head *bh)
1197 {
1198         clear_buffer_dirty(bh);
1199         if (bh->b_assoc_map) {
1200                 struct address_space *buffer_mapping = bh->b_page->mapping;
1201
1202                 spin_lock(&buffer_mapping->private_lock);
1203                 list_del_init(&bh->b_assoc_buffers);
1204                 bh->b_assoc_map = NULL;
1205                 spin_unlock(&buffer_mapping->private_lock);
1206         }
1207         __brelse(bh);
1208 }
1209 EXPORT_SYMBOL(__bforget);
1210
1211 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1212 {
1213         lock_buffer(bh);
1214         if (buffer_uptodate(bh)) {
1215                 unlock_buffer(bh);
1216                 return bh;
1217         } else {
1218                 get_bh(bh);
1219                 bh->b_end_io = end_buffer_read_sync;
1220                 submit_bh(READ, bh);
1221                 wait_on_buffer(bh);
1222                 if (buffer_uptodate(bh))
1223                         return bh;
1224         }
1225         brelse(bh);
1226         return NULL;
1227 }
1228
1229 /*
1230  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1231  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1232  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1233  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1234  * CPU's LRUs at the same time.
1235  *
1236  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1237  * sb_find_get_block().
1238  *
1239  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1240  * a local interrupt disable for that.
1241  */
1242
1243 #define BH_LRU_SIZE     8
1244
1245 struct bh_lru {
1246         struct buffer_head *bhs[BH_LRU_SIZE];
1247 };
1248
1249 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1250
1251 #ifdef CONFIG_SMP
1252 #define bh_lru_lock()   local_irq_disable()
1253 #define bh_lru_unlock() local_irq_enable()
1254 #else
1255 #define bh_lru_lock()   preempt_disable()
1256 #define bh_lru_unlock() preempt_enable()
1257 #endif
1258
1259 static inline void check_irqs_on(void)
1260 {
1261 #ifdef irqs_disabled
1262         BUG_ON(irqs_disabled());
1263 #endif
1264 }
1265
1266 /*
1267  * The LRU management algorithm is dopey-but-simple.  Sorry.
1268  */
1269 static void bh_lru_install(struct buffer_head *bh)
1270 {
1271         struct buffer_head *evictee = NULL;
1272
1273         check_irqs_on();
1274         bh_lru_lock();
1275         if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1276                 struct buffer_head *bhs[BH_LRU_SIZE];
1277                 int in;
1278                 int out = 0;
1279
1280                 get_bh(bh);
1281                 bhs[out++] = bh;
1282                 for (in = 0; in < BH_LRU_SIZE; in++) {
1283                         struct buffer_head *bh2 =
1284                                 __this_cpu_read(bh_lrus.bhs[in]);
1285
1286                         if (bh2 == bh) {
1287                                 __brelse(bh2);
1288                         } else {
1289                                 if (out >= BH_LRU_SIZE) {
1290                                         BUG_ON(evictee != NULL);
1291                                         evictee = bh2;
1292                                 } else {
1293                                         bhs[out++] = bh2;
1294                                 }
1295                         }
1296                 }
1297                 while (out < BH_LRU_SIZE)
1298                         bhs[out++] = NULL;
1299                 memcpy(__this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1300         }
1301         bh_lru_unlock();
1302
1303         if (evictee)
1304                 __brelse(evictee);
1305 }
1306
1307 /*
1308  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1309  */
1310 static struct buffer_head *
1311 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1312 {
1313         struct buffer_head *ret = NULL;
1314         unsigned int i;
1315
1316         check_irqs_on();
1317         bh_lru_lock();
1318         for (i = 0; i < BH_LRU_SIZE; i++) {
1319                 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1320
1321                 if (bh && bh->b_bdev == bdev &&
1322                                 bh->b_blocknr == block && bh->b_size == size) {
1323                         if (i) {
1324                                 while (i) {
1325                                         __this_cpu_write(bh_lrus.bhs[i],
1326                                                 __this_cpu_read(bh_lrus.bhs[i - 1]));
1327                                         i--;
1328                                 }
1329                                 __this_cpu_write(bh_lrus.bhs[0], bh);
1330                         }
1331                         get_bh(bh);
1332                         ret = bh;
1333                         break;
1334                 }
1335         }
1336         bh_lru_unlock();
1337         return ret;
1338 }
1339
1340 /*
1341  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1342  * it in the LRU and mark it as accessed.  If it is not present then return
1343  * NULL
1344  */
1345 struct buffer_head *
1346 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1347 {
1348         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1349
1350         if (bh == NULL) {
1351                 bh = __find_get_block_slow(bdev, block);
1352                 if (bh)
1353                         bh_lru_install(bh);
1354         }
1355         if (bh)
1356                 touch_buffer(bh);
1357         return bh;
1358 }
1359 EXPORT_SYMBOL(__find_get_block);
1360
1361 /*
1362  * __getblk will locate (and, if necessary, create) the buffer_head
1363  * which corresponds to the passed block_device, block and size. The
1364  * returned buffer has its reference count incremented.
1365  *
1366  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1367  * illegal block number, __getblk() will happily return a buffer_head
1368  * which represents the non-existent block.  Very weird.
1369  *
1370  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1371  * attempt is failing.  FIXME, perhaps?
1372  */
1373 struct buffer_head *
1374 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1375 {
1376         struct buffer_head *bh = __find_get_block(bdev, block, size);
1377
1378         might_sleep();
1379         if (bh == NULL)
1380                 bh = __getblk_slow(bdev, block, size);
1381         return bh;
1382 }
1383 EXPORT_SYMBOL(__getblk);
1384
1385 /*
1386  * Do async read-ahead on a buffer..
1387  */
1388 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1389 {
1390         struct buffer_head *bh = __getblk(bdev, block, size);
1391         if (likely(bh)) {
1392                 ll_rw_block(READA, 1, &bh);
1393                 brelse(bh);
1394         }
1395 }
1396 EXPORT_SYMBOL(__breadahead);
1397
1398 /**
1399  *  __bread() - reads a specified block and returns the bh
1400  *  @bdev: the block_device to read from
1401  *  @block: number of block
1402  *  @size: size (in bytes) to read
1403  * 
1404  *  Reads a specified block, and returns buffer head that contains it.
1405  *  It returns NULL if the block was unreadable.
1406  */
1407 struct buffer_head *
1408 __bread(struct block_device *bdev, sector_t block, unsigned size)
1409 {
1410         struct buffer_head *bh = __getblk(bdev, block, size);
1411
1412         if (likely(bh) && !buffer_uptodate(bh))
1413                 bh = __bread_slow(bh);
1414         return bh;
1415 }
1416 EXPORT_SYMBOL(__bread);
1417
1418 /*
1419  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1420  * This doesn't race because it runs in each cpu either in irq
1421  * or with preempt disabled.
1422  */
1423 static void invalidate_bh_lru(void *arg)
1424 {
1425         struct bh_lru *b = &get_cpu_var(bh_lrus);
1426         int i;
1427
1428         for (i = 0; i < BH_LRU_SIZE; i++) {
1429                 brelse(b->bhs[i]);
1430                 b->bhs[i] = NULL;
1431         }
1432         put_cpu_var(bh_lrus);
1433 }
1434         
1435 void invalidate_bh_lrus(void)
1436 {
1437         on_each_cpu(invalidate_bh_lru, NULL, 1);
1438 }
1439 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1440
1441 void set_bh_page(struct buffer_head *bh,
1442                 struct page *page, unsigned long offset)
1443 {
1444         bh->b_page = page;
1445         BUG_ON(offset >= PAGE_SIZE);
1446         if (PageHighMem(page))
1447                 /*
1448                  * This catches illegal uses and preserves the offset:
1449                  */
1450                 bh->b_data = (char *)(0 + offset);
1451         else
1452                 bh->b_data = page_address(page) + offset;
1453 }
1454 EXPORT_SYMBOL(set_bh_page);
1455
1456 /*
1457  * Called when truncating a buffer on a page completely.
1458  */
1459 static void discard_buffer(struct buffer_head * bh)
1460 {
1461         lock_buffer(bh);
1462         clear_buffer_dirty(bh);
1463         bh->b_bdev = NULL;
1464         clear_buffer_mapped(bh);
1465         clear_buffer_req(bh);
1466         clear_buffer_new(bh);
1467         clear_buffer_delay(bh);
1468         clear_buffer_unwritten(bh);
1469         unlock_buffer(bh);
1470 }
1471
1472 /**
1473  * block_invalidatepage - invalidate part of all of a buffer-backed page
1474  *
1475  * @page: the page which is affected
1476  * @offset: the index of the truncation point
1477  *
1478  * block_invalidatepage() is called when all or part of the page has become
1479  * invalidatedby a truncate operation.
