Cleanup after commit 585d3bc06f4ca57f975a5a1f698f65a45ea66225
[linux-3.10.git] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/capability.h>
28 #include <linux/blkdev.h>
29 #include <linux/file.h>
30 #include <linux/quotaops.h>
31 #include <linux/highmem.h>
32 #include <linux/module.h>
33 #include <linux/writeback.h>
34 #include <linux/hash.h>
35 #include <linux/suspend.h>
36 #include <linux/buffer_head.h>
37 #include <linux/task_io_accounting_ops.h>
38 #include <linux/bio.h>
39 #include <linux/notifier.h>
40 #include <linux/cpu.h>
41 #include <linux/bitops.h>
42 #include <linux/mpage.h>
43 #include <linux/bit_spinlock.h>
44
45 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
46
47 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
48
49 inline void
50 init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
51 {
52         bh->b_end_io = handler;
53         bh->b_private = private;
54 }
55
56 static int sync_buffer(void *word)
57 {
58         struct block_device *bd;
59         struct buffer_head *bh
60                 = container_of(word, struct buffer_head, b_state);
61
62         smp_mb();
63         bd = bh->b_bdev;
64         if (bd)
65                 blk_run_address_space(bd->bd_inode->i_mapping);
66         io_schedule();
67         return 0;
68 }
69
70 void __lock_buffer(struct buffer_head *bh)
71 {
72         wait_on_bit_lock(&bh->b_state, BH_Lock, sync_buffer,
73                                                         TASK_UNINTERRUPTIBLE);
74 }
75 EXPORT_SYMBOL(__lock_buffer);
76
77 void unlock_buffer(struct buffer_head *bh)
78 {
79         clear_bit_unlock(BH_Lock, &bh->b_state);
80         smp_mb__after_clear_bit();
81         wake_up_bit(&bh->b_state, BH_Lock);
82 }
83
84 /*
85  * Block until a buffer comes unlocked.  This doesn't stop it
86  * from becoming locked again - you have to lock it yourself
87  * if you want to preserve its state.
88  */
89 void __wait_on_buffer(struct buffer_head * bh)
90 {
91         wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
92 }
93
94 static void
95 __clear_page_buffers(struct page *page)
96 {
97         ClearPagePrivate(page);
98         set_page_private(page, 0);
99         page_cache_release(page);
100 }
101
102
103 static int quiet_error(struct buffer_head *bh)
104 {
105         if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
106                 return 0;
107         return 1;
108 }
109
110
111 static void buffer_io_error(struct buffer_head *bh)
112 {
113         char b[BDEVNAME_SIZE];
114         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
115                         bdevname(bh->b_bdev, b),
116                         (unsigned long long)bh->b_blocknr);
117 }
118
119 /*
120  * End-of-IO handler helper function which does not touch the bh after
121  * unlocking it.
122  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
123  * a race there is benign: unlock_buffer() only use the bh's address for
124  * hashing after unlocking the buffer, so it doesn't actually touch the bh
125  * itself.
126  */
127 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
128 {
129         if (uptodate) {
130                 set_buffer_uptodate(bh);
131         } else {
132                 /* This happens, due to failed READA attempts. */
133                 clear_buffer_uptodate(bh);
134         }
135         unlock_buffer(bh);
136 }
137
138 /*
139  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
140  * unlock the buffer. This is what ll_rw_block uses too.
141  */
142 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
143 {
144         __end_buffer_read_notouch(bh, uptodate);
145         put_bh(bh);
146 }
147
148 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
149 {
150         char b[BDEVNAME_SIZE];
151
152         if (uptodate) {
153                 set_buffer_uptodate(bh);
154         } else {
155                 if (!buffer_eopnotsupp(bh) && !quiet_error(bh)) {
156                         buffer_io_error(bh);
157                         printk(KERN_WARNING "lost page write due to "
158                                         "I/O error on %s\n",
159                                        bdevname(bh->b_bdev, b));
160                 }
161                 set_buffer_write_io_error(bh);
162                 clear_buffer_uptodate(bh);
163         }
164         unlock_buffer(bh);
165         put_bh(bh);
166 }
167
168 /*
169  * Various filesystems appear to want __find_get_block to be non-blocking.
170  * But it's the page lock which protects the buffers.  To get around this,
171  * we get exclusion from try_to_free_buffers with the blockdev mapping's
172  * private_lock.
173  *
174  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
175  * may be quite high.  This code could TryLock the page, and if that
176  * succeeds, there is no need to take private_lock. (But if
177  * private_lock is contended then so is mapping->tree_lock).
178  */
179 static struct buffer_head *
180 __find_get_block_slow(struct block_device *bdev, sector_t block)
181 {
182         struct inode *bd_inode = bdev->bd_inode;
183         struct address_space *bd_mapping = bd_inode->i_mapping;
184         struct buffer_head *ret = NULL;
185         pgoff_t index;
186         struct buffer_head *bh;
187         struct buffer_head *head;
188         struct page *page;
189         int all_mapped = 1;
190
191         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
192         page = find_get_page(bd_mapping, index);
193         if (!page)
194                 goto out;
195
196         spin_lock(&bd_mapping->private_lock);
197         if (!page_has_buffers(page))
198                 goto out_unlock;
199         head = page_buffers(page);
200         bh = head;
201         do {
202                 if (bh->b_blocknr == block) {
203                         ret = bh;
204                         get_bh(bh);
205                         goto out_unlock;
206                 }
207                 if (!buffer_mapped(bh))
208                         all_mapped = 0;
209                 bh = bh->b_this_page;
210         } while (bh != head);
211
212         /* we might be here because some of the buffers on this page are
213          * not mapped.  This is due to various races between
214          * file io on the block device and getblk.  It gets dealt with
215          * elsewhere, don't buffer_error if we had some unmapped buffers
216          */
217         if (all_mapped) {
218                 printk("__find_get_block_slow() failed. "
219                         "block=%llu, b_blocknr=%llu\n",
220                         (unsigned long long)block,
221                         (unsigned long long)bh->b_blocknr);
222                 printk("b_state=0x%08lx, b_size=%zu\n",
223                         bh->b_state, bh->b_size);
224                 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
225         }
226 out_unlock:
227         spin_unlock(&bd_mapping->private_lock);
228         page_cache_release(page);
229 out:
230         return ret;
231 }
232
233 /* If invalidate_buffers() will trash dirty buffers, it means some kind
234    of fs corruption is going on. Trashing dirty data always imply losing
235    information that was supposed to be just stored on the physical layer
236    by the user.
237
238    Thus invalidate_buffers in general usage is not allwowed to trash
239    dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
240    be preserved.  These buffers are simply skipped.
241   
242    We also skip buffers which are still in use.  For example this can
243    happen if a userspace program is reading the block device.
244
245    NOTE: In the case where the user removed a removable-media-disk even if
246    there's still dirty data not synced on disk (due a bug in the device driver
247    or due an error of the user), by not destroying the dirty buffers we could
248    generate corruption also on the next media inserted, thus a parameter is
249    necessary to handle this case in the most safe way possible (trying
250    to not corrupt also the new disk inserted with the data belonging to
251    the old now corrupted disk). Also for the ramdisk the natural thing
252    to do in order to release the ramdisk memory is to destroy dirty buffers.
253
254    These are two special cases. Normal usage imply the device driver
255    to issue a sync on the device (without waiting I/O completion) and
256    then an invalidate_buffers call that doesn't trash dirty buffers.
257
258    For handling cache coherency with the blkdev pagecache the 'update' case
259    is been introduced. It is needed to re-read from disk any pinned
260    buffer. NOTE: re-reading from disk is destructive so we can do it only
261    when we assume nobody is changing the buffercache under our I/O and when
262    we think the disk contains more recent information than the buffercache.
263    The update == 1 pass marks the buffers we need to update, the update == 2
264    pass does the actual I/O. */
265 void invalidate_bdev(struct block_device *bdev)
266 {
267         struct address_space *mapping = bdev->bd_inode->i_mapping;
268
269         if (mapping->nrpages == 0)
270                 return;
271
272         invalidate_bh_lrus();
273         invalidate_mapping_pages(mapping, 0, -1);
274 }
275
276 /*
277  * Kick pdflush then try to free up some ZONE_NORMAL memory.
278  */
279 static void free_more_memory(void)
280 {
281         struct zone *zone;
282         int nid;
283
284         wakeup_pdflush(1024);
285         yield();
286
287         for_each_online_node(nid) {
288                 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
289                                                 gfp_zone(GFP_NOFS), NULL,
290                                                 &zone);
291                 if (zone)
292                         try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
293                                                 GFP_NOFS);
294         }
295 }
296
297 /*
298  * I/O completion handler for block_read_full_page() - pages
299  * which come unlocked at the end of I/O.
300  */
301 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
302 {
303         unsigned long flags;
304         struct buffer_head *first;
305         struct buffer_head *tmp;
306         struct page *page;
307         int page_uptodate = 1;
308
309         BUG_ON(!buffer_async_read(bh));
310
311         page = bh->b_page;
312         if (uptodate) {
313                 set_buffer_uptodate(bh);
314         } else {
315                 clear_buffer_uptodate(bh);
316                 if (!quiet_error(bh))
317                         buffer_io_error(bh);
318                 SetPageError(page);
319         }
320
321         /*
322          * Be _very_ careful from here on. Bad things can happen if
323          * two buffer heads end IO at almost the same time and both
324          * decide that the page is now completely done.
325          */
326         first = page_buffers(page);
327         local_irq_save(flags);
328         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
329         clear_buffer_async_read(bh);
330         unlock_buffer(bh);
331         tmp = bh;
332         do {
333                 if (!buffer_uptodate(tmp))
334                         page_uptodate = 0;
335                 if (buffer_async_read(tmp)) {
336                         BUG_ON(!buffer_locked(tmp));
337                         goto still_busy;
338                 }
339                 tmp = tmp->b_this_page;
340         } while (tmp != bh);
341         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
342         local_irq_restore(flags);
343
344         /*
345          * If none of the buffers had errors and they are all
346          * uptodate then we can set the page uptodate.
347          */
348         if (page_uptodate && !PageError(page))
349                 SetPageUptodate(page);
350         unlock_page(page);
351         return;
352
353 still_busy:
354         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
355         local_irq_restore(flags);
356         return;
357 }
358
359 /*
360  * Completion handler for block_write_full_page() - pages which are unlocked
361  * during I/O, and which have PageWriteback cleared upon I/O completion.
362  */
363 static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
364 {
365         char b[BDEVNAME_SIZE];
366         unsigned long flags;
367         struct buffer_head *first;
368         struct buffer_head *tmp;
369         struct page *page;
370
371         BUG_ON(!buffer_async_write(bh));
372
373         page = bh->b_page;
374         if (uptodate) {
375                 set_buffer_uptodate(bh);
376         } else {
377                 if (!quiet_error(bh)) {
378                         buffer_io_error(bh);
379                         printk(KERN_WARNING "lost page write due to "
380                                         "I/O error on %s\n",
381                                bdevname(bh->b_bdev, b));
382                 }
383                 set_bit(AS_EIO, &page->mapping->flags);
384                 set_buffer_write_io_error(bh);
385                 clear_buffer_uptodate(bh);
386                 SetPageError(page);
387         }
388
389         first = page_buffers(page);
390         local_irq_save(flags);
391         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
392
393         clear_buffer_async_write(bh);
394         unlock_buffer(bh);
395         tmp = bh->b_this_page;
396         while (tmp != bh) {
397                 if (buffer_async_write(tmp)) {
398                         BUG_ON(!buffer_locked(tmp));
399                         goto still_busy;
400                 }
401                 tmp = tmp->b_this_page;
402         }
403         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
404         local_irq_restore(flags);
405         end_page_writeback(page);
406         return;
407
408 still_busy:
409         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
410         local_irq_restore(flags);
411         return;
412 }
413
414 /*
415  * If a page's buffers are under async readin (end_buffer_async_read
416  * completion) then there is a possibility that another thread of
417  * control could lock one of the buffers after it has completed
418  * but while some of the other buffers have not completed.  This
419  * locked buffer would confuse end_buffer_async_read() into not unlocking
420  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
421  * that this buffer is not under async I/O.
422  *
423  * The page comes unlocked when it has no locked buffer_async buffers
424  * left.
425  *
426  * PageLocked prevents anyone starting new async I/O reads any of
427  * the buffers.
428  *
429  * PageWriteback is used to prevent simultaneous writeout of the same
430  * page.
431  *
432  * PageLocked prevents anyone from starting writeback of a page which is
433  * under read I/O (PageWriteback is only ever set against a locked page).
434  */
435 static void mark_buffer_async_read(struct buffer_head *bh)
436 {
437         bh->b_end_io = end_buffer_async_read;
438         set_buffer_async_read(bh);
439 }
440
441 void mark_buffer_async_write(struct buffer_head *bh)
442 {
443         bh->b_end_io = end_buffer_async_write;
444         set_buffer_async_write(bh);
445 }
446 EXPORT_SYMBOL(mark_buffer_async_write);
447
448
449 /*
450  * fs/buffer.c contains helper functions for buffer-backed address space's
451  * fsync functions.  A common requirement for buffer-based filesystems is
452  * that certain data from the backing blockdev needs to be written out for
453  * a successful fsync().  For example, ext2 indirect blocks need to be
454  * written back and waited upon before fsync() returns.
