ext4: use i_size_read in ext4_unaligned_aio()
[linux-3.10.git] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/capability.h>
28 #include <linux/blkdev.h>
29 #include <linux/file.h>
30 #include <linux/quotaops.h>
31 #include <linux/highmem.h>
32 #include <linux/export.h>
33 #include <linux/writeback.h>
34 #include <linux/hash.h>
35 #include <linux/suspend.h>
36 #include <linux/buffer_head.h>
37 #include <linux/task_io_accounting_ops.h>
38 #include <linux/bio.h>
39 #include <linux/notifier.h>
40 #include <linux/cpu.h>
41 #include <linux/bitops.h>
42 #include <linux/mpage.h>
43 #include <linux/bit_spinlock.h>
44 #include <trace/events/block.h>
45
46 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
47
48 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
49
50 void 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 inline void touch_buffer(struct buffer_head *bh)
58 {
59         trace_block_touch_buffer(bh);
60         mark_page_accessed(bh->b_page);
61 }
62 EXPORT_SYMBOL(touch_buffer);
63
64 static int sleep_on_buffer(void *word)
65 {
66         io_schedule();
67         return 0;
68 }
69
70 void __lock_buffer(struct buffer_head *bh)
71 {
72         wait_on_bit_lock(&bh->b_state, BH_Lock, sleep_on_buffer,
73                                                         TASK_UNINTERRUPTIBLE);
74 }
75 EXPORT_SYMBOL(__lock_buffer);
76
77 void unlock_buffer(struct buffer_head *bh)
78 {
79         clear_bit_unlock(BH_Lock, &bh->b_state);
80         smp_mb__after_clear_bit();
81         wake_up_bit(&bh->b_state, BH_Lock);
82 }
83 EXPORT_SYMBOL(unlock_buffer);
84
85 /*
86  * Block until a buffer comes unlocked.  This doesn't stop it
87  * from becoming locked again - you have to lock it yourself
88  * if you want to preserve its state.
89  */
90 void __wait_on_buffer(struct buffer_head * bh)
91 {
92         wait_on_bit(&bh->b_state, BH_Lock, sleep_on_buffer, TASK_UNINTERRUPTIBLE);
93 }
94 EXPORT_SYMBOL(__wait_on_buffer);
95
96 static void
97 __clear_page_buffers(struct page *page)
98 {
99         ClearPagePrivate(page);
100         set_page_private(page, 0);
101         page_cache_release(page);
102 }
103
104
105 static int quiet_error(struct buffer_head *bh)
106 {
107         if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
108                 return 0;
109         return 1;
110 }
111
112
113 static void buffer_io_error(struct buffer_head *bh)
114 {
115         char b[BDEVNAME_SIZE];
116         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
117                         bdevname(bh->b_bdev, b),
118                         (unsigned long long)bh->b_blocknr);
119 }
120
121 /*
122  * End-of-IO handler helper function which does not touch the bh after
123  * unlocking it.
124  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
125  * a race there is benign: unlock_buffer() only use the bh's address for
126  * hashing after unlocking the buffer, so it doesn't actually touch the bh
127  * itself.
128  */
129 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
130 {
131         if (uptodate) {
132                 set_buffer_uptodate(bh);
133         } else {
134                 /* This happens, due to failed READA attempts. */
135                 clear_buffer_uptodate(bh);
136         }
137         unlock_buffer(bh);
138 }
139
140 /*
141  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
142  * unlock the buffer. This is what ll_rw_block uses too.
143  */
144 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
145 {
146         __end_buffer_read_notouch(bh, uptodate);
147         put_bh(bh);
148 }
149 EXPORT_SYMBOL(end_buffer_read_sync);
150
151 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
152 {
153         char b[BDEVNAME_SIZE];
154
155         if (uptodate) {
156                 set_buffer_uptodate(bh);
157         } else {
158                 if (!quiet_error(bh)) {
159                         buffer_io_error(bh);
160                         printk(KERN_WARNING "lost page write due to "
161                                         "I/O error on %s\n",
162                                        bdevname(bh->b_bdev, b));
163                 }
164                 set_buffer_write_io_error(bh);
165                 clear_buffer_uptodate(bh);
166         }
167         unlock_buffer(bh);
168         put_bh(bh);
169 }
170 EXPORT_SYMBOL(end_buffer_write_sync);
171
172 /*
173  * Various filesystems appear to want __find_get_block to be non-blocking.
174  * But it's the page lock which protects the buffers.  To get around this,
175  * we get exclusion from try_to_free_buffers with the blockdev mapping's
176  * private_lock.
177  *
178  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
179  * may be quite high.  This code could TryLock the page, and if that
180  * succeeds, there is no need to take private_lock. (But if
181  * private_lock is contended then so is mapping->tree_lock).
182  */
183 static struct buffer_head *
184 __find_get_block_slow(struct block_device *bdev, sector_t block)
185 {
186         struct inode *bd_inode = bdev->bd_inode;
187         struct address_space *bd_mapping = bd_inode->i_mapping;
188         struct buffer_head *ret = NULL;
189         pgoff_t index;
190         struct buffer_head *bh;
191         struct buffer_head *head;
192         struct page *page;
193         int all_mapped = 1;
194
195         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
196         page = find_get_page(bd_mapping, index);
197         if (!page)
198                 goto out;
199
200         spin_lock(&bd_mapping->private_lock);
201         if (!page_has_buffers(page))
202                 goto out_unlock;
203         head = page_buffers(page);
204         bh = head;
205         do {
206                 if (!buffer_mapped(bh))
207                         all_mapped = 0;
208                 else if (bh->b_blocknr == block) {
209                         ret = bh;
210                         get_bh(bh);
211                         goto out_unlock;
212                 }
213                 bh = bh->b_this_page;
214         } while (bh != head);
215
216         /* we might be here because some of the buffers on this page are
217          * not mapped.  This is due to various races between
218          * file io on the block device and getblk.  It gets dealt with
219          * elsewhere, don't buffer_error if we had some unmapped buffers
220          */
221         if (all_mapped) {
222                 char b[BDEVNAME_SIZE];
223
224                 printk("__find_get_block_slow() failed. "
225                         "block=%llu, b_blocknr=%llu\n",
226                         (unsigned long long)block,
227                         (unsigned long long)bh->b_blocknr);
228                 printk("b_state=0x%08lx, b_size=%zu\n",
229                         bh->b_state, bh->b_size);
230                 printk("device %s blocksize: %d\n", bdevname(bdev, b),
231                         1 << bd_inode->i_blkbits);
232         }
233 out_unlock:
234         spin_unlock(&bd_mapping->private_lock);
235         page_cache_release(page);
236 out:
237         return ret;
238 }
239
240 /*
241  * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
242  */
243 static void free_more_memory(void)
244 {
245         struct zone *zone;
246         int nid;
247
248         wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
249         yield();
250
251         for_each_online_node(nid) {
252                 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
253                                                 gfp_zone(GFP_NOFS), NULL,
254                                                 &zone);
255                 if (zone)
256                         try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
257                                                 GFP_NOFS, NULL);
258         }
259 }
260
261 /*
262  * I/O completion handler for block_read_full_page() - pages
263  * which come unlocked at the end of I/O.
264  */
265 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
266 {
267         unsigned long flags;
268         struct buffer_head *first;
269         struct buffer_head *tmp;
270         struct page *page;
271         int page_uptodate = 1;
272
273         BUG_ON(!buffer_async_read(bh));
274
275         page = bh->b_page;
276         if (uptodate) {
277                 set_buffer_uptodate(bh);
278         } else {
279                 clear_buffer_uptodate(bh);
280                 if (!quiet_error(bh))
281                         buffer_io_error(bh);
282                 SetPageError(page);
283         }
284
285         /*
286          * Be _very_ careful from here on. Bad things can happen if
287          * two buffer heads end IO at almost the same time and both
288          * decide that the page is now completely done.