1480  *
1481  * block_invalidatepage() does not have to release all buffers, but it must
1482  * ensure that no dirty buffer is left outside @offset and that no I/O
1483  * is underway against any of the blocks which are outside the truncation
1484  * point.  Because the caller is about to free (and possibly reuse) those
1485  * blocks on-disk.
1486  */
1487 void block_invalidatepage(struct page *page, unsigned long offset)
1488 {
1489         struct buffer_head *head, *bh, *next;
1490         unsigned int curr_off = 0;
1491
1492         BUG_ON(!PageLocked(page));
1493         if (!page_has_buffers(page))
1494                 goto out;
1495
1496         head = page_buffers(page);
1497         bh = head;
1498         do {
1499                 unsigned int next_off = curr_off + bh->b_size;
1500                 next = bh->b_this_page;
1501
1502                 /*
1503                  * is this block fully invalidated?
1504                  */
1505                 if (offset <= curr_off)
1506                         discard_buffer(bh);
1507                 curr_off = next_off;
1508                 bh = next;
1509         } while (bh != head);
1510
1511         /*
1512          * We release buffers only if the entire page is being invalidated.
1513          * The get_block cached value has been unconditionally invalidated,
1514          * so real IO is not possible anymore.
1515          */
1516         if (offset == 0)
1517                 try_to_release_page(page, 0);
1518 out:
1519         return;
1520 }
1521 EXPORT_SYMBOL(block_invalidatepage);
1522
1523 /*
1524  * We attach and possibly dirty the buffers atomically wrt
1525  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1526  * is already excluded via the page lock.
1527  */
1528 void create_empty_buffers(struct page *page,
1529                         unsigned long blocksize, unsigned long b_state)
1530 {
1531         struct buffer_head *bh, *head, *tail;
1532
1533         head = alloc_page_buffers(page, blocksize, 1);
1534         bh = head;
1535         do {
1536                 bh->b_state |= b_state;
1537                 tail = bh;
1538                 bh = bh->b_this_page;
1539         } while (bh);
1540         tail->b_this_page = head;
1541
1542         spin_lock(&page->mapping->private_lock);
1543         if (PageUptodate(page) || PageDirty(page)) {
1544                 bh = head;
1545                 do {
1546                         if (PageDirty(page))
1547                                 set_buffer_dirty(bh);
1548                         if (PageUptodate(page))
1549                                 set_buffer_uptodate(bh);
1550                         bh = bh->b_this_page;
1551                 } while (bh != head);
1552         }
1553         attach_page_buffers(page, head);
1554         spin_unlock(&page->mapping->private_lock);
1555 }
1556 EXPORT_SYMBOL(create_empty_buffers);
1557
1558 /*
1559  * We are taking a block for data and we don't want any output from any
1560  * buffer-cache aliases starting from return from that function and
1561  * until the moment when something will explicitly mark the buffer
1562  * dirty (hopefully that will not happen until we will free that block ;-)
1563  * We don't even need to mark it not-uptodate - nobody can expect
1564  * anything from a newly allocated buffer anyway. We used to used
1565  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1566  * don't want to mark the alias unmapped, for example - it would confuse
1567  * anyone who might pick it with bread() afterwards...
1568  *
1569  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1570  * be writeout I/O going on against recently-freed buffers.  We don't
1571  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1572  * only if we really need to.  That happens here.
1573  */
1574 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1575 {
1576         struct buffer_head *old_bh;
1577
1578         might_sleep();
1579
1580         old_bh = __find_get_block_slow(bdev, block);
1581         if (old_bh) {
1582                 clear_buffer_dirty(old_bh);
1583                 wait_on_buffer(old_bh);
1584                 clear_buffer_req(old_bh);
1585                 __brelse(old_bh);
1586         }
1587 }
1588 EXPORT_SYMBOL(unmap_underlying_metadata);
1589
1590 /*
1591  * NOTE! All mapped/uptodate combinations are valid:
1592  *
1593  *      Mapped  Uptodate        Meaning
1594  *
1595  *      No      No              "unknown" - must do get_block()
1596  *      No      Yes             "hole" - zero-filled
1597  *      Yes     No              "allocated" - allocated on disk, not read in
1598  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1599  *
1600  * "Dirty" is valid only with the last case (mapped+uptodate).
1601  */
1602
1603 /*
1604  * While block_write_full_page is writing back the dirty buffers under
1605  * the page lock, whoever dirtied the buffers may decide to clean them
1606  * again at any time.  We handle that by only looking at the buffer
1607  * state inside lock_buffer().
1608  *
1609  * If block_write_full_page() is called for regular writeback
1610  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1611  * locked buffer.   This only can happen if someone has written the buffer
1612  * directly, with submit_bh().  At the address_space level PageWriteback
1613  * prevents this contention from occurring.
1614  *
1615  * If block_write_full_page() is called with wbc->sync_mode ==
1616  * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1617  * causes the writes to be flagged as synchronous writes.
1618  */
1619 static int __block_write_full_page(struct inode *inode, struct page *page,
1620                         get_block_t *get_block, struct writeback_control *wbc,
1621                         bh_end_io_t *handler)
1622 {
1623         int err;
1624         sector_t block;
1625         sector_t last_block;
1626         struct buffer_head *bh, *head;
1627         const unsigned blocksize = 1 << inode->i_blkbits;
1628         int nr_underway = 0;
1629         int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1630                         WRITE_SYNC : WRITE);
1631
1632         BUG_ON(!PageLocked(page));
1633
1634         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1635
1636         if (!page_has_buffers(page)) {
1637                 create_empty_buffers(page, blocksize,
1638                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1639         }
1640
1641         /*
1642          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1643          * here, and the (potentially unmapped) buffers may become dirty at
1644          * any time.  If a buffer becomes dirty here after we've inspected it
1645          * then we just miss that fact, and the page stays dirty.
1646          *
1647          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1648          * handle that here by just cleaning them.
1649          */
1650
1651         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1652         head = page_buffers(page);
1653         bh = head;
1654
1655         /*
1656          * Get all the dirty buffers mapped to disk addresses and
1657          * handle any aliases from the underlying blockdev's mapping.
1658          */
1659         do {
1660                 if (block > last_block) {
1661                         /*
1662                          * mapped buffers outside i_size will occur, because
1663                          * this page can be outside i_size when there is a
1664                          * truncate in progress.
1665                          */
1666                         /*
1667                          * The buffer was zeroed by block_write_full_page()
1668                          */
1669                         clear_buffer_dirty(bh);
1670                         set_buffer_uptodate(bh);
1671                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1672                            buffer_dirty(bh)) {
1673                         WARN_ON(bh->b_size != blocksize);
1674                         err = get_block(inode, block, bh, 1);
1675                         if (err)
1676                                 goto recover;
1677                         clear_buffer_delay(bh);
1678                         if (buffer_new(bh)) {
1679                                 /* blockdev mappings never come here */
1680                                 clear_buffer_new(bh);
1681                                 unmap_underlying_metadata(bh->b_bdev,
1682                                                         bh->b_blocknr);
1683                         }
1684                 }
1685                 bh = bh->b_this_page;
1686                 block++;
1687         } while (bh != head);
1688
1689         do {
1690                 if (!buffer_mapped(bh))
1691                         continue;
1692                 /*
1693                  * If it's a fully non-blocking write attempt and we cannot
1694                  * lock the buffer then redirty the page.  Note that this can
1695                  * potentially cause a busy-wait loop from writeback threads
1696                  * and kswapd activity, but those code paths have their own
1697                  * higher-level throttling.
1698                  */
1699                 if (wbc->sync_mode != WB_SYNC_NONE) {
1700                         lock_buffer(bh);
1701                 } else if (!trylock_buffer(bh)) {
1702                         redirty_page_for_writepage(wbc, page);
1703                         continue;
1704                 }
1705                 if (test_clear_buffer_dirty(bh)) {
1706                         mark_buffer_async_write_endio(bh, handler);
1707                 } else {
1708                         unlock_buffer(bh);
1709                 }
1710         } while ((bh = bh->b_this_page) != head);
1711
1712         /*
1713          * The page and its buffers are protected by PageWriteback(), so we can
1714          * drop the bh refcounts early.