455  *
456  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
457  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
458  * management of a list of dependent buffers at ->i_mapping->private_list.
459  *
460  * Locking is a little subtle: try_to_free_buffers() will remove buffers
461  * from their controlling inode's queue when they are being freed.  But
462  * try_to_free_buffers() will be operating against the *blockdev* mapping
463  * at the time, not against the S_ISREG file which depends on those buffers.
464  * So the locking for private_list is via the private_lock in the address_space
465  * which backs the buffers.  Which is different from the address_space 
466  * against which the buffers are listed.  So for a particular address_space,
467  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
468  * mapping->private_list will always be protected by the backing blockdev's
469  * ->private_lock.
470  *
471  * Which introduces a requirement: all buffers on an address_space's
472  * ->private_list must be from the same address_space: the blockdev's.
473  *
474  * address_spaces which do not place buffers at ->private_list via these
475  * utility functions are free to use private_lock and private_list for
476  * whatever they want.  The only requirement is that list_empty(private_list)
477  * be true at clear_inode() time.
478  *
479  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
480  * filesystems should do that.  invalidate_inode_buffers() should just go
481  * BUG_ON(!list_empty).
482  *
483  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
484  * take an address_space, not an inode.  And it should be called
485  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
486  * queued up.
487  *
488  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
489  * list if it is already on a list.  Because if the buffer is on a list,
490  * it *must* already be on the right one.  If not, the filesystem is being
491  * silly.  This will save a ton of locking.  But first we have to ensure
492  * that buffers are taken *off* the old inode's list when they are freed
493  * (presumably in truncate).  That requires careful auditing of all
494  * filesystems (do it inside bforget()).  It could also be done by bringing
495  * b_inode back.
496  */
497
498 /*
499  * The buffer's backing address_space's private_lock must be held
500  */
501 static void __remove_assoc_queue(struct buffer_head *bh)
502 {
503         list_del_init(&bh->b_assoc_buffers);
504         WARN_ON(!bh->b_assoc_map);
505         if (buffer_write_io_error(bh))
506                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
507         bh->b_assoc_map = NULL;
508 }
509
510 int inode_has_buffers(struct inode *inode)
511 {
512         return !list_empty(&inode->i_data.private_list);
513 }
514
515 /*
516  * osync is designed to support O_SYNC io.  It waits synchronously for
517  * all already-submitted IO to complete, but does not queue any new
518  * writes to the disk.
519  *
520  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
521  * you dirty the buffers, and then use osync_inode_buffers to wait for
522  * completion.  Any other dirty buffers which are not yet queued for
523  * write will not be flushed to disk by the osync.
524  */
525 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
526 {
527         struct buffer_head *bh;
528         struct list_head *p;
529         int err = 0;
530
531         spin_lock(lock);
532 repeat:
533         list_for_each_prev(p, list) {
534                 bh = BH_ENTRY(p);
535                 if (buffer_locked(bh)) {
536                         get_bh(bh);
537                         spin_unlock(lock);
538                         wait_on_buffer(bh);
539                         if (!buffer_uptodate(bh))
540                                 err = -EIO;
541                         brelse(bh);
542                         spin_lock(lock);
543                         goto repeat;
544                 }
545         }
546         spin_unlock(lock);
547         return err;
548 }
549
550 /**
551  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
552  * @mapping: the mapping which wants those buffers written
553  *
554  * Starts I/O against the buffers at mapping->private_list, and waits upon
555  * that I/O.
556  *
557  * Basically, this is a convenience function for fsync().
558  * @mapping is a file or directory which needs those buffers to be written for
559  * a successful fsync().
560  */
561 int sync_mapping_buffers(struct address_space *mapping)
562 {
563         struct address_space *buffer_mapping = mapping->assoc_mapping;
564
565         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
566                 return 0;
567
568         return fsync_buffers_list(&buffer_mapping->private_lock,
569                                         &mapping->private_list);
570 }
571 EXPORT_SYMBOL(sync_mapping_buffers);
572
573 /*
574  * Called when we've recently written block `bblock', and it is known that
575  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
576  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
577  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
578  */
579 void write_boundary_block(struct block_device *bdev,
580                         sector_t bblock, unsigned blocksize)
581 {
582         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
583         if (bh) {
584                 if (buffer_dirty(bh))
585                         ll_rw_block(WRITE, 1, &bh);
586                 put_bh(bh);
587         }
588 }
589
590 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
591 {
592         struct address_space *mapping = inode->i_mapping;
593         struct address_space *buffer_mapping = bh->b_page->mapping;
594
595         mark_buffer_dirty(bh);
596         if (!mapping->assoc_mapping) {
597                 mapping->assoc_mapping = buffer_mapping;
598         } else {
599                 BUG_ON(mapping->assoc_mapping != buffer_mapping);
600         }
601         if (!bh->b_assoc_map) {
602                 spin_lock(&buffer_mapping->private_lock);
603                 list_move_tail(&bh->b_assoc_buffers,
604                                 &mapping->private_list);
605                 bh->b_assoc_map = mapping;
606                 spin_unlock(&buffer_mapping->private_lock);
607         }
608 }
609 EXPORT_SYMBOL(mark_buffer_dirty_inode);
610
611 /*
612  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
613  * dirty.
614  *
615  * If warn is true, then emit a warning if the page is not uptodate and has
616  * not been truncated.
617  */
618 static void __set_page_dirty(struct page *page,
619                 struct address_space *mapping, int warn)
620 {
621         spin_lock_irq(&mapping->tree_lock);
622         if (page->mapping) {    /* Race with truncate? */
623                 WARN_ON_ONCE(warn && !PageUptodate(page));
624
625                 if (mapping_cap_account_dirty(mapping)) {
626                         __inc_zone_page_state(page, NR_FILE_DIRTY);
627                         __inc_bdi_stat(mapping->backing_dev_info,
628                                         BDI_RECLAIMABLE);
629                         task_dirty_inc(current);
630                         task_io_account_write(PAGE_CACHE_SIZE);
631                 }
632                 radix_tree_tag_set(&mapping->page_tree,
633                                 page_index(page), PAGECACHE_TAG_DIRTY);
634         }
635         spin_unlock_irq(&mapping->tree_lock);
636         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
637 }
638
639 /*
640  * Add a page to the dirty page list.
641  *
642  * It is a sad fact of life that this function is called from several places
643  * deeply under spinlocking.  It may not sleep.
644  *
645  * If the page has buffers, the uptodate buffers are set dirty, to preserve
646  * dirty-state coherency between the page and the buffers.  It the page does
647  * not have buffers then when they are later attached they will all be set
648  * dirty.
649  *
650  * The buffers are dirtied before the page is dirtied.  There's a small race
651  * window in which a writepage caller may see the page cleanness but not the
652  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
653  * before the buffers, a concurrent writepage caller could clear the page dirty
654  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
655  * page on the dirty page list.
656  *
657  * We use private_lock to lock against try_to_free_buffers while using the
658  * page's buffer list.  Also use this to protect against clean buffers being
659  * added to the page after it was set dirty.
660  *
661  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
662  * address_space though.
663  */
664 int __set_page_dirty_buffers(struct page *page)
665 {
666         int newly_dirty;
667         struct address_space *mapping = page_mapping(page);
668
669         if (unlikely(!mapping))
670                 return !TestSetPageDirty(page);
671
672         spin_lock(&mapping->private_lock);
673         if (page_has_buffers(page)) {
674                 struct buffer_head *head = page_buffers(page);
675                 struct buffer_head *bh = head;
676
677                 do {
678                         set_buffer_dirty(bh);
679                         bh = bh->b_this_page;
680                 } while (bh != head);
681         }
682         newly_dirty = !TestSetPageDirty(page);
683         spin_unlock(&mapping->private_lock);
684
685         if (newly_dirty)
686                 __set_page_dirty(page, mapping, 1);
687         return newly_dirty;
688 }
689 EXPORT_SYMBOL(__set_page_dirty_buffers);
690
691 /*
692  * Write out and wait upon a list of buffers.
693  *
694  * We have conflicting pressures: we want to make sure that all
695  * initially dirty buffers get waited on, but that any subsequently
696  * dirtied buffers don't.  After all, we don't want fsync to last
697  * forever if somebody is actively writing to the file.
698  *
699  * Do this in two main stages: first we copy dirty buffers to a
700  * temporary inode list, queueing the writes as we go.  Then we clean
701  * up, waiting for those writes to complete.
702  * 
703  * During this second stage, any subsequent updates to the file may end
704  * up refiling the buffer on the original inode's dirty list again, so
705  * there is a chance we will end up with a buffer queued for write but
706  * not yet completed on that list.  So, as a final cleanup we go through
707  * the osync code to catch these locked, dirty buffers without requeuing
708  * any newly dirty buffers for write.
709  */
710 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
711 {
712         struct buffer_head *bh;
713         struct list_head tmp;
714         struct address_space *mapping;
715         int err = 0, err2;
716
717         INIT_LIST_HEAD(&tmp);
718
719         spin_lock(lock);
720         while (!list_empty(list)) {
721                 bh = BH_ENTRY(list->next);
722                 mapping = bh->b_assoc_map;
723                 __remove_assoc_queue(bh);
724                 /* Avoid race with mark_buffer_dirty_inode() which does
725                  * a lockless check and we rely on seeing the dirty bit */
726                 smp_mb();
727                 if (buffer_dirty(bh) || buffer_locked(bh)) {
728                         list_add(&bh->b_assoc_buffers, &tmp);
729                         bh->b_assoc_map = mapping;
730                         if (buffer_dirty(bh)) {
731                                 get_bh(bh);
732                                 spin_unlock(lock);
733                                 /*
734                                  * Ensure any pending I/O completes so that
735                                  * ll_rw_block() actually writes the current
736                                  * contents - it is a noop if I/O is still in
737                                  * flight on potentially older contents.
738                                  */
739                                 ll_rw_block(SWRITE_SYNC, 1, &bh);
740                                 brelse(bh);
741                                 spin_lock(lock);
742                         }
743                 }
744         }
745
746         while (!list_empty(&tmp)) {
747                 bh = BH_ENTRY(tmp.prev);
748                 get_bh(bh);
749                 mapping = bh->b_assoc_map;
750                 __remove_assoc_queue(bh);
751                 /* Avoid race with mark_buffer_dirty_inode() which does
752                  * a lockless check and we rely on seeing the dirty bit */
753                 smp_mb();
754                 if (buffer_dirty(bh)) {
755                         list_add(&bh->b_assoc_buffers,
756                                  &mapping->private_list);
757                         bh->b_assoc_map = mapping;
758                 }
759                 spin_unlock(lock);
760                 wait_on_buffer(bh);
761                 if (!buffer_uptodate(bh))
762                         err = -EIO;
763                 brelse(bh);
764                 spin_lock(lock);
765         }
766         
767         spin_unlock(lock);
768         err2 = osync_buffers_list(lock, list);
769         if (err)
770                 return err;
771         else
772                 return err2;
773 }
774
775 /*
776  * Invalidate any and all dirty buffers on a given inode.  We are
777  * probably unmounting the fs, but that doesn't mean we have already
778  * done a sync().  Just drop the buffers from the inode list.
779  *
780  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
781  * assumes that all the buffers are against the blockdev.  Not true
782  * for reiserfs.
783  */
784 void invalidate_inode_buffers(struct inode *inode)
785 {
786         if (inode_has_buffers(inode)) {
787                 struct address_space *mapping = &inode->i_data;
788                 struct list_head *list = &mapping->private_list;
789                 struct address_space *buffer_mapping = mapping->assoc_mapping;
790
791                 spin_lock(&buffer_mapping->private_lock);
792                 while (!list_empty(list))
793                         __remove_assoc_queue(BH_ENTRY(list->next));
794                 spin_unlock(&buffer_mapping->private_lock);
795         }
796 }
797 EXPORT_SYMBOL(invalidate_inode_buffers);
798
799 /*
800  * Remove any clean buffers from the inode's buffer list.  This is called
801  * when we're trying to free the inode itself.  Those buffers can pin it.
802  *
803  * Returns true if all buffers were removed.
804  */
805 int remove_inode_buffers(struct inode *inode)
806 {
807         int ret = 1;
808
809         if (inode_has_buffers(inode)) {
810                 struct address_space *mapping = &inode->i_data;
811                 struct list_head *list = &mapping->private_list;
812                 struct address_space *buffer_mapping = mapping->assoc_mapping;
813
814                 spin_lock(&buffer_mapping->private_lock);
815                 while (!list_empty(list)) {
816                         struct buffer_head *bh = BH_ENTRY(list->next);
817                         if (buffer_dirty(bh)) {
818                                 ret = 0;
819                                 break;
820                         }
821                         __remove_assoc_queue(bh);
822                 }
823                 spin_unlock(&buffer_mapping->private_lock);
824         }
825         return ret;
826 }
827
828 /*
829  * Create the appropriate buffers when given a page for data area and
830  * the size of each buffer.. Use the bh->b_this_page linked list to
831  * follow the buffers created.  Return NULL if unable to create more
832  * buffers.
833  *
834  * The retry flag is used to differentiate async IO (paging, swapping)
835  * which may not fail from ordinary buffer allocations.