289          */
290         first = page_buffers(page);
291         local_irq_save(flags);
292         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
293         clear_buffer_async_read(bh);
294         unlock_buffer(bh);
295         tmp = bh;
296         do {
297                 if (!buffer_uptodate(tmp))
298                         page_uptodate = 0;
299                 if (buffer_async_read(tmp)) {
300                         BUG_ON(!buffer_locked(tmp));
301                         goto still_busy;
302                 }
303                 tmp = tmp->b_this_page;
304         } while (tmp != bh);
305         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
306         local_irq_restore(flags);
307
308         /*
309          * If none of the buffers had errors and they are all
310          * uptodate then we can set the page uptodate.
311          */
312         if (page_uptodate && !PageError(page))
313                 SetPageUptodate(page);
314         unlock_page(page);
315         return;
316
317 still_busy:
318         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
319         local_irq_restore(flags);
320         return;
321 }
322
323 /*
324  * Completion handler for block_write_full_page() - pages which are unlocked
325  * during I/O, and which have PageWriteback cleared upon I/O completion.
326  */
327 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
328 {
329         char b[BDEVNAME_SIZE];
330         unsigned long flags;
331         struct buffer_head *first;
332         struct buffer_head *tmp;
333         struct page *page;
334
335         BUG_ON(!buffer_async_write(bh));
336
337         page = bh->b_page;
338         if (uptodate) {
339                 set_buffer_uptodate(bh);
340         } else {
341                 if (!quiet_error(bh)) {
342                         buffer_io_error(bh);
343                         printk(KERN_WARNING "lost page write due to "
344                                         "I/O error on %s\n",
345                                bdevname(bh->b_bdev, b));
346                 }
347                 set_bit(AS_EIO, &page->mapping->flags);
348                 set_buffer_write_io_error(bh);
349                 clear_buffer_uptodate(bh);
350                 SetPageError(page);
351         }
352
353         first = page_buffers(page);
354         local_irq_save(flags);
355         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
356
357         clear_buffer_async_write(bh);
358         unlock_buffer(bh);
359         tmp = bh->b_this_page;
360         while (tmp != bh) {
361                 if (buffer_async_write(tmp)) {
362                         BUG_ON(!buffer_locked(tmp));
363                         goto still_busy;
364                 }
365                 tmp = tmp->b_this_page;
366         }
367         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
368         local_irq_restore(flags);
369         end_page_writeback(page);
370         return;
371
372 still_busy:
373         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
374         local_irq_restore(flags);
375         return;
376 }
377 EXPORT_SYMBOL(end_buffer_async_write);
378
379 /*
380  * If a page's buffers are under async readin (end_buffer_async_read
381  * completion) then there is a possibility that another thread of
382  * control could lock one of the buffers after it has completed
383  * but while some of the other buffers have not completed.  This
384  * locked buffer would confuse end_buffer_async_read() into not unlocking
385  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
386  * that this buffer is not under async I/O.
387  *
388  * The page comes unlocked when it has no locked buffer_async buffers
389  * left.
390  *
391  * PageLocked prevents anyone starting new async I/O reads any of
392  * the buffers.
393  *
394  * PageWriteback is used to prevent simultaneous writeout of the same
395  * page.
396  *
397  * PageLocked prevents anyone from starting writeback of a page which is
398  * under read I/O (PageWriteback is only ever set against a locked page).
399  */
400 static void mark_buffer_async_read(struct buffer_head *bh)
401 {
402         bh->b_end_io = end_buffer_async_read;
403         set_buffer_async_read(bh);
404 }
405
406 static void mark_buffer_async_write_endio(struct buffer_head *bh,
407                                           bh_end_io_t *handler)
408 {
409         bh->b_end_io = handler;
410         set_buffer_async_write(bh);
411 }
412
413 void mark_buffer_async_write(struct buffer_head *bh)
414 {
415         mark_buffer_async_write_endio(bh, end_buffer_async_write);
416 }
417 EXPORT_SYMBOL(mark_buffer_async_write);
418
419
420 /*
421  * fs/buffer.c contains helper functions for buffer-backed address space's
422  * fsync functions.  A common requirement for buffer-based filesystems is
423  * that certain data from the backing blockdev needs to be written out for
424  * a successful fsync().  For example, ext2 indirect blocks need to be
425  * written back and waited upon before fsync() returns.
426  *
427  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
428  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
429  * management of a list of dependent buffers at ->i_mapping->private_list.
430  *
431  * Locking is a little subtle: try_to_free_buffers() will remove buffers
432  * from their controlling inode's queue when they are being freed.  But
433  * try_to_free_buffers() will be operating against the *blockdev* mapping
434  * at the time, not against the S_ISREG file which depends on those buffers.
435  * So the locking for private_list is via the private_lock in the address_space
436  * which backs the buffers.  Which is different from the address_space 
437  * against which the buffers are listed.  So for a particular address_space,
438  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
439  * mapping->private_list will always be protected by the backing blockdev's
440  * ->private_lock.
441  *
442  * Which introduces a requirement: all buffers on an address_space's
443  * ->private_list must be from the same address_space: the blockdev's.
444  *
445  * address_spaces which do not place buffers at ->private_list via these
446  * utility functions are free to use private_lock and private_list for
447  * whatever they want.  The only requirement is that list_empty(private_list)
448  * be true at clear_inode() time.
449  *
450  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
451  * filesystems should do that.  invalidate_inode_buffers() should just go
452  * BUG_ON(!list_empty).
453  *
454  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
455  * take an address_space, not an inode.  And it should be called
456  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
457  * queued up.
458  *
459  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
460  * list if it is already on a list.  Because if the buffer is on a list,
461  * it *must* already be on the right one.  If not, the filesystem is being
462  * silly.  This will save a ton of locking.  But first we have to ensure
463  * that buffers are taken *off* the old inode's list when they are freed
464  * (presumably in truncate).  That requires careful auditing of all
465  * filesystems (do it inside bforget()).  It could also be done by bringing
466  * b_inode back.
467  */
468
469 /*
470  * The buffer's backing address_space's private_lock must be held
471  */
472 static void __remove_assoc_queue(struct buffer_head *bh)
473 {
474         list_del_init(&bh->b_assoc_buffers);
475         WARN_ON(!bh->b_assoc_map);
476         if (buffer_write_io_error(bh))
477                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
478         bh->b_assoc_map = NULL;
479 }
480
481 int inode_has_buffers(struct inode *inode)
482 {
483         return !list_empty(&inode->i_data.private_list);
484 }
485
486 /*
487  * osync is designed to support O_SYNC io.  It waits synchronously for
488  * all already-submitted IO to complete, but does not queue any new
489  * writes to the disk.
490  *
491  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
492  * you dirty the buffers, and then use osync_inode_buffers to wait for
493  * completion.  Any other dirty buffers which are not yet queued for
494  * write will not be flushed to disk by the osync.
495  */
496 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
497 {
498         struct buffer_head *bh;
499         struct list_head *p;
500         int err = 0;
501
502         spin_lock(lock);
503 repeat:
504         list_for_each_prev(p, list) {
505                 bh = BH_ENTRY(p);
506                 if (buffer_locked(bh)) {
507                         get_bh(bh);
508                         spin_unlock(lock);
509                         wait_on_buffer(bh);
510                         if (!buffer_uptodate(bh))
511                                 err = -EIO;
512                         brelse(bh);
513                         spin_lock(lock);
514                         goto repeat;
515                 }
516         }
517         spin_unlock(lock);
518         return err;
519 }
520
521 static void do_thaw_one(struct super_block *sb, void *unused)
522 {
523         char b[BDEVNAME_SIZE];
524         while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
525                 printk(KERN_WARNING "Emergency Thaw on %s\n",
526                        bdevname(sb->s_bdev, b));
527 }
528
529 static void do_thaw_all(struct work_struct *work)
530 {
531         iterate_supers(do_thaw_one, NULL);
532         kfree(work);
533         printk(KERN_WARNING "Emergency Thaw complete\n");
534 }
535
536 /**
537  * emergency_thaw_all -- forcibly thaw every frozen filesystem
538  *
539  * Used for emergency unfreeze of all filesystems via SysRq
540  */
541 void emergency_thaw_all(void)
542 {
543         struct work_struct *work;
544
545         work = kmalloc(sizeof(*work), GFP_ATOMIC);
546         if (work) {
547                 INIT_WORK(work, do_thaw_all);
548                 schedule_work(work);
549         }
550 }
551
552 /**
553  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
554  * @mapping: the mapping which wants those buffers written
555  *
556  * Starts I/O against the buffers at mapping->private_list, and waits upon
557  * that I/O.