1715          */
1716         BUG_ON(PageWriteback(page));
1717         set_page_writeback(page);
1718
1719         do {
1720                 struct buffer_head *next = bh->b_this_page;
1721                 if (buffer_async_write(bh)) {
1722                         submit_bh(write_op, bh);
1723                         nr_underway++;
1724                 }
1725                 bh = next;
1726         } while (bh != head);
1727         unlock_page(page);
1728
1729         err = 0;
1730 done:
1731         if (nr_underway == 0) {
1732                 /*
1733                  * The page was marked dirty, but the buffers were
1734                  * clean.  Someone wrote them back by hand with
1735                  * ll_rw_block/submit_bh.  A rare case.
1736                  */
1737                 end_page_writeback(page);
1738
1739                 /*
1740                  * The page and buffer_heads can be released at any time from
1741                  * here on.
1742                  */
1743         }
1744         return err;
1745
1746 recover:
1747         /*
1748          * ENOSPC, or some other error.  We may already have added some
1749          * blocks to the file, so we need to write these out to avoid
1750          * exposing stale data.
1751          * The page is currently locked and not marked for writeback
1752          */
1753         bh = head;
1754         /* Recovery: lock and submit the mapped buffers */
1755         do {
1756                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1757                     !buffer_delay(bh)) {
1758                         lock_buffer(bh);
1759                         mark_buffer_async_write_endio(bh, handler);
1760                 } else {
1761                         /*
1762                          * The buffer may have been set dirty during
1763                          * attachment to a dirty page.
1764                          */
1765                         clear_buffer_dirty(bh);
1766                 }
1767         } while ((bh = bh->b_this_page) != head);
1768         SetPageError(page);
1769         BUG_ON(PageWriteback(page));
1770         mapping_set_error(page->mapping, err);
1771         set_page_writeback(page);
1772         do {
1773                 struct buffer_head *next = bh->b_this_page;
1774                 if (buffer_async_write(bh)) {
1775                         clear_buffer_dirty(bh);
1776                         submit_bh(write_op, bh);
1777                         nr_underway++;
1778                 }
1779                 bh = next;
1780         } while (bh != head);
1781         unlock_page(page);
1782         goto done;
1783 }
1784
1785 /*
1786  * If a page has any new buffers, zero them out here, and mark them uptodate
1787  * and dirty so they'll be written out (in order to prevent uninitialised
1788  * block data from leaking). And clear the new bit.
1789  */
1790 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1791 {
1792         unsigned int block_start, block_end;
1793         struct buffer_head *head, *bh;
1794
1795         BUG_ON(!PageLocked(page));
1796         if (!page_has_buffers(page))
1797                 return;
1798
1799         bh = head = page_buffers(page);
1800         block_start = 0;
1801         do {
1802                 block_end = block_start + bh->b_size;
1803
1804                 if (buffer_new(bh)) {
1805                         if (block_end > from && block_start < to) {
1806                                 if (!PageUptodate(page)) {
1807                                         unsigned start, size;
1808
1809                                         start = max(from, block_start);
1810                                         size = min(to, block_end) - start;
1811
1812                                         zero_user(page, start, size);
1813                                         set_buffer_uptodate(bh);
1814                                 }
1815
1816                                 clear_buffer_new(bh);
1817                                 mark_buffer_dirty(bh);
1818                         }
1819                 }
1820
1821                 block_start = block_end;
1822                 bh = bh->b_this_page;
1823         } while (bh != head);
1824 }
1825 EXPORT_SYMBOL(page_zero_new_buffers);
1826
1827 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1828                 get_block_t *get_block)
1829 {
1830         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1831         unsigned to = from + len;
1832         struct inode *inode = page->mapping->host;
1833         unsigned block_start, block_end;
1834         sector_t block;
1835         int err = 0;
1836         unsigned blocksize, bbits;
1837         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1838
1839         BUG_ON(!PageLocked(page));
1840         BUG_ON(from > PAGE_CACHE_SIZE);
1841         BUG_ON(to > PAGE_CACHE_SIZE);
1842         BUG_ON(from > to);
1843
1844         blocksize = 1 << inode->i_blkbits;
1845         if (!page_has_buffers(page))
1846                 create_empty_buffers(page, blocksize, 0);
1847         head = page_buffers(page);
1848
1849         bbits = inode->i_blkbits;
1850         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1851
1852         for(bh = head, block_start = 0; bh != head || !block_start;
1853             block++, block_start=block_end, bh = bh->b_this_page) {
1854                 block_end = block_start + blocksize;
1855                 if (block_end <= from || block_start >= to) {
1856                         if (PageUptodate(page)) {
1857                                 if (!buffer_uptodate(bh))
1858                                         set_buffer_uptodate(bh);
1859                         }
1860                         continue;
1861                 }
1862                 if (buffer_new(bh))
1863                         clear_buffer_new(bh);
1864                 if (!buffer_mapped(bh)) {
1865                         WARN_ON(bh->b_size != blocksize);
1866                         err = get_block(inode, block, bh, 1);
1867                         if (err)
1868                                 break;
1869                         if (buffer_new(bh)) {
1870                                 unmap_underlying_metadata(bh->b_bdev,
1871                                                         bh->b_blocknr);
1872                                 if (PageUptodate(page)) {
1873                                         clear_buffer_new(bh);
1874                                         set_buffer_uptodate(bh);
1875                                         mark_buffer_dirty(bh);
1876                                         continue;
1877                                 }
1878                                 if (block_end > to || block_start < from)
1879                                         zero_user_segments(page,
1880                                                 to, block_end,
1881                                                 block_start, from);
1882                                 continue;
1883                         }
1884                 }
1885                 if (PageUptodate(page)) {
1886                         if (!buffer_uptodate(bh))
1887                                 set_buffer_uptodate(bh);
1888                         continue; 
1889                 }
1890                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1891                     !buffer_unwritten(bh) &&
1892                      (block_start < from || block_end > to)) {
1893                         ll_rw_block(READ, 1, &bh);
1894                         *wait_bh++=bh;
1895                 }
1896         }
1897         /*
1898          * If we issued read requests - let them complete.
1899          */
1900         while(wait_bh > wait) {
1901                 wait_on_buffer(*--wait_bh);
1902                 if (!buffer_uptodate(*wait_bh))
1903                         err = -EIO;
1904         }
1905         if (unlikely(err)) {
1906                 page_zero_new_buffers(page, from, to);
1907                 ClearPageUptodate(page);
1908         }
1909         return err;
1910 }
1911 EXPORT_SYMBOL(__block_write_begin);
1912
1913 static int __block_commit_write(struct inode *inode, struct page *page,
1914                 unsigned from, unsigned to)
1915 {
1916         unsigned block_start, block_end;
1917         int partial = 0;
1918         unsigned blocksize;
1919         struct buffer_head *bh, *head;
1920
1921         blocksize = 1 << inode->i_blkbits;
1922
1923         for(bh = head = page_buffers(page), block_start = 0;
1924             bh != head || !block_start;
1925             block_start=block_end, bh = bh->b_this_page) {
1926                 block_end = block_start + blocksize;
1927                 if (block_end <= from || block_start >= to) {
1928                         if (!buffer_uptodate(bh))
1929                                 partial = 1;
1930                 } else {
1931                         set_buffer_uptodate(bh);
1932                         mark_buffer_dirty(bh);
1933                 }
1934                 clear_buffer_new(bh);
1935         }
1936
1937         /*
1938          * If this is a partial write which happened to make all buffers
1939          * uptodate then we can optimize away a bogus readpage() for
1940          * the next read(). Here we 'discover' whether the page went
1941          * uptodate as a result of this (potentially partial) write.
1942          */
1943         if (!partial)
1944                 SetPageUptodate(page);
1945         return 0;
1946 }
1947
1948 /*
1949  * block_write_begin takes care of the basic task of block allocation and
1950  * bringing partial write blocks uptodate first.
1951  *
1952  * The filesystem needs to handle block truncation upon failure.
1953  */
1954 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
1955                 unsigned flags, struct page **pagep, get_block_t *get_block)
1956 {
1957         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1958         struct page *page;
1959         int status;
1960
1961         page = grab_cache_page_write_begin(mapping, index, flags);
1962         if (!page)
1963                 return -ENOMEM;
1964
1965         status = __block_write_begin(page, pos, len, get_block);
1966         if (unlikely(status)) {
1967                 unlock_page(page);
1968                 page_cache_release(page);
1969                 page = NULL;
1970         }
1971
1972         *pagep = page;
1973         return status;
1974 }
1975 EXPORT_SYMBOL(block_write_begin);
1976
1977 int block_write_end(struct file *file, struct address_space *mapping,
1978                         loff_t pos, unsigned len, unsigned copied,
1979                         struct page *page, void *fsdata)
1980 {
1981         struct inode *inode = mapping->host;
1982         unsigned start;
1983
1984         start = pos & (PAGE_CACHE_SIZE - 1);
1985
1986         if (unlikely(copied < len)) {
1987                 /*
1988                  * The buffers that were written will now be uptodate, so we
1989                  * don't have to worry about a readpage reading them and
1990                  * overwriting a partial write. However if we have encountered
1991                  * a short write and only partially written into a buffer, it
1992                  * will not be marked uptodate, so a readpage might come in and
1993                  * destroy our partial write.
