hfsplus: creation of hidden dir on mount can fail
[linux-2.6.git] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/capability.h>
28 #include <linux/blkdev.h>
29 #include <linux/file.h>
30 #include <linux/quotaops.h>
31 #include <linux/highmem.h>
32 #include <linux/module.h>
33 #include <linux/writeback.h>
34 #include <linux/hash.h>
35 #include <linux/suspend.h>
36 #include <linux/buffer_head.h>
37 #include <linux/task_io_accounting_ops.h>
38 #include <linux/bio.h>
39 #include <linux/notifier.h>
40 #include <linux/cpu.h>
41 #include <linux/bitops.h>
42 #include <linux/mpage.h>
43 #include <linux/bit_spinlock.h>
44
45 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
46
47 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
48
49 inline void
50 init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
51 {
52         bh->b_end_io = handler;
53         bh->b_private = private;
54 }
55 EXPORT_SYMBOL(init_buffer);
56
57 static int sleep_on_buffer(void *word)
58 {
59         io_schedule();
60         return 0;
61 }
62
63 void __lock_buffer(struct buffer_head *bh)
64 {
65         wait_on_bit_lock(&bh->b_state, BH_Lock, sleep_on_buffer,
66                                                         TASK_UNINTERRUPTIBLE);
67 }
68 EXPORT_SYMBOL(__lock_buffer);
69
70 void unlock_buffer(struct buffer_head *bh)
71 {
72         clear_bit_unlock(BH_Lock, &bh->b_state);
73         smp_mb__after_clear_bit();
74         wake_up_bit(&bh->b_state, BH_Lock);
75 }
76 EXPORT_SYMBOL(unlock_buffer);
77
78 /*
79  * Block until a buffer comes unlocked.  This doesn't stop it
80  * from becoming locked again - you have to lock it yourself
81  * if you want to preserve its state.
82  */
83 void __wait_on_buffer(struct buffer_head * bh)
84 {
85         wait_on_bit(&bh->b_state, BH_Lock, sleep_on_buffer, TASK_UNINTERRUPTIBLE);
86 }
87 EXPORT_SYMBOL(__wait_on_buffer);
88
89 static void
90 __clear_page_buffers(struct page *page)
91 {
92         ClearPagePrivate(page);
93         set_page_private(page, 0);
94         page_cache_release(page);
95 }
96
97
98 static int quiet_error(struct buffer_head *bh)
99 {
100         if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
101                 return 0;
102         return 1;
103 }
104
105
106 static void buffer_io_error(struct buffer_head *bh)
107 {
108         char b[BDEVNAME_SIZE];
109         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
110                         bdevname(bh->b_bdev, b),
111                         (unsigned long long)bh->b_blocknr);
112 }
113
114 /*
115  * End-of-IO handler helper function which does not touch the bh after
116  * unlocking it.
117  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
118  * a race there is benign: unlock_buffer() only use the bh's address for
119  * hashing after unlocking the buffer, so it doesn't actually touch the bh
120  * itself.
121  */
122 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
123 {
124         if (uptodate) {
125                 set_buffer_uptodate(bh);
126         } else {
127                 /* This happens, due to failed READA attempts. */
128                 clear_buffer_uptodate(bh);
129         }
130         unlock_buffer(bh);
131 }
132
133 /*
134  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
135  * unlock the buffer. This is what ll_rw_block uses too.
136  */
137 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
138 {
139         __end_buffer_read_notouch(bh, uptodate);
140         put_bh(bh);
141 }
142 EXPORT_SYMBOL(end_buffer_read_sync);
143
144 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
145 {
146         char b[BDEVNAME_SIZE];
147
148         if (uptodate) {
149                 set_buffer_uptodate(bh);
150         } else {
151                 if (!quiet_error(bh)) {
152                         buffer_io_error(bh);
153                         printk(KERN_WARNING "lost page write due to "
154                                         "I/O error on %s\n",
155                                        bdevname(bh->b_bdev, b));
156                 }
157                 set_buffer_write_io_error(bh);
158                 clear_buffer_uptodate(bh);
159         }
160         unlock_buffer(bh);
161         put_bh(bh);
162 }
163 EXPORT_SYMBOL(end_buffer_write_sync);
164
165 /*
166  * Various filesystems appear to want __find_get_block to be non-blocking.
167  * But it's the page lock which protects the buffers.  To get around this,
168  * we get exclusion from try_to_free_buffers with the blockdev mapping's
169  * private_lock.
170  *
171  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
172  * may be quite high.  This code could TryLock the page, and if that
173  * succeeds, there is no need to take private_lock. (But if
174  * private_lock is contended then so is mapping->tree_lock).
175  */
176 static struct buffer_head *
177 __find_get_block_slow(struct block_device *bdev, sector_t block)
178 {
179         struct inode *bd_inode = bdev->bd_inode;
180         struct address_space *bd_mapping = bd_inode->i_mapping;
181         struct buffer_head *ret = NULL;
182         pgoff_t index;
183         struct buffer_head *bh;
184         struct buffer_head *head;
185         struct page *page;
186         int all_mapped = 1;
187
188         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
189         page = find_get_page(bd_mapping, index);
190         if (!page)
191                 goto out;
192
193         spin_lock(&bd_mapping->private_lock);
194         if (!page_has_buffers(page))
195                 goto out_unlock;
196         head = page_buffers(page);
197         bh = head;
198         do {
199                 if (!buffer_mapped(bh))
200                         all_mapped = 0;
201                 else if (bh->b_blocknr == block) {
202                         ret = bh;
203                         get_bh(bh);
204                         goto out_unlock;
205                 }
206                 bh = bh->b_this_page;
207         } while (bh != head);
208
209         /* we might be here because some of the buffers on this page are
210          * not mapped.  This is due to various races between
211          * file io on the block device and getblk.  It gets dealt with
212          * elsewhere, don't buffer_error if we had some unmapped buffers
213          */
214         if (all_mapped) {
215                 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 void
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
925         do {
926                 if (!buffer_mapped(bh)) {
927                         init_buffer(bh, NULL, NULL);
928                         bh->b_bdev = bdev;
929                         bh->b_blocknr = block;
930                         if (uptodate)
931                                 set_buffer_uptodate(bh);
932                         set_buffer_mapped(bh);
933                 }
934                 block++;
935                 bh = bh->b_this_page;
936         } while (bh != head);
937 }
938
939 /*
940  * Create the page-cache page that contains the requested block.
941  *
942  * This is user purely for blockdev mappings.
943  */
944 static struct page *
945 grow_dev_page(struct block_device *bdev, sector_t block,
946                 pgoff_t index, int size)
947 {
948         struct inode *inode = bdev->bd_inode;
949         struct page *page;
950         struct buffer_head *bh;
951
952         page = find_or_create_page(inode->i_mapping, index,
953                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
954         if (!page)
955                 return NULL;
956
957         BUG_ON(!PageLocked(page));
958
959         if (page_has_buffers(page)) {
960                 bh = page_buffers(page);
961                 if (bh->b_size == size) {
962                         init_page_buffers(page, bdev, block, size);
963                         return page;
964                 }
965                 if (!try_to_free_buffers(page))
966                         goto failed;
967         }
968
969         /*
970          * Allocate some buffers for this page
971          */
972         bh = alloc_page_buffers(page, size, 0);
973         if (!bh)
974                 goto failed;
975
976         /*
977          * Link the page to the buffers and initialise them.  Take the
978          * lock to be atomic wrt __find_get_block(), which does not
979          * run under the page lock.
980          */
981         spin_lock(&inode->i_mapping->private_lock);
982         link_dev_buffers(page, bh);
983         init_page_buffers(page, bdev, block, size);
984         spin_unlock(&inode->i_mapping->private_lock);
985         return page;
986
987 failed:
988         BUG();
989         unlock_page(page);
990         page_cache_release(page);
991         return NULL;
992 }
993
994 /*
995  * Create buffers for the specified block device block's page.  If
996  * that page was dirty, the buffers are set dirty also.
997  */
998 static int
999 grow_buffers(struct block_device *bdev, sector_t block, int size)
1000 {
1001         struct page *page;
1002         pgoff_t index;
1003         int sizebits;
1004
1005         sizebits = -1;
1006         do {
1007                 sizebits++;
1008         } while ((size << sizebits) < PAGE_SIZE);
1009
1010         index = block >> sizebits;
1011
1012         /*
1013          * Check for a block which wants to lie outside our maximum possible
1014          * pagecache index.  (this comparison is done using sector_t types).
1015          */
1016         if (unlikely(index != block >> sizebits)) {
1017                 char b[BDEVNAME_SIZE];
1018
1019                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1020                         "device %s\n",
1021                         __func__, (unsigned long long)block,
1022                         bdevname(bdev, b));
1023                 return -EIO;
1024         }
1025         block = index << sizebits;
1026         /* Create a page with the proper size buffers.. */
1027         page = grow_dev_page(bdev, block, index, size);
1028         if (!page)
1029                 return 0;
1030         unlock_page(page);
1031         page_cache_release(page);
1032         return 1;
1033 }
1034
1035 static struct buffer_head *
1036 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1037 {
1038         /* Size must be multiple of hard sectorsize */
1039         if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1040                         (size < 512 || size > PAGE_SIZE))) {
1041                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1042                                         size);
1043                 printk(KERN_ERR "logical block size: %d\n",
1044                                         bdev_logical_block_size(bdev));
1045
1046                 dump_stack();
1047                 return NULL;
1048         }
1049
1050         for (;;) {
1051                 struct buffer_head * bh;
1052                 int ret;
1053
1054                 bh = __find_get_block(bdev, block, size);
1055                 if (bh)
1056                         return bh;
1057
1058                 ret = grow_buffers(bdev, block, size);
1059                 if (ret < 0)
1060                         return NULL;
1061                 if (ret == 0)
1062                         free_more_memory();
1063         }
1064 }
1065
1066 /*
1067  * The relationship between dirty buffers and dirty pages:
1068  *
1069  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1070  * the page is tagged dirty in its radix tree.
1071  *
1072  * At all times, the dirtiness of the buffers represents the dirtiness of
1073  * subsections of the page.  If the page has buffers, the page dirty bit is
1074  * merely a hint about the true dirty state.
1075  *
1076  * When a page is set dirty in its entirety, all its buffers are marked dirty
1077  * (if the page has buffers).