558  *
559  * Basically, this is a convenience function for fsync().
560  * @mapping is a file or directory which needs those buffers to be written for
561  * a successful fsync().
562  */
563 int sync_mapping_buffers(struct address_space *mapping)
564 {
565         struct address_space *buffer_mapping = mapping->private_data;
566
567         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
568                 return 0;
569
570         return fsync_buffers_list(&buffer_mapping->private_lock,
571                                         &mapping->private_list);
572 }
573 EXPORT_SYMBOL(sync_mapping_buffers);
574
575 /*
576  * Called when we've recently written block `bblock', and it is known that
577  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
578  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
579  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
580  */
581 void write_boundary_block(struct block_device *bdev,
582                         sector_t bblock, unsigned blocksize)
583 {
584         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
585         if (bh) {
586                 if (buffer_dirty(bh))
587                         ll_rw_block(WRITE, 1, &bh);
588                 put_bh(bh);
589         }
590 }
591
592 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
593 {
594         struct address_space *mapping = inode->i_mapping;
595         struct address_space *buffer_mapping = bh->b_page->mapping;
596
597         mark_buffer_dirty(bh);
598         if (!mapping->private_data) {
599                 mapping->private_data = buffer_mapping;
600         } else {
601                 BUG_ON(mapping->private_data != buffer_mapping);
602         }
603         if (!bh->b_assoc_map) {
604                 spin_lock(&buffer_mapping->private_lock);
605                 list_move_tail(&bh->b_assoc_buffers,
606                                 &mapping->private_list);
607                 bh->b_assoc_map = mapping;
608                 spin_unlock(&buffer_mapping->private_lock);
609         }
610 }
611 EXPORT_SYMBOL(mark_buffer_dirty_inode);
612
613 /*
614  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
615  * dirty.
616  *
617  * If warn is true, then emit a warning if the page is not uptodate and has
618  * not been truncated.
619  */
620 static void __set_page_dirty(struct page *page,
621                 struct address_space *mapping, int warn)
622 {
623         unsigned long flags;
624
625         spin_lock_irqsave(&mapping->tree_lock, flags);
626         if (page->mapping) {    /* Race with truncate? */
627                 WARN_ON_ONCE(warn && !PageUptodate(page));
628                 account_page_dirtied(page, mapping);
629                 radix_tree_tag_set(&mapping->page_tree,
630                                 page_index(page), PAGECACHE_TAG_DIRTY);
631         }
632         spin_unlock_irqrestore(&mapping->tree_lock, flags);
633         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
634 }
635
636 /*
637  * Add a page to the dirty page list.
638  *
639  * It is a sad fact of life that this function is called from several places
640  * deeply under spinlocking.  It may not sleep.
641  *
642  * If the page has buffers, the uptodate buffers are set dirty, to preserve
643  * dirty-state coherency between the page and the buffers.  It the page does
644  * not have buffers then when they are later attached they will all be set
645  * dirty.
646  *
647  * The buffers are dirtied before the page is dirtied.  There's a small race
648  * window in which a writepage caller may see the page cleanness but not the
649  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
650  * before the buffers, a concurrent writepage caller could clear the page dirty
651  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
652  * page on the dirty page list.
653  *
654  * We use private_lock to lock against try_to_free_buffers while using the
655  * page's buffer list.  Also use this to protect against clean buffers being
656  * added to the page after it was set dirty.
657  *
658  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
659  * address_space though.
660  */
661 int __set_page_dirty_buffers(struct page *page)
662 {
663         int newly_dirty;
664         struct address_space *mapping = page_mapping(page);
665
666         if (unlikely(!mapping))
667                 return !TestSetPageDirty(page);
668
669         spin_lock(&mapping->private_lock);
670         if (page_has_buffers(page)) {
671                 struct buffer_head *head = page_buffers(page);
672                 struct buffer_head *bh = head;
673
674                 do {
675                         set_buffer_dirty(bh);
676                         bh = bh->b_this_page;
677                 } while (bh != head);
678         }
679         newly_dirty = !TestSetPageDirty(page);
680         spin_unlock(&mapping->private_lock);
681
682         if (newly_dirty)
683                 __set_page_dirty(page, mapping, 1);
684         return newly_dirty;
685 }
686 EXPORT_SYMBOL(__set_page_dirty_buffers);
687
688 /*
689  * Write out and wait upon a list of buffers.
690  *
691  * We have conflicting pressures: we want to make sure that all
692  * initially dirty buffers get waited on, but that any subsequently
693  * dirtied buffers don't.  After all, we don't want fsync to last
694  * forever if somebody is actively writing to the file.
695  *
696  * Do this in two main stages: first we copy dirty buffers to a
697  * temporary inode list, queueing the writes as we go.  Then we clean
698  * up, waiting for those writes to complete.
699  * 
700  * During this second stage, any subsequent updates to the file may end
701  * up refiling the buffer on the original inode's dirty list again, so
702  * there is a chance we will end up with a buffer queued for write but
703  * not yet completed on that list.  So, as a final cleanup we go through
704  * the osync code to catch these locked, dirty buffers without requeuing
705  * any newly dirty buffers for write.
706  */
707 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
708 {
709         struct buffer_head *bh;
710         struct list_head tmp;
711         struct address_space *mapping;
712         int err = 0, err2;
713         struct blk_plug plug;
714
715         INIT_LIST_HEAD(&tmp);
716         blk_start_plug(&plug);
717
718         spin_lock(lock);
719         while (!list_empty(list)) {
720                 bh = BH_ENTRY(list->next);
721                 mapping = bh->b_assoc_map;
722                 __remove_assoc_queue(bh);
723                 /* Avoid race with mark_buffer_dirty_inode() which does
724                  * a lockless check and we rely on seeing the dirty bit */
725                 smp_mb();
726                 if (buffer_dirty(bh) || buffer_locked(bh)) {
727                         list_add(&bh->b_assoc_buffers, &tmp);
728                         bh->b_assoc_map = mapping;
729                         if (buffer_dirty(bh)) {
730                                 get_bh(bh);
731                                 spin_unlock(lock);
732                                 /*
733                                  * Ensure any pending I/O completes so that
734                                  * write_dirty_buffer() actually writes the
735                                  * current contents - it is a noop if I/O is
736                                  * still in flight on potentially older
737                                  * contents.
738                                  */
739                                 write_dirty_buffer(bh, WRITE_SYNC);
740
741                                 /*
742                                  * Kick off IO for the previous mapping. Note
743                                  * that we will not run the very last mapping,
744                                  * wait_on_buffer() will do that for us
745                                  * through sync_buffer().
746                                  */
747                                 brelse(bh);
748                                 spin_lock(lock);
749                         }
750                 }
751         }
752
753         spin_unlock(lock);
754         blk_finish_plug(&plug);
755         spin_lock(lock);
756
757         while (!list_empty(&tmp)) {
758                 bh = BH_ENTRY(tmp.prev);
759                 get_bh(bh);
760                 mapping = bh->b_assoc_map;
761                 __remove_assoc_queue(bh);
762                 /* Avoid race with mark_buffer_dirty_inode() which does
763                  * a lockless check and we rely on seeing the dirty bit */
764                 smp_mb();
765                 if (buffer_dirty(bh)) {
766                         list_add(&bh->b_assoc_buffers,
767                                  &mapping->private_list);
768                         bh->b_assoc_map = mapping;
769                 }
770                 spin_unlock(lock);
771                 wait_on_buffer(bh);
772                 if (!buffer_uptodate(bh))
773                         err = -EIO;
774                 brelse(bh);
775                 spin_lock(lock);
776         }
777         
778         spin_unlock(lock);
779         err2 = osync_buffers_list(lock, list);
780         if (err)
781                 return err;
782         else
783                 return err2;
784 }
785
786 /*
787  * Invalidate any and all dirty buffers on a given inode.  We are
788  * probably unmounting the fs, but that doesn't mean we have already
789  * done a sync().  Just drop the buffers from the inode list.
790  *
791  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
792  * assumes that all the buffers are against the blockdev.  Not true
793  * for reiserfs.