1994                  *
1995                  * Do the simplest thing, and just treat any short write to a
1996                  * non uptodate page as a zero-length write, and force the
1997                  * caller to redo the whole thing.
1998                  */
1999                 if (!PageUptodate(page))
2000                         copied = 0;
2001
2002                 page_zero_new_buffers(page, start+copied, start+len);
2003         }
2004         flush_dcache_page(page);
2005
2006         /* This could be a short (even 0-length) commit */
2007         __block_commit_write(inode, page, start, start+copied);
2008
2009         return copied;
2010 }
2011 EXPORT_SYMBOL(block_write_end);
2012
2013 int generic_write_end(struct file *file, struct address_space *mapping,
2014                         loff_t pos, unsigned len, unsigned copied,
2015                         struct page *page, void *fsdata)
2016 {
2017         struct inode *inode = mapping->host;
2018         int i_size_changed = 0;
2019
2020         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2021
2022         /*
2023          * No need to use i_size_read() here, the i_size
2024          * cannot change under us because we hold i_mutex.
2025          *
2026          * But it's important to update i_size while still holding page lock:
2027          * page writeout could otherwise come in and zero beyond i_size.
2028          */
2029         if (pos+copied > inode->i_size) {
2030                 i_size_write(inode, pos+copied);
2031                 i_size_changed = 1;
2032         }
2033
2034         unlock_page(page);
2035         page_cache_release(page);
2036
2037         /*
2038          * Don't mark the inode dirty under page lock. First, it unnecessarily
2039          * makes the holding time of page lock longer. Second, it forces lock
2040          * ordering of page lock and transaction start for journaling
2041          * filesystems.
2042          */
2043         if (i_size_changed)
2044                 mark_inode_dirty(inode);
2045
2046         return copied;
2047 }
2048 EXPORT_SYMBOL(generic_write_end);
2049
2050 /*
2051  * block_is_partially_uptodate checks whether buffers within a page are
2052  * uptodate or not.
2053  *
2054  * Returns true if all buffers which correspond to a file portion
2055  * we want to read are uptodate.
2056  */
2057 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2058                                         unsigned long from)
2059 {
2060         struct inode *inode = page->mapping->host;
2061         unsigned block_start, block_end, blocksize;
2062         unsigned to;
2063         struct buffer_head *bh, *head;
2064         int ret = 1;
2065
2066         if (!page_has_buffers(page))
2067                 return 0;
2068
2069         blocksize = 1 << inode->i_blkbits;
2070         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2071         to = from + to;
2072         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2073                 return 0;
2074
2075         head = page_buffers(page);
2076         bh = head;
2077         block_start = 0;
2078         do {
2079                 block_end = block_start + blocksize;
2080                 if (block_end > from && block_start < to) {
2081                         if (!buffer_uptodate(bh)) {
2082                                 ret = 0;
2083                                 break;
2084                         }
2085                         if (block_end >= to)
2086                                 break;
2087                 }
2088                 block_start = block_end;
2089                 bh = bh->b_this_page;
2090         } while (bh != head);
2091
2092         return ret;
2093 }
2094 EXPORT_SYMBOL(block_is_partially_uptodate);
2095
2096 /*
2097  * Generic "read page" function for block devices that have the normal
2098  * get_block functionality. This is most of the block device filesystems.
2099  * Reads the page asynchronously --- the unlock_buffer() and
2100  * set/clear_buffer_uptodate() functions propagate buffer state into the
2101  * page struct once IO has completed.
2102  */
2103 int block_read_full_page(struct page *page, get_block_t *get_block)
2104 {
2105         struct inode *inode = page->mapping->host;
2106         sector_t iblock, lblock;
2107         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2108         unsigned int blocksize;
2109         int nr, i;
2110         int fully_mapped = 1;
2111
2112         BUG_ON(!PageLocked(page));
2113         blocksize = 1 << inode->i_blkbits;
2114         if (!page_has_buffers(page))
2115                 create_empty_buffers(page, blocksize, 0);
2116         head = page_buffers(page);
2117
2118         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2119         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2120         bh = head;
2121         nr = 0;
2122         i = 0;
2123
2124         do {
2125                 if (buffer_uptodate(bh))
2126                         continue;
2127
2128                 if (!buffer_mapped(bh)) {
2129                         int err = 0;
2130
2131                         fully_mapped = 0;
2132                         if (iblock < lblock) {
2133                                 WARN_ON(bh->b_size != blocksize);
2134                                 err = get_block(inode, iblock, bh, 0);
2135                                 if (err)
2136                                         SetPageError(page);
2137                         }
2138                         if (!buffer_mapped(bh)) {
2139                                 zero_user(page, i * blocksize, blocksize);
2140                                 if (!err)
2141                                         set_buffer_uptodate(bh);
2142                                 continue;
2143                         }
2144                         /*
2145                          * get_block() might have updated the buffer
2146                          * synchronously
2147                          */
2148                         if (buffer_uptodate(bh))
2149                                 continue;
2150                 }
2151                 arr[nr++] = bh;
2152         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2153
2154         if (fully_mapped)
2155                 SetPageMappedToDisk(page);
2156
2157         if (!nr) {
2158                 /*
2159                  * All buffers are uptodate - we can set the page uptodate
2160                  * as well. But not if get_block() returned an error.
2161                  */
2162                 if (!PageError(page))
2163                         SetPageUptodate(page);
2164                 unlock_page(page);
2165                 return 0;
2166         }
2167
2168         /* Stage two: lock the buffers */
2169         for (i = 0; i < nr; i++) {
2170                 bh = arr[i];
2171                 lock_buffer(bh);
2172                 mark_buffer_async_read(bh);
2173         }
2174
2175         /*
2176          * Stage 3: start the IO.  Check for uptodateness
2177          * inside the buffer lock in case another process reading
2178          * the underlying blockdev brought it uptodate (the sct fix).
2179          */
2180         for (i = 0; i < nr; i++) {
2181                 bh = arr[i];
2182                 if (buffer_uptodate(bh))
2183                         end_buffer_async_read(bh, 1);
2184                 else
2185                         submit_bh(READ, bh);
2186         }
2187         return 0;
2188 }
2189 EXPORT_SYMBOL(block_read_full_page);
2190
2191 /* utility function for filesystems that need to do work on expanding
2192  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2193  * deal with the hole.  
2194  */
2195 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2196 {
2197         struct address_space *mapping = inode->i_mapping;
2198         struct page *page;
2199         void *fsdata;
2200         int err;
2201
2202         err = inode_newsize_ok(inode, size);
2203         if (err)
2204                 goto out;
2205
2206         err = pagecache_write_begin(NULL, mapping, size, 0,
2207                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2208                                 &page, &fsdata);
2209         if (err)
2210                 goto out;
2211
2212         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2213         BUG_ON(err > 0);
2214
2215 out:
2216         return err;
2217 }
2218 EXPORT_SYMBOL(generic_cont_expand_simple);
2219
2220 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2221                             loff_t pos, loff_t *bytes)
2222 {
2223         struct inode *inode = mapping->host;
2224         unsigned blocksize = 1 << inode->i_blkbits;
2225         struct page *page;
2226         void *fsdata;
2227         pgoff_t index, curidx;
2228         loff_t curpos;
2229         unsigned zerofrom, offset, len;
2230         int err = 0;
2231
2232         index = pos >> PAGE_CACHE_SHIFT;
2233         offset = pos & ~PAGE_CACHE_MASK;
2234
2235         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2236                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2237                 if (zerofrom & (blocksize-1)) {
2238                         *bytes |= (blocksize-1);
2239                         (*bytes)++;
2240                 }
2241                 len = PAGE_CACHE_SIZE - zerofrom;
2242
2243                 err = pagecache_write_begin(file, mapping, curpos, len,
2244                                                 AOP_FLAG_UNINTERRUPTIBLE,
2245                                                 &page, &fsdata);
2246                 if (err)
2247                         goto out;
2248                 zero_user(page, zerofrom, len);
2249                 err = pagecache_write_end(file, mapping, curpos, len, len,
2250                                                 page, fsdata);
2251                 if (err < 0)
2252                         goto out;
2253                 BUG_ON(err != len);
2254                 err = 0;
2255
2256                 balance_dirty_pages_ratelimited(mapping);
2257         }
2258
2259         /* page covers the boundary, find the boundary offset */
2260         if (index == curidx) {
2261                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2262                 /* if we will expand the thing last block will be filled */
2263                 if (offset <= zerofrom) {
2264                         goto out;
2265                 }
2266                 if (zerofrom & (blocksize-1)) {
2267                         *bytes |= (blocksize-1);
2268                         (*bytes)++;
2269                 }
2270                 len = offset - zerofrom;
2271
2272                 err = pagecache_write_begin(file, mapping, curpos, len,
2273                                                 AOP_FLAG_UNINTERRUPTIBLE,
2274                                                 &page, &fsdata);
2275                 if (err)
2276                         goto out;
2277                 zero_user(page, zerofrom, len);
2278                 err = pagecache_write_end(file, mapping, curpos, len, len,
2279                                                 page, fsdata);
2280                 if (err < 0)
2281                         goto out;
2282                 BUG_ON(err != len);
2283                 err = 0;
2284         }
2285 out:
2286         return err;
2287 }
2288
2289 /*
2290  * For moronic filesystems that do not allow holes in file.
