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