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