836  */
837 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
838                 int retry)
839 {
840         struct buffer_head *bh, *head;
841         long offset;
842
843 try_again:
844         head = NULL;
845         offset = PAGE_SIZE;
846         while ((offset -= size) >= 0) {
847                 bh = alloc_buffer_head(GFP_NOFS);
848                 if (!bh)
849                         goto no_grow;
850
851                 bh->b_bdev = NULL;
852                 bh->b_this_page = head;
853                 bh->b_blocknr = -1;
854                 head = bh;
855
856                 bh->b_state = 0;
857                 atomic_set(&bh->b_count, 0);
858                 bh->b_private = NULL;
859                 bh->b_size = size;
860
861                 /* Link the buffer to its page */
862                 set_bh_page(bh, page, offset);
863
864                 init_buffer(bh, NULL, NULL);
865         }
866         return head;
867 /*
868  * In case anything failed, we just free everything we got.
869  */
870 no_grow:
871         if (head) {
872                 do {
873                         bh = head;
874                         head = head->b_this_page;
875                         free_buffer_head(bh);
876                 } while (head);
877         }
878
879         /*
880          * Return failure for non-async IO requests.  Async IO requests
881          * are not allowed to fail, so we have to wait until buffer heads
882          * become available.  But we don't want tasks sleeping with 
883          * partially complete buffers, so all were released above.
884          */
885         if (!retry)
886                 return NULL;
887
888         /* We're _really_ low on memory. Now we just
889          * wait for old buffer heads to become free due to
890          * finishing IO.  Since this is an async request and
891          * the reserve list is empty, we're sure there are 
892          * async buffer heads in use.
893          */
894         free_more_memory();
895         goto try_again;
896 }
897 EXPORT_SYMBOL_GPL(alloc_page_buffers);
898
899 static inline void
900 link_dev_buffers(struct page *page, struct buffer_head *head)
901 {
902         struct buffer_head *bh, *tail;
903
904         bh = head;
905         do {
906                 tail = bh;
907                 bh = bh->b_this_page;
908         } while (bh);
909         tail->b_this_page = head;
910         attach_page_buffers(page, head);
911 }
912
913 /*
914  * Initialise the state of a blockdev page's buffers.
915  */ 
916 static void
917 init_page_buffers(struct page *page, struct block_device *bdev,
918                         sector_t block, int size)
919 {
920         struct buffer_head *head = page_buffers(page);
921         struct buffer_head *bh = head;
922         int uptodate = PageUptodate(page);
923
924         do {
925                 if (!buffer_mapped(bh)) {
926                         init_buffer(bh, NULL, NULL);
927                         bh->b_bdev = bdev;
928                         bh->b_blocknr = block;
929                         if (uptodate)
930                                 set_buffer_uptodate(bh);
931                         set_buffer_mapped(bh);
932                 }
933                 block++;
934                 bh = bh->b_this_page;
935         } while (bh != head);
936 }
937
938 /*
939  * Create the page-cache page that contains the requested block.
940  *
941  * This is user purely for blockdev mappings.
942  */
943 static struct page *
944 grow_dev_page(struct block_device *bdev, sector_t block,
945                 pgoff_t index, int size)
946 {
947         struct inode *inode = bdev->bd_inode;
948         struct page *page;
949         struct buffer_head *bh;
950
951         page = find_or_create_page(inode->i_mapping, index,
952                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
953         if (!page)
954                 return NULL;
955
956         BUG_ON(!PageLocked(page));
957
958         if (page_has_buffers(page)) {
959                 bh = page_buffers(page);
960                 if (bh->b_size == size) {
961                         init_page_buffers(page, bdev, block, size);
962                         return page;
963                 }
964                 if (!try_to_free_buffers(page))
965                         goto failed;
966         }
967
968         /*
969          * Allocate some buffers for this page
970          */
971         bh = alloc_page_buffers(page, size, 0);
972         if (!bh)
973                 goto failed;
974
975         /*
976          * Link the page to the buffers and initialise them.  Take the
977          * lock to be atomic wrt __find_get_block(), which does not
978          * run under the page lock.
979          */
980         spin_lock(&inode->i_mapping->private_lock);
981         link_dev_buffers(page, bh);
982         init_page_buffers(page, bdev, block, size);
983         spin_unlock(&inode->i_mapping->private_lock);
984         return page;
985
986 failed:
987         BUG();
988         unlock_page(page);
989         page_cache_release(page);
990         return NULL;
991 }
992
993 /*
994  * Create buffers for the specified block device block's page.  If
995  * that page was dirty, the buffers are set dirty also.
996  */
997 static int
998 grow_buffers(struct block_device *bdev, sector_t block, int size)
999 {
1000         struct page *page;
1001         pgoff_t index;
1002         int sizebits;
1003
1004         sizebits = -1;
1005         do {
1006                 sizebits++;
1007         } while ((size << sizebits) < PAGE_SIZE);
1008
1009         index = block >> sizebits;
1010
1011         /*
1012          * Check for a block which wants to lie outside our maximum possible
1013          * pagecache index.  (this comparison is done using sector_t types).
1014          */
1015         if (unlikely(index != block >> sizebits)) {
1016                 char b[BDEVNAME_SIZE];
1017
1018                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1019                         "device %s\n",
1020                         __func__, (unsigned long long)block,
1021                         bdevname(bdev, b));
1022                 return -EIO;
1023         }
1024         block = index << sizebits;
1025         /* Create a page with the proper size buffers.. */
1026         page = grow_dev_page(bdev, block, index, size);
1027         if (!page)
1028                 return 0;
1029         unlock_page(page);
1030         page_cache_release(page);
1031         return 1;
1032 }
1033
1034 static struct buffer_head *
1035 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1036 {
1037         /* Size must be multiple of hard sectorsize */
1038         if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1039                         (size < 512 || size > PAGE_SIZE))) {
1040                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1041                                         size);
1042                 printk(KERN_ERR "hardsect size: %d\n",
1043                                         bdev_hardsect_size(bdev));
1044
1045                 dump_stack();
1046                 return NULL;
1047         }
1048
1049         for (;;) {
1050                 struct buffer_head * bh;
1051                 int ret;
1052
1053                 bh = __find_get_block(bdev, block, size);
1054                 if (bh)
1055                         return bh;
1056
1057                 ret = grow_buffers(bdev, block, size);
1058                 if (ret < 0)
1059                         return NULL;
1060                 if (ret == 0)
1061                         free_more_memory();
1062         }
1063 }
1064
1065 /*
1066  * The relationship between dirty buffers and dirty pages:
1067  *
1068  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1069  * the page is tagged dirty in its radix tree.
1070  *
1071  * At all times, the dirtiness of the buffers represents the dirtiness of
1072  * subsections of the page.  If the page has buffers, the page dirty bit is
1073  * merely a hint about the true dirty state.
1074  *
1075  * When a page is set dirty in its entirety, all its buffers are marked dirty
1076  * (if the page has buffers).
1077  *
1078  * When a buffer is marked dirty, its page is dirtied, but the page's other
1079  * buffers are not.
1080  *
1081  * Also.  When blockdev buffers are explicitly read with bread(), they
1082  * individually become uptodate.  But their backing page remains not
1083  * uptodate - even if all of its buffers are uptodate.  A subsequent
1084  * block_read_full_page() against that page will discover all the uptodate
1085  * buffers, will set the page uptodate and will perform no I/O.
1086  */
1087
1088 /**
1089  * mark_buffer_dirty - mark a buffer_head as needing writeout
1090  * @bh: the buffer_head to mark dirty
1091  *
1092  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1093  * backing page dirty, then tag the page as dirty in its address_space's radix
1094  * tree and then attach the address_space's inode to its superblock's dirty
1095  * inode list.
1096  *
1097  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1098  * mapping->tree_lock and the global inode_lock.
1099  */
1100 void mark_buffer_dirty(struct buffer_head *bh)
1101 {
1102         WARN_ON_ONCE(!buffer_uptodate(bh));
1103
1104         /*
1105          * Very *carefully* optimize the it-is-already-dirty case.
1106          *
1107          * Don't let the final "is it dirty" escape to before we
1108          * perhaps modified the buffer.
1109          */
1110         if (buffer_dirty(bh)) {
1111                 smp_mb();
1112                 if (buffer_dirty(bh))
1113                         return;
1114         }
1115
1116         if (!test_set_buffer_dirty(bh)) {
1117                 struct page *page = bh->b_page;
1118                 if (!TestSetPageDirty(page))
1119                         __set_page_dirty(page, page_mapping(page), 0);
1120         }
1121 }
1122
1123 /*
1124  * Decrement a buffer_head's reference count.  If all buffers against a page
1125  * have zero reference count, are clean and unlocked, and if the page is clean
1126  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1127  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1128  * a page but it ends up not being freed, and buffers may later be reattached).
1129  */
1130 void __brelse(struct buffer_head * buf)
1131 {
1132         if (atomic_read(&buf->b_count)) {
1133                 put_bh(buf);
1134                 return;
1135         }
1136         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1137 }
1138
1139 /*
1140  * bforget() is like brelse(), except it discards any
1141  * potentially dirty data.
1142  */
1143 void __bforget(struct buffer_head *bh)
1144 {
1145         clear_buffer_dirty(bh);
1146         if (bh->b_assoc_map) {
1147                 struct address_space *buffer_mapping = bh->b_page->mapping;
1148
1149                 spin_lock(&buffer_mapping->private_lock);
1150                 list_del_init(&bh->b_assoc_buffers);
1151                 bh->b_assoc_map = NULL;
1152                 spin_unlock(&buffer_mapping->private_lock);
1153         }
1154         __brelse(bh);
1155 }
1156
1157 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1158 {
1159         lock_buffer(bh);
1160         if (buffer_uptodate(bh)) {
1161                 unlock_buffer(bh);
1162                 return bh;
1163         } else {
1164                 get_bh(bh);
1165                 bh->b_end_io = end_buffer_read_sync;
1166                 submit_bh(READ, bh);
1167                 wait_on_buffer(bh);
1168                 if (buffer_uptodate(bh))
1169                         return bh;
1170         }
1171         brelse(bh);
1172         return NULL;
1173 }
1174
1175 /*
1176  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1177  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1178  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1179  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1180  * CPU's LRUs at the same time.
1181  *
1182  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1183  * sb_find_get_block().
1184  *
1185  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1186  * a local interrupt disable for that.
1187  */
1188
1189 #define BH_LRU_SIZE     8
1190
1191 struct bh_lru {
1192         struct buffer_head *bhs[BH_LRU_SIZE];
1193 };
1194
1195 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1196
1197 #ifdef CONFIG_SMP
1198 #define bh_lru_lock()   local_irq_disable()
1199 #define bh_lru_unlock() local_irq_enable()
1200 #else
1201 #define bh_lru_lock()   preempt_disable()
1202 #define bh_lru_unlock() preempt_enable()
1203 #endif
1204
1205 static inline void check_irqs_on(void)
1206 {
1207 #ifdef irqs_disabled
1208         BUG_ON(irqs_disabled());
1209 #endif
1210 }
1211
1212 /*
1213  * The LRU management algorithm is dopey-but-simple.  Sorry.
1214  */
1215 static void bh_lru_install(struct buffer_head *bh)
1216 {
1217         struct buffer_head *evictee = NULL;
1218         struct bh_lru *lru;
1219
1220         check_irqs_on();
1221         bh_lru_lock();
1222         lru = &__get_cpu_var(bh_lrus);
1223         if (lru->bhs[0] != bh) {
1224                 struct buffer_head *bhs[BH_LRU_SIZE];
1225                 int in;
1226                 int out = 0;
1227
1228                 get_bh(bh);
1229                 bhs[out++] = bh;
1230                 for (in = 0; in < BH_LRU_SIZE; in++) {
1231                         struct buffer_head *bh2 = lru->bhs[in];
1232
1233                         if (bh2 == bh) {
1234                                 __brelse(bh2);
1235                         } else {
1236                                 if (out >= BH_LRU_SIZE) {
1237                                         BUG_ON(evictee != NULL);
1238                                         evictee = bh2;
1239                                 } else {
1240                                         bhs[out++] = bh2;
1241                                 }
1242                         }
1243                 }
1244                 while (out < BH_LRU_SIZE)
1245                         bhs[out++] = NULL;
1246                 memcpy(lru->bhs, bhs, sizeof(bhs));
1247         }
1248         bh_lru_unlock();
1249
1250         if (evictee)
1251                 __brelse(evictee);
1252 }
1253
1254 /*
1255  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1256  */
1257 static struct buffer_head *
1258 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1259 {
1260         struct buffer_head *ret = NULL;
1261         struct bh_lru *lru;
1262         unsigned int i;
1263
1264         check_irqs_on();
1265         bh_lru_lock();
1266         lru = &__get_cpu_var(bh_lrus);
1267         for (i = 0; i < BH_LRU_SIZE; i++) {
1268                 struct buffer_head *bh = lru->bhs[i];
1269
1270                 if (bh && bh->b_bdev == bdev &&
1271                                 bh->b_blocknr == block && bh->b_size == size) {
1272                         if (i) {
1273                                 while (i) {
1274                                         lru->bhs[i] = lru->bhs[i - 1];
1275                                         i--;
1276                                 }
1277                                 lru->bhs[0] = bh;
1278                         }
1279                         get_bh(bh);
1280                         ret = bh;
1281                         break;
1282                 }
1283         }
1284         bh_lru_unlock();
1285         return ret;
1286 }
1287
1288 /*
1289  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1290  * it in the LRU and mark it as accessed.  If it is not present then return
1291  * NULL
1292  */
1293 struct buffer_head *
1294 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1295 {
1296         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1297
1298         if (bh == NULL) {
1299                 bh = __find_get_block_slow(bdev, block);
1300                 if (bh)
1301                         bh_lru_install(bh);
1302         }
1303         if (bh)
1304                 touch_buffer(bh);
1305         return bh;
1306 }
1307 EXPORT_SYMBOL(__find_get_block);
1308
1309 /*
1310  * __getblk will locate (and, if necessary, create) the buffer_head
1311  * which corresponds to the passed block_device, block and size. The
1312  * returned buffer has its reference count incremented.