794  */
795 void invalidate_inode_buffers(struct inode *inode)
796 {
797         if (inode_has_buffers(inode)) {
798                 struct address_space *mapping = &inode->i_data;
799                 struct list_head *list = &mapping->private_list;
800                 struct address_space *buffer_mapping = mapping->private_data;
801
802                 spin_lock(&buffer_mapping->private_lock);
803                 while (!list_empty(list))
804                         __remove_assoc_queue(BH_ENTRY(list->next));
805                 spin_unlock(&buffer_mapping->private_lock);
806         }
807 }
808 EXPORT_SYMBOL(invalidate_inode_buffers);
809
810 /*
811  * Remove any clean buffers from the inode's buffer list.  This is called
812  * when we're trying to free the inode itself.  Those buffers can pin it.
813  *
814  * Returns true if all buffers were removed.
815  */
816 int remove_inode_buffers(struct inode *inode)
817 {
818         int ret = 1;
819
820         if (inode_has_buffers(inode)) {
821                 struct address_space *mapping = &inode->i_data;
822                 struct list_head *list = &mapping->private_list;
823                 struct address_space *buffer_mapping = mapping->private_data;
824
825                 spin_lock(&buffer_mapping->private_lock);
826                 while (!list_empty(list)) {
827                         struct buffer_head *bh = BH_ENTRY(list->next);
828                         if (buffer_dirty(bh)) {
829                                 ret = 0;
830                                 break;
831                         }
832                         __remove_assoc_queue(bh);
833                 }
834                 spin_unlock(&buffer_mapping->private_lock);
835         }
836         return ret;
837 }
838
839 /*
840  * Create the appropriate buffers when given a page for data area and
841  * the size of each buffer.. Use the bh->b_this_page linked list to
842  * follow the buffers created.  Return NULL if unable to create more
843  * buffers.
844  *
845  * The retry flag is used to differentiate async IO (paging, swapping)
846  * which may not fail from ordinary buffer allocations.
847  */
848 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
849                 int retry)
850 {
851         struct buffer_head *bh, *head;
852         long offset;
853
854 try_again:
855         head = NULL;
856         offset = PAGE_SIZE;
857         while ((offset -= size) >= 0) {
858                 bh = alloc_buffer_head(GFP_NOFS);
859                 if (!bh)
860                         goto no_grow;
861
862                 bh->b_this_page = head;
863                 bh->b_blocknr = -1;
864                 head = bh;
865
866                 bh->b_size = size;
867
868                 /* Link the buffer to its page */
869                 set_bh_page(bh, page, offset);
870         }
871         return head;
872 /*
873  * In case anything failed, we just free everything we got.
874  */
875 no_grow:
876         if (head) {
877                 do {
878                         bh = head;
879                         head = head->b_this_page;
880                         free_buffer_head(bh);
881                 } while (head);
882         }
883
884         /*
885          * Return failure for non-async IO requests.  Async IO requests
886          * are not allowed to fail, so we have to wait until buffer heads
887          * become available.  But we don't want tasks sleeping with 
888          * partially complete buffers, so all were released above.
889          */
890         if (!retry)
891                 return NULL;
892
893         /* We're _really_ low on memory. Now we just
894          * wait for old buffer heads to become free due to
895          * finishing IO.  Since this is an async request and
896          * the reserve list is empty, we're sure there are 
897          * async buffer heads in use.
898          */
899         free_more_memory();
900         goto try_again;
901 }
902 EXPORT_SYMBOL_GPL(alloc_page_buffers);
903
904 static inline void
905 link_dev_buffers(struct page *page, struct buffer_head *head)
906 {
907         struct buffer_head *bh, *tail;
908
909         bh = head;
910         do {
911                 tail = bh;
912                 bh = bh->b_this_page;
913         } while (bh);
914         tail->b_this_page = head;
915         attach_page_buffers(page, head);
916 }
917
918 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
919 {
920         sector_t retval = ~((sector_t)0);
921         loff_t sz = i_size_read(bdev->bd_inode);
922
923         if (sz) {
924                 unsigned int sizebits = blksize_bits(size);
925                 retval = (sz >> sizebits);
926         }
927         return retval;
928 }
929
930 /*
931  * Initialise the state of a blockdev page's buffers.
932  */ 
933 static sector_t
934 init_page_buffers(struct page *page, struct block_device *bdev,
935                         sector_t block, int size)
936 {
937         struct buffer_head *head = page_buffers(page);
938         struct buffer_head *bh = head;
939         int uptodate = PageUptodate(page);
940         sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
941
942         do {
943                 if (!buffer_mapped(bh)) {
944                         init_buffer(bh, NULL, NULL);
945                         bh->b_bdev = bdev;
946                         bh->b_blocknr = block;
947                         if (uptodate)
948                                 set_buffer_uptodate(bh);
949                         if (block < end_block)
950                                 set_buffer_mapped(bh);
951                 }
952                 block++;
953                 bh = bh->b_this_page;
954         } while (bh != head);
955
956         /*
957          * Caller needs to validate requested block against end of device.
958          */
959         return end_block;
960 }
961
962 /*
963  * Create the page-cache page that contains the requested block.
964  *
965  * This is used purely for blockdev mappings.
966  */
967 static int
968 grow_dev_page(struct block_device *bdev, sector_t block,
969                 pgoff_t index, int size, int sizebits)
970 {
971         struct inode *inode = bdev->bd_inode;
972         struct page *page;
973         struct buffer_head *bh;
974         sector_t end_block;
975         int ret = 0;            /* Will call free_more_memory() */
976
977         page = find_or_create_page(inode->i_mapping, index,
978                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
979         if (!page)
980                 return ret;
981
982         BUG_ON(!PageLocked(page));
983
984         if (page_has_buffers(page)) {
985                 bh = page_buffers(page);
986                 if (bh->b_size == size) {
987                         end_block = init_page_buffers(page, bdev,
988                                                 index << sizebits, size);
989                         goto done;
990                 }
991                 if (!try_to_free_buffers(page))
992                         goto failed;
993         }
994
995         /*
996          * Allocate some buffers for this page
997          */
998         bh = alloc_page_buffers(page, size, 0);
999         if (!bh)
1000                 goto failed;
1001
1002         /*
1003          * Link the page to the buffers and initialise them.  Take the
1004          * lock to be atomic wrt __find_get_block(), which does not
1005          * run under the page lock.
1006          */
1007         spin_lock(&inode->i_mapping->private_lock);
1008         link_dev_buffers(page, bh);
1009         end_block = init_page_buffers(page, bdev, index << sizebits, size);
1010         spin_unlock(&inode->i_mapping->private_lock);
1011 done:
1012         ret = (block < end_block) ? 1 : -ENXIO;
1013 failed:
1014         unlock_page(page);
1015         page_cache_release(page);
1016         return ret;
1017 }
1018
1019 /*
1020  * Create buffers for the specified block device block's page.  If
1021  * that page was dirty, the buffers are set dirty also.
1022  */
1023 static int
1024 grow_buffers(struct block_device *bdev, sector_t block, int size)
1025 {
1026         pgoff_t index;
1027         int sizebits;
1028
1029         sizebits = -1;
1030         do {
1031                 sizebits++;
1032         } while ((size << sizebits) < PAGE_SIZE);
1033
1034         index = block >> sizebits;
1035
1036         /*
1037          * Check for a block which wants to lie outside our maximum possible
1038          * pagecache index.  (this comparison is done using sector_t types).
1039          */
1040         if (unlikely(index != block >> sizebits)) {
1041                 char b[BDEVNAME_SIZE];
1042
1043                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1044                         "device %s\n",
1045                         __func__, (unsigned long long)block,
1046                         bdevname(bdev, b));
1047                 return -EIO;
1048         }
1049
1050         /* Create a page with the proper size buffers.. */
1051         return grow_dev_page(bdev, block, index, size, sizebits);
1052 }
1053
1054 static struct buffer_head *
1055 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1056 {
1057         /* Size must be multiple of hard sectorsize */
1058         if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1059                         (size < 512 || size > PAGE_SIZE))) {
1060                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1061                                         size);
1062                 printk(KERN_ERR "logical block size: %d\n",
1063                                         bdev_logical_block_size(bdev));
1064
1065                 dump_stack();
1066                 return NULL;
1067         }
1068
1069         for (;;) {
1070                 struct buffer_head *bh;
1071                 int ret;
1072
1073                 bh = __find_get_block(bdev, block, size);
1074                 if (bh)
1075                         return bh;
1076
1077                 ret = grow_buffers(bdev, block, size);
1078                 if (ret < 0)
1079                         return NULL;
1080                 if (ret == 0)
1081                         free_more_memory();
1082         }
1083 }
1084
1085 /*
1086  * The relationship between dirty buffers and dirty pages:
1087  *
1088  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1089  * the page is tagged dirty in its radix tree.