1078  *
1079  * When a buffer is marked dirty, its page is dirtied, but the page's other
1080  * buffers are not.
1081  *
1082  * Also.  When blockdev buffers are explicitly read with bread(), they
1083  * individually become uptodate.  But their backing page remains not
1084  * uptodate - even if all of its buffers are uptodate.  A subsequent
1085  * block_read_full_page() against that page will discover all the uptodate
1086  * buffers, will set the page uptodate and will perform no I/O.
1087  */
1088
1089 /**
1090  * mark_buffer_dirty - mark a buffer_head as needing writeout
1091  * @bh: the buffer_head to mark dirty
1092  *
1093  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1094  * backing page dirty, then tag the page as dirty in its address_space's radix
1095  * tree and then attach the address_space's inode to its superblock's dirty
1096  * inode list.
1097  *
1098  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1099  * mapping->tree_lock and mapping->host->i_lock.
1100  */
1101 void mark_buffer_dirty(struct buffer_head *bh)
1102 {
1103         WARN_ON_ONCE(!buffer_uptodate(bh));
1104
1105         /*
1106          * Very *carefully* optimize the it-is-already-dirty case.
1107          *
1108          * Don't let the final "is it dirty" escape to before we
1109          * perhaps modified the buffer.
1110          */
1111         if (buffer_dirty(bh)) {
1112                 smp_mb();
1113                 if (buffer_dirty(bh))
1114                         return;
1115         }
1116
1117         if (!test_set_buffer_dirty(bh)) {
1118                 struct page *page = bh->b_page;
1119                 if (!TestSetPageDirty(page)) {
1120                         struct address_space *mapping = page_mapping(page);
1121                         if (mapping)
1122                                 __set_page_dirty(page, mapping, 0);
1123                 }
1124         }
1125 }
1126 EXPORT_SYMBOL(mark_buffer_dirty);
1127
1128 /*
1129  * Decrement a buffer_head's reference count.  If all buffers against a page
1130  * have zero reference count, are clean and unlocked, and if the page is clean
1131  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1132  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1133  * a page but it ends up not being freed, and buffers may later be reattached).
1134  */
1135 void __brelse(struct buffer_head * buf)
1136 {
1137         if (atomic_read(&buf->b_count)) {
1138                 put_bh(buf);
1139                 return;
1140         }
1141         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1142 }
1143 EXPORT_SYMBOL(__brelse);
1144
1145 /*
1146  * bforget() is like brelse(), except it discards any
1147  * potentially dirty data.
1148  */
1149 void __bforget(struct buffer_head *bh)
1150 {
1151         clear_buffer_dirty(bh);
1152         if (bh->b_assoc_map) {
1153                 struct address_space *buffer_mapping = bh->b_page->mapping;
1154
1155                 spin_lock(&buffer_mapping->private_lock);
1156                 list_del_init(&bh->b_assoc_buffers);
1157                 bh->b_assoc_map = NULL;
1158                 spin_unlock(&buffer_mapping->private_lock);
1159         }
1160         __brelse(bh);
1161 }
1162 EXPORT_SYMBOL(__bforget);
1163
1164 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1165 {
1166         lock_buffer(bh);
1167         if (buffer_uptodate(bh)) {
1168                 unlock_buffer(bh);
1169                 return bh;
1170         } else {
1171                 get_bh(bh);
1172                 bh->b_end_io = end_buffer_read_sync;
1173                 submit_bh(READ, bh);
1174                 wait_on_buffer(bh);
1175                 if (buffer_uptodate(bh))
1176                         return bh;
1177         }
1178         brelse(bh);
1179         return NULL;
1180 }
1181
1182 /*
1183  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1184  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1185  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1186  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1187  * CPU's LRUs at the same time.
1188  *
1189  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1190  * sb_find_get_block().
1191  *
1192  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1193  * a local interrupt disable for that.
1194  */
1195
1196 #define BH_LRU_SIZE     8
1197
1198 struct bh_lru {
1199         struct buffer_head *bhs[BH_LRU_SIZE];
1200 };
1201
1202 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1203
1204 #ifdef CONFIG_SMP
1205 #define bh_lru_lock()   local_irq_disable()
1206 #define bh_lru_unlock() local_irq_enable()
1207 #else
1208 #define bh_lru_lock()   preempt_disable()
1209 #define bh_lru_unlock() preempt_enable()
1210 #endif
1211
1212 static inline void check_irqs_on(void)
1213 {
1214 #ifdef irqs_disabled
1215         BUG_ON(irqs_disabled());
1216 #endif
1217 }
1218
1219 /*
1220  * The LRU management algorithm is dopey-but-simple.  Sorry.
1221  */
1222 static void bh_lru_install(struct buffer_head *bh)
1223 {
1224         struct buffer_head *evictee = NULL;
1225
1226         check_irqs_on();
1227         bh_lru_lock();
1228         if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1229                 struct buffer_head *bhs[BH_LRU_SIZE];
1230                 int in;
1231                 int out = 0;
1232
1233                 get_bh(bh);
1234                 bhs[out++] = bh;
1235                 for (in = 0; in < BH_LRU_SIZE; in++) {
1236                         struct buffer_head *bh2 =
1237                                 __this_cpu_read(bh_lrus.bhs[in]);
1238
1239                         if (bh2 == bh) {
1240                                 __brelse(bh2);
1241                         } else {
1242                                 if (out >= BH_LRU_SIZE) {
1243                                         BUG_ON(evictee != NULL);
1244                                         evictee = bh2;
1245                                 } else {
1246                                         bhs[out++] = bh2;
1247                                 }
1248                         }
1249                 }
1250                 while (out < BH_LRU_SIZE)
1251                         bhs[out++] = NULL;
1252                 memcpy(__this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1253         }
1254         bh_lru_unlock();
1255
1256         if (evictee)
1257                 __brelse(evictee);
1258 }
1259
1260 /*
1261  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1262  */
1263 static struct buffer_head *
1264 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1265 {
1266         struct buffer_head *ret = NULL;
1267         unsigned int i;
1268
1269         check_irqs_on();
1270         bh_lru_lock();
1271         for (i = 0; i < BH_LRU_SIZE; i++) {
1272                 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1273
1274                 if (bh && bh->b_bdev == bdev &&
1275                                 bh->b_blocknr == block && bh->b_size == size) {
1276                         if (i) {
1277                                 while (i) {
1278                                         __this_cpu_write(bh_lrus.bhs[i],
1279                                                 __this_cpu_read(bh_lrus.bhs[i - 1]));
1280                                         i--;
1281                                 }
1282                                 __this_cpu_write(bh_lrus.bhs[0], bh);
1283                         }
1284                         get_bh(bh);
1285                         ret = bh;
1286                         break;
1287                 }
1288         }
1289         bh_lru_unlock();
1290         return ret;
1291 }
1292
1293 /*
1294  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1295  * it in the LRU and mark it as accessed.  If it is not present then return
1296  * NULL
1297  */
1298 struct buffer_head *
1299 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1300 {
1301         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1302
1303         if (bh == NULL) {
1304                 bh = __find_get_block_slow(bdev, block);
1305                 if (bh)
1306                         bh_lru_install(bh);
1307         }
1308         if (bh)
1309                 touch_buffer(bh);
1310         return bh;
1311 }
1312 EXPORT_SYMBOL(__find_get_block);
1313
1314 /*
1315  * __getblk will locate (and, if necessary, create) the buffer_head
1316  * which corresponds to the passed block_device, block and size. The
1317  * returned buffer has its reference count incremented.
1318  *
1319  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1320  * illegal block number, __getblk() will happily return a buffer_head
1321  * which represents the non-existent block.  Very weird.
1322  *
1323  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1324  * attempt is failing.  FIXME, perhaps?
1325  */
1326 struct buffer_head *
1327 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1328 {
1329         struct buffer_head *bh = __find_get_block(bdev, block, size);
1330
1331         might_sleep();
1332         if (bh == NULL)
1333                 bh = __getblk_slow(bdev, block, size);
1334         return bh;
1335 }
1336 EXPORT_SYMBOL(__getblk);
1337
1338 /*
1339  * Do async read-ahead on a buffer..
1340  */
1341 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1342 {
1343         struct buffer_head *bh = __getblk(bdev, block, size);
1344         if (likely(bh)) {
1345                 ll_rw_block(READA, 1, &bh);
1346                 brelse(bh);
1347         }
1348 }
1349 EXPORT_SYMBOL(__breadahead);
1350
1351 /**
1352  *  __bread() - reads a specified block and returns the bh
1353  *  @bdev: the block_device to read from
1354  *  @block: number of block
1355  *  @size: size (in bytes) to read
1356  * 
1357  *  Reads a specified block, and returns buffer head that contains it.
1358  *  It returns NULL if the block was unreadable.
1359  */
1360 struct buffer_head *
1361 __bread(struct block_device *bdev, sector_t block, unsigned size)
1362 {
1363         struct buffer_head *bh = __getblk(bdev, block, size);
1364
1365         if (likely(bh) && !buffer_uptodate(bh))
1366                 bh = __bread_slow(bh);
1367         return bh;
1368 }
1369 EXPORT_SYMBOL(__bread);
1370
1371 /*
1372  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1373  * This doesn't race because it runs in each cpu either in irq
1374  * or with preempt disabled.
1375  */
1376 static void invalidate_bh_lru(void *arg)
1377 {
1378         struct bh_lru *b = &get_cpu_var(bh_lrus);
1379         int i;
1380
1381         for (i = 0; i < BH_LRU_SIZE; i++) {
1382                 brelse(b->bhs[i]);
1383                 b->bhs[i] = NULL;
1384         }
1385         put_cpu_var(bh_lrus);
1386 }
1387         
1388 void invalidate_bh_lrus(void)
1389 {
1390         on_each_cpu(invalidate_bh_lru, NULL, 1);
1391 }
1392 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1393
1394 void set_bh_page(struct buffer_head *bh,
1395                 struct page *page, unsigned long offset)
1396 {
1397         bh->b_page = page;
1398         BUG_ON(offset >= PAGE_SIZE);
1399         if (PageHighMem(page))
1400                 /*
1401                  * This catches illegal uses and preserves the offset:
1402                  */
1403                 bh->b_data = (char *)(0 + offset);
1404         else
1405                 bh->b_data = page_address(page) + offset;
1406 }
1407 EXPORT_SYMBOL(set_bh_page);
1408
1409 /*
1410  * Called when truncating a buffer on a page completely.