2291  * We may have to extend the file.
2292  */
2293 int cont_write_begin(struct file *file, struct address_space *mapping,
2294                         loff_t pos, unsigned len, unsigned flags,
2295                         struct page **pagep, void **fsdata,
2296                         get_block_t *get_block, loff_t *bytes)
2297 {
2298         struct inode *inode = mapping->host;
2299         unsigned blocksize = 1 << inode->i_blkbits;
2300         unsigned zerofrom;
2301         int err;
2302
2303         err = cont_expand_zero(file, mapping, pos, bytes);
2304         if (err)
2305                 return err;
2306
2307         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2308         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2309                 *bytes |= (blocksize-1);
2310                 (*bytes)++;
2311         }
2312
2313         return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2314 }
2315 EXPORT_SYMBOL(cont_write_begin);
2316
2317 int block_commit_write(struct page *page, unsigned from, unsigned to)
2318 {
2319         struct inode *inode = page->mapping->host;
2320         __block_commit_write(inode,page,from,to);
2321         return 0;
2322 }
2323 EXPORT_SYMBOL(block_commit_write);
2324
2325 /*
2326  * block_page_mkwrite() is not allowed to change the file size as it gets
2327  * called from a page fault handler when a page is first dirtied. Hence we must
2328  * be careful to check for EOF conditions here. We set the page up correctly
2329  * for a written page which means we get ENOSPC checking when writing into
2330  * holes and correct delalloc and unwritten extent mapping on filesystems that
2331  * support these features.
2332  *
2333  * We are not allowed to take the i_mutex here so we have to play games to
2334  * protect against truncate races as the page could now be beyond EOF.  Because
2335  * truncate writes the inode size before removing pages, once we have the
2336  * page lock we can determine safely if the page is beyond EOF. If it is not
2337  * beyond EOF, then the page is guaranteed safe against truncation until we
2338  * unlock the page.
2339  */
2340 int
2341 block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2342                    get_block_t get_block)
2343 {
2344         struct page *page = vmf->page;
2345         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2346         unsigned long end;
2347         loff_t size;
2348         int ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
2349
2350         lock_page(page);
2351         size = i_size_read(inode);
2352         if ((page->mapping != inode->i_mapping) ||
2353             (page_offset(page) > size)) {
2354                 /* page got truncated out from underneath us */
2355                 unlock_page(page);
2356                 goto out;
2357         }
2358
2359         /* page is wholly or partially inside EOF */
2360         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2361                 end = size & ~PAGE_CACHE_MASK;
2362         else
2363                 end = PAGE_CACHE_SIZE;
2364
2365         ret = __block_write_begin(page, 0, end, get_block);
2366         if (!ret)
2367                 ret = block_commit_write(page, 0, end);
2368
2369         if (unlikely(ret)) {
2370                 unlock_page(page);
2371                 if (ret == -ENOMEM)
2372                         ret = VM_FAULT_OOM;
2373                 else /* -ENOSPC, -EIO, etc */
2374                         ret = VM_FAULT_SIGBUS;
2375         } else
2376                 ret = VM_FAULT_LOCKED;
2377
2378 out:
2379         return ret;
2380 }
2381 EXPORT_SYMBOL(block_page_mkwrite);
2382
2383 /*
2384  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2385  * immediately, while under the page lock.  So it needs a special end_io
2386  * handler which does not touch the bh after unlocking it.
2387  */
2388 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2389 {
2390         __end_buffer_read_notouch(bh, uptodate);
2391 }
2392
2393 /*
2394  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2395  * the page (converting it to circular linked list and taking care of page
2396  * dirty races).
2397  */
2398 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2399 {
2400         struct buffer_head *bh;
2401
2402         BUG_ON(!PageLocked(page));
2403
2404         spin_lock(&page->mapping->private_lock);
2405         bh = head;
2406         do {
2407                 if (PageDirty(page))
2408                         set_buffer_dirty(bh);
2409                 if (!bh->b_this_page)
2410                         bh->b_this_page = head;
2411                 bh = bh->b_this_page;
2412         } while (bh != head);
2413         attach_page_buffers(page, head);
2414         spin_unlock(&page->mapping->private_lock);
2415 }
2416
2417 /*
2418  * On entry, the page is fully not uptodate.
2419  * On exit the page is fully uptodate in the areas outside (from,to)
2420  * The filesystem needs to handle block truncation upon failure.
2421  */
2422 int nobh_write_begin(struct address_space *mapping,
2423                         loff_t pos, unsigned len, unsigned flags,
2424                         struct page **pagep, void **fsdata,
2425                         get_block_t *get_block)
2426 {
2427         struct inode *inode = mapping->host;
2428         const unsigned blkbits = inode->i_blkbits;
2429         const unsigned blocksize = 1 << blkbits;
2430         struct buffer_head *head, *bh;
2431         struct page *page;
2432         pgoff_t index;
2433         unsigned from, to;
2434         unsigned block_in_page;
2435         unsigned block_start, block_end;
2436         sector_t block_in_file;
2437         int nr_reads = 0;
2438         int ret = 0;
2439         int is_mapped_to_disk = 1;
2440
2441         index = pos >> PAGE_CACHE_SHIFT;
2442         from = pos & (PAGE_CACHE_SIZE - 1);
2443         to = from + len;
2444
2445         page = grab_cache_page_write_begin(mapping, index, flags);
2446         if (!page)
2447                 return -ENOMEM;
2448         *pagep = page;
2449         *fsdata = NULL;
2450
2451         if (page_has_buffers(page)) {
2452                 ret = __block_write_begin(page, pos, len, get_block);
2453                 if (unlikely(ret))
2454                         goto out_release;
2455                 return ret;
2456         }
2457
2458         if (PageMappedToDisk(page))
2459                 return 0;
2460
2461         /*
2462          * Allocate buffers so that we can keep track of state, and potentially
2463          * attach them to the page if an error occurs. In the common case of
2464          * no error, they will just be freed again without ever being attached
2465          * to the page (which is all OK, because we're under the page lock).
2466          *
2467          * Be careful: the buffer linked list is a NULL terminated one, rather
2468          * than the circular one we're used to.
2469          */
2470         head = alloc_page_buffers(page, blocksize, 0);
2471         if (!head) {
2472                 ret = -ENOMEM;
2473                 goto out_release;
2474         }
2475
2476         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2477
2478         /*
2479          * We loop across all blocks in the page, whether or not they are
2480          * part of the affected region.  This is so we can discover if the
2481          * page is fully mapped-to-disk.
2482          */
2483         for (block_start = 0, block_in_page = 0, bh = head;
2484                   block_start < PAGE_CACHE_SIZE;
2485                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2486                 int create;
2487
2488                 block_end = block_start + blocksize;
2489                 bh->b_state = 0;
2490                 create = 1;
2491                 if (block_start >= to)
2492                         create = 0;
2493                 ret = get_block(inode, block_in_file + block_in_page,
2494                                         bh, create);
2495                 if (ret)
2496                         goto failed;
2497                 if (!buffer_mapped(bh))
2498                         is_mapped_to_disk = 0;
2499                 if (buffer_new(bh))
2500                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2501                 if (PageUptodate(page)) {
2502                         set_buffer_uptodate(bh);
2503                         continue;
2504                 }
2505                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2506                         zero_user_segments(page, block_start, from,
2507                                                         to, block_end);
2508                         continue;
2509                 }
2510                 if (buffer_uptodate(bh))
2511                         continue;       /* reiserfs does this */
2512                 if (block_start < from || block_end > to) {
2513                         lock_buffer(bh);
2514                         bh->b_end_io = end_buffer_read_nobh;
2515                         submit_bh(READ, bh);
2516                         nr_reads++;
2517                 }
2518         }
2519
2520         if (nr_reads) {
2521                 /*
2522                  * The page is locked, so these buffers are protected from
2523                  * any VM or truncate activity.  Hence we don't need to care
2524                  * for the buffer_head refcounts.