1090  *
1091  * At all times, the dirtiness of the buffers represents the dirtiness of
1092  * subsections of the page.  If the page has buffers, the page dirty bit is
1093  * merely a hint about the true dirty state.
1094  *
1095  * When a page is set dirty in its entirety, all its buffers are marked dirty
1096  * (if the page has buffers).
1097  *
1098  * When a buffer is marked dirty, its page is dirtied, but the page's other
1099  * buffers are not.
1100  *
1101  * Also.  When blockdev buffers are explicitly read with bread(), they
1102  * individually become uptodate.  But their backing page remains not
1103  * uptodate - even if all of its buffers are uptodate.  A subsequent
1104  * block_read_full_page() against that page will discover all the uptodate
1105  * buffers, will set the page uptodate and will perform no I/O.
1106  */
1107
1108 /**
1109  * mark_buffer_dirty - mark a buffer_head as needing writeout
1110  * @bh: the buffer_head to mark dirty
1111  *
1112  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1113  * backing page dirty, then tag the page as dirty in its address_space's radix
1114  * tree and then attach the address_space's inode to its superblock's dirty
1115  * inode list.
1116  *
1117  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1118  * mapping->tree_lock and mapping->host->i_lock.
1119  */
1120 void mark_buffer_dirty(struct buffer_head *bh)
1121 {
1122         WARN_ON_ONCE(!buffer_uptodate(bh));
1123
1124         trace_block_dirty_buffer(bh);
1125
1126         /*
1127          * Very *carefully* optimize the it-is-already-dirty case.
1128          *
1129          * Don't let the final "is it dirty" escape to before we
1130          * perhaps modified the buffer.
1131          */
1132         if (buffer_dirty(bh)) {
1133                 smp_mb();
1134                 if (buffer_dirty(bh))
1135                         return;
1136         }
1137
1138         if (!test_set_buffer_dirty(bh)) {
1139                 struct page *page = bh->b_page;
1140                 if (!TestSetPageDirty(page)) {
1141                         struct address_space *mapping = page_mapping(page);
1142                         if (mapping)
1143                                 __set_page_dirty(page, mapping, 0);
1144                 }
1145         }
1146 }
1147 EXPORT_SYMBOL(mark_buffer_dirty);
1148
1149 /*
1150  * Decrement a buffer_head's reference count.  If all buffers against a page
1151  * have zero reference count, are clean and unlocked, and if the page is clean
1152  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1153  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1154  * a page but it ends up not being freed, and buffers may later be reattached).
1155  */
1156 void __brelse(struct buffer_head * buf)
1157 {
1158         if (atomic_read(&buf->b_count)) {
1159                 put_bh(buf);
1160                 return;
1161         }
1162         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1163 }
1164 EXPORT_SYMBOL(__brelse);
1165
1166 /*
1167  * bforget() is like brelse(), except it discards any
1168  * potentially dirty data.
1169  */
1170 void __bforget(struct buffer_head *bh)
1171 {
1172         clear_buffer_dirty(bh);
1173         if (bh->b_assoc_map) {
1174                 struct address_space *buffer_mapping = bh->b_page->mapping;
1175
1176                 spin_lock(&buffer_mapping->private_lock);
1177                 list_del_init(&bh->b_assoc_buffers);
1178                 bh->b_assoc_map = NULL;
1179                 spin_unlock(&buffer_mapping->private_lock);
1180         }
1181         __brelse(bh);
1182 }
1183 EXPORT_SYMBOL(__bforget);
1184
1185 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1186 {
1187         lock_buffer(bh);
1188         if (buffer_uptodate(bh)) {
1189                 unlock_buffer(bh);
1190                 return bh;
1191         } else {
1192                 get_bh(bh);
1193                 bh->b_end_io = end_buffer_read_sync;
1194                 submit_bh(READ, bh);
1195                 wait_on_buffer(bh);
1196                 if (buffer_uptodate(bh))
1197                         return bh;
1198         }
1199         brelse(bh);
1200         return NULL;
1201 }
1202
1203 /*
1204  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1205  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1206  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1207  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1208  * CPU's LRUs at the same time.
1209  *
1210  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1211  * sb_find_get_block().
1212  *
1213  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1214  * a local interrupt disable for that.
1215  */
1216
1217 #define BH_LRU_SIZE     8
1218
1219 struct bh_lru {
1220         struct buffer_head *bhs[BH_LRU_SIZE];
1221 };
1222
1223 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1224
1225 #ifdef CONFIG_SMP
1226 #define bh_lru_lock()   local_irq_disable()
1227 #define bh_lru_unlock() local_irq_enable()
1228 #else
1229 #define bh_lru_lock()   preempt_disable()
1230 #define bh_lru_unlock() preempt_enable()
1231 #endif
1232
1233 static inline void check_irqs_on(void)
1234 {
1235 #ifdef irqs_disabled
1236         BUG_ON(irqs_disabled());
1237 #endif
1238 }
1239
1240 /*
1241  * The LRU management algorithm is dopey-but-simple.  Sorry.
1242  */
1243 static void bh_lru_install(struct buffer_head *bh)
1244 {
1245         struct buffer_head *evictee = NULL;
1246
1247         check_irqs_on();
1248         bh_lru_lock();
1249         if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1250                 struct buffer_head *bhs[BH_LRU_SIZE];
1251                 int in;
1252                 int out = 0;
1253
1254                 get_bh(bh);
1255                 bhs[out++] = bh;
1256                 for (in = 0; in < BH_LRU_SIZE; in++) {
1257                         struct buffer_head *bh2 =
1258                                 __this_cpu_read(bh_lrus.bhs[in]);
1259
1260                         if (bh2 == bh) {
1261                                 __brelse(bh2);
1262                         } else {
1263                                 if (out >= BH_LRU_SIZE) {
1264                                         BUG_ON(evictee != NULL);
1265                                         evictee = bh2;
1266                                 } else {
1267                                         bhs[out++] = bh2;
1268                                 }
1269                         }
1270                 }
1271                 while (out < BH_LRU_SIZE)
1272                         bhs[out++] = NULL;
1273                 memcpy(__this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1274         }
1275         bh_lru_unlock();
1276
1277         if (evictee)
1278                 __brelse(evictee);
1279 }
1280
1281 /*
1282  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1283  */
1284 static struct buffer_head *
1285 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1286 {
1287         struct buffer_head *ret = NULL;
1288         unsigned int i;
1289
1290         check_irqs_on();
1291         bh_lru_lock();
1292         for (i = 0; i < BH_LRU_SIZE; i++) {
1293                 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1294
1295                 if (bh && bh->b_bdev == bdev &&
1296                                 bh->b_blocknr == block && bh->b_size == size) {
1297                         if (i) {
1298                                 while (i) {
1299                                         __this_cpu_write(bh_lrus.bhs[i],
1300                                                 __this_cpu_read(bh_lrus.bhs[i - 1]));
1301                                         i--;
1302                                 }
1303                                 __this_cpu_write(bh_lrus.bhs[0], bh);
1304                         }
1305                         get_bh(bh);
1306                         ret = bh;
1307                         break;
1308                 }
1309         }
1310         bh_lru_unlock();
1311         return ret;
1312 }
1313
1314 /*
1315  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1316  * it in the LRU and mark it as accessed.  If it is not present then return
1317  * NULL
1318  */
1319 struct buffer_head *
1320 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1321 {
1322         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1323
1324         if (bh == NULL) {
1325                 bh = __find_get_block_slow(bdev, block);
1326                 if (bh)
1327                         bh_lru_install(bh);
1328         }
1329         if (bh)
1330                 touch_buffer(bh);
1331         return bh;
1332 }
1333 EXPORT_SYMBOL(__find_get_block);
1334
1335 /*
1336  * __getblk will locate (and, if necessary, create) the buffer_head
1337  * which corresponds to the passed block_device, block and size. The
1338  * returned buffer has its reference count incremented.