1411  */
1412 static void discard_buffer(struct buffer_head * bh)
1413 {
1414         lock_buffer(bh);
1415         clear_buffer_dirty(bh);
1416         bh->b_bdev = NULL;
1417         clear_buffer_mapped(bh);
1418         clear_buffer_req(bh);
1419         clear_buffer_new(bh);
1420         clear_buffer_delay(bh);
1421         clear_buffer_unwritten(bh);
1422         unlock_buffer(bh);
1423 }
1424
1425 /**
1426  * block_invalidatepage - invalidate part or all of a buffer-backed page
1427  *
1428  * @page: the page which is affected
1429  * @offset: the index of the truncation point
1430  *
1431  * block_invalidatepage() is called when all or part of the page has become
1432  * invalidated by a truncate operation.
1433  *
1434  * block_invalidatepage() does not have to release all buffers, but it must
1435  * ensure that no dirty buffer is left outside @offset and that no I/O
1436  * is underway against any of the blocks which are outside the truncation
1437  * point.  Because the caller is about to free (and possibly reuse) those
1438  * blocks on-disk.
1439  */
1440 void block_invalidatepage(struct page *page, unsigned long offset)
1441 {
1442         struct buffer_head *head, *bh, *next;
1443         unsigned int curr_off = 0;
1444
1445         BUG_ON(!PageLocked(page));
1446         if (!page_has_buffers(page))
1447                 goto out;
1448
1449         head = page_buffers(page);
1450         bh = head;
1451         do {
1452                 unsigned int next_off = curr_off + bh->b_size;
1453                 next = bh->b_this_page;
1454
1455                 /*
1456                  * is this block fully invalidated?
1457                  */
1458                 if (offset <= curr_off)
1459                         discard_buffer(bh);
1460                 curr_off = next_off;
1461                 bh = next;
1462         } while (bh != head);
1463
1464         /*
1465          * We release buffers only if the entire page is being invalidated.
1466          * The get_block cached value has been unconditionally invalidated,
1467          * so real IO is not possible anymore.
1468          */
1469         if (offset == 0)
1470                 try_to_release_page(page, 0);
1471 out:
1472         return;
1473 }
1474 EXPORT_SYMBOL(block_invalidatepage);
1475
1476 /*
1477  * We attach and possibly dirty the buffers atomically wrt
1478  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1479  * is already excluded via the page lock.
1480  */
1481 void create_empty_buffers(struct page *page,
1482                         unsigned long blocksize, unsigned long b_state)
1483 {
1484         struct buffer_head *bh, *head, *tail;
1485
1486         head = alloc_page_buffers(page, blocksize, 1);
1487         bh = head;
1488         do {
1489                 bh->b_state |= b_state;
1490                 tail = bh;
1491                 bh = bh->b_this_page;
1492         } while (bh);
1493         tail->b_this_page = head;
1494
1495         spin_lock(&page->mapping->private_lock);
1496         if (PageUptodate(page) || PageDirty(page)) {
1497                 bh = head;
1498                 do {
1499                         if (PageDirty(page))
1500                                 set_buffer_dirty(bh);
1501                         if (PageUptodate(page))
1502                                 set_buffer_uptodate(bh);
1503                         bh = bh->b_this_page;
1504                 } while (bh != head);
1505         }
1506         attach_page_buffers(page, head);
1507         spin_unlock(&page->mapping->private_lock);
1508 }
1509 EXPORT_SYMBOL(create_empty_buffers);
1510
1511 /*
1512  * We are taking a block for data and we don't want any output from any
1513  * buffer-cache aliases starting from return from that function and
1514  * until the moment when something will explicitly mark the buffer
1515  * dirty (hopefully that will not happen until we will free that block ;-)
1516  * We don't even need to mark it not-uptodate - nobody can expect
1517  * anything from a newly allocated buffer anyway. We used to used
1518  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1519  * don't want to mark the alias unmapped, for example - it would confuse
1520  * anyone who might pick it with bread() afterwards...
1521  *
1522  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1523  * be writeout I/O going on against recently-freed buffers.  We don't
1524  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1525  * only if we really need to.  That happens here.
1526  */
1527 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1528 {
1529         struct buffer_head *old_bh;
1530
1531         might_sleep();
1532
1533         old_bh = __find_get_block_slow(bdev, block);
1534         if (old_bh) {
1535                 clear_buffer_dirty(old_bh);
1536                 wait_on_buffer(old_bh);
1537                 clear_buffer_req(old_bh);
1538                 __brelse(old_bh);
1539         }
1540 }
1541 EXPORT_SYMBOL(unmap_underlying_metadata);
1542
1543 /*
1544  * NOTE! All mapped/uptodate combinations are valid:
1545  *
1546  *      Mapped  Uptodate        Meaning
1547  *
1548  *      No      No              "unknown" - must do get_block()
1549  *      No      Yes             "hole" - zero-filled
1550  *      Yes     No              "allocated" - allocated on disk, not read in
1551  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1552  *
1553  * "Dirty" is valid only with the last case (mapped+uptodate).
1554  */
1555
1556 /*
1557  * While block_write_full_page is writing back the dirty buffers under
1558  * the page lock, whoever dirtied the buffers may decide to clean them
1559  * again at any time.  We handle that by only looking at the buffer
1560  * state inside lock_buffer().
1561  *
1562  * If block_write_full_page() is called for regular writeback
1563  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1564  * locked buffer.   This only can happen if someone has written the buffer
1565  * directly, with submit_bh().  At the address_space level PageWriteback
1566  * prevents this contention from occurring.
1567  *
1568  * If block_write_full_page() is called with wbc->sync_mode ==
1569  * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1570  * causes the writes to be flagged as synchronous writes.
1571  */
1572 static int __block_write_full_page(struct inode *inode, struct page *page,
1573                         get_block_t *get_block, struct writeback_control *wbc,
1574                         bh_end_io_t *handler)
1575 {
1576         int err;
1577         sector_t block;
1578         sector_t last_block;
1579         struct buffer_head *bh, *head;
1580         const unsigned blocksize = 1 << inode->i_blkbits;
1581         int nr_underway = 0;
1582         int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1583                         WRITE_SYNC : WRITE);
1584
1585         BUG_ON(!PageLocked(page));
1586
1587         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1588
1589         if (!page_has_buffers(page)) {
1590                 create_empty_buffers(page, blocksize,
1591                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1592         }
1593
1594         /*
1595          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1596          * here, and the (potentially unmapped) buffers may become dirty at
1597          * any time.  If a buffer becomes dirty here after we've inspected it
1598          * then we just miss that fact, and the page stays dirty.
1599          *
1600          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1601          * handle that here by just cleaning them.
1602          */
1603
1604         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1605         head = page_buffers(page);
1606         bh = head;
1607
1608         /*
1609          * Get all the dirty buffers mapped to disk addresses and
1610          * handle any aliases from the underlying blockdev's mapping.
1611          */
1612         do {
1613                 if (block > last_block) {
1614                         /*
1615                          * mapped buffers outside i_size will occur, because
1616                          * this page can be outside i_size when there is a
1617                          * truncate in progress.
1618                          */
1619                         /*
1620                          * The buffer was zeroed by block_write_full_page()
1621                          */
1622                         clear_buffer_dirty(bh);
1623                         set_buffer_uptodate(bh);
1624                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1625                            buffer_dirty(bh)) {
1626                         WARN_ON(bh->b_size != blocksize);
1627                         err = get_block(inode, block, bh, 1);
1628                         if (err)
1629                                 goto recover;
1630                         clear_buffer_delay(bh);
1631                         if (buffer_new(bh)) {
1632                                 /* blockdev mappings never come here */
1633                                 clear_buffer_new(bh);
1634                                 unmap_underlying_metadata(bh->b_bdev,
1635                                                         bh->b_blocknr);
1636                         }
1637                 }
1638                 bh = bh->b_this_page;
1639                 block++;
1640         } while (bh != head);
1641
1642         do {
1643                 if (!buffer_mapped(bh))
1644                         continue;
1645                 /*
1646                  * If it's a fully non-blocking write attempt and we cannot
1647                  * lock the buffer then redirty the page.  Note that this can
1648                  * potentially cause a busy-wait loop from writeback threads
1649                  * and kswapd activity, but those code paths have their own
1650                  * higher-level throttling.
1651                  */
1652                 if (wbc->sync_mode != WB_SYNC_NONE) {
1653                         lock_buffer(bh);
1654                 } else if (!trylock_buffer(bh)) {
1655                         redirty_page_for_writepage(wbc, page);
1656                         continue;
1657                 }
1658                 if (test_clear_buffer_dirty(bh)) {
1659                         mark_buffer_async_write_endio(bh, handler);
1660                 } else {
1661                         unlock_buffer(bh);
1662                 }
1663         } while ((bh = bh->b_this_page) != head);
1664
1665         /*
1666          * The page and its buffers are protected by PageWriteback(), so we can
1667          * drop the bh refcounts early.
1668          */
1669         BUG_ON(PageWriteback(page));
1670         set_page_writeback(page);
1671
1672         do {
1673                 struct buffer_head *next = bh->b_this_page;
1674                 if (buffer_async_write(bh)) {
1675                         submit_bh(write_op, bh);
1676                         nr_underway++;
1677                 }
1678                 bh = next;
1679         } while (bh != head);
1680         unlock_page(page);
1681
1682         err = 0;
1683 done:
1684         if (nr_underway == 0) {
1685                 /*
1686                  * The page was marked dirty, but the buffers were
1687                  * clean.  Someone wrote them back by hand with
1688                  * ll_rw_block/submit_bh.  A rare case.
1689                  */
1690                 end_page_writeback(page);
1691
1692                 /*
1693                  * The page and buffer_heads can be released at any time from
1694                  * here on.
1695                  */
1696         }
1697         return err;
1698
1699 recover:
1700         /*
1701          * ENOSPC, or some other error.  We may already have added some
1702          * blocks to the file, so we need to write these out to avoid
1703          * exposing stale data.
1704          * The page is currently locked and not marked for writeback
1705          */
1706         bh = head;
1707         /* Recovery: lock and submit the mapped buffers */
1708         do {
1709                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1710                     !buffer_delay(bh)) {
1711                         lock_buffer(bh);
1712                         mark_buffer_async_write_endio(bh, handler);
1713                 } else {
1714                         /*
1715                          * The buffer may have been set dirty during
1716                          * attachment to a dirty page.