2525                  */
2526                 for (bh = head; bh; bh = bh->b_this_page) {
2527                         wait_on_buffer(bh);
2528                         if (!buffer_uptodate(bh))
2529                                 ret = -EIO;
2530                 }
2531                 if (ret)
2532                         goto failed;
2533         }
2534
2535         if (is_mapped_to_disk)
2536                 SetPageMappedToDisk(page);
2537
2538         *fsdata = head; /* to be released by nobh_write_end */
2539
2540         return 0;
2541
2542 failed:
2543         BUG_ON(!ret);
2544         /*
2545          * Error recovery is a bit difficult. We need to zero out blocks that
2546          * were newly allocated, and dirty them to ensure they get written out.
2547          * Buffers need to be attached to the page at this point, otherwise
2548          * the handling of potential IO errors during writeout would be hard
2549          * (could try doing synchronous writeout, but what if that fails too?)
2550          */
2551         attach_nobh_buffers(page, head);
2552         page_zero_new_buffers(page, from, to);
2553
2554 out_release:
2555         unlock_page(page);
2556         page_cache_release(page);
2557         *pagep = NULL;
2558
2559         return ret;
2560 }
2561 EXPORT_SYMBOL(nobh_write_begin);
2562
2563 int nobh_write_end(struct file *file, struct address_space *mapping,
2564                         loff_t pos, unsigned len, unsigned copied,
2565                         struct page *page, void *fsdata)
2566 {
2567         struct inode *inode = page->mapping->host;
2568         struct buffer_head *head = fsdata;
2569         struct buffer_head *bh;
2570         BUG_ON(fsdata != NULL && page_has_buffers(page));
2571
2572         if (unlikely(copied < len) && head)
2573                 attach_nobh_buffers(page, head);
2574         if (page_has_buffers(page))
2575                 return generic_write_end(file, mapping, pos, len,
2576                                         copied, page, fsdata);
2577
2578         SetPageUptodate(page);
2579         set_page_dirty(page);
2580         if (pos+copied > inode->i_size) {
2581                 i_size_write(inode, pos+copied);
2582                 mark_inode_dirty(inode);
2583         }
2584
2585         unlock_page(page);
2586         page_cache_release(page);
2587
2588         while (head) {
2589                 bh = head;
2590                 head = head->b_this_page;
2591                 free_buffer_head(bh);
2592         }
2593
2594         return copied;
2595 }
2596 EXPORT_SYMBOL(nobh_write_end);
2597
2598 /*
2599  * nobh_writepage() - based on block_full_write_page() except
2600  * that it tries to operate without attaching bufferheads to
2601  * the page.
2602  */
2603 int nobh_writepage(struct page *page, get_block_t *get_block,
2604                         struct writeback_control *wbc)
2605 {
2606         struct inode * const inode = page->mapping->host;
2607         loff_t i_size = i_size_read(inode);
2608         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2609         unsigned offset;
2610         int ret;
2611
2612         /* Is the page fully inside i_size? */
2613         if (page->index < end_index)
2614                 goto out;
2615
2616         /* Is the page fully outside i_size? (truncate in progress) */
2617         offset = i_size & (PAGE_CACHE_SIZE-1);
2618         if (page->index >= end_index+1 || !offset) {
2619                 /*
2620                  * The page may have dirty, unmapped buffers.  For example,
2621                  * they may have been added in ext3_writepage().  Make them
2622                  * freeable here, so the page does not leak.
2623                  */
2624 #if 0
2625                 /* Not really sure about this  - do we need this ? */
2626                 if (page->mapping->a_ops->invalidatepage)
2627                         page->mapping->a_ops->invalidatepage(page, offset);
2628 #endif
2629                 unlock_page(page);
2630                 return 0; /* don't care */
2631         }
2632
2633         /*
2634          * The page straddles i_size.  It must be zeroed out on each and every
2635          * writepage invocation because it may be mmapped.  "A file is mapped
2636          * in multiples of the page size.  For a file that is not a multiple of
2637          * the  page size, the remaining memory is zeroed when mapped, and
2638          * writes to that region are not written out to the file."
2639          */
2640         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2641 out:
2642         ret = mpage_writepage(page, get_block, wbc);
2643         if (ret == -EAGAIN)
2644                 ret = __block_write_full_page(inode, page, get_block, wbc,
2645                                               end_buffer_async_write);
2646         return ret;
2647 }
2648 EXPORT_SYMBOL(nobh_writepage);
2649
2650 int nobh_truncate_page(struct address_space *mapping,
2651                         loff_t from, get_block_t *get_block)
2652 {
2653         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2654         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2655         unsigned blocksize;
2656         sector_t iblock;
2657         unsigned length, pos;
2658         struct inode *inode = mapping->host;
2659         struct page *page;
2660         struct buffer_head map_bh;
2661         int err;
2662
2663         blocksize = 1 << inode->i_blkbits;
2664         length = offset & (blocksize - 1);
2665
2666         /* Block boundary? Nothing to do */
2667         if (!length)
2668                 return 0;
2669
2670         length = blocksize - length;
2671         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2672
2673         page = grab_cache_page(mapping, index);
2674         err = -ENOMEM;
2675         if (!page)
2676                 goto out;
2677
2678         if (page_has_buffers(page)) {
2679 has_buffers:
2680                 unlock_page(page);
2681                 page_cache_release(page);
2682                 return block_truncate_page(mapping, from, get_block);
2683         }
2684
2685         /* Find the buffer that contains "offset" */
2686         pos = blocksize;
2687         while (offset >= pos) {
2688                 iblock++;
2689                 pos += blocksize;
2690         }
2691
2692         map_bh.b_size = blocksize;
2693         map_bh.b_state = 0;
2694         err = get_block(inode, iblock, &map_bh, 0);
2695         if (err)
2696                 goto unlock;
2697         /* unmapped? It's a hole - nothing to do */
2698         if (!buffer_mapped(&map_bh))
2699                 goto unlock;
2700
2701         /* Ok, it's mapped. Make sure it's up-to-date */
2702         if (!PageUptodate(page)) {
2703                 err = mapping->a_ops->readpage(NULL, page);
2704                 if (err) {
2705                         page_cache_release(page);
2706                         goto out;
2707                 }
2708                 lock_page(page);
2709                 if (!PageUptodate(page)) {
2710                         err = -EIO;
2711                         goto unlock;
2712                 }
2713                 if (page_has_buffers(page))
2714                         goto has_buffers;
2715         }
2716         zero_user(page, offset, length);
2717         set_page_dirty(page);
2718         err = 0;
2719
2720 unlock:
2721         unlock_page(page);
2722         page_cache_release(page);
2723 out:
2724         return err;
2725 }
2726 EXPORT_SYMBOL(nobh_truncate_page);
2727
2728 int block_truncate_page(struct address_space *mapping,
2729                         loff_t from, get_block_t *get_block)
2730 {
2731         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2732         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2733         unsigned blocksize;
2734         sector_t iblock;
2735         unsigned length, pos;
2736         struct inode *inode = mapping->host;
2737         struct page *page;
2738         struct buffer_head *bh;
2739         int err;
2740
2741         blocksize = 1 << inode->i_blkbits;
2742         length = offset & (blocksize - 1);
2743
2744         /* Block boundary? Nothing to do */
2745         if (!length)
2746                 return 0;
2747
2748         length = blocksize - length;
2749         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2750         
2751         page = grab_cache_page(mapping, index);
2752         err = -ENOMEM;
2753         if (!page)
2754                 goto out;
2755
2756         if (!page_has_buffers(page))
2757                 create_empty_buffers(page, blocksize, 0);
2758
2759         /* Find the buffer that contains "offset" */
2760         bh = page_buffers(page);
2761         pos = blocksize;
2762         while (offset >= pos) {
2763                 bh = bh->b_this_page;
2764                 iblock++;
2765                 pos += blocksize;
2766         }
2767
2768         err = 0;
2769         if (!buffer_mapped(bh)) {
2770                 WARN_ON(bh->b_size != blocksize);
2771                 err = get_block(inode, iblock, bh, 0);
2772                 if (err)
2773                         goto unlock;
2774                 /* unmapped? It's a hole - nothing to do */
2775                 if (!buffer_mapped(bh))
2776                         goto unlock;
2777         }
2778
2779         /* Ok, it's mapped. Make sure it's up-to-date */
2780         if (PageUptodate(page))
2781                 set_buffer_uptodate(bh);
2782
2783         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2784                 err = -EIO;
2785                 ll_rw_block(READ, 1, &bh);
2786                 wait_on_buffer(bh);
2787                 /* Uhhuh. Read error. Complain and punt. */
2788                 if (!buffer_uptodate(bh))
2789                         goto unlock;
2790         }
2791
2792         zero_user(page, offset, length);
2793         mark_buffer_dirty(bh);
2794         err = 0;
2795
2796 unlock:
2797         unlock_page(page);
2798         page_cache_release(page);
2799 out:
2800         return err;
2801 }
2802 EXPORT_SYMBOL(block_truncate_page);
2803
2804 /*
2805  * The generic ->writepage function for buffer-backed address_spaces
2806  * this form passes in the end_io handler used to finish the IO.