1313  *
1314  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1315  * illegal block number, __getblk() will happily return a buffer_head
1316  * which represents the non-existent block.  Very weird.
1317  *
1318  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1319  * attempt is failing.  FIXME, perhaps?
1320  */
1321 struct buffer_head *
1322 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1323 {
1324         struct buffer_head *bh = __find_get_block(bdev, block, size);
1325
1326         might_sleep();
1327         if (bh == NULL)
1328                 bh = __getblk_slow(bdev, block, size);
1329         return bh;
1330 }
1331 EXPORT_SYMBOL(__getblk);
1332
1333 /*
1334  * Do async read-ahead on a buffer..
1335  */
1336 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1337 {
1338         struct buffer_head *bh = __getblk(bdev, block, size);
1339         if (likely(bh)) {
1340                 ll_rw_block(READA, 1, &bh);
1341                 brelse(bh);
1342         }
1343 }
1344 EXPORT_SYMBOL(__breadahead);
1345
1346 /**
1347  *  __bread() - reads a specified block and returns the bh
1348  *  @bdev: the block_device to read from
1349  *  @block: number of block
1350  *  @size: size (in bytes) to read
1351  * 
1352  *  Reads a specified block, and returns buffer head that contains it.
1353  *  It returns NULL if the block was unreadable.
1354  */
1355 struct buffer_head *
1356 __bread(struct block_device *bdev, sector_t block, unsigned size)
1357 {
1358         struct buffer_head *bh = __getblk(bdev, block, size);
1359
1360         if (likely(bh) && !buffer_uptodate(bh))
1361                 bh = __bread_slow(bh);
1362         return bh;
1363 }
1364 EXPORT_SYMBOL(__bread);
1365
1366 /*
1367  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1368  * This doesn't race because it runs in each cpu either in irq
1369  * or with preempt disabled.
1370  */
1371 static void invalidate_bh_lru(void *arg)
1372 {
1373         struct bh_lru *b = &get_cpu_var(bh_lrus);
1374         int i;
1375
1376         for (i = 0; i < BH_LRU_SIZE; i++) {
1377                 brelse(b->bhs[i]);
1378                 b->bhs[i] = NULL;
1379         }
1380         put_cpu_var(bh_lrus);
1381 }
1382         
1383 void invalidate_bh_lrus(void)
1384 {
1385         on_each_cpu(invalidate_bh_lru, NULL, 1);
1386 }
1387 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1388
1389 void set_bh_page(struct buffer_head *bh,
1390                 struct page *page, unsigned long offset)
1391 {
1392         bh->b_page = page;
1393         BUG_ON(offset >= PAGE_SIZE);
1394         if (PageHighMem(page))
1395                 /*
1396                  * This catches illegal uses and preserves the offset:
1397                  */
1398                 bh->b_data = (char *)(0 + offset);
1399         else
1400                 bh->b_data = page_address(page) + offset;
1401 }
1402 EXPORT_SYMBOL(set_bh_page);
1403
1404 /*
1405  * Called when truncating a buffer on a page completely.
1406  */
1407 static void discard_buffer(struct buffer_head * bh)
1408 {
1409         lock_buffer(bh);
1410         clear_buffer_dirty(bh);
1411         bh->b_bdev = NULL;
1412         clear_buffer_mapped(bh);
1413         clear_buffer_req(bh);
1414         clear_buffer_new(bh);
1415         clear_buffer_delay(bh);
1416         clear_buffer_unwritten(bh);
1417         unlock_buffer(bh);
1418 }
1419
1420 /**
1421  * block_invalidatepage - invalidate part of all of a buffer-backed page
1422  *
1423  * @page: the page which is affected
1424  * @offset: the index of the truncation point
1425  *
1426  * block_invalidatepage() is called when all or part of the page has become
1427  * invalidatedby a truncate operation.
1428  *
1429  * block_invalidatepage() does not have to release all buffers, but it must
1430  * ensure that no dirty buffer is left outside @offset and that no I/O
1431  * is underway against any of the blocks which are outside the truncation
1432  * point.  Because the caller is about to free (and possibly reuse) those
1433  * blocks on-disk.
1434  */
1435 void block_invalidatepage(struct page *page, unsigned long offset)
1436 {
1437         struct buffer_head *head, *bh, *next;
1438         unsigned int curr_off = 0;
1439
1440         BUG_ON(!PageLocked(page));
1441         if (!page_has_buffers(page))
1442                 goto out;
1443
1444         head = page_buffers(page);
1445         bh = head;
1446         do {
1447                 unsigned int next_off = curr_off + bh->b_size;
1448                 next = bh->b_this_page;
1449
1450                 /*
1451                  * is this block fully invalidated?
1452                  */
1453                 if (offset <= curr_off)
1454                         discard_buffer(bh);
1455                 curr_off = next_off;
1456                 bh = next;
1457         } while (bh != head);
1458
1459         /*
1460          * We release buffers only if the entire page is being invalidated.
1461          * The get_block cached value has been unconditionally invalidated,
1462          * so real IO is not possible anymore.
1463          */
1464         if (offset == 0)
1465                 try_to_release_page(page, 0);
1466 out:
1467         return;
1468 }
1469 EXPORT_SYMBOL(block_invalidatepage);
1470
1471 /*
1472  * We attach and possibly dirty the buffers atomically wrt
1473  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1474  * is already excluded via the page lock.
1475  */
1476 void create_empty_buffers(struct page *page,
1477                         unsigned long blocksize, unsigned long b_state)
1478 {
1479         struct buffer_head *bh, *head, *tail;
1480
1481         head = alloc_page_buffers(page, blocksize, 1);
1482         bh = head;
1483         do {
1484                 bh->b_state |= b_state;
1485                 tail = bh;
1486                 bh = bh->b_this_page;
1487         } while (bh);
1488         tail->b_this_page = head;
1489
1490         spin_lock(&page->mapping->private_lock);
1491         if (PageUptodate(page) || PageDirty(page)) {
1492                 bh = head;
1493                 do {
1494                         if (PageDirty(page))
1495                                 set_buffer_dirty(bh);
1496                         if (PageUptodate(page))
1497                                 set_buffer_uptodate(bh);
1498                         bh = bh->b_this_page;
1499                 } while (bh != head);
1500         }
1501         attach_page_buffers(page, head);
1502         spin_unlock(&page->mapping->private_lock);
1503 }
1504 EXPORT_SYMBOL(create_empty_buffers);
1505
1506 /*
1507  * We are taking a block for data and we don't want any output from any
1508  * buffer-cache aliases starting from return from that function and
1509  * until the moment when something will explicitly mark the buffer
1510  * dirty (hopefully that will not happen until we will free that block ;-)
1511  * We don't even need to mark it not-uptodate - nobody can expect
1512  * anything from a newly allocated buffer anyway. We used to used
1513  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1514  * don't want to mark the alias unmapped, for example - it would confuse
1515  * anyone who might pick it with bread() afterwards...
1516  *
1517  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1518  * be writeout I/O going on against recently-freed buffers.  We don't
1519  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1520  * only if we really need to.  That happens here.
1521  */
1522 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1523 {
1524         struct buffer_head *old_bh;
1525
1526         might_sleep();
1527
1528         old_bh = __find_get_block_slow(bdev, block);
1529         if (old_bh) {
1530                 clear_buffer_dirty(old_bh);
1531                 wait_on_buffer(old_bh);
1532                 clear_buffer_req(old_bh);
1533                 __brelse(old_bh);
1534         }
1535 }
1536 EXPORT_SYMBOL(unmap_underlying_metadata);
1537
1538 /*
1539  * NOTE! All mapped/uptodate combinations are valid:
1540  *
1541  *      Mapped  Uptodate        Meaning
1542  *
1543  *      No      No              "unknown" - must do get_block()
1544  *      No      Yes             "hole" - zero-filled
1545  *      Yes     No              "allocated" - allocated on disk, not read in
1546  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1547  *
1548  * "Dirty" is valid only with the last case (mapped+uptodate).
1549  */
1550
1551 /*
1552  * While block_write_full_page is writing back the dirty buffers under
1553  * the page lock, whoever dirtied the buffers may decide to clean them
1554  * again at any time.  We handle that by only looking at the buffer
1555  * state inside lock_buffer().
1556  *
1557  * If block_write_full_page() is called for regular writeback
1558  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1559  * locked buffer.   This only can happen if someone has written the buffer
1560  * directly, with submit_bh().  At the address_space level PageWriteback
1561  * prevents this contention from occurring.
1562  */
1563 static int __block_write_full_page(struct inode *inode, struct page *page,
1564                         get_block_t *get_block, struct writeback_control *wbc)
1565 {
1566         int err;
1567         sector_t block;
1568         sector_t last_block;
1569         struct buffer_head *bh, *head;
1570         const unsigned blocksize = 1 << inode->i_blkbits;
1571         int nr_underway = 0;
1572
1573         BUG_ON(!PageLocked(page));
1574
1575         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1576
1577         if (!page_has_buffers(page)) {
1578                 create_empty_buffers(page, blocksize,
1579                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1580         }
1581
1582         /*
1583          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1584          * here, and the (potentially unmapped) buffers may become dirty at
1585          * any time.  If a buffer becomes dirty here after we've inspected it
1586          * then we just miss that fact, and the page stays dirty.
1587          *
1588          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1589          * handle that here by just cleaning them.
1590          */
1591
1592         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1593         head = page_buffers(page);
1594         bh = head;
1595
1596         /*
1597          * Get all the dirty buffers mapped to disk addresses and
1598          * handle any aliases from the underlying blockdev's mapping.
1599          */
1600         do {
1601                 if (block > last_block) {
1602                         /*
1603                          * mapped buffers outside i_size will occur, because
1604                          * this page can be outside i_size when there is a
1605                          * truncate in progress.
1606                          */
1607                         /*
1608                          * The buffer was zeroed by block_write_full_page()
1609                          */
1610                         clear_buffer_dirty(bh);
1611                         set_buffer_uptodate(bh);
1612                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1613                            buffer_dirty(bh)) {
1614                         WARN_ON(bh->b_size != blocksize);
1615                         err = get_block(inode, block, bh, 1);
1616                         if (err)
1617                                 goto recover;
1618                         clear_buffer_delay(bh);
1619                         if (buffer_new(bh)) {
1620                                 /* blockdev mappings never come here */
1621                                 clear_buffer_new(bh);
1622                                 unmap_underlying_metadata(bh->b_bdev,
1623                                                         bh->b_blocknr);
1624                         }
1625                 }
1626                 bh = bh->b_this_page;
1627                 block++;
1628         } while (bh != head);
1629
1630         do {
1631                 if (!buffer_mapped(bh))
1632                         continue;
1633                 /*
1634                  * If it's a fully non-blocking write attempt and we cannot
1635                  * lock the buffer then redirty the page.  Note that this can
1636                  * potentially cause a busy-wait loop from pdflush and kswapd
1637                  * activity, but those code paths have their own higher-level
1638                  * throttling.
1639                  */
1640                 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1641                         lock_buffer(bh);
1642                 } else if (!trylock_buffer(bh)) {
1643                         redirty_page_for_writepage(wbc, page);
1644                         continue;
1645                 }
1646                 if (test_clear_buffer_dirty(bh)) {
1647                         mark_buffer_async_write(bh);
1648                 } else {
1649                         unlock_buffer(bh);
1650                 }
1651         } while ((bh = bh->b_this_page) != head);
1652
1653         /*
1654          * The page and its buffers are protected by PageWriteback(), so we can
1655          * drop the bh refcounts early.
1656          */
1657         BUG_ON(PageWriteback(page));
1658         set_page_writeback(page);
1659
1660         do {
1661                 struct buffer_head *next = bh->b_this_page;
1662                 if (buffer_async_write(bh)) {
1663                         submit_bh(WRITE, bh);
1664                         nr_underway++;
1665                 }
1666                 bh = next;
1667         } while (bh != head);
1668         unlock_page(page);
1669
1670         err = 0;
1671 done:
1672         if (nr_underway == 0) {
1673                 /*
1674                  * The page was marked dirty, but the buffers were
1675                  * clean.  Someone wrote them back by hand with
1676                  * ll_rw_block/submit_bh.  A rare case.
1677                  */
1678                 end_page_writeback(page);
1679
1680                 /*
1681                  * The page and buffer_heads can be released at any time from
1682                  * here on.
1683                  */
1684         }
1685         return err;
1686
1687 recover:
1688         /*
1689          * ENOSPC, or some other error.  We may already have added some
1690          * blocks to the file, so we need to write these out to avoid
1691          * exposing stale data.