1339  *
1340  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1341  * attempt is failing.  FIXME, perhaps?
1342  */
1343 struct buffer_head *
1344 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1345 {
1346         struct buffer_head *bh = __find_get_block(bdev, block, size);
1347
1348         might_sleep();
1349         if (bh == NULL)
1350                 bh = __getblk_slow(bdev, block, size);
1351         return bh;
1352 }
1353 EXPORT_SYMBOL(__getblk);
1354
1355 /*
1356  * Do async read-ahead on a buffer..
1357  */
1358 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1359 {
1360         struct buffer_head *bh = __getblk(bdev, block, size);
1361         if (likely(bh)) {
1362                 ll_rw_block(READA, 1, &bh);
1363                 brelse(bh);
1364         }
1365 }
1366 EXPORT_SYMBOL(__breadahead);
1367
1368 /**
1369  *  __bread() - reads a specified block and returns the bh
1370  *  @bdev: the block_device to read from
1371  *  @block: number of block
1372  *  @size: size (in bytes) to read
1373  * 
1374  *  Reads a specified block, and returns buffer head that contains it.
1375  *  It returns NULL if the block was unreadable.
1376  */
1377 struct buffer_head *
1378 __bread(struct block_device *bdev, sector_t block, unsigned size)
1379 {
1380         struct buffer_head *bh = __getblk(bdev, block, size);
1381
1382         if (likely(bh) && !buffer_uptodate(bh))
1383                 bh = __bread_slow(bh);
1384         return bh;
1385 }
1386 EXPORT_SYMBOL(__bread);
1387
1388 /*
1389  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1390  * This doesn't race because it runs in each cpu either in irq
1391  * or with preempt disabled.
1392  */
1393 static void invalidate_bh_lru(void *arg)
1394 {
1395         struct bh_lru *b = &get_cpu_var(bh_lrus);
1396         int i;
1397
1398         for (i = 0; i < BH_LRU_SIZE; i++) {
1399                 brelse(b->bhs[i]);
1400                 b->bhs[i] = NULL;
1401         }
1402         put_cpu_var(bh_lrus);
1403 }
1404
1405 static bool has_bh_in_lru(int cpu, void *dummy)
1406 {
1407         struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1408         int i;
1409         
1410         for (i = 0; i < BH_LRU_SIZE; i++) {
1411                 if (b->bhs[i])
1412                         return 1;
1413         }
1414
1415         return 0;
1416 }
1417
1418 void invalidate_bh_lrus(void)
1419 {
1420         on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1421 }
1422 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1423
1424 void set_bh_page(struct buffer_head *bh,
1425                 struct page *page, unsigned long offset)
1426 {
1427         bh->b_page = page;
1428         BUG_ON(offset >= PAGE_SIZE);
1429         if (PageHighMem(page))
1430                 /*
1431                  * This catches illegal uses and preserves the offset:
1432                  */
1433                 bh->b_data = (char *)(0 + offset);
1434         else
1435                 bh->b_data = page_address(page) + offset;
1436 }
1437 EXPORT_SYMBOL(set_bh_page);
1438
1439 /*
1440  * Called when truncating a buffer on a page completely.
1441  */
1442 static void discard_buffer(struct buffer_head * bh)
1443 {
1444         lock_buffer(bh);
1445         clear_buffer_dirty(bh);
1446         bh->b_bdev = NULL;
1447         clear_buffer_mapped(bh);
1448         clear_buffer_req(bh);
1449         clear_buffer_new(bh);
1450         clear_buffer_delay(bh);
1451         clear_buffer_unwritten(bh);
1452         unlock_buffer(bh);
1453 }
1454
1455 /**
1456  * block_invalidatepage - invalidate part or all of a buffer-backed page
1457  *
1458  * @page: the page which is affected
1459  * @offset: the index of the truncation point
1460  *
1461  * block_invalidatepage() is called when all or part of the page has become
1462  * invalidated by a truncate operation.
1463  *
1464  * block_invalidatepage() does not have to release all buffers, but it must
1465  * ensure that no dirty buffer is left outside @offset and that no I/O
1466  * is underway against any of the blocks which are outside the truncation
1467  * point.  Because the caller is about to free (and possibly reuse) those
1468  * blocks on-disk.
1469  */
1470 void block_invalidatepage(struct page *page, unsigned long offset)
1471 {
1472         struct buffer_head *head, *bh, *next;
1473         unsigned int curr_off = 0;
1474
1475         BUG_ON(!PageLocked(page));
1476         if (!page_has_buffers(page))
1477                 goto out;
1478
1479         head = page_buffers(page);
1480         bh = head;
1481         do {
1482                 unsigned int next_off = curr_off + bh->b_size;
1483                 next = bh->b_this_page;
1484
1485                 /*
1486                  * is this block fully invalidated?
1487                  */
1488                 if (offset <= curr_off)
1489                         discard_buffer(bh);
1490                 curr_off = next_off;
1491                 bh = next;
1492         } while (bh != head);
1493
1494         /*
1495          * We release buffers only if the entire page is being invalidated.
1496          * The get_block cached value has been unconditionally invalidated,
1497          * so real IO is not possible anymore.
1498          */
1499         if (offset == 0)
1500                 try_to_release_page(page, 0);
1501 out:
1502         return;
1503 }
1504 EXPORT_SYMBOL(block_invalidatepage);
1505
1506 /*
1507  * We attach and possibly dirty the buffers atomically wrt
1508  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1509  * is already excluded via the page lock.
1510  */
1511 void create_empty_buffers(struct page *page,
1512                         unsigned long blocksize, unsigned long b_state)
1513 {
1514         struct buffer_head *bh, *head, *tail;
1515
1516         head = alloc_page_buffers(page, blocksize, 1);
1517         bh = head;
1518         do {
1519                 bh->b_state |= b_state;
1520                 tail = bh;
1521                 bh = bh->b_this_page;
1522         } while (bh);
1523         tail->b_this_page = head;
1524
1525         spin_lock(&page->mapping->private_lock);
1526         if (PageUptodate(page) || PageDirty(page)) {
1527                 bh = head;
1528                 do {
1529                         if (PageDirty(page))
1530                                 set_buffer_dirty(bh);
1531                         if (PageUptodate(page))
1532                                 set_buffer_uptodate(bh);
1533                         bh = bh->b_this_page;
1534                 } while (bh != head);
1535         }
1536         attach_page_buffers(page, head);
1537         spin_unlock(&page->mapping->private_lock);
1538 }
1539 EXPORT_SYMBOL(create_empty_buffers);
1540
1541 /*
1542  * We are taking a block for data and we don't want any output from any
1543  * buffer-cache aliases starting from return from that function and
1544  * until the moment when something will explicitly mark the buffer
1545  * dirty (hopefully that will not happen until we will free that block ;-)
1546  * We don't even need to mark it not-uptodate - nobody can expect
1547  * anything from a newly allocated buffer anyway. We used to used
1548  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1549  * don't want to mark the alias unmapped, for example - it would confuse
1550  * anyone who might pick it with bread() afterwards...
1551  *
1552  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1553  * be writeout I/O going on against recently-freed buffers.  We don't
1554  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1555  * only if we really need to.  That happens here.
1556  */
1557 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1558 {
1559         struct buffer_head *old_bh;
1560
1561         might_sleep();
1562
1563         old_bh = __find_get_block_slow(bdev, block);
1564         if (old_bh) {
1565                 clear_buffer_dirty(old_bh);
1566                 wait_on_buffer(old_bh);
1567                 clear_buffer_req(old_bh);
1568                 __brelse(old_bh);
1569         }
1570 }
1571 EXPORT_SYMBOL(unmap_underlying_metadata);
1572
1573 /*
1574  * Size is a power-of-two in the range 512..PAGE_SIZE,
1575  * and the case we care about most is PAGE_SIZE.