2807  */
2808 int block_write_full_page_endio(struct page *page, get_block_t *get_block,
2809                         struct writeback_control *wbc, bh_end_io_t *handler)
2810 {
2811         struct inode * const inode = page->mapping->host;
2812         loff_t i_size = i_size_read(inode);
2813         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2814         unsigned offset;
2815
2816         /* Is the page fully inside i_size? */
2817         if (page->index < end_index)
2818                 return __block_write_full_page(inode, page, get_block, wbc,
2819                                                handler);
2820
2821         /* Is the page fully outside i_size? (truncate in progress) */
2822         offset = i_size & (PAGE_CACHE_SIZE-1);
2823         if (page->index >= end_index+1 || !offset) {
2824                 /*
2825                  * The page may have dirty, unmapped buffers.  For example,
2826                  * they may have been added in ext3_writepage().  Make them
2827                  * freeable here, so the page does not leak.
2828                  */
2829                 do_invalidatepage(page, 0);
2830                 unlock_page(page);
2831                 return 0; /* don't care */
2832         }
2833
2834         /*
2835          * The page straddles i_size.  It must be zeroed out on each and every
2836          * writepage invocation because it may be mmapped.  "A file is mapped
2837          * in multiples of the page size.  For a file that is not a multiple of
2838          * the  page size, the remaining memory is zeroed when mapped, and
2839          * writes to that region are not written out to the file."
2840          */
2841         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2842         return __block_write_full_page(inode, page, get_block, wbc, handler);
2843 }
2844 EXPORT_SYMBOL(block_write_full_page_endio);
2845
2846 /*
2847  * The generic ->writepage function for buffer-backed address_spaces
2848  */
2849 int block_write_full_page(struct page *page, get_block_t *get_block,
2850                         struct writeback_control *wbc)
2851 {
2852         return block_write_full_page_endio(page, get_block, wbc,
2853                                            end_buffer_async_write);
2854 }
2855 EXPORT_SYMBOL(block_write_full_page);
2856
2857 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2858                             get_block_t *get_block)
2859 {
2860         struct buffer_head tmp;
2861         struct inode *inode = mapping->host;
2862         tmp.b_state = 0;
2863         tmp.b_blocknr = 0;
2864         tmp.b_size = 1 << inode->i_blkbits;
2865         get_block(inode, block, &tmp, 0);
2866         return tmp.b_blocknr;
2867 }
2868 EXPORT_SYMBOL(generic_block_bmap);
2869
2870 static void end_bio_bh_io_sync(struct bio *bio, int err)
2871 {
2872         struct buffer_head *bh = bio->bi_private;
2873
2874         if (err == -EOPNOTSUPP) {
2875                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2876         }
2877
2878         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2879                 set_bit(BH_Quiet, &bh->b_state);
2880
2881         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2882         bio_put(bio);
2883 }
2884
2885 int submit_bh(int rw, struct buffer_head * bh)
2886 {
2887         struct bio *bio;
2888         int ret = 0;
2889
2890         BUG_ON(!buffer_locked(bh));
2891         BUG_ON(!buffer_mapped(bh));
2892         BUG_ON(!bh->b_end_io);
2893         BUG_ON(buffer_delay(bh));
2894         BUG_ON(buffer_unwritten(bh));
2895
2896         /*
2897          * Only clear out a write error when rewriting
2898          */
2899         if (test_set_buffer_req(bh) && (rw & WRITE))
2900                 clear_buffer_write_io_error(bh);
2901
2902         /*
2903          * from here on down, it's all bio -- do the initial mapping,
2904          * submit_bio -> generic_make_request may further map this bio around
2905          */
2906         bio = bio_alloc(GFP_NOIO, 1);
2907
2908         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2909         bio->bi_bdev = bh->b_bdev;
2910         bio->bi_io_vec[0].bv_page = bh->b_page;
2911         bio->bi_io_vec[0].bv_len = bh->b_size;
2912         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2913
2914         bio->bi_vcnt = 1;
2915         bio->bi_idx = 0;
2916         bio->bi_size = bh->b_size;
2917
2918         bio->bi_end_io = end_bio_bh_io_sync;
2919         bio->bi_private = bh;
2920
2921         bio_get(bio);
2922         submit_bio(rw, bio);
2923
2924         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2925                 ret = -EOPNOTSUPP;
2926
2927         bio_put(bio);
2928         return ret;
2929 }
2930 EXPORT_SYMBOL(submit_bh);
2931
2932 /**
2933  * ll_rw_block: low-level access to block devices (DEPRECATED)
2934  * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
2935  * @nr: number of &struct buffer_heads in the array
2936  * @bhs: array of pointers to &struct buffer_head
2937  *
2938  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2939  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2940  * %READA option is described in the documentation for generic_make_request()
2941  * which ll_rw_block() calls.
2942  *
2943  * This function drops any buffer that it cannot get a lock on (with the
2944  * BH_Lock state bit), any buffer that appears to be clean when doing a write
2945  * request, and any buffer that appears to be up-to-date when doing read
2946  * request.  Further it marks as clean buffers that are processed for
2947  * writing (the buffer cache won't assume that they are actually clean
2948  * until the buffer gets unlocked).
2949  *
2950  * ll_rw_block sets b_end_io to simple completion handler that marks
2951  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2952  * any waiters. 
2953  *
2954  * All of the buffers must be for the same device, and must also be a
2955  * multiple of the current approved size for the device.
2956  */
2957 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2958 {
2959         int i;
2960
2961         for (i = 0; i < nr; i++) {
2962                 struct buffer_head *bh = bhs[i];
2963
2964                 if (!trylock_buffer(bh))
2965                         continue;
2966                 if (rw == WRITE) {
2967                         if (test_clear_buffer_dirty(bh)) {
2968                                 bh->b_end_io = end_buffer_write_sync;
2969                                 get_bh(bh);
2970                                 submit_bh(WRITE, bh);
2971                                 continue;
2972                         }
2973                 } else {
2974                         if (!buffer_uptodate(bh)) {
2975                                 bh->b_end_io = end_buffer_read_sync;
2976                                 get_bh(bh);
2977                                 submit_bh(rw, bh);
2978                                 continue;
2979                         }
2980                 }
2981                 unlock_buffer(bh);
2982         }
2983 }
2984 EXPORT_SYMBOL(ll_rw_block);
2985
2986 void write_dirty_buffer(struct buffer_head *bh, int rw)
2987 {
2988         lock_buffer(bh);
2989         if (!test_clear_buffer_dirty(bh)) {
2990                 unlock_buffer(bh);
2991                 return;
2992         }
2993         bh->b_end_io = end_buffer_write_sync;
2994         get_bh(bh);
2995         submit_bh(rw, bh);
2996 }
2997 EXPORT_SYMBOL(write_dirty_buffer);
2998
2999 /*
3000  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3001  * and then start new I/O and then wait upon it.  The caller must have a ref on
3002  * the buffer_head.
3003  */
3004 int __sync_dirty_buffer(struct buffer_head *bh, int rw)
3005 {
3006         int ret = 0;
3007
3008         WARN_ON(atomic_read(&bh->b_count) < 1);
3009         lock_buffer(bh);
3010         if (test_clear_buffer_dirty(bh)) {
3011                 get_bh(bh);
3012                 bh->b_end_io = end_buffer_write_sync;
3013                 ret = submit_bh(rw, bh);
3014                 wait_on_buffer(bh);
3015                 if (!ret && !buffer_uptodate(bh))
3016                         ret = -EIO;
3017         } else {
3018                 unlock_buffer(bh);
3019         }
3020         return ret;
3021 }
3022 EXPORT_SYMBOL(__sync_dirty_buffer);
3023
3024 int sync_dirty_buffer(struct buffer_head *bh)
3025 {
3026         return __sync_dirty_buffer(bh, WRITE_SYNC);
3027 }
3028 EXPORT_SYMBOL(sync_dirty_buffer);
3029
3030 /*
3031  * try_to_free_buffers() checks if all the buffers on this particular page
3032  * are unused, and releases them if so.