1692          * The page is currently locked and not marked for writeback
1693          */
1694         bh = head;
1695         /* Recovery: lock and submit the mapped buffers */
1696         do {
1697                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1698                     !buffer_delay(bh)) {
1699                         lock_buffer(bh);
1700                         mark_buffer_async_write(bh);
1701                 } else {
1702                         /*
1703                          * The buffer may have been set dirty during
1704                          * attachment to a dirty page.
1705                          */
1706                         clear_buffer_dirty(bh);
1707                 }
1708         } while ((bh = bh->b_this_page) != head);
1709         SetPageError(page);
1710         BUG_ON(PageWriteback(page));
1711         mapping_set_error(page->mapping, err);
1712         set_page_writeback(page);
1713         do {
1714                 struct buffer_head *next = bh->b_this_page;
1715                 if (buffer_async_write(bh)) {
1716                         clear_buffer_dirty(bh);
1717                         submit_bh(WRITE, bh);
1718                         nr_underway++;
1719                 }
1720                 bh = next;
1721         } while (bh != head);
1722         unlock_page(page);
1723         goto done;
1724 }
1725
1726 /*
1727  * If a page has any new buffers, zero them out here, and mark them uptodate
1728  * and dirty so they'll be written out (in order to prevent uninitialised
1729  * block data from leaking). And clear the new bit.
1730  */
1731 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1732 {
1733         unsigned int block_start, block_end;
1734         struct buffer_head *head, *bh;
1735
1736         BUG_ON(!PageLocked(page));
1737         if (!page_has_buffers(page))
1738                 return;
1739
1740         bh = head = page_buffers(page);
1741         block_start = 0;
1742         do {
1743                 block_end = block_start + bh->b_size;
1744
1745                 if (buffer_new(bh)) {
1746                         if (block_end > from && block_start < to) {
1747                                 if (!PageUptodate(page)) {
1748                                         unsigned start, size;
1749
1750                                         start = max(from, block_start);
1751                                         size = min(to, block_end) - start;
1752
1753                                         zero_user(page, start, size);
1754                                         set_buffer_uptodate(bh);
1755                                 }
1756
1757                                 clear_buffer_new(bh);
1758                                 mark_buffer_dirty(bh);
1759                         }
1760                 }
1761
1762                 block_start = block_end;
1763                 bh = bh->b_this_page;
1764         } while (bh != head);
1765 }
1766 EXPORT_SYMBOL(page_zero_new_buffers);
1767
1768 static int __block_prepare_write(struct inode *inode, struct page *page,
1769                 unsigned from, unsigned to, get_block_t *get_block)
1770 {
1771         unsigned block_start, block_end;
1772         sector_t block;
1773         int err = 0;
1774         unsigned blocksize, bbits;
1775         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1776
1777         BUG_ON(!PageLocked(page));
1778         BUG_ON(from > PAGE_CACHE_SIZE);
1779         BUG_ON(to > PAGE_CACHE_SIZE);
1780         BUG_ON(from > to);
1781
1782         blocksize = 1 << inode->i_blkbits;
1783         if (!page_has_buffers(page))
1784                 create_empty_buffers(page, blocksize, 0);
1785         head = page_buffers(page);
1786
1787         bbits = inode->i_blkbits;
1788         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1789
1790         for(bh = head, block_start = 0; bh != head || !block_start;
1791             block++, block_start=block_end, bh = bh->b_this_page) {
1792                 block_end = block_start + blocksize;
1793                 if (block_end <= from || block_start >= to) {
1794                         if (PageUptodate(page)) {
1795                                 if (!buffer_uptodate(bh))
1796                                         set_buffer_uptodate(bh);
1797                         }
1798                         continue;
1799                 }
1800                 if (buffer_new(bh))
1801                         clear_buffer_new(bh);
1802                 if (!buffer_mapped(bh)) {
1803                         WARN_ON(bh->b_size != blocksize);
1804                         err = get_block(inode, block, bh, 1);
1805                         if (err)
1806                                 break;
1807                         if (buffer_new(bh)) {
1808                                 unmap_underlying_metadata(bh->b_bdev,
1809                                                         bh->b_blocknr);
1810                                 if (PageUptodate(page)) {
1811                                         clear_buffer_new(bh);
1812                                         set_buffer_uptodate(bh);
1813                                         mark_buffer_dirty(bh);
1814                                         continue;
1815                                 }
1816                                 if (block_end > to || block_start < from)
1817                                         zero_user_segments(page,
1818                                                 to, block_end,
1819                                                 block_start, from);
1820                                 continue;
1821                         }
1822                 }
1823                 if (PageUptodate(page)) {
1824                         if (!buffer_uptodate(bh))
1825                                 set_buffer_uptodate(bh);
1826                         continue; 
1827                 }
1828                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1829                     !buffer_unwritten(bh) &&
1830                      (block_start < from || block_end > to)) {
1831                         ll_rw_block(READ, 1, &bh);
1832                         *wait_bh++=bh;
1833                 }
1834         }
1835         /*
1836          * If we issued read requests - let them complete.
1837          */
1838         while(wait_bh > wait) {
1839                 wait_on_buffer(*--wait_bh);
1840                 if (!buffer_uptodate(*wait_bh))
1841                         err = -EIO;
1842         }
1843         if (unlikely(err))
1844                 page_zero_new_buffers(page, from, to);
1845         return err;
1846 }
1847
1848 static int __block_commit_write(struct inode *inode, struct page *page,
1849                 unsigned from, unsigned to)
1850 {
1851         unsigned block_start, block_end;
1852         int partial = 0;
1853         unsigned blocksize;
1854         struct buffer_head *bh, *head;
1855
1856         blocksize = 1 << inode->i_blkbits;
1857
1858         for(bh = head = page_buffers(page), block_start = 0;
1859             bh != head || !block_start;
1860             block_start=block_end, bh = bh->b_this_page) {
1861                 block_end = block_start + blocksize;
1862                 if (block_end <= from || block_start >= to) {
1863                         if (!buffer_uptodate(bh))
1864                                 partial = 1;
1865                 } else {
1866                         set_buffer_uptodate(bh);
1867                         mark_buffer_dirty(bh);
1868                 }
1869                 clear_buffer_new(bh);
1870         }
1871
1872         /*
1873          * If this is a partial write which happened to make all buffers
1874          * uptodate then we can optimize away a bogus readpage() for
1875          * the next read(). Here we 'discover' whether the page went
1876          * uptodate as a result of this (potentially partial) write.
1877          */
1878         if (!partial)
1879                 SetPageUptodate(page);
1880         return 0;
1881 }
1882
1883 /*
1884  * block_write_begin takes care of the basic task of block allocation and
1885  * bringing partial write blocks uptodate first.
1886  *
1887  * If *pagep is not NULL, then block_write_begin uses the locked page
1888  * at *pagep rather than allocating its own. In this case, the page will
1889  * not be unlocked or deallocated on failure.
1890  */
1891 int block_write_begin(struct file *file, struct address_space *mapping,
1892                         loff_t pos, unsigned len, unsigned flags,
1893                         struct page **pagep, void **fsdata,
1894                         get_block_t *get_block)
1895 {
1896         struct inode *inode = mapping->host;
1897         int status = 0;
1898         struct page *page;
1899         pgoff_t index;
1900         unsigned start, end;
1901         int ownpage = 0;
1902
1903         index = pos >> PAGE_CACHE_SHIFT;
1904         start = pos & (PAGE_CACHE_SIZE - 1);
1905         end = start + len;
1906
1907         page = *pagep;
1908         if (page == NULL) {
1909                 ownpage = 1;
1910                 page = grab_cache_page_write_begin(mapping, index, flags);
1911                 if (!page) {
1912                         status = -ENOMEM;
1913                         goto out;
1914                 }
1915                 *pagep = page;
1916         } else
1917                 BUG_ON(!PageLocked(page));
1918
1919         status = __block_prepare_write(inode, page, start, end, get_block);
1920         if (unlikely(status)) {
1921                 ClearPageUptodate(page);
1922
1923                 if (ownpage) {
1924                         unlock_page(page);
1925                         page_cache_release(page);
1926                         *pagep = NULL;
1927
1928                         /*
1929                          * prepare_write() may have instantiated a few blocks
1930                          * outside i_size.  Trim these off again. Don't need
1931                          * i_size_read because we hold i_mutex.
1932                          */
1933                         if (pos + len > inode->i_size)
1934                                 vmtruncate(inode, inode->i_size);
1935                 }
1936         }
1937
1938 out:
1939         return status;
1940 }
1941 EXPORT_SYMBOL(block_write_begin);
1942
1943 int block_write_end(struct file *file, struct address_space *mapping,
1944                         loff_t pos, unsigned len, unsigned copied,
1945                         struct page *page, void *fsdata)
1946 {
1947         struct inode *inode = mapping->host;
1948         unsigned start;
1949
1950         start = pos & (PAGE_CACHE_SIZE - 1);
1951
1952         if (unlikely(copied < len)) {
1953                 /*
1954                  * The buffers that were written will now be uptodate, so we
1955                  * don't have to worry about a readpage reading them and
1956                  * overwriting a partial write. However if we have encountered
1957                  * a short write and only partially written into a buffer, it
1958                  * will not be marked uptodate, so a readpage might come in and
1959                  * destroy our partial write.
1960                  *
1961                  * Do the simplest thing, and just treat any short write to a
1962                  * non uptodate page as a zero-length write, and force the
1963                  * caller to redo the whole thing.
1964                  */
1965                 if (!PageUptodate(page))
1966                         copied = 0;
1967
1968                 page_zero_new_buffers(page, start+copied, start+len);
1969         }
1970         flush_dcache_page(page);
1971
1972         /* This could be a short (even 0-length) commit */
1973         __block_commit_write(inode, page, start, start+copied);
1974
1975         return copied;
1976 }
1977 EXPORT_SYMBOL(block_write_end);
1978
1979 int generic_write_end(struct file *file, struct address_space *mapping,
1980                         loff_t pos, unsigned len, unsigned copied,
1981                         struct page *page, void *fsdata)
1982 {
1983         struct inode *inode = mapping->host;
1984         int i_size_changed = 0;
1985
1986         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1987
1988         /*
1989          * No need to use i_size_read() here, the i_size
1990          * cannot change under us because we hold i_mutex.
1991          *
1992          * But it's important to update i_size while still holding page lock:
1993          * page writeout could otherwise come in and zero beyond i_size.
1994          */
1995         if (pos+copied > inode->i_size) {
1996                 i_size_write(inode, pos+copied);
1997                 i_size_changed = 1;
1998         }
1999
2000         unlock_page(page);
2001         page_cache_release(page);
2002
2003         /*
2004          * Don't mark the inode dirty under page lock. First, it unnecessarily
2005          * makes the holding time of page lock longer. Second, it forces lock
2006          * ordering of page lock and transaction start for journaling
2007          * filesystems.
2008          */
2009         if (i_size_changed)
2010                 mark_inode_dirty(inode);
2011
2012         return copied;
2013 }
2014 EXPORT_SYMBOL(generic_write_end);
2015
2016 /*
2017  * block_is_partially_uptodate checks whether buffers within a page are
2018  * uptodate or not.
2019  *
2020  * Returns true if all buffers which correspond to a file portion
2021  * we want to read are uptodate.
2022  */
2023 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2024                                         unsigned long from)
2025 {
2026         struct inode *inode = page->mapping->host;
2027         unsigned block_start, block_end, blocksize;
2028         unsigned to;
2029         struct buffer_head *bh, *head;
2030         int ret = 1;
2031
2032         if (!page_has_buffers(page))
2033                 return 0;
2034
2035         blocksize = 1 << inode->i_blkbits;
2036         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2037         to = from + to;
2038         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2039                 return 0;
2040
2041         head = page_buffers(page);
2042         bh = head;
2043         block_start = 0;
2044         do {
2045                 block_end = block_start + blocksize;
2046                 if (block_end > from && block_start < to) {
2047                         if (!buffer_uptodate(bh)) {
2048                                 ret = 0;
2049                                 break;
2050                         }
2051                         if (block_end >= to)
2052                                 break;
2053                 }
2054                 block_start = block_end;
2055                 bh = bh->b_this_page;
2056         } while (bh != head);
2057
2058         return ret;
2059 }
2060 EXPORT_SYMBOL(block_is_partially_uptodate);
2061
2062 /*
2063  * Generic "read page" function for block devices that have the normal
2064  * get_block functionality. This is most of the block device filesystems.
2065  * Reads the page asynchronously --- the unlock_buffer() and
2066  * set/clear_buffer_uptodate() functions propagate buffer state into the
2067  * page struct once IO has completed.