1717                          */
1718                         clear_buffer_dirty(bh);
1719                 }
1720         } while ((bh = bh->b_this_page) != head);
1721         SetPageError(page);
1722         BUG_ON(PageWriteback(page));
1723         mapping_set_error(page->mapping, err);
1724         set_page_writeback(page);
1725         do {
1726                 struct buffer_head *next = bh->b_this_page;
1727                 if (buffer_async_write(bh)) {
1728                         clear_buffer_dirty(bh);
1729                         submit_bh(write_op, bh);
1730                         nr_underway++;
1731                 }
1732                 bh = next;
1733         } while (bh != head);
1734         unlock_page(page);
1735         goto done;
1736 }
1737
1738 /*
1739  * If a page has any new buffers, zero them out here, and mark them uptodate
1740  * and dirty so they'll be written out (in order to prevent uninitialised
1741  * block data from leaking). And clear the new bit.
1742  */
1743 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1744 {
1745         unsigned int block_start, block_end;
1746         struct buffer_head *head, *bh;
1747
1748         BUG_ON(!PageLocked(page));
1749         if (!page_has_buffers(page))
1750                 return;
1751
1752         bh = head = page_buffers(page);
1753         block_start = 0;
1754         do {
1755                 block_end = block_start + bh->b_size;
1756
1757                 if (buffer_new(bh)) {
1758                         if (block_end > from && block_start < to) {
1759                                 if (!PageUptodate(page)) {
1760                                         unsigned start, size;
1761
1762                                         start = max(from, block_start);
1763                                         size = min(to, block_end) - start;
1764
1765                                         zero_user(page, start, size);
1766                                         set_buffer_uptodate(bh);
1767                                 }
1768
1769                                 clear_buffer_new(bh);
1770                                 mark_buffer_dirty(bh);
1771                         }
1772                 }
1773
1774                 block_start = block_end;
1775                 bh = bh->b_this_page;
1776         } while (bh != head);
1777 }
1778 EXPORT_SYMBOL(page_zero_new_buffers);
1779
1780 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1781                 get_block_t *get_block)
1782 {
1783         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1784         unsigned to = from + len;
1785         struct inode *inode = page->mapping->host;
1786         unsigned block_start, block_end;
1787         sector_t block;
1788         int err = 0;
1789         unsigned blocksize, bbits;
1790         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1791
1792         BUG_ON(!PageLocked(page));
1793         BUG_ON(from > PAGE_CACHE_SIZE);
1794         BUG_ON(to > PAGE_CACHE_SIZE);
1795         BUG_ON(from > to);
1796
1797         blocksize = 1 << inode->i_blkbits;
1798         if (!page_has_buffers(page))
1799                 create_empty_buffers(page, blocksize, 0);
1800         head = page_buffers(page);
1801
1802         bbits = inode->i_blkbits;
1803         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1804
1805         for(bh = head, block_start = 0; bh != head || !block_start;
1806             block++, block_start=block_end, bh = bh->b_this_page) {
1807                 block_end = block_start + blocksize;
1808                 if (block_end <= from || block_start >= to) {
1809                         if (PageUptodate(page)) {
1810                                 if (!buffer_uptodate(bh))
1811                                         set_buffer_uptodate(bh);
1812                         }
1813                         continue;
1814                 }
1815                 if (buffer_new(bh))
1816                         clear_buffer_new(bh);
1817                 if (!buffer_mapped(bh)) {
1818                         WARN_ON(bh->b_size != blocksize);
1819                         err = get_block(inode, block, bh, 1);
1820                         if (err)
1821                                 break;
1822                         if (buffer_new(bh)) {
1823                                 unmap_underlying_metadata(bh->b_bdev,
1824                                                         bh->b_blocknr);
1825                                 if (PageUptodate(page)) {
1826                                         clear_buffer_new(bh);
1827                                         set_buffer_uptodate(bh);
1828                                         mark_buffer_dirty(bh);
1829                                         continue;
1830                                 }
1831                                 if (block_end > to || block_start < from)
1832                                         zero_user_segments(page,
1833                                                 to, block_end,
1834                                                 block_start, from);
1835                                 continue;
1836                         }
1837                 }
1838                 if (PageUptodate(page)) {
1839                         if (!buffer_uptodate(bh))
1840                                 set_buffer_uptodate(bh);
1841                         continue; 
1842                 }
1843                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1844                     !buffer_unwritten(bh) &&
1845                      (block_start < from || block_end > to)) {
1846                         ll_rw_block(READ, 1, &bh);
1847                         *wait_bh++=bh;
1848                 }
1849         }
1850         /*
1851          * If we issued read requests - let them complete.
1852          */
1853         while(wait_bh > wait) {
1854                 wait_on_buffer(*--wait_bh);
1855                 if (!buffer_uptodate(*wait_bh))
1856                         err = -EIO;
1857         }
1858         if (unlikely(err))
1859                 page_zero_new_buffers(page, from, to);
1860         return err;
1861 }
1862 EXPORT_SYMBOL(__block_write_begin);
1863
1864 static int __block_commit_write(struct inode *inode, struct page *page,
1865                 unsigned from, unsigned to)
1866 {
1867         unsigned block_start, block_end;
1868         int partial = 0;
1869         unsigned blocksize;
1870         struct buffer_head *bh, *head;
1871
1872         blocksize = 1 << inode->i_blkbits;
1873
1874         for(bh = head = page_buffers(page), block_start = 0;
1875             bh != head || !block_start;
1876             block_start=block_end, bh = bh->b_this_page) {
1877                 block_end = block_start + blocksize;
1878                 if (block_end <= from || block_start >= to) {
1879                         if (!buffer_uptodate(bh))
1880                                 partial = 1;
1881                 } else {
1882                         set_buffer_uptodate(bh);
1883                         mark_buffer_dirty(bh);
1884                 }
1885                 clear_buffer_new(bh);
1886         }
1887
1888         /*
1889          * If this is a partial write which happened to make all buffers
1890          * uptodate then we can optimize away a bogus readpage() for
1891          * the next read(). Here we 'discover' whether the page went
1892          * uptodate as a result of this (potentially partial) write.
1893          */
1894         if (!partial)
1895                 SetPageUptodate(page);
1896         return 0;
1897 }
1898
1899 /*
1900  * block_write_begin takes care of the basic task of block allocation and
1901  * bringing partial write blocks uptodate first.
1902  *
1903  * The filesystem needs to handle block truncation upon failure.
1904  */
1905 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
1906                 unsigned flags, struct page **pagep, get_block_t *get_block)
1907 {
1908         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1909         struct page *page;
1910         int status;
1911
1912         page = grab_cache_page_write_begin(mapping, index, flags);
1913         if (!page)
1914                 return -ENOMEM;
1915
1916         status = __block_write_begin(page, pos, len, get_block);
1917         if (unlikely(status)) {
1918                 unlock_page(page);
1919                 page_cache_release(page);
1920                 page = NULL;
1921         }
1922
1923         *pagep = page;
1924         return status;
1925 }
1926 EXPORT_SYMBOL(block_write_begin);
1927
1928 int block_write_end(struct file *file, struct address_space *mapping,
1929                         loff_t pos, unsigned len, unsigned copied,
1930                         struct page *page, void *fsdata)
1931 {
1932         struct inode *inode = mapping->host;
1933         unsigned start;
1934
1935         start = pos & (PAGE_CACHE_SIZE - 1);
1936
1937         if (unlikely(copied < len)) {
1938                 /*
1939                  * The buffers that were written will now be uptodate, so we
1940                  * don't have to worry about a readpage reading them and
1941                  * overwriting a partial write. However if we have encountered
1942                  * a short write and only partially written into a buffer, it
1943                  * will not be marked uptodate, so a readpage might come in and
1944                  * destroy our partial write.
1945                  *
1946                  * Do the simplest thing, and just treat any short write to a
1947                  * non uptodate page as a zero-length write, and force the
1948                  * caller to redo the whole thing.
1949                  */
1950                 if (!PageUptodate(page))
1951                         copied = 0;
1952
1953                 page_zero_new_buffers(page, start+copied, start+len);
1954         }
1955         flush_dcache_page(page);
1956
1957         /* This could be a short (even 0-length) commit */
1958         __block_commit_write(inode, page, start, start+copied);
1959
1960         return copied;
1961 }
1962 EXPORT_SYMBOL(block_write_end);
1963
1964 int generic_write_end(struct file *file, struct address_space *mapping,
1965                         loff_t pos, unsigned len, unsigned copied,
1966                         struct page *page, void *fsdata)
1967 {
1968         struct inode *inode = mapping->host;
1969         int i_size_changed = 0;
1970
1971         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1972
1973         /*
1974          * No need to use i_size_read() here, the i_size
1975          * cannot change under us because we hold i_mutex.
1976          *
1977          * But it's important to update i_size while still holding page lock:
1978          * page writeout could otherwise come in and zero beyond i_size.
1979          */
1980         if (pos+copied > inode->i_size) {
1981                 i_size_write(inode, pos+copied);
1982                 i_size_changed = 1;
1983         }
1984
1985         unlock_page(page);
1986         page_cache_release(page);
1987
1988         /*
1989          * Don't mark the inode dirty under page lock. First, it unnecessarily
1990          * makes the holding time of page lock longer. Second, it forces lock
1991          * ordering of page lock and transaction start for journaling
1992          * filesystems.
1993          */
1994         if (i_size_changed)
1995                 mark_inode_dirty(inode);
1996
1997         return copied;
1998 }
1999 EXPORT_SYMBOL(generic_write_end);
2000
2001 /*
2002  * block_is_partially_uptodate checks whether buffers within a page are
2003  * uptodate or not.
2004  *
2005  * Returns true if all buffers which correspond to a file portion
2006  * we want to read are uptodate.