1576  *
1577  * So this *could* possibly be written with those
1578  * constraints in mind (relevant mostly if some
1579  * architecture has a slow bit-scan instruction)
1580  */
1581 static inline int block_size_bits(unsigned int blocksize)
1582 {
1583         return ilog2(blocksize);
1584 }
1585
1586 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1587 {
1588         BUG_ON(!PageLocked(page));
1589
1590         if (!page_has_buffers(page))
1591                 create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1592         return page_buffers(page);
1593 }
1594
1595 /*
1596  * NOTE! All mapped/uptodate combinations are valid:
1597  *
1598  *      Mapped  Uptodate        Meaning
1599  *
1600  *      No      No              "unknown" - must do get_block()
1601  *      No      Yes             "hole" - zero-filled
1602  *      Yes     No              "allocated" - allocated on disk, not read in
1603  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1604  *
1605  * "Dirty" is valid only with the last case (mapped+uptodate).
1606  */
1607
1608 /*
1609  * While block_write_full_page is writing back the dirty buffers under
1610  * the page lock, whoever dirtied the buffers may decide to clean them
1611  * again at any time.  We handle that by only looking at the buffer
1612  * state inside lock_buffer().
1613  *
1614  * If block_write_full_page() is called for regular writeback
1615  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1616  * locked buffer.   This only can happen if someone has written the buffer
1617  * directly, with submit_bh().  At the address_space level PageWriteback
1618  * prevents this contention from occurring.
1619  *
1620  * If block_write_full_page() is called with wbc->sync_mode ==
1621  * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1622  * causes the writes to be flagged as synchronous writes.
1623  */
1624 static int __block_write_full_page(struct inode *inode, struct page *page,
1625                         get_block_t *get_block, struct writeback_control *wbc,
1626                         bh_end_io_t *handler)
1627 {
1628         int err;
1629         sector_t block;
1630         sector_t last_block;
1631         struct buffer_head *bh, *head;
1632         unsigned int blocksize, bbits;
1633         int nr_underway = 0;
1634         int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1635                         WRITE_SYNC : WRITE);
1636
1637         head = create_page_buffers(page, inode,
1638                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1639
1640         /*
1641          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1642          * here, and the (potentially unmapped) buffers may become dirty at
1643          * any time.  If a buffer becomes dirty here after we've inspected it
1644          * then we just miss that fact, and the page stays dirty.
1645          *
1646          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1647          * handle that here by just cleaning them.
1648          */
1649
1650         bh = head;
1651         blocksize = bh->b_size;
1652         bbits = block_size_bits(blocksize);
1653
1654         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1655         last_block = (i_size_read(inode) - 1) >> bbits;
1656
1657         /*
1658          * Get all the dirty buffers mapped to disk addresses and
1659          * handle any aliases from the underlying blockdev's mapping.
1660          */
1661         do {
1662                 if (block > last_block) {
1663                         /*
1664                          * mapped buffers outside i_size will occur, because
1665                          * this page can be outside i_size when there is a
1666                          * truncate in progress.
1667                          */
1668                         /*
1669                          * The buffer was zeroed by block_write_full_page()
1670                          */
1671                         clear_buffer_dirty(bh);
1672                         set_buffer_uptodate(bh);
1673                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1674                            buffer_dirty(bh)) {
1675                         WARN_ON(bh->b_size != blocksize);
1676                         err = get_block(inode, block, bh, 1);
1677                         if (err)
1678                                 goto recover;
1679                         clear_buffer_delay(bh);
1680                         if (buffer_new(bh)) {
1681                                 /* blockdev mappings never come here */
1682                                 clear_buffer_new(bh);
1683                                 unmap_underlying_metadata(bh->b_bdev,
1684                                                         bh->b_blocknr);
1685                         }
1686                 }
1687                 bh = bh->b_this_page;
1688                 block++;
1689         } while (bh != head);
1690
1691         do {
1692                 if (!buffer_mapped(bh))
1693                         continue;
1694                 /*
1695                  * If it's a fully non-blocking write attempt and we cannot
1696                  * lock the buffer then redirty the page.  Note that this can
1697                  * potentially cause a busy-wait loop from writeback threads
1698                  * and kswapd activity, but those code paths have their own
1699                  * higher-level throttling.
1700                  */
1701                 if (wbc->sync_mode != WB_SYNC_NONE) {
1702                         lock_buffer(bh);
1703                 } else if (!trylock_buffer(bh)) {
1704                         redirty_page_for_writepage(wbc, page);
1705                         continue;
1706                 }
1707                 if (test_clear_buffer_dirty(bh)) {
1708                         mark_buffer_async_write_endio(bh, handler);
1709                 } else {
1710                         unlock_buffer(bh);
1711                 }
1712         } while ((bh = bh->b_this_page) != head);
1713
1714         /*
1715          * The page and its buffers are protected by PageWriteback(), so we can
1716          * drop the bh refcounts early.
1717          */
1718         BUG_ON(PageWriteback(page));
1719         set_page_writeback(page);
1720
1721         do {
1722                 struct buffer_head *next = bh->b_this_page;
1723                 if (buffer_async_write(bh)) {
1724                         submit_bh(write_op, bh);
1725                         nr_underway++;
1726                 }
1727                 bh = next;
1728         } while (bh != head);
1729         unlock_page(page);
1730
1731         err = 0;
1732 done:
1733         if (nr_underway == 0) {
1734                 /*
1735                  * The page was marked dirty, but the buffers were
1736                  * clean.  Someone wrote them back by hand with
1737                  * ll_rw_block/submit_bh.  A rare case.
1738                  */
1739                 end_page_writeback(page);
1740
1741                 /*
1742                  * The page and buffer_heads can be released at any time from
1743                  * here on.
1744                  */
1745         }
1746         return err;
1747
1748 recover:
1749         /*
1750          * ENOSPC, or some other error.  We may already have added some
1751          * blocks to the file, so we need to write these out to avoid
1752          * exposing stale data.
1753          * The page is currently locked and not marked for writeback
1754          */
1755         bh = head;
1756         /* Recovery: lock and submit the mapped buffers */
1757         do {
1758                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1759                     !buffer_delay(bh)) {
1760                         lock_buffer(bh);
1761                         mark_buffer_async_write_endio(bh, handler);
1762                 } else {
1763                         /*
1764                          * The buffer may have been set dirty during
1765                          * attachment to a dirty page.
1766                          */
1767                         clear_buffer_dirty(bh);
1768                 }
1769         } while ((bh = bh->b_this_page) != head);
1770         SetPageError(page);
1771         BUG_ON(PageWriteback(page));
1772         mapping_set_error(page->mapping, err);
1773         set_page_writeback(page);
1774         do {
1775                 struct buffer_head *next = bh->b_this_page;
1776                 if (buffer_async_write(bh)) {
1777                         clear_buffer_dirty(bh);
1778                         submit_bh(write_op, bh);
1779                         nr_underway++;
1780                 }
1781                 bh = next;
1782         } while (bh != head);
1783         unlock_page(page);
1784         goto done;
1785 }
1786
1787 /*
1788  * If a page has any new buffers, zero them out here, and mark them uptodate
1789  * and dirty so they'll be written out (in order to prevent uninitialised
1790  * block data from leaking). And clear the new bit.