2068  */
2069 int block_read_full_page(struct page *page, get_block_t *get_block)
2070 {
2071         struct inode *inode = page->mapping->host;
2072         sector_t iblock, lblock;
2073         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2074         unsigned int blocksize;
2075         int nr, i;
2076         int fully_mapped = 1;
2077
2078         BUG_ON(!PageLocked(page));
2079         blocksize = 1 << inode->i_blkbits;
2080         if (!page_has_buffers(page))
2081                 create_empty_buffers(page, blocksize, 0);
2082         head = page_buffers(page);
2083
2084         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2085         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2086         bh = head;
2087         nr = 0;
2088         i = 0;
2089
2090         do {
2091                 if (buffer_uptodate(bh))
2092                         continue;
2093
2094                 if (!buffer_mapped(bh)) {
2095                         int err = 0;
2096
2097                         fully_mapped = 0;
2098                         if (iblock < lblock) {
2099                                 WARN_ON(bh->b_size != blocksize);
2100                                 err = get_block(inode, iblock, bh, 0);
2101                                 if (err)
2102                                         SetPageError(page);
2103                         }
2104                         if (!buffer_mapped(bh)) {
2105                                 zero_user(page, i * blocksize, blocksize);
2106                                 if (!err)
2107                                         set_buffer_uptodate(bh);
2108                                 continue;
2109                         }
2110                         /*
2111                          * get_block() might have updated the buffer
2112                          * synchronously
2113                          */
2114                         if (buffer_uptodate(bh))
2115                                 continue;
2116                 }
2117                 arr[nr++] = bh;
2118         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2119
2120         if (fully_mapped)
2121                 SetPageMappedToDisk(page);
2122
2123         if (!nr) {
2124                 /*
2125                  * All buffers are uptodate - we can set the page uptodate
2126                  * as well. But not if get_block() returned an error.
2127                  */
2128                 if (!PageError(page))
2129                         SetPageUptodate(page);
2130                 unlock_page(page);
2131                 return 0;
2132         }
2133
2134         /* Stage two: lock the buffers */
2135         for (i = 0; i < nr; i++) {
2136                 bh = arr[i];
2137                 lock_buffer(bh);
2138                 mark_buffer_async_read(bh);
2139         }
2140
2141         /*
2142          * Stage 3: start the IO.  Check for uptodateness
2143          * inside the buffer lock in case another process reading
2144          * the underlying blockdev brought it uptodate (the sct fix).
2145          */
2146         for (i = 0; i < nr; i++) {
2147                 bh = arr[i];
2148                 if (buffer_uptodate(bh))
2149                         end_buffer_async_read(bh, 1);
2150                 else
2151                         submit_bh(READ, bh);
2152         }
2153         return 0;
2154 }
2155
2156 /* utility function for filesystems that need to do work on expanding
2157  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2158  * deal with the hole.  
2159  */
2160 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2161 {
2162         struct address_space *mapping = inode->i_mapping;
2163         struct page *page;
2164         void *fsdata;
2165         unsigned long limit;
2166         int err;
2167
2168         err = -EFBIG;
2169         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2170         if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2171                 send_sig(SIGXFSZ, current, 0);
2172                 goto out;
2173         }
2174         if (size > inode->i_sb->s_maxbytes)
2175                 goto out;
2176
2177         err = pagecache_write_begin(NULL, mapping, size, 0,
2178                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2179                                 &page, &fsdata);
2180         if (err)
2181                 goto out;
2182
2183         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2184         BUG_ON(err > 0);
2185
2186 out:
2187         return err;
2188 }
2189
2190 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2191                             loff_t pos, loff_t *bytes)
2192 {
2193         struct inode *inode = mapping->host;
2194         unsigned blocksize = 1 << inode->i_blkbits;
2195         struct page *page;
2196         void *fsdata;
2197         pgoff_t index, curidx;
2198         loff_t curpos;
2199         unsigned zerofrom, offset, len;
2200         int err = 0;
2201
2202         index = pos >> PAGE_CACHE_SHIFT;
2203         offset = pos & ~PAGE_CACHE_MASK;
2204
2205         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2206                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2207                 if (zerofrom & (blocksize-1)) {
2208                         *bytes |= (blocksize-1);
2209                         (*bytes)++;
2210                 }
2211                 len = PAGE_CACHE_SIZE - zerofrom;
2212
2213                 err = pagecache_write_begin(file, mapping, curpos, len,
2214                                                 AOP_FLAG_UNINTERRUPTIBLE,
2215                                                 &page, &fsdata);
2216                 if (err)
2217                         goto out;
2218                 zero_user(page, zerofrom, len);
2219                 err = pagecache_write_end(file, mapping, curpos, len, len,
2220                                                 page, fsdata);
2221                 if (err < 0)
2222                         goto out;
2223                 BUG_ON(err != len);
2224                 err = 0;
2225
2226                 balance_dirty_pages_ratelimited(mapping);
2227         }
2228
2229         /* page covers the boundary, find the boundary offset */
2230         if (index == curidx) {
2231                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2232                 /* if we will expand the thing last block will be filled */
2233                 if (offset <= zerofrom) {
2234                         goto out;
2235                 }
2236                 if (zerofrom & (blocksize-1)) {
2237                         *bytes |= (blocksize-1);
2238                         (*bytes)++;
2239                 }
2240                 len = offset - zerofrom;
2241
2242                 err = pagecache_write_begin(file, mapping, curpos, len,
2243                                                 AOP_FLAG_UNINTERRUPTIBLE,
2244                                                 &page, &fsdata);
2245                 if (err)
2246                         goto out;
2247                 zero_user(page, zerofrom, len);
2248                 err = pagecache_write_end(file, mapping, curpos, len, len,
2249                                                 page, fsdata);
2250                 if (err < 0)
2251                         goto out;
2252                 BUG_ON(err != len);
2253                 err = 0;
2254         }
2255 out:
2256         return err;
2257 }
2258
2259 /*
2260  * For moronic filesystems that do not allow holes in file.
2261  * We may have to extend the file.
2262  */
2263 int cont_write_begin(struct file *file, struct address_space *mapping,
2264                         loff_t pos, unsigned len, unsigned flags,
2265                         struct page **pagep, void **fsdata,
2266                         get_block_t *get_block, loff_t *bytes)
2267 {
2268         struct inode *inode = mapping->host;
2269         unsigned blocksize = 1 << inode->i_blkbits;
2270         unsigned zerofrom;
2271         int err;
2272
2273         err = cont_expand_zero(file, mapping, pos, bytes);
2274         if (err)
2275                 goto out;
2276
2277         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2278         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2279                 *bytes |= (blocksize-1);
2280                 (*bytes)++;
2281         }
2282
2283         *pagep = NULL;
2284         err = block_write_begin(file, mapping, pos, len,
2285                                 flags, pagep, fsdata, get_block);
2286 out:
2287         return err;
2288 }
2289
2290 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2291                         get_block_t *get_block)
2292 {
2293         struct inode *inode = page->mapping->host;
2294         int err = __block_prepare_write(inode, page, from, to, get_block);
2295         if (err)
2296                 ClearPageUptodate(page);
2297         return err;
2298 }
2299
2300 int block_commit_write(struct page *page, unsigned from, unsigned to)
2301 {
2302         struct inode *inode = page->mapping->host;
2303         __block_commit_write(inode,page,from,to);
2304         return 0;
2305 }
2306
2307 /*
2308  * block_page_mkwrite() is not allowed to change the file size as it gets
2309  * called from a page fault handler when a page is first dirtied. Hence we must
2310  * be careful to check for EOF conditions here. We set the page up correctly
2311  * for a written page which means we get ENOSPC checking when writing into
2312  * holes and correct delalloc and unwritten extent mapping on filesystems that
2313  * support these features.
2314  *
2315  * We are not allowed to take the i_mutex here so we have to play games to
2316  * protect against truncate races as the page could now be beyond EOF.  Because
2317  * vmtruncate() writes the inode size before removing pages, once we have the
2318  * page lock we can determine safely if the page is beyond EOF. If it is not
2319  * beyond EOF, then the page is guaranteed safe against truncation until we
2320  * unlock the page.
2321  */
2322 int
2323 block_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2324                    get_block_t get_block)
2325 {
2326         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2327         unsigned long end;
2328         loff_t size;
2329         int ret = -EINVAL;
2330
2331         lock_page(page);
2332         size = i_size_read(inode);
2333         if ((page->mapping != inode->i_mapping) ||
2334             (page_offset(page) > size)) {
2335                 /* page got truncated out from underneath us */
2336                 goto out_unlock;
2337         }
2338
2339         /* page is wholly or partially inside EOF */
2340         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2341                 end = size & ~PAGE_CACHE_MASK;
2342         else
2343                 end = PAGE_CACHE_SIZE;
2344
2345         ret = block_prepare_write(page, 0, end, get_block);
2346         if (!ret)
2347                 ret = block_commit_write(page, 0, end);
2348
2349 out_unlock:
2350         unlock_page(page);
2351         return ret;
2352 }
2353
2354 /*
2355  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2356  * immediately, while under the page lock.  So it needs a special end_io
2357  * handler which does not touch the bh after unlocking it.
2358  */
2359 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2360 {
2361         __end_buffer_read_notouch(bh, uptodate);
2362 }
2363
2364 /*
2365  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2366  * the page (converting it to circular linked list and taking care of page
2367  * dirty races).
2368  */
2369 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2370 {
2371         struct buffer_head *bh;
2372
2373         BUG_ON(!PageLocked(page));
2374
2375         spin_lock(&page->mapping->private_lock);
2376         bh = head;
2377         do {
2378                 if (PageDirty(page))
2379                         set_buffer_dirty(bh);
2380                 if (!bh->b_this_page)
2381                         bh->b_this_page = head;
2382                 bh = bh->b_this_page;
2383         } while (bh != head);
2384         attach_page_buffers(page, head);
2385         spin_unlock(&page->mapping->private_lock);
2386 }
2387
2388 /*
2389  * On entry, the page is fully not uptodate.
2390  * On exit the page is fully uptodate in the areas outside (from,to)
2391  */
2392 int nobh_write_begin(struct file *file, struct address_space *mapping,
2393                         loff_t pos, unsigned len, unsigned flags,
2394                         struct page **pagep, void **fsdata,
2395                         get_block_t *get_block)
2396 {
2397         struct inode *inode = mapping->host;
2398         const unsigned blkbits = inode->i_blkbits;
2399         const unsigned blocksize = 1 << blkbits;
2400         struct buffer_head *head, *bh;
2401         struct page *page;
2402         pgoff_t index;
2403         unsigned from, to;
2404         unsigned block_in_page;
2405         unsigned block_start, block_end;
2406         sector_t block_in_file;
2407         int nr_reads = 0;
2408         int ret = 0;
2409         int is_mapped_to_disk = 1;
2410
2411         index = pos >> PAGE_CACHE_SHIFT;
2412         from = pos & (PAGE_CACHE_SIZE - 1);
2413         to = from + len;
2414
2415         page = grab_cache_page_write_begin(mapping, index, flags);
2416         if (!page)
2417                 return -ENOMEM;
2418         *pagep = page;
2419         *fsdata = NULL;
2420
2421         if (page_has_buffers(page)) {
2422                 unlock_page(page);
2423                 page_cache_release(page);
2424                 *pagep = NULL;
2425                 return block_write_begin(file, mapping, pos, len, flags, pagep,
2426                                         fsdata, get_block);
2427         }
2428
2429         if (PageMappedToDisk(page))
2430                 return 0;
2431
2432         /*
2433          * Allocate buffers so that we can keep track of state, and potentially
2434          * attach them to the page if an error occurs. In the common case of
2435          * no error, they will just be freed again without ever being attached
2436          * to the page (which is all OK, because we're under the page lock).
2437          *
2438          * Be careful: the buffer linked list is a NULL terminated one, rather
2439          * than the circular one we're used to.
2440          */
2441         head = alloc_page_buffers(page, blocksize, 0);
2442         if (!head) {
2443                 ret = -ENOMEM;
2444                 goto out_release;
2445         }
2446
2447         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2448
2449         /*
2450          * We loop across all blocks in the page, whether or not they are
2451          * part of the affected region.  This is so we can discover if the
2452          * page is fully mapped-to-disk.
2453          */
2454         for (block_start = 0, block_in_page = 0, bh = head;
2455                   block_start < PAGE_CACHE_SIZE;
2456                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2457                 int create;
2458
2459                 block_end = block_start + blocksize;
2460                 bh->b_state = 0;
2461                 create = 1;
2462                 if (block_start >= to)
2463                         create = 0;
2464                 ret = get_block(inode, block_in_file + block_in_page,
2465                                         bh, create);
2466                 if (ret)
2467                         goto failed;
2468                 if (!buffer_mapped(bh))
2469                         is_mapped_to_disk = 0;
2470                 if (buffer_new(bh))
2471                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2472                 if (PageUptodate(page)) {
2473                         set_buffer_uptodate(bh);
2474                         continue;
2475                 }
2476                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2477                         zero_user_segments(page, block_start, from,
2478                                                         to, block_end);
2479                         continue;
2480                 }
2481                 if (buffer_uptodate(bh))
2482                         continue;       /* reiserfs does this */
2483                 if (block_start < from || block_end > to) {
2484                         lock_buffer(bh);
2485                         bh->b_end_io = end_buffer_read_nobh;
2486                         submit_bh(READ, bh);
2487                         nr_reads++;
2488                 }
2489         }
2490
2491         if (nr_reads) {
2492                 /*
2493                  * The page is locked, so these buffers are protected from
2494                  * any VM or truncate activity.  Hence we don't need to care
2495                  * for the buffer_head refcounts.