2007  */
2008 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2009                                         unsigned long from)
2010 {
2011         struct inode *inode = page->mapping->host;
2012         unsigned block_start, block_end, blocksize;
2013         unsigned to;
2014         struct buffer_head *bh, *head;
2015         int ret = 1;
2016
2017         if (!page_has_buffers(page))
2018                 return 0;
2019
2020         blocksize = 1 << inode->i_blkbits;
2021         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2022         to = from + to;
2023         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2024                 return 0;
2025
2026         head = page_buffers(page);
2027         bh = head;
2028         block_start = 0;
2029         do {
2030                 block_end = block_start + blocksize;
2031                 if (block_end > from && block_start < to) {
2032                         if (!buffer_uptodate(bh)) {
2033                                 ret = 0;
2034                                 break;
2035                         }
2036                         if (block_end >= to)
2037                                 break;
2038                 }
2039                 block_start = block_end;
2040                 bh = bh->b_this_page;
2041         } while (bh != head);
2042
2043         return ret;
2044 }
2045 EXPORT_SYMBOL(block_is_partially_uptodate);
2046
2047 /*
2048  * Generic "read page" function for block devices that have the normal
2049  * get_block functionality. This is most of the block device filesystems.
2050  * Reads the page asynchronously --- the unlock_buffer() and
2051  * set/clear_buffer_uptodate() functions propagate buffer state into the
2052  * page struct once IO has completed.
2053  */
2054 int block_read_full_page(struct page *page, get_block_t *get_block)
2055 {
2056         struct inode *inode = page->mapping->host;
2057         sector_t iblock, lblock;
2058         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2059         unsigned int blocksize;
2060         int nr, i;
2061         int fully_mapped = 1;
2062
2063         BUG_ON(!PageLocked(page));
2064         blocksize = 1 << inode->i_blkbits;
2065         if (!page_has_buffers(page))
2066                 create_empty_buffers(page, blocksize, 0);
2067         head = page_buffers(page);
2068
2069         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2070         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2071         bh = head;
2072         nr = 0;
2073         i = 0;
2074
2075         do {
2076                 if (buffer_uptodate(bh))
2077                         continue;
2078
2079                 if (!buffer_mapped(bh)) {
2080                         int err = 0;
2081
2082                         fully_mapped = 0;
2083                         if (iblock < lblock) {
2084                                 WARN_ON(bh->b_size != blocksize);
2085                                 err = get_block(inode, iblock, bh, 0);
2086                                 if (err)
2087                                         SetPageError(page);
2088                         }
2089                         if (!buffer_mapped(bh)) {
2090                                 zero_user(page, i * blocksize, blocksize);
2091                                 if (!err)
2092                                         set_buffer_uptodate(bh);
2093                                 continue;
2094                         }
2095                         /*
2096                          * get_block() might have updated the buffer
2097                          * synchronously
2098                          */
2099                         if (buffer_uptodate(bh))
2100                                 continue;
2101                 }
2102                 arr[nr++] = bh;
2103         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2104
2105         if (fully_mapped)
2106                 SetPageMappedToDisk(page);
2107
2108         if (!nr) {
2109                 /*
2110                  * All buffers are uptodate - we can set the page uptodate
2111                  * as well. But not if get_block() returned an error.
2112                  */
2113                 if (!PageError(page))
2114                         SetPageUptodate(page);
2115                 unlock_page(page);
2116                 return 0;
2117         }
2118
2119         /* Stage two: lock the buffers */
2120         for (i = 0; i < nr; i++) {
2121                 bh = arr[i];
2122                 lock_buffer(bh);
2123                 mark_buffer_async_read(bh);
2124         }
2125
2126         /*
2127          * Stage 3: start the IO.  Check for uptodateness
2128          * inside the buffer lock in case another process reading
2129          * the underlying blockdev brought it uptodate (the sct fix).
2130          */
2131         for (i = 0; i < nr; i++) {
2132                 bh = arr[i];
2133                 if (buffer_uptodate(bh))
2134                         end_buffer_async_read(bh, 1);
2135                 else
2136                         submit_bh(READ, bh);
2137         }
2138         return 0;
2139 }
2140 EXPORT_SYMBOL(block_read_full_page);
2141
2142 /* utility function for filesystems that need to do work on expanding
2143  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2144  * deal with the hole.  
2145  */
2146 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2147 {
2148         struct address_space *mapping = inode->i_mapping;
2149         struct page *page;
2150         void *fsdata;
2151         int err;
2152
2153         err = inode_newsize_ok(inode, size);
2154         if (err)
2155                 goto out;
2156
2157         err = pagecache_write_begin(NULL, mapping, size, 0,
2158                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2159                                 &page, &fsdata);
2160         if (err)
2161                 goto out;
2162
2163         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2164         BUG_ON(err > 0);
2165
2166 out:
2167         return err;
2168 }
2169 EXPORT_SYMBOL(generic_cont_expand_simple);
2170
2171 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2172                             loff_t pos, loff_t *bytes)
2173 {
2174         struct inode *inode = mapping->host;
2175         unsigned blocksize = 1 << inode->i_blkbits;
2176         struct page *page;
2177         void *fsdata;
2178         pgoff_t index, curidx;
2179         loff_t curpos;
2180         unsigned zerofrom, offset, len;
2181         int err = 0;
2182
2183         index = pos >> PAGE_CACHE_SHIFT;
2184         offset = pos & ~PAGE_CACHE_MASK;
2185
2186         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2187                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2188                 if (zerofrom & (blocksize-1)) {
2189                         *bytes |= (blocksize-1);
2190                         (*bytes)++;
2191                 }
2192                 len = PAGE_CACHE_SIZE - zerofrom;
2193
2194                 err = pagecache_write_begin(file, mapping, curpos, len,
2195                                                 AOP_FLAG_UNINTERRUPTIBLE,
2196                                                 &page, &fsdata);
2197                 if (err)
2198                         goto out;
2199                 zero_user(page, zerofrom, len);
2200                 err = pagecache_write_end(file, mapping, curpos, len, len,
2201                                                 page, fsdata);
2202                 if (err < 0)
2203                         goto out;
2204                 BUG_ON(err != len);
2205                 err = 0;
2206
2207                 balance_dirty_pages_ratelimited(mapping);
2208         }
2209
2210         /* page covers the boundary, find the boundary offset */
2211         if (index == curidx) {
2212                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2213                 /* if we will expand the thing last block will be filled */
2214                 if (offset <= zerofrom) {
2215                         goto out;
2216                 }
2217                 if (zerofrom & (blocksize-1)) {
2218                         *bytes |= (blocksize-1);
2219                         (*bytes)++;
2220                 }
2221                 len = offset - zerofrom;
2222
2223                 err = pagecache_write_begin(file, mapping, curpos, len,
2224                                                 AOP_FLAG_UNINTERRUPTIBLE,
2225                                                 &page, &fsdata);
2226                 if (err)
2227                         goto out;
2228                 zero_user(page, zerofrom, len);
2229                 err = pagecache_write_end(file, mapping, curpos, len, len,
2230                                                 page, fsdata);
2231                 if (err < 0)
2232                         goto out;
2233                 BUG_ON(err != len);
2234                 err = 0;
2235         }
2236 out:
2237         return err;
2238 }
2239
2240 /*
2241  * For moronic filesystems that do not allow holes in file.
2242  * We may have to extend the file.
2243  */
2244 int cont_write_begin(struct file *file, struct address_space *mapping,
2245                         loff_t pos, unsigned len, unsigned flags,
2246                         struct page **pagep, void **fsdata,
2247                         get_block_t *get_block, loff_t *bytes)
2248 {
2249         struct inode *inode = mapping->host;
2250         unsigned blocksize = 1 << inode->i_blkbits;
2251         unsigned zerofrom;
2252         int err;
2253
2254         err = cont_expand_zero(file, mapping, pos, bytes);
2255         if (err)
2256                 return err;
2257
2258         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2259         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2260                 *bytes |= (blocksize-1);
2261                 (*bytes)++;
2262         }
2263
2264         return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2265 }
2266 EXPORT_SYMBOL(cont_write_begin);
2267
2268 int block_commit_write(struct page *page, unsigned from, unsigned to)
2269 {
2270         struct inode *inode = page->mapping->host;
2271         __block_commit_write(inode,page,from,to);
2272         return 0;
2273 }
2274 EXPORT_SYMBOL(block_commit_write);
2275
2276 /*
2277  * block_page_mkwrite() is not allowed to change the file size as it gets
2278  * called from a page fault handler when a page is first dirtied. Hence we must
2279  * be careful to check for EOF conditions here. We set the page up correctly
2280  * for a written page which means we get ENOSPC checking when writing into
2281  * holes and correct delalloc and unwritten extent mapping on filesystems that
2282  * support these features.
2283  *
2284  * We are not allowed to take the i_mutex here so we have to play games to
2285  * protect against truncate races as the page could now be beyond EOF.  Because
2286  * truncate writes the inode size before removing pages, once we have the
2287  * page lock we can determine safely if the page is beyond EOF. If it is not
2288  * beyond EOF, then the page is guaranteed safe against truncation until we
2289  * unlock the page.
2290  *
2291  * Direct callers of this function should call vfs_check_frozen() so that page
2292  * fault does not busyloop until the fs is thawed.
2293  */
2294 int __block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2295                          get_block_t get_block)
2296 {
2297         struct page *page = vmf->page;
2298         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2299         unsigned long end;
2300         loff_t size;
2301         int ret;
2302
2303         lock_page(page);
2304         size = i_size_read(inode);
2305         if ((page->mapping != inode->i_mapping) ||
2306             (page_offset(page) > size)) {
2307                 /* We overload EFAULT to mean page got truncated */
2308                 ret = -EFAULT;
2309                 goto out_unlock;
2310         }
2311
2312         /* page is wholly or partially inside EOF */
2313         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2314                 end = size & ~PAGE_CACHE_MASK;
2315         else
2316                 end = PAGE_CACHE_SIZE;
2317
2318         ret = __block_write_begin(page, 0, end, get_block);
2319         if (!ret)
2320                 ret = block_commit_write(page, 0, end);
2321
2322         if (unlikely(ret < 0))
2323                 goto out_unlock;
2324         /*
2325          * Freezing in progress? We check after the page is marked dirty and
2326          * with page lock held so if the test here fails, we are sure freezing
2327          * code will wait during syncing until the page fault is done - at that
2328          * point page will be dirty and unlocked so freezing code will write it
2329          * and writeprotect it again.