3033  *
3034  * Exclusion against try_to_free_buffers may be obtained by either
3035  * locking the page or by holding its mapping's private_lock.
3036  *
3037  * If the page is dirty but all the buffers are clean then we need to
3038  * be sure to mark the page clean as well.  This is because the page
3039  * may be against a block device, and a later reattachment of buffers
3040  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3041  * filesystem data on the same device.
3042  *
3043  * The same applies to regular filesystem pages: if all the buffers are
3044  * clean then we set the page clean and proceed.  To do that, we require
3045  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3046  * private_lock.
3047  *
3048  * try_to_free_buffers() is non-blocking.
3049  */
3050 static inline int buffer_busy(struct buffer_head *bh)
3051 {
3052         return atomic_read(&bh->b_count) |
3053                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3054 }
3055
3056 static int
3057 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3058 {
3059         struct buffer_head *head = page_buffers(page);
3060         struct buffer_head *bh;
3061
3062         bh = head;
3063         do {
3064                 if (buffer_write_io_error(bh) && page->mapping)
3065                         set_bit(AS_EIO, &page->mapping->flags);
3066                 if (buffer_busy(bh))
3067                         goto failed;
3068                 bh = bh->b_this_page;
3069         } while (bh != head);
3070
3071         do {
3072                 struct buffer_head *next = bh->b_this_page;
3073
3074                 if (bh->b_assoc_map)
3075                         __remove_assoc_queue(bh);
3076                 bh = next;
3077         } while (bh != head);
3078         *buffers_to_free = head;
3079         __clear_page_buffers(page);
3080         return 1;
3081 failed:
3082         return 0;
3083 }
3084
3085 int try_to_free_buffers(struct page *page)
3086 {
3087         struct address_space * const mapping = page->mapping;
3088         struct buffer_head *buffers_to_free = NULL;
3089         int ret = 0;
3090
3091         BUG_ON(!PageLocked(page));
3092         if (PageWriteback(page))
3093                 return 0;
3094
3095         if (mapping == NULL) {          /* can this still happen? */
3096                 ret = drop_buffers(page, &buffers_to_free);
3097                 goto out;
3098         }
3099
3100         spin_lock(&mapping->private_lock);
3101         ret = drop_buffers(page, &buffers_to_free);
3102
3103         /*
3104          * If the filesystem writes its buffers by hand (eg ext3)
3105          * then we can have clean buffers against a dirty page.  We
3106          * clean the page here; otherwise the VM will never notice
3107          * that the filesystem did any IO at all.
3108          *
3109          * Also, during truncate, discard_buffer will have marked all
3110          * the page's buffers clean.  We discover that here and clean
3111          * the page also.
3112          *
3113          * private_lock must be held over this entire operation in order
3114          * to synchronise against __set_page_dirty_buffers and prevent the
3115          * dirty bit from being lost.
3116          */
3117         if (ret)
3118                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3119         spin_unlock(&mapping->private_lock);
3120 out:
3121         if (buffers_to_free) {
3122                 struct buffer_head *bh = buffers_to_free;
3123
3124                 do {
3125                         struct buffer_head *next = bh->b_this_page;
3126                         free_buffer_head(bh);
3127                         bh = next;
3128                 } while (bh != buffers_to_free);
3129         }
3130         return ret;
3131 }
3132 EXPORT_SYMBOL(try_to_free_buffers);
3133
3134 /*
3135  * There are no bdflush tunables left.  But distributions are
3136  * still running obsolete flush daemons, so we terminate them here.
3137  *
3138  * Use of bdflush() is deprecated and will be removed in a future kernel.
3139  * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3140  */
3141 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3142 {
3143         static int msg_count;
3144
3145         if (!capable(CAP_SYS_ADMIN))
3146                 return -EPERM;
3147
3148         if (msg_count < 5) {
3149                 msg_count++;
3150                 printk(KERN_INFO
3151                         "warning: process `%s' used the obsolete bdflush"
3152                         " system call\n", current->comm);
3153                 printk(KERN_INFO "Fix your initscripts?\n");
3154         }
3155
3156         if (func == 1)
3157                 do_exit(0);
3158         return 0;
3159 }
3160
3161 /*
3162  * Buffer-head allocation
3163  */
3164 static struct kmem_cache *bh_cachep;
3165
3166 /*
3167  * Once the number of bh's in the machine exceeds this level, we start
3168  * stripping them in writeback.
3169  */
3170 static int max_buffer_heads;
3171
3172 int buffer_heads_over_limit;
3173
3174 struct bh_accounting {
3175         int nr;                 /* Number of live bh's */
3176         int ratelimit;          /* Limit cacheline bouncing */
3177 };
3178
3179 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3180
3181 static void recalc_bh_state(void)
3182 {
3183         int i;
3184         int tot = 0;
3185
3186         if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3187                 return;
3188         __this_cpu_write(bh_accounting.ratelimit, 0);
3189         for_each_online_cpu(i)
3190                 tot += per_cpu(bh_accounting, i).nr;
3191         buffer_heads_over_limit = (tot > max_buffer_heads);
3192 }
3193
3194 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3195 {
3196         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3197         if (ret) {
3198                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3199                 preempt_disable();
3200                 __this_cpu_inc(bh_accounting.nr);
3201                 recalc_bh_state();
3202                 preempt_enable();
3203         }
3204         return ret;
3205 }
3206 EXPORT_SYMBOL(alloc_buffer_head);
3207
3208 void free_buffer_head(struct buffer_head *bh)
3209 {
3210         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3211         kmem_cache_free(bh_cachep, bh);
3212         preempt_disable();
3213         __this_cpu_dec(bh_accounting.nr);
3214         recalc_bh_state();
3215         preempt_enable();
3216 }
3217 EXPORT_SYMBOL(free_buffer_head);
3218
3219 static void buffer_exit_cpu(int cpu)
3220 {
3221         int i;
3222         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3223
3224         for (i = 0; i < BH_LRU_SIZE; i++) {
3225                 brelse(b->bhs[i]);
3226                 b->bhs[i] = NULL;
3227         }
3228         this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3229         per_cpu(bh_accounting, cpu).nr = 0;
3230 }
3231
3232 static int buffer_cpu_notify(struct notifier_block *self,
3233                               unsigned long action, void *hcpu)
3234 {
3235         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3236                 buffer_exit_cpu((unsigned long)hcpu);
3237         return NOTIFY_OK;
3238 }
3239
3240 /**
3241  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3242  * @bh: struct buffer_head
3243  *
3244  * Return true if the buffer is up-to-date and false,
3245  * with the buffer locked, if not.
3246  */
3247 int bh_uptodate_or_lock(struct buffer_head *bh)
3248 {
3249         if (!buffer_uptodate(bh)) {
3250                 lock_buffer(bh);
3251                 if (!buffer_uptodate(bh))
3252                         return 0;
3253                 unlock_buffer(bh);
3254         }
3255         return 1;
3256 }
3257 EXPORT_SYMBOL(bh_uptodate_or_lock);
3258
3259 /**
3260  * bh_submit_read - Submit a locked buffer for reading
3261  * @bh: struct buffer_head
3262  *
3263  * Returns zero on success and -EIO on error.
3264  */
3265 int bh_submit_read(struct buffer_head *bh)
3266 {
3267         BUG_ON(!buffer_locked(bh));
3268
3269         if (buffer_uptodate(bh)) {
3270                 unlock_buffer(bh);
3271                 return 0;
3272         }
3273
3274         get_bh(bh);
3275         bh->b_end_io = end_buffer_read_sync;
3276         submit_bh(READ, bh);
3277         wait_on_buffer(bh);
3278         if (buffer_uptodate(bh))
3279                 return 0;
3280         return -EIO;
3281 }
3282 EXPORT_SYMBOL(bh_submit_read);
3283
3284 void __init buffer_init(void)
3285 {
3286         int nrpages;
3287
3288         bh_cachep = kmem_cache_create("buffer_head",
3289                         sizeof(struct buffer_head), 0,
3290                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3291                                 SLAB_MEM_SPREAD),
3292                                 NULL);
3293
3294         /*
3295          * Limit the bh occupancy to 10% of ZONE_NORMAL
3296          */
3297         nrpages = (nr_free_buffer_pages() * 10) / 100;
3298         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3299         hotcpu_notifier(buffer_cpu_notify, 0);
3300 }