2496                  */
2497                 for (bh = head; bh; bh = bh->b_this_page) {
2498                         wait_on_buffer(bh);
2499                         if (!buffer_uptodate(bh))
2500                                 ret = -EIO;
2501                 }
2502                 if (ret)
2503                         goto failed;
2504         }
2505
2506         if (is_mapped_to_disk)
2507                 SetPageMappedToDisk(page);
2508
2509         *fsdata = head; /* to be released by nobh_write_end */
2510
2511         return 0;
2512
2513 failed:
2514         BUG_ON(!ret);
2515         /*
2516          * Error recovery is a bit difficult. We need to zero out blocks that
2517          * were newly allocated, and dirty them to ensure they get written out.
2518          * Buffers need to be attached to the page at this point, otherwise
2519          * the handling of potential IO errors during writeout would be hard
2520          * (could try doing synchronous writeout, but what if that fails too?)
2521          */
2522         attach_nobh_buffers(page, head);
2523         page_zero_new_buffers(page, from, to);
2524
2525 out_release:
2526         unlock_page(page);
2527         page_cache_release(page);
2528         *pagep = NULL;
2529
2530         if (pos + len > inode->i_size)
2531                 vmtruncate(inode, inode->i_size);
2532
2533         return ret;
2534 }
2535 EXPORT_SYMBOL(nobh_write_begin);
2536
2537 int nobh_write_end(struct file *file, struct address_space *mapping,
2538                         loff_t pos, unsigned len, unsigned copied,
2539                         struct page *page, void *fsdata)
2540 {
2541         struct inode *inode = page->mapping->host;
2542         struct buffer_head *head = fsdata;
2543         struct buffer_head *bh;
2544         BUG_ON(fsdata != NULL && page_has_buffers(page));
2545
2546         if (unlikely(copied < len) && head)
2547                 attach_nobh_buffers(page, head);
2548         if (page_has_buffers(page))
2549                 return generic_write_end(file, mapping, pos, len,
2550                                         copied, page, fsdata);
2551
2552         SetPageUptodate(page);
2553         set_page_dirty(page);
2554         if (pos+copied > inode->i_size) {
2555                 i_size_write(inode, pos+copied);
2556                 mark_inode_dirty(inode);
2557         }
2558
2559         unlock_page(page);
2560         page_cache_release(page);
2561
2562         while (head) {
2563                 bh = head;
2564                 head = head->b_this_page;
2565                 free_buffer_head(bh);
2566         }
2567
2568         return copied;
2569 }
2570 EXPORT_SYMBOL(nobh_write_end);
2571
2572 /*
2573  * nobh_writepage() - based on block_full_write_page() except
2574  * that it tries to operate without attaching bufferheads to
2575  * the page.
2576  */
2577 int nobh_writepage(struct page *page, get_block_t *get_block,
2578                         struct writeback_control *wbc)
2579 {
2580         struct inode * const inode = page->mapping->host;
2581         loff_t i_size = i_size_read(inode);
2582         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2583         unsigned offset;
2584         int ret;
2585
2586         /* Is the page fully inside i_size? */
2587         if (page->index < end_index)
2588                 goto out;
2589
2590         /* Is the page fully outside i_size? (truncate in progress) */
2591         offset = i_size & (PAGE_CACHE_SIZE-1);
2592         if (page->index >= end_index+1 || !offset) {
2593                 /*
2594                  * The page may have dirty, unmapped buffers.  For example,
2595                  * they may have been added in ext3_writepage().  Make them
2596                  * freeable here, so the page does not leak.
2597                  */
2598 #if 0
2599                 /* Not really sure about this  - do we need this ? */
2600                 if (page->mapping->a_ops->invalidatepage)
2601                         page->mapping->a_ops->invalidatepage(page, offset);
2602 #endif
2603                 unlock_page(page);
2604                 return 0; /* don't care */
2605         }
2606
2607         /*
2608          * The page straddles i_size.  It must be zeroed out on each and every
2609          * writepage invocation because it may be mmapped.  "A file is mapped
2610          * in multiples of the page size.  For a file that is not a multiple of
2611          * the  page size, the remaining memory is zeroed when mapped, and
2612          * writes to that region are not written out to the file."
2613          */
2614         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2615 out:
2616         ret = mpage_writepage(page, get_block, wbc);
2617         if (ret == -EAGAIN)
2618                 ret = __block_write_full_page(inode, page, get_block, wbc);
2619         return ret;
2620 }
2621 EXPORT_SYMBOL(nobh_writepage);
2622
2623 int nobh_truncate_page(struct address_space *mapping,
2624                         loff_t from, get_block_t *get_block)
2625 {
2626         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2627         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2628         unsigned blocksize;
2629         sector_t iblock;
2630         unsigned length, pos;
2631         struct inode *inode = mapping->host;
2632         struct page *page;
2633         struct buffer_head map_bh;
2634         int err;
2635
2636         blocksize = 1 << inode->i_blkbits;
2637         length = offset & (blocksize - 1);
2638
2639         /* Block boundary? Nothing to do */
2640         if (!length)
2641                 return 0;
2642
2643         length = blocksize - length;
2644         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2645
2646         page = grab_cache_page(mapping, index);
2647         err = -ENOMEM;
2648         if (!page)
2649                 goto out;
2650
2651         if (page_has_buffers(page)) {
2652 has_buffers:
2653                 unlock_page(page);
2654                 page_cache_release(page);
2655                 return block_truncate_page(mapping, from, get_block);
2656         }
2657
2658         /* Find the buffer that contains "offset" */
2659         pos = blocksize;
2660         while (offset >= pos) {
2661                 iblock++;
2662                 pos += blocksize;
2663         }
2664
2665         err = get_block(inode, iblock, &map_bh, 0);
2666         if (err)
2667                 goto unlock;
2668         /* unmapped? It's a hole - nothing to do */
2669         if (!buffer_mapped(&map_bh))
2670                 goto unlock;
2671
2672         /* Ok, it's mapped. Make sure it's up-to-date */
2673         if (!PageUptodate(page)) {
2674                 err = mapping->a_ops->readpage(NULL, page);
2675                 if (err) {
2676                         page_cache_release(page);
2677                         goto out;
2678                 }
2679                 lock_page(page);
2680                 if (!PageUptodate(page)) {
2681                         err = -EIO;
2682                         goto unlock;
2683                 }
2684                 if (page_has_buffers(page))
2685                         goto has_buffers;
2686         }
2687         zero_user(page, offset, length);
2688         set_page_dirty(page);
2689         err = 0;
2690
2691 unlock:
2692         unlock_page(page);
2693         page_cache_release(page);
2694 out:
2695         return err;
2696 }
2697 EXPORT_SYMBOL(nobh_truncate_page);
2698
2699 int block_truncate_page(struct address_space *mapping,
2700                         loff_t from, get_block_t *get_block)
2701 {
2702         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2703         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2704         unsigned blocksize;
2705         sector_t iblock;
2706         unsigned length, pos;
2707         struct inode *inode = mapping->host;
2708         struct page *page;
2709         struct buffer_head *bh;
2710         int err;
2711
2712         blocksize = 1 << inode->i_blkbits;
2713         length = offset & (blocksize - 1);
2714
2715         /* Block boundary? Nothing to do */
2716         if (!length)
2717                 return 0;
2718
2719         length = blocksize - length;
2720         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2721         
2722         page = grab_cache_page(mapping, index);
2723         err = -ENOMEM;
2724         if (!page)
2725                 goto out;
2726
2727         if (!page_has_buffers(page))
2728                 create_empty_buffers(page, blocksize, 0);
2729
2730         /* Find the buffer that contains "offset" */
2731         bh = page_buffers(page);
2732         pos = blocksize;
2733         while (offset >= pos) {
2734                 bh = bh->b_this_page;
2735                 iblock++;
2736                 pos += blocksize;
2737         }
2738
2739         err = 0;
2740         if (!buffer_mapped(bh)) {
2741                 WARN_ON(bh->b_size != blocksize);
2742                 err = get_block(inode, iblock, bh, 0);
2743                 if (err)
2744                         goto unlock;
2745                 /* unmapped? It's a hole - nothing to do */
2746                 if (!buffer_mapped(bh))
2747                         goto unlock;
2748         }
2749
2750         /* Ok, it's mapped. Make sure it's up-to-date */
2751         if (PageUptodate(page))
2752                 set_buffer_uptodate(bh);
2753
2754         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2755                 err = -EIO;
2756                 ll_rw_block(READ, 1, &bh);
2757                 wait_on_buffer(bh);
2758                 /* Uhhuh. Read error. Complain and punt. */
2759                 if (!buffer_uptodate(bh))
2760                         goto unlock;
2761         }
2762
2763         zero_user(page, offset, length);
2764         mark_buffer_dirty(bh);
2765         err = 0;
2766
2767 unlock:
2768         unlock_page(page);
2769         page_cache_release(page);
2770 out:
2771         return err;
2772 }
2773
2774 /*
2775  * The generic ->writepage function for buffer-backed address_spaces
2776  */
2777 int block_write_full_page(struct page *page, get_block_t *get_block,
2778                         struct writeback_control *wbc)
2779 {
2780         struct inode * const inode = page->mapping->host;
2781         loff_t i_size = i_size_read(inode);
2782         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2783         unsigned offset;
2784
2785         /* Is the page fully inside i_size? */
2786         if (page->index < end_index)
2787                 return __block_write_full_page(inode, page, get_block, wbc);
2788
2789         /* Is the page fully outside i_size? (truncate in progress) */
2790         offset = i_size & (PAGE_CACHE_SIZE-1);
2791         if (page->index >= end_index+1 || !offset) {
2792                 /*
2793                  * The page may have dirty, unmapped buffers.  For example,
2794                  * they may have been added in ext3_writepage().  Make them
2795                  * freeable here, so the page does not leak.
2796                  */
2797                 do_invalidatepage(page, 0);
2798                 unlock_page(page);
2799                 return 0; /* don't care */
2800         }
2801
2802         /*
2803          * The page straddles i_size.  It must be zeroed out on each and every
2804          * writepage invokation because it may be mmapped.  "A file is mapped
2805          * in multiples of the page size.  For a file that is not a multiple of
2806          * the  page size, the remaining memory is zeroed when mapped, and
2807          * writes to that region are not written out to the file."
2808          */
2809         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2810         return __block_write_full_page(inode, page, get_block, wbc);
2811 }
2812
2813 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2814                             get_block_t *get_block)
2815 {
2816         struct buffer_head tmp;
2817         struct inode *inode = mapping->host;
2818         tmp.b_state = 0;
2819         tmp.b_blocknr = 0;
2820         tmp.b_size = 1 << inode->i_blkbits;
2821         get_block(inode, block, &tmp, 0);
2822         return tmp.b_blocknr;
2823 }
2824
2825 static void end_bio_bh_io_sync(struct bio *bio, int err)
2826 {
2827         struct buffer_head *bh = bio->bi_private;
2828
2829         if (err == -EOPNOTSUPP) {
2830                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2831                 set_bit(BH_Eopnotsupp, &bh->b_state);
2832         }
2833
2834         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2835                 set_bit(BH_Quiet, &bh->b_state);
2836
2837         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2838         bio_put(bio);
2839 }
2840
2841 int submit_bh(int rw, struct buffer_head * bh)
2842 {
2843         struct bio *bio;
2844         int ret = 0;
2845
2846         BUG_ON(!buffer_locked(bh));
2847         BUG_ON(!buffer_mapped(bh));
2848         BUG_ON(!bh->b_end_io);
2849
2850         /*
2851          * Mask in barrier bit for a write (could be either a WRITE or a
2852          * WRITE_SYNC
2853          */
2854         if (buffer_ordered(bh) && (rw & WRITE))
2855                 rw |= WRITE_BARRIER;
2856
2857         /*
2858          * Only clear out a write error when rewriting
2859          */
2860         if (test_set_buffer_req(bh) && (rw & WRITE))
2861                 clear_buffer_write_io_error(bh);
2862
2863         /*
2864          * from here on down, it's all bio -- do the initial mapping,
2865          * submit_bio -> generic_make_request may further map this bio around
2866          */
2867         bio = bio_alloc(GFP_NOIO, 1);
2868
2869         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2870         bio->bi_bdev = bh->b_bdev;
2871         bio->bi_io_vec[0].bv_page = bh->b_page;
2872         bio->bi_io_vec[0].bv_len = bh->b_size;
2873         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2874
2875         bio->bi_vcnt = 1;
2876         bio->bi_idx = 0;
2877         bio->bi_size = bh->b_size;
2878
2879         bio->bi_end_io = end_bio_bh_io_sync;
2880         bio->bi_private = bh;
2881
2882         bio_get(bio);
2883         submit_bio(rw, bio);
2884
2885         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2886                 ret = -EOPNOTSUPP;
2887
2888         bio_put(bio);
2889         return ret;
2890 }
2891
2892 /**
2893  * ll_rw_block: low-level access to block devices (DEPRECATED)
2894  * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
2895  * @nr: number of &struct buffer_heads in the array
2896  * @bhs: array of pointers to &struct buffer_head
2897  *
2898  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2899  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2900  * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2901  * are sent to disk. The fourth %READA option is described in the documentation
2902  * for generic_make_request() which ll_rw_block() calls.
2903  *
2904  * This function drops any buffer that it cannot get a lock on (with the
2905  * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
2906  * clean when doing a write request, and any buffer that appears to be
2907  * up-to-date when doing read request.  Further it marks as clean buffers that
2908  * are processed for writing (the buffer cache won't assume that they are
2909  * actually clean until the buffer gets unlocked).