2330          */
2331         set_page_dirty(page);
2332         if (inode->i_sb->s_frozen != SB_UNFROZEN) {
2333                 ret = -EAGAIN;
2334                 goto out_unlock;
2335         }
2336         wait_on_page_writeback(page);
2337         return 0;
2338 out_unlock:
2339         unlock_page(page);
2340         return ret;
2341 }
2342 EXPORT_SYMBOL(__block_page_mkwrite);
2343
2344 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2345                    get_block_t get_block)
2346 {
2347         int ret;
2348         struct super_block *sb = vma->vm_file->f_path.dentry->d_inode->i_sb;
2349
2350         /*
2351          * This check is racy but catches the common case. The check in
2352          * __block_page_mkwrite() is reliable.
2353          */
2354         vfs_check_frozen(sb, SB_FREEZE_WRITE);
2355         ret = __block_page_mkwrite(vma, vmf, get_block);
2356         return block_page_mkwrite_return(ret);
2357 }
2358 EXPORT_SYMBOL(block_page_mkwrite);
2359
2360 /*
2361  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2362  * immediately, while under the page lock.  So it needs a special end_io
2363  * handler which does not touch the bh after unlocking it.
2364  */
2365 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2366 {
2367         __end_buffer_read_notouch(bh, uptodate);
2368 }
2369
2370 /*
2371  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2372  * the page (converting it to circular linked list and taking care of page
2373  * dirty races).
2374  */
2375 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2376 {
2377         struct buffer_head *bh;
2378
2379         BUG_ON(!PageLocked(page));
2380
2381         spin_lock(&page->mapping->private_lock);
2382         bh = head;
2383         do {
2384                 if (PageDirty(page))
2385                         set_buffer_dirty(bh);
2386                 if (!bh->b_this_page)
2387                         bh->b_this_page = head;
2388                 bh = bh->b_this_page;
2389         } while (bh != head);
2390         attach_page_buffers(page, head);
2391         spin_unlock(&page->mapping->private_lock);
2392 }
2393
2394 /*
2395  * On entry, the page is fully not uptodate.
2396  * On exit the page is fully uptodate in the areas outside (from,to)
2397  * The filesystem needs to handle block truncation upon failure.
2398  */
2399 int nobh_write_begin(struct address_space *mapping,
2400                         loff_t pos, unsigned len, unsigned flags,
2401                         struct page **pagep, void **fsdata,
2402                         get_block_t *get_block)
2403 {
2404         struct inode *inode = mapping->host;
2405         const unsigned blkbits = inode->i_blkbits;
2406         const unsigned blocksize = 1 << blkbits;
2407         struct buffer_head *head, *bh;
2408         struct page *page;
2409         pgoff_t index;
2410         unsigned from, to;
2411         unsigned block_in_page;
2412         unsigned block_start, block_end;
2413         sector_t block_in_file;
2414         int nr_reads = 0;
2415         int ret = 0;
2416         int is_mapped_to_disk = 1;
2417
2418         index = pos >> PAGE_CACHE_SHIFT;
2419         from = pos & (PAGE_CACHE_SIZE - 1);
2420         to = from + len;
2421
2422         page = grab_cache_page_write_begin(mapping, index, flags);
2423         if (!page)
2424                 return -ENOMEM;
2425         *pagep = page;
2426         *fsdata = NULL;
2427
2428         if (page_has_buffers(page)) {
2429                 ret = __block_write_begin(page, pos, len, get_block);
2430                 if (unlikely(ret))
2431                         goto out_release;
2432                 return ret;
2433         }
2434
2435         if (PageMappedToDisk(page))
2436                 return 0;
2437
2438         /*
2439          * Allocate buffers so that we can keep track of state, and potentially
2440          * attach them to the page if an error occurs. In the common case of
2441          * no error, they will just be freed again without ever being attached
2442          * to the page (which is all OK, because we're under the page lock).
2443          *
2444          * Be careful: the buffer linked list is a NULL terminated one, rather
2445          * than the circular one we're used to.
2446          */
2447         head = alloc_page_buffers(page, blocksize, 0);
2448         if (!head) {
2449                 ret = -ENOMEM;
2450                 goto out_release;
2451         }
2452
2453         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2454
2455         /*
2456          * We loop across all blocks in the page, whether or not they are
2457          * part of the affected region.  This is so we can discover if the
2458          * page is fully mapped-to-disk.
2459          */
2460         for (block_start = 0, block_in_page = 0, bh = head;
2461                   block_start < PAGE_CACHE_SIZE;
2462                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2463                 int create;
2464
2465                 block_end = block_start + blocksize;
2466                 bh->b_state = 0;
2467                 create = 1;
2468                 if (block_start >= to)
2469                         create = 0;
2470                 ret = get_block(inode, block_in_file + block_in_page,
2471                                         bh, create);
2472                 if (ret)
2473                         goto failed;
2474                 if (!buffer_mapped(bh))
2475                         is_mapped_to_disk = 0;
2476                 if (buffer_new(bh))
2477                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2478                 if (PageUptodate(page)) {
2479                         set_buffer_uptodate(bh);
2480                         continue;
2481                 }
2482                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2483                         zero_user_segments(page, block_start, from,
2484                                                         to, block_end);
2485                         continue;
2486                 }
2487                 if (buffer_uptodate(bh))
2488                         continue;       /* reiserfs does this */
2489                 if (block_start < from || block_end > to) {
2490                         lock_buffer(bh);
2491                         bh->b_end_io = end_buffer_read_nobh;
2492                         submit_bh(READ, bh);
2493                         nr_reads++;
2494                 }
2495         }
2496
2497         if (nr_reads) {
2498                 /*
2499                  * The page is locked, so these buffers are protected from
2500                  * any VM or truncate activity.  Hence we don't need to care
2501                  * for the buffer_head refcounts.
2502                  */
2503                 for (bh = head; bh; bh = bh->b_this_page) {
2504                         wait_on_buffer(bh);
2505                         if (!buffer_uptodate(bh))
2506                                 ret = -EIO;
2507                 }
2508                 if (ret)
2509                         goto failed;
2510         }
2511
2512         if (is_mapped_to_disk)
2513                 SetPageMappedToDisk(page);
2514
2515         *fsdata = head; /* to be released by nobh_write_end */
2516
2517         return 0;
2518
2519 failed:
2520         BUG_ON(!ret);
2521         /*
2522          * Error recovery is a bit difficult. We need to zero out blocks that
2523          * were newly allocated, and dirty them to ensure they get written out.
2524          * Buffers need to be attached to the page at this point, otherwise
2525          * the handling of potential IO errors during writeout would be hard
2526          * (could try doing synchronous writeout, but what if that fails too?)
2527          */
2528         attach_nobh_buffers(page, head);
2529         page_zero_new_buffers(page, from, to);
2530
2531 out_release:
2532         unlock_page(page);
2533         page_cache_release(page);
2534         *pagep = NULL;
2535
2536         return ret;
2537 }
2538 EXPORT_SYMBOL(nobh_write_begin);
2539
2540 int nobh_write_end(struct file *file, struct address_space *mapping,
2541                         loff_t pos, unsigned len, unsigned copied,
2542                         struct page *page, void *fsdata)
2543 {
2544         struct inode *inode = page->mapping->host;
2545         struct buffer_head *head = fsdata;
2546         struct buffer_head *bh;
2547         BUG_ON(fsdata != NULL && page_has_buffers(page));
2548
2549         if (unlikely(copied < len) && head)
2550                 attach_nobh_buffers(page, head);
2551         if (page_has_buffers(page))
2552                 return generic_write_end(file, mapping, pos, len,
2553                                         copied, page, fsdata);
2554
2555         SetPageUptodate(page);
2556         set_page_dirty(page);
2557         if (pos+copied > inode->i_size) {
2558                 i_size_write(inode, pos+copied);
2559                 mark_inode_dirty(inode);
2560         }
2561
2562         unlock_page(page);
2563         page_cache_release(page);
2564
2565         while (head) {
2566                 bh = head;
2567                 head = head->b_this_page;
2568                 free_buffer_head(bh);
2569         }
2570
2571         return copied;
2572 }
2573 EXPORT_SYMBOL(nobh_write_end);
2574
2575 /*
2576  * nobh_writepage() - based on block_full_write_page() except
2577  * that it tries to operate without attaching bufferheads to
2578  * the page.
2579  */
2580 int nobh_writepage(struct page *page, get_block_t *get_block,
2581                         struct writeback_control *wbc)
2582 {
2583         struct inode * const inode = page->mapping->host;
2584         loff_t i_size = i_size_read(inode);
2585         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2586         unsigned offset;
2587         int ret;
2588
2589         /* Is the page fully inside i_size? */
2590         if (page->index < end_index)
2591                 goto out;
2592
2593         /* Is the page fully outside i_size? (truncate in progress) */
2594         offset = i_size & (PAGE_CACHE_SIZE-1);
2595         if (page->index >= end_index+1 || !offset) {
2596                 /*
2597                  * The page may have dirty, unmapped buffers.  For example,
2598                  * they may have been added in ext3_writepage().  Make them
2599                  * freeable here, so the page does not leak.
2600                  */
2601 #if 0
2602                 /* Not really sure about this  - do we need this ? */
2603                 if (page->mapping->a_ops->invalidatepage)
2604                         page->mapping->a_ops->invalidatepage(page, offset);
2605 #endif
2606                 unlock_page(page);
2607                 return 0; /* don't care */
2608         }
2609
2610         /*
2611          * The page straddles i_size.  It must be zeroed out on each and every
2612          * writepage invocation because it may be mmapped.  "A file is mapped
2613          * in multiples of the page size.  For a file that is not a multiple of
2614          * the  page size, the remaining memory is zeroed when mapped, and
2615          * writes to that region are not written out to the file."
2616          */
2617         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2618 out:
2619         ret = mpage_writepage(page, get_block, wbc);
2620         if (ret == -EAGAIN)
2621                 ret = __block_write_full_page(inode, page, get_block, wbc,
2622                                               end_buffer_async_write);
2623         return ret;
2624 }
2625 EXPORT_SYMBOL(nobh_writepage);
2626
2627 int nobh_truncate_page(struct address_space *mapping,
2628                         loff_t from, get_block_t *get_block)
2629 {
2630         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2631         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2632         unsigned blocksize;
2633         sector_t iblock;
2634         unsigned length, pos;
2635         struct inode *inode = mapping->host;
2636         struct page *page;
2637         struct buffer_head map_bh;
2638         int err;
2639
2640         blocksize = 1 << inode->i_blkbits;
2641         length = offset & (blocksize - 1);
2642
2643         /* Block boundary? Nothing to do */
2644         if (!length)
2645                 return 0;
2646
2647         length = blocksize - length;
2648         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2649
2650         page = grab_cache_page(mapping, index);
2651         err = -ENOMEM;
2652         if (!page)
2653                 goto out;
2654
2655         if (page_has_buffers(page)) {
2656 has_buffers:
2657                 unlock_page(page);
2658                 page_cache_release(page);
2659                 return block_truncate_page(mapping, from, get_block);
2660         }
2661
2662         /* Find the buffer that contains "offset" */
2663         pos = blocksize;
2664         while (offset >= pos) {
2665                 iblock++;
2666                 pos += blocksize;
2667         }
2668
2669         map_bh.b_size = blocksize;
2670         map_bh.b_state = 0;
2671         err = get_block(inode, iblock, &map_bh, 0);
2672         if (err)
2673                 goto unlock;
2674         /* unmapped? It's a hole - nothing to do */
2675         if (!buffer_mapped(&map_bh))
2676                 goto unlock;
2677
2678         /* Ok, it's mapped. Make sure it's up-to-date */
2679         if (!PageUptodate(page)) {
2680                 err = mapping->a_ops->readpage(NULL, page);
2681                 if (err) {
2682                         page_cache_release(page);
2683                         goto out;
2684                 }
2685                 lock_page(page);
2686                 if (!PageUptodate(page)) {
2687                         err = -EIO;
2688                         goto unlock;
2689                 }
2690                 if (page_has_buffers(page))
2691                         goto has_buffers;
2692         }
2693         zero_user(page, offset, length);
2694         set_page_dirty(page);
2695         err = 0;
2696
2697 unlock:
2698         unlock_page(page);
2699         page_cache_release(page);
2700 out:
2701         return err;
2702 }
2703 EXPORT_SYMBOL(nobh_truncate_page);
2704
2705 int block_truncate_page(struct address_space *mapping,
2706                         loff_t from, get_block_t *get_block)
2707 {
2708         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2709         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2710         unsigned blocksize;
2711         sector_t iblock;
2712         unsigned length, pos;
2713         struct inode *inode = mapping->host;
2714         struct page *page;
2715         struct buffer_head *bh;
2716         int err;
2717
2718         blocksize = 1 << inode->i_blkbits;
2719         length = offset & (blocksize - 1);
2720
2721         /* Block boundary? Nothing to do */
2722         if (!length)
2723                 return 0;
2724
2725         length = blocksize - length;
2726         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2727         
2728         page = grab_cache_page(mapping, index);
2729         err = -ENOMEM;
2730         if (!page)
2731                 goto out;
2732
2733         if (!page_has_buffers(page))
2734                 create_empty_buffers(page, blocksize, 0);
2735
2736         /* Find the buffer that contains "offset" */
2737         bh = page_buffers(page);
2738         pos = blocksize;
2739         while (offset >= pos) {
2740                 bh = bh->b_this_page;
2741                 iblock++;
2742                 pos += blocksize;
2743         }
2744
2745         err = 0;
2746         if (!buffer_mapped(bh)) {
2747                 WARN_ON(bh->b_size != blocksize);
2748                 err = get_block(inode, iblock, bh, 0);
2749                 if (err)
2750                         goto unlock;
2751                 /* unmapped? It's a hole - nothing to do */
2752                 if (!buffer_mapped(bh))
2753                         goto unlock;
2754         }
2755
2756         /* Ok, it's mapped. Make sure it's up-to-date */
2757         if (PageUptodate(page))
2758                 set_buffer_uptodate(bh);
2759
2760         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2761                 err = -EIO;
2762                 ll_rw_block(READ, 1, &bh);
2763                 wait_on_buffer(bh);
2764                 /* Uhhuh. Read error. Complain and punt. */
2765                 if (!buffer_uptodate(bh))
2766                         goto unlock;
2767         }
2768
2769         zero_user(page, offset, length);
2770         mark_buffer_dirty(bh);
2771         err = 0;
2772
2773 unlock:
2774         unlock_page(page);
2775         page_cache_release(page);
2776 out:
2777         return err;
2778 }
2779 EXPORT_SYMBOL(block_truncate_page);
2780
2781 /*
2782  * The generic ->writepage function for buffer-backed address_spaces
2783  * this form passes in the end_io handler used to finish the IO.
2784  */
2785 int block_write_full_page_endio(struct page *page, get_block_t *get_block,
2786                         struct writeback_control *wbc, bh_end_io_t *handler)
2787 {
2788         struct inode * const inode = page->mapping->host;
2789         loff_t i_size = i_size_read(inode);
2790         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2791         unsigned offset;
2792
2793         /* Is the page fully inside i_size? */
2794         if (page->index < end_index)
2795                 return __block_write_full_page(inode, page, get_block, wbc,
2796                                                handler);
2797
2798         /* Is the page fully outside i_size? (truncate in progress) */
2799         offset = i_size & (PAGE_CACHE_SIZE-1);
2800         if (page->index >= end_index+1 || !offset) {
2801                 /*
2802                  * The page may have dirty, unmapped buffers.  For example,
2803                  * they may have been added in ext3_writepage().  Make them
2804                  * freeable here, so the page does not leak.
2805                  */
2806                 do_invalidatepage(page, 0);
2807                 unlock_page(page);
2808                 return 0; /* don't care */
2809         }
2810
2811         /*
2812          * The page straddles i_size.  It must be zeroed out on each and every
2813          * writepage invocation because it may be mmapped.  "A file is mapped
2814          * in multiples of the page size.  For a file that is not a multiple of
2815          * the  page size, the remaining memory is zeroed when mapped, and
2816          * writes to that region are not written out to the file."
2817          */
2818         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2819         return __block_write_full_page(inode, page, get_block, wbc, handler);
2820 }
2821 EXPORT_SYMBOL(block_write_full_page_endio);
2822
2823 /*
2824  * The generic ->writepage function for buffer-backed address_spaces
2825  */
2826 int block_write_full_page(struct page *page, get_block_t *get_block,
2827                         struct writeback_control *wbc)
2828 {
2829         return block_write_full_page_endio(page, get_block, wbc,
2830                                            end_buffer_async_write);
2831 }
2832 EXPORT_SYMBOL(block_write_full_page);
2833
2834 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2835                             get_block_t *get_block)
2836 {
2837         struct buffer_head tmp;
2838         struct inode *inode = mapping->host;
2839         tmp.b_state = 0;
2840         tmp.b_blocknr = 0;
2841         tmp.b_size = 1 << inode->i_blkbits;
2842         get_block(inode, block, &tmp, 0);
2843         return tmp.b_blocknr;
2844 }
2845 EXPORT_SYMBOL(generic_block_bmap);
2846
2847 static void end_bio_bh_io_sync(struct bio *bio, int err)
2848 {
2849         struct buffer_head *bh = bio->bi_private;
2850
2851         if (err == -EOPNOTSUPP) {
2852                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2853         }
2854
2855         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2856                 set_bit(BH_Quiet, &bh->b_state);
2857
2858         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2859         bio_put(bio);
2860 }
2861
2862 int submit_bh(int rw, struct buffer_head * bh)
2863 {
2864         struct bio *bio;
2865         int ret = 0;
2866
2867         BUG_ON(!buffer_locked(bh));
2868         BUG_ON(!buffer_mapped(bh));
2869         BUG_ON(!bh->b_end_io);
2870         BUG_ON(buffer_delay(bh));
2871         BUG_ON(buffer_unwritten(bh));
2872
2873         /*
2874          * Only clear out a write error when rewriting
2875          */
2876         if (test_set_buffer_req(bh) && (rw & WRITE))
2877                 clear_buffer_write_io_error(bh);
2878
2879         /*
2880          * from here on down, it's all bio -- do the initial mapping,
2881          * submit_bio -> generic_make_request may further map this bio around
2882          */
2883         bio = bio_alloc(GFP_NOIO, 1);
2884
2885         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2886         bio->bi_bdev = bh->b_bdev;
2887         bio->bi_io_vec[0].bv_page = bh->b_page;
2888         bio->bi_io_vec[0].bv_len = bh->b_size;
2889         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2890
2891         bio->bi_vcnt = 1;
2892         bio->bi_idx = 0;
2893         bio->bi_size = bh->b_size;
2894
2895         bio->bi_end_io = end_bio_bh_io_sync;
2896         bio->bi_private = bh;
2897
2898         bio_get(bio);
2899         submit_bio(rw, bio);
2900
2901         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2902                 ret = -EOPNOTSUPP;
2903
2904         bio_put(bio);
2905         return ret;
2906 }
2907 EXPORT_SYMBOL(submit_bh);
2908
2909 /**
2910  * ll_rw_block: low-level access to block devices (DEPRECATED)
2911  * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
2912  * @nr: number of &struct buffer_heads in the array
2913  * @bhs: array of pointers to &struct buffer_head
2914  *
2915  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2916  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2917  * %READA option is described in the documentation for generic_make_request()
2918  * which ll_rw_block() calls.
2919  *
2920  * This function drops any buffer that it cannot get a lock on (with the
2921  * BH_Lock state bit), any buffer that appears to be clean when doing a write
2922  * request, and any buffer that appears to be up-to-date when doing read
2923  * request.  Further it marks as clean buffers that are processed for
2924  * writing (the buffer cache won't assume that they are actually clean
2925  * until the buffer gets unlocked).
2926  *
2927  * ll_rw_block sets b_end_io to simple completion handler that marks
2928  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2929  * any waiters. 
2930  *
2931  * All of the buffers must be for the same device, and must also be a
2932  * multiple of the current approved size for the device.