2910  *
2911  * ll_rw_block sets b_end_io to simple completion handler that marks
2912  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2913  * any waiters. 
2914  *
2915  * All of the buffers must be for the same device, and must also be a
2916  * multiple of the current approved size for the device.
2917  */
2918 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2919 {
2920         int i;
2921
2922         for (i = 0; i < nr; i++) {
2923                 struct buffer_head *bh = bhs[i];
2924
2925                 if (rw == SWRITE || rw == SWRITE_SYNC)
2926                         lock_buffer(bh);
2927                 else if (!trylock_buffer(bh))
2928                         continue;
2929
2930                 if (rw == WRITE || rw == SWRITE || rw == SWRITE_SYNC) {
2931                         if (test_clear_buffer_dirty(bh)) {
2932                                 bh->b_end_io = end_buffer_write_sync;
2933                                 get_bh(bh);
2934                                 if (rw == SWRITE_SYNC)
2935                                         submit_bh(WRITE_SYNC, bh);
2936                                 else
2937                                         submit_bh(WRITE, bh);
2938                                 continue;
2939                         }
2940                 } else {
2941                         if (!buffer_uptodate(bh)) {
2942                                 bh->b_end_io = end_buffer_read_sync;
2943                                 get_bh(bh);
2944                                 submit_bh(rw, bh);
2945                                 continue;
2946                         }
2947                 }
2948                 unlock_buffer(bh);
2949         }
2950 }
2951
2952 /*
2953  * For a data-integrity writeout, we need to wait upon any in-progress I/O
2954  * and then start new I/O and then wait upon it.  The caller must have a ref on
2955  * the buffer_head.
2956  */
2957 int sync_dirty_buffer(struct buffer_head *bh)
2958 {
2959         int ret = 0;
2960
2961         WARN_ON(atomic_read(&bh->b_count) < 1);
2962         lock_buffer(bh);
2963         if (test_clear_buffer_dirty(bh)) {
2964                 get_bh(bh);
2965                 bh->b_end_io = end_buffer_write_sync;
2966                 ret = submit_bh(WRITE, bh);
2967                 wait_on_buffer(bh);
2968                 if (buffer_eopnotsupp(bh)) {
2969                         clear_buffer_eopnotsupp(bh);
2970                         ret = -EOPNOTSUPP;
2971                 }
2972                 if (!ret && !buffer_uptodate(bh))
2973                         ret = -EIO;
2974         } else {
2975                 unlock_buffer(bh);
2976         }
2977         return ret;
2978 }
2979
2980 /*
2981  * try_to_free_buffers() checks if all the buffers on this particular page
2982  * are unused, and releases them if so.
2983  *
2984  * Exclusion against try_to_free_buffers may be obtained by either
2985  * locking the page or by holding its mapping's private_lock.
2986  *
2987  * If the page is dirty but all the buffers are clean then we need to
2988  * be sure to mark the page clean as well.  This is because the page
2989  * may be against a block device, and a later reattachment of buffers
2990  * to a dirty page will set *all* buffers dirty.  Which would corrupt
2991  * filesystem data on the same device.
2992  *
2993  * The same applies to regular filesystem pages: if all the buffers are
2994  * clean then we set the page clean and proceed.  To do that, we require
2995  * total exclusion from __set_page_dirty_buffers().  That is obtained with
2996  * private_lock.
2997  *
2998  * try_to_free_buffers() is non-blocking.
2999  */
3000 static inline int buffer_busy(struct buffer_head *bh)
3001 {
3002         return atomic_read(&bh->b_count) |
3003                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3004 }
3005
3006 static int
3007 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3008 {
3009         struct buffer_head *head = page_buffers(page);
3010         struct buffer_head *bh;
3011
3012         bh = head;
3013         do {
3014                 if (buffer_write_io_error(bh) && page->mapping)
3015                         set_bit(AS_EIO, &page->mapping->flags);
3016                 if (buffer_busy(bh))
3017                         goto failed;
3018                 bh = bh->b_this_page;
3019         } while (bh != head);
3020
3021         do {
3022                 struct buffer_head *next = bh->b_this_page;
3023
3024                 if (bh->b_assoc_map)
3025                         __remove_assoc_queue(bh);
3026                 bh = next;
3027         } while (bh != head);
3028         *buffers_to_free = head;
3029         __clear_page_buffers(page);
3030         return 1;
3031 failed:
3032         return 0;
3033 }
3034
3035 int try_to_free_buffers(struct page *page)
3036 {
3037         struct address_space * const mapping = page->mapping;
3038         struct buffer_head *buffers_to_free = NULL;
3039         int ret = 0;
3040
3041         BUG_ON(!PageLocked(page));
3042         if (PageWriteback(page))
3043                 return 0;
3044
3045         if (mapping == NULL) {          /* can this still happen? */
3046                 ret = drop_buffers(page, &buffers_to_free);
3047                 goto out;
3048         }
3049
3050         spin_lock(&mapping->private_lock);
3051         ret = drop_buffers(page, &buffers_to_free);
3052
3053         /*
3054          * If the filesystem writes its buffers by hand (eg ext3)
3055          * then we can have clean buffers against a dirty page.  We
3056          * clean the page here; otherwise the VM will never notice
3057          * that the filesystem did any IO at all.
3058          *
3059          * Also, during truncate, discard_buffer will have marked all
3060          * the page's buffers clean.  We discover that here and clean
3061          * the page also.
3062          *
3063          * private_lock must be held over this entire operation in order
3064          * to synchronise against __set_page_dirty_buffers and prevent the
3065          * dirty bit from being lost.
3066          */
3067         if (ret)
3068                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3069         spin_unlock(&mapping->private_lock);
3070 out:
3071         if (buffers_to_free) {
3072                 struct buffer_head *bh = buffers_to_free;
3073
3074                 do {
3075                         struct buffer_head *next = bh->b_this_page;
3076                         free_buffer_head(bh);
3077                         bh = next;
3078                 } while (bh != buffers_to_free);
3079         }
3080         return ret;
3081 }
3082 EXPORT_SYMBOL(try_to_free_buffers);
3083
3084 void block_sync_page(struct page *page)
3085 {
3086         struct address_space *mapping;
3087
3088         smp_mb();
3089         mapping = page_mapping(page);
3090         if (mapping)
3091                 blk_run_backing_dev(mapping->backing_dev_info, page);
3092 }
3093
3094 /*
3095  * There are no bdflush tunables left.  But distributions are
3096  * still running obsolete flush daemons, so we terminate them here.
3097  *
3098  * Use of bdflush() is deprecated and will be removed in a future kernel.
3099  * The `pdflush' kernel threads fully replace bdflush daemons and this call.
3100  */
3101 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3102 {
3103         static int msg_count;
3104
3105         if (!capable(CAP_SYS_ADMIN))
3106                 return -EPERM;
3107
3108         if (msg_count < 5) {
3109                 msg_count++;
3110                 printk(KERN_INFO
3111                         "warning: process `%s' used the obsolete bdflush"
3112                         " system call\n", current->comm);
3113                 printk(KERN_INFO "Fix your initscripts?\n");
3114         }
3115
3116         if (func == 1)
3117                 do_exit(0);
3118         return 0;
3119 }
3120
3121 /*
3122  * Buffer-head allocation
3123  */
3124 static struct kmem_cache *bh_cachep;
3125
3126 /*
3127  * Once the number of bh's in the machine exceeds this level, we start
3128  * stripping them in writeback.
3129  */
3130 static int max_buffer_heads;
3131
3132 int buffer_heads_over_limit;
3133
3134 struct bh_accounting {
3135         int nr;                 /* Number of live bh's */
3136         int ratelimit;          /* Limit cacheline bouncing */
3137 };
3138
3139 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3140
3141 static void recalc_bh_state(void)
3142 {
3143         int i;
3144         int tot = 0;
3145
3146         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3147                 return;
3148         __get_cpu_var(bh_accounting).ratelimit = 0;
3149         for_each_online_cpu(i)
3150                 tot += per_cpu(bh_accounting, i).nr;
3151         buffer_heads_over_limit = (tot > max_buffer_heads);
3152 }
3153         
3154 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3155 {
3156         struct buffer_head *ret = kmem_cache_alloc(bh_cachep, gfp_flags);
3157         if (ret) {
3158                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3159                 get_cpu_var(bh_accounting).nr++;
3160                 recalc_bh_state();
3161                 put_cpu_var(bh_accounting);
3162         }
3163         return ret;
3164 }
3165 EXPORT_SYMBOL(alloc_buffer_head);
3166
3167 void free_buffer_head(struct buffer_head *bh)
3168 {
3169         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3170         kmem_cache_free(bh_cachep, bh);
3171         get_cpu_var(bh_accounting).nr--;
3172         recalc_bh_state();
3173         put_cpu_var(bh_accounting);
3174 }
3175 EXPORT_SYMBOL(free_buffer_head);
3176
3177 static void buffer_exit_cpu(int cpu)
3178 {
3179         int i;
3180         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3181
3182         for (i = 0; i < BH_LRU_SIZE; i++) {
3183                 brelse(b->bhs[i]);
3184                 b->bhs[i] = NULL;
3185         }
3186         get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3187         per_cpu(bh_accounting, cpu).nr = 0;
3188         put_cpu_var(bh_accounting);
3189 }
3190
3191 static int buffer_cpu_notify(struct notifier_block *self,
3192                               unsigned long action, void *hcpu)
3193 {
3194         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3195                 buffer_exit_cpu((unsigned long)hcpu);
3196         return NOTIFY_OK;
3197 }
3198
3199 /**
3200  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3201  * @bh: struct buffer_head
3202  *
3203  * Return true if the buffer is up-to-date and false,
3204  * with the buffer locked, if not.
3205  */
3206 int bh_uptodate_or_lock(struct buffer_head *bh)
3207 {
3208         if (!buffer_uptodate(bh)) {
3209                 lock_buffer(bh);
3210                 if (!buffer_uptodate(bh))
3211                         return 0;
3212                 unlock_buffer(bh);
3213         }
3214         return 1;
3215 }
3216 EXPORT_SYMBOL(bh_uptodate_or_lock);
3217
3218 /**
3219  * bh_submit_read - Submit a locked buffer for reading
3220  * @bh: struct buffer_head
3221  *
3222  * Returns zero on success and -EIO on error.
3223  */
3224 int bh_submit_read(struct buffer_head *bh)
3225 {
3226         BUG_ON(!buffer_locked(bh));
3227
3228         if (buffer_uptodate(bh)) {
3229                 unlock_buffer(bh);
3230                 return 0;
3231         }
3232
3233         get_bh(bh);
3234         bh->b_end_io = end_buffer_read_sync;
3235         submit_bh(READ, bh);
3236         wait_on_buffer(bh);
3237         if (buffer_uptodate(bh))
3238                 return 0;
3239         return -EIO;
3240 }
3241 EXPORT_SYMBOL(bh_submit_read);
3242
3243 static void
3244 init_buffer_head(void *data)
3245 {
3246         struct buffer_head *bh = data;
3247
3248         memset(bh, 0, sizeof(*bh));
3249         INIT_LIST_HEAD(&bh->b_assoc_buffers);
3250 }
3251
3252 void __init buffer_init(void)
3253 {
3254         int nrpages;
3255
3256         bh_cachep = kmem_cache_create("buffer_head",
3257                         sizeof(struct buffer_head), 0,
3258                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3259                                 SLAB_MEM_SPREAD),
3260                                 init_buffer_head);
3261
3262         /*
3263          * Limit the bh occupancy to 10% of ZONE_NORMAL
3264          */
3265         nrpages = (nr_free_buffer_pages() * 10) / 100;
3266         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3267         hotcpu_notifier(buffer_cpu_notify, 0);
3268 }
3269
3270 EXPORT_SYMBOL(__bforget);
3271 EXPORT_SYMBOL(__brelse);
3272 EXPORT_SYMBOL(__wait_on_buffer);
3273 EXPORT_SYMBOL(block_commit_write);
3274 EXPORT_SYMBOL(block_prepare_write);
3275 EXPORT_SYMBOL(block_page_mkwrite);
3276 EXPORT_SYMBOL(block_read_full_page);
3277 EXPORT_SYMBOL(block_sync_page);
3278 EXPORT_SYMBOL(block_truncate_page);
3279 EXPORT_SYMBOL(block_write_full_page);
3280 EXPORT_SYMBOL(cont_write_begin);
3281 EXPORT_SYMBOL(end_buffer_read_sync);
3282 EXPORT_SYMBOL(end_buffer_write_sync);
3283 EXPORT_SYMBOL(file_fsync);
3284 EXPORT_SYMBOL(generic_block_bmap);
3285 EXPORT_SYMBOL(generic_cont_expand_simple);
3286 EXPORT_SYMBOL(init_buffer);
3287 EXPORT_SYMBOL(invalidate_bdev);
3288 EXPORT_SYMBOL(ll_rw_block);
3289 EXPORT_SYMBOL(mark_buffer_dirty);
3290 EXPORT_SYMBOL(submit_bh);
3291 EXPORT_SYMBOL(sync_dirty_buffer);
3292 EXPORT_SYMBOL(unlock_buffer);