2933  */
2934 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2935 {
2936         int i;
2937
2938         for (i = 0; i < nr; i++) {
2939                 struct buffer_head *bh = bhs[i];
2940
2941                 if (!trylock_buffer(bh))
2942                         continue;
2943                 if (rw == WRITE) {
2944                         if (test_clear_buffer_dirty(bh)) {
2945                                 bh->b_end_io = end_buffer_write_sync;
2946                                 get_bh(bh);
2947                                 submit_bh(WRITE, bh);
2948                                 continue;
2949                         }
2950                 } else {
2951                         if (!buffer_uptodate(bh)) {
2952                                 bh->b_end_io = end_buffer_read_sync;
2953                                 get_bh(bh);
2954                                 submit_bh(rw, bh);
2955                                 continue;
2956                         }
2957                 }
2958                 unlock_buffer(bh);
2959         }
2960 }
2961 EXPORT_SYMBOL(ll_rw_block);
2962
2963 void write_dirty_buffer(struct buffer_head *bh, int rw)
2964 {
2965         lock_buffer(bh);
2966         if (!test_clear_buffer_dirty(bh)) {
2967                 unlock_buffer(bh);
2968                 return;
2969         }
2970         bh->b_end_io = end_buffer_write_sync;
2971         get_bh(bh);
2972         submit_bh(rw, bh);
2973 }
2974 EXPORT_SYMBOL(write_dirty_buffer);
2975
2976 /*
2977  * For a data-integrity writeout, we need to wait upon any in-progress I/O
2978  * and then start new I/O and then wait upon it.  The caller must have a ref on
2979  * the buffer_head.
2980  */
2981 int __sync_dirty_buffer(struct buffer_head *bh, int rw)
2982 {
2983         int ret = 0;
2984
2985         WARN_ON(atomic_read(&bh->b_count) < 1);
2986         lock_buffer(bh);
2987         if (test_clear_buffer_dirty(bh)) {
2988                 get_bh(bh);
2989                 bh->b_end_io = end_buffer_write_sync;
2990                 ret = submit_bh(rw, bh);
2991                 wait_on_buffer(bh);
2992                 if (!ret && !buffer_uptodate(bh))
2993                         ret = -EIO;
2994         } else {
2995                 unlock_buffer(bh);
2996         }
2997         return ret;
2998 }
2999 EXPORT_SYMBOL(__sync_dirty_buffer);
3000
3001 int sync_dirty_buffer(struct buffer_head *bh)
3002 {
3003         return __sync_dirty_buffer(bh, WRITE_SYNC);
3004 }
3005 EXPORT_SYMBOL(sync_dirty_buffer);
3006
3007 /*
3008  * try_to_free_buffers() checks if all the buffers on this particular page
3009  * are unused, and releases them if so.
3010  *
3011  * Exclusion against try_to_free_buffers may be obtained by either
3012  * locking the page or by holding its mapping's private_lock.
3013  *
3014  * If the page is dirty but all the buffers are clean then we need to
3015  * be sure to mark the page clean as well.  This is because the page
3016  * may be against a block device, and a later reattachment of buffers
3017  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3018  * filesystem data on the same device.
3019  *
3020  * The same applies to regular filesystem pages: if all the buffers are
3021  * clean then we set the page clean and proceed.  To do that, we require
3022  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3023  * private_lock.
3024  *
3025  * try_to_free_buffers() is non-blocking.
3026  */
3027 static inline int buffer_busy(struct buffer_head *bh)
3028 {
3029         return atomic_read(&bh->b_count) |
3030                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3031 }
3032
3033 static int
3034 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3035 {
3036         struct buffer_head *head = page_buffers(page);
3037         struct buffer_head *bh;
3038
3039         bh = head;
3040         do {
3041                 if (buffer_write_io_error(bh) && page->mapping)
3042                         set_bit(AS_EIO, &page->mapping->flags);
3043                 if (buffer_busy(bh))
3044                         goto failed;
3045                 bh = bh->b_this_page;
3046         } while (bh != head);
3047
3048         do {
3049                 struct buffer_head *next = bh->b_this_page;
3050
3051                 if (bh->b_assoc_map)
3052                         __remove_assoc_queue(bh);
3053                 bh = next;
3054         } while (bh != head);
3055         *buffers_to_free = head;
3056         __clear_page_buffers(page);
3057         return 1;
3058 failed:
3059         return 0;
3060 }
3061
3062 int try_to_free_buffers(struct page *page)
3063 {
3064         struct address_space * const mapping = page->mapping;
3065         struct buffer_head *buffers_to_free = NULL;
3066         int ret = 0;
3067
3068         BUG_ON(!PageLocked(page));
3069         if (PageWriteback(page))
3070                 return 0;
3071
3072         if (mapping == NULL) {          /* can this still happen? */
3073                 ret = drop_buffers(page, &buffers_to_free);
3074                 goto out;
3075         }
3076
3077         spin_lock(&mapping->private_lock);
3078         ret = drop_buffers(page, &buffers_to_free);
3079
3080         /*
3081          * If the filesystem writes its buffers by hand (eg ext3)
3082          * then we can have clean buffers against a dirty page.  We
3083          * clean the page here; otherwise the VM will never notice
3084          * that the filesystem did any IO at all.
3085          *
3086          * Also, during truncate, discard_buffer will have marked all
3087          * the page's buffers clean.  We discover that here and clean
3088          * the page also.
3089          *
3090          * private_lock must be held over this entire operation in order
3091          * to synchronise against __set_page_dirty_buffers and prevent the
3092          * dirty bit from being lost.
3093          */
3094         if (ret)
3095                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3096         spin_unlock(&mapping->private_lock);
3097 out:
3098         if (buffers_to_free) {
3099                 struct buffer_head *bh = buffers_to_free;
3100
3101                 do {
3102                         struct buffer_head *next = bh->b_this_page;
3103                         free_buffer_head(bh);
3104                         bh = next;
3105                 } while (bh != buffers_to_free);
3106         }
3107         return ret;
3108 }
3109 EXPORT_SYMBOL(try_to_free_buffers);
3110
3111 /*
3112  * There are no bdflush tunables left.  But distributions are
3113  * still running obsolete flush daemons, so we terminate them here.
3114  *
3115  * Use of bdflush() is deprecated and will be removed in a future kernel.
3116  * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3117  */
3118 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3119 {
3120         static int msg_count;
3121
3122         if (!capable(CAP_SYS_ADMIN))
3123                 return -EPERM;
3124
3125         if (msg_count < 5) {
3126                 msg_count++;
3127                 printk(KERN_INFO
3128                         "warning: process `%s' used the obsolete bdflush"
3129                         " system call\n", current->comm);
3130                 printk(KERN_INFO "Fix your initscripts?\n");
3131         }
3132
3133         if (func == 1)
3134                 do_exit(0);
3135         return 0;
3136 }
3137
3138 /*
3139  * Buffer-head allocation
3140  */
3141 static struct kmem_cache *bh_cachep;
3142
3143 /*
3144  * Once the number of bh's in the machine exceeds this level, we start
3145  * stripping them in writeback.
3146  */
3147 static int max_buffer_heads;
3148
3149 int buffer_heads_over_limit;
3150
3151 struct bh_accounting {
3152         int nr;                 /* Number of live bh's */
3153         int ratelimit;          /* Limit cacheline bouncing */
3154 };
3155
3156 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3157
3158 static void recalc_bh_state(void)
3159 {
3160         int i;
3161         int tot = 0;
3162
3163         if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3164                 return;
3165         __this_cpu_write(bh_accounting.ratelimit, 0);
3166         for_each_online_cpu(i)
3167                 tot += per_cpu(bh_accounting, i).nr;
3168         buffer_heads_over_limit = (tot > max_buffer_heads);
3169 }
3170
3171 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3172 {
3173         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3174         if (ret) {
3175                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3176                 preempt_disable();
3177                 __this_cpu_inc(bh_accounting.nr);
3178                 recalc_bh_state();
3179                 preempt_enable();
3180         }
3181         return ret;
3182 }
3183 EXPORT_SYMBOL(alloc_buffer_head);
3184
3185 void free_buffer_head(struct buffer_head *bh)
3186 {
3187         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3188         kmem_cache_free(bh_cachep, bh);
3189         preempt_disable();
3190         __this_cpu_dec(bh_accounting.nr);
3191         recalc_bh_state();
3192         preempt_enable();
3193 }
3194 EXPORT_SYMBOL(free_buffer_head);
3195
3196 static void buffer_exit_cpu(int cpu)
3197 {
3198         int i;
3199         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3200
3201         for (i = 0; i < BH_LRU_SIZE; i++) {
3202                 brelse(b->bhs[i]);
3203                 b->bhs[i] = NULL;
3204         }
3205         this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3206         per_cpu(bh_accounting, cpu).nr = 0;
3207 }
3208
3209 static int buffer_cpu_notify(struct notifier_block *self,
3210                               unsigned long action, void *hcpu)
3211 {
3212         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3213                 buffer_exit_cpu((unsigned long)hcpu);
3214         return NOTIFY_OK;
3215 }
3216
3217 /**
3218  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3219  * @bh: struct buffer_head
3220  *
3221  * Return true if the buffer is up-to-date and false,
3222  * with the buffer locked, if not.
3223  */
3224 int bh_uptodate_or_lock(struct buffer_head *bh)
3225 {
3226         if (!buffer_uptodate(bh)) {
3227                 lock_buffer(bh);
3228                 if (!buffer_uptodate(bh))
3229                         return 0;
3230                 unlock_buffer(bh);
3231         }
3232         return 1;
3233 }
3234 EXPORT_SYMBOL(bh_uptodate_or_lock);
3235
3236 /**
3237  * bh_submit_read - Submit a locked buffer for reading
3238  * @bh: struct buffer_head
3239  *
3240  * Returns zero on success and -EIO on error.
3241  */
3242 int bh_submit_read(struct buffer_head *bh)
3243 {
3244         BUG_ON(!buffer_locked(bh));
3245
3246         if (buffer_uptodate(bh)) {
3247                 unlock_buffer(bh);
3248                 return 0;
3249         }
3250
3251         get_bh(bh);
3252         bh->b_end_io = end_buffer_read_sync;
3253         submit_bh(READ, bh);
3254         wait_on_buffer(bh);
3255         if (buffer_uptodate(bh))
3256                 return 0;
3257         return -EIO;
3258 }
3259 EXPORT_SYMBOL(bh_submit_read);
3260
3261 void __init buffer_init(void)
3262 {
3263         int nrpages;
3264
3265         bh_cachep = kmem_cache_create("buffer_head",
3266                         sizeof(struct buffer_head), 0,
3267                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3268                                 SLAB_MEM_SPREAD),
3269                                 NULL);
3270
3271         /*
3272          * Limit the bh occupancy to 10% of ZONE_NORMAL
3273          */
3274         nrpages = (nr_free_buffer_pages() * 10) / 100;
3275         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3276         hotcpu_notifier(buffer_cpu_notify, 0);
3277 }