1791  */
1792 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1793 {
1794         unsigned int block_start, block_end;
1795         struct buffer_head *head, *bh;
1796
1797         BUG_ON(!PageLocked(page));
1798         if (!page_has_buffers(page))
1799                 return;
1800
1801         bh = head = page_buffers(page);
1802         block_start = 0;
1803         do {
1804                 block_end = block_start + bh->b_size;
1805
1806                 if (buffer_new(bh)) {
1807                         if (block_end > from && block_start < to) {
1808                                 if (!PageUptodate(page)) {
1809                                         unsigned start, size;
1810
1811                                         start = max(from, block_start);
1812                                         size = min(to, block_end) - start;
1813
1814                                         zero_user(page, start, size);
1815                                         set_buffer_uptodate(bh);
1816                                 }
1817
1818                                 clear_buffer_new(bh);
1819                                 mark_buffer_dirty(bh);
1820                         }
1821                 }
1822
1823                 block_start = block_end;
1824                 bh = bh->b_this_page;
1825         } while (bh != head);
1826 }
1827 EXPORT_SYMBOL(page_zero_new_buffers);
1828
1829 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1830                 get_block_t *get_block)
1831 {
1832         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1833         unsigned to = from + len;
1834         struct inode *inode = page->mapping->host;
1835         unsigned block_start, block_end;
1836         sector_t block;
1837         int err = 0;
1838         unsigned blocksize, bbits;
1839         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1840
1841         BUG_ON(!PageLocked(page));
1842         BUG_ON(from > PAGE_CACHE_SIZE);
1843         BUG_ON(to > PAGE_CACHE_SIZE);
1844         BUG_ON(from > to);
1845
1846         head = create_page_buffers(page, inode, 0);
1847         blocksize = head->b_size;
1848         bbits = block_size_bits(blocksize);
1849
1850         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1851
1852         for(bh = head, block_start = 0; bh != head || !block_start;
1853             block++, block_start=block_end, bh = bh->b_this_page) {
1854                 block_end = block_start + blocksize;
1855                 if (block_end <= from || block_start >= to) {
1856                         if (PageUptodate(page)) {
1857                                 if (!buffer_uptodate(bh))
1858                                         set_buffer_uptodate(bh);
1859                         }
1860                         continue;
1861                 }
1862                 if (buffer_new(bh))
1863                         clear_buffer_new(bh);
1864                 if (!buffer_mapped(bh)) {
1865                         WARN_ON(bh->b_size != blocksize);
1866                         err = get_block(inode, block, bh, 1);
1867                         if (err)
1868                                 break;
1869                         if (buffer_new(bh)) {
1870                                 unmap_underlying_metadata(bh->b_bdev,
1871                                                         bh->b_blocknr);
1872                                 if (PageUptodate(page)) {
1873                                         clear_buffer_new(bh);
1874                                         set_buffer_uptodate(bh);
1875                                         mark_buffer_dirty(bh);
1876                                         continue;
1877                                 }
1878                                 if (block_end > to || block_start < from)
1879                                         zero_user_segments(page,
1880                                                 to, block_end,
1881                                                 block_start, from);
1882                                 continue;
1883                         }
1884                 }
1885                 if (PageUptodate(page)) {
1886                         if (!buffer_uptodate(bh))
1887                                 set_buffer_uptodate(bh);
1888                         continue; 
1889                 }
1890                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1891                     !buffer_unwritten(bh) &&
1892                      (block_start < from || block_end > to)) {
1893                         ll_rw_block(READ, 1, &bh);
1894                         *wait_bh++=bh;
1895                 }
1896         }
1897         /*
1898          * If we issued read requests - let them complete.
1899          */
1900         while(wait_bh > wait) {
1901                 wait_on_buffer(*--wait_bh);
1902                 if (!buffer_uptodate(*wait_bh))
1903                         err = -EIO;
1904         }
1905         if (unlikely(err))
1906                 page_zero_new_buffers(page, from, to);
1907         return err;
1908 }
1909 EXPORT_SYMBOL(__block_write_begin);
1910
1911 static int __block_commit_write(struct inode *inode, struct page *page,
1912                 unsigned from, unsigned to)
1913 {
1914         unsigned block_start, block_end;
1915         int partial = 0;
1916         unsigned blocksize;
1917         struct buffer_head *bh, *head;
1918
1919         bh = head = page_buffers(page);
1920         blocksize = bh->b_size;
1921
1922         block_start = 0;
1923         do {
1924                 block_end = block_start + blocksize;
1925                 if (block_end <= from || block_start >= to) {
1926                         if (!buffer_uptodate(bh))
1927                                 partial = 1;
1928                 } else {
1929                         set_buffer_uptodate(bh);
1930                         mark_buffer_dirty(bh);
1931                 }
1932                 clear_buffer_new(bh);
1933
1934                 block_start = block_end;
1935                 bh = bh->b_this_page;
1936         } while (bh != head);
1937
1938         /*
1939          * If this is a partial write which happened to make all buffers
1940          * uptodate then we can optimize away a bogus readpage() for
1941          * the next read(). Here we 'discover' whether the page went
1942          * uptodate as a result of this (potentially partial) write.
1943          */
1944         if (!partial)
1945                 SetPageUptodate(page);
1946         return 0;
1947 }
1948
1949 /*
1950  * block_write_begin takes care of the basic task of block allocation and
1951  * bringing partial write blocks uptodate first.
1952  *
1953  * The filesystem needs to handle block truncation upon failure.
1954  */
1955 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
1956                 unsigned flags, struct page **pagep, get_block_t *get_block)
1957 {
1958         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1959         struct page *page;
1960         int status;
1961
1962         page = grab_cache_page_write_begin(mapping, index, flags);
1963         if (!page)
1964                 return -ENOMEM;
1965
1966         status = __block_write_begin(page, pos, len, get_block);
1967         if (unlikely(status)) {
1968                 unlock_page(page);
1969                 page_cache_release(page);
1970                 page = NULL;
1971         }
1972
1973         *pagep = page;
1974         return status;
1975 }
1976 EXPORT_SYMBOL(block_write_begin);
1977
1978 int block_write_end(struct file *file, struct address_space *mapping,
1979                         loff_t pos, unsigned len, unsigned copied,
1980                         struct page *page, void *fsdata)
1981 {
1982         struct inode *inode = mapping->host;
1983         unsigned start;
1984
1985         start = pos & (PAGE_CACHE_SIZE - 1);
1986
1987         if (unlikely(copied < len)) {
1988                 /*
1989                  * The buffers that were written will now be uptodate, so we
1990                  * don't have to worry about a readpage reading them and
1991                  * overwriting a partial write. However if we have encountered
1992                  * a short write and only partially written into a buffer, it
1993                  * will not be marked uptodate, so a readpage might come in and
1994                  * destroy our partial write.
1995                  *
1996                  * Do the simplest thing, and just treat any short write to a
1997                  * non uptodate page as a zero-length write, and force the
1998                  * caller to redo the whole thing.
1999                  */
2000                 if (!PageUptodate(page))
2001                         copied = 0;
2002
2003                 page_zero_new_buffers(page, start+copied, start+len);
2004         }
2005         flush_dcache_page(page);
2006
2007         /* This could be a short (even 0-length) commit */
2008         __block_commit_write(inode, page, start, start+copied);
2009
2010         return copied;
2011 }
2012 EXPORT_SYMBOL(block_write_end);
2013
2014 int generic_write_end(struct file *file, struct address_space *mapping,
2015                         loff_t pos, unsigned len, unsigned copied,
2016                         struct page *page, void *fsdata)
2017 {
2018         struct inode *inode = mapping->host;
2019         int i_size_changed = 0;
2020
2021         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2022
2023         /*
2024          * No need to use i_size_read() here, the i_size
2025          * cannot change under us because we hold i_mutex.
2026          *
2027          * But it's important to update i_size while still holding page lock:
2028          * page writeout could otherwise come in and zero beyond i_size.
2029          */
2030         if (pos+copied > inode->i_size) {
2031                 i_size_write(inode, pos+copied);
2032                 i_size_changed = 1;
2033         }
2034
2035         unlock_page(page);
2036         page_cache_release(page);
2037
2038         /*
2039          * Don't mark the inode dirty under page lock. First, it unnecessarily
2040          * makes the holding time of page lock longer. Second, it forces lock
2041          * ordering of page lock and transaction start for journaling
2042          * filesystems.
2043          */
2044         if (i_size_changed)
2045                 mark_inode_dirty(inode);
2046
2047         return copied;
2048 }
2049 EXPORT_SYMBOL(generic_write_end);
2050
2051 /*
2052  * block_is_partially_uptodate checks whether buffers within a page are
2053  * uptodate or not.
2054  *
2055  * Returns true if all buffers which correspond to a file portion
2056  * we want to read are uptodate.
2057  */
2058 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2059                                         unsigned long from)
2060 {
2061         unsigned block_start, block_end, blocksize;
2062         unsigned to;
2063         struct buffer_head *bh, *head;
2064         int ret = 1;
2065
2066         if (!page_has_buffers(page))
2067                 return 0;
2068
2069         head = page_buffers(page);
2070         blocksize = head->b_size;
2071         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2072         to = from + to;
2073         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2074                 return 0;
2075
2076         bh = head;
2077         block_start = 0;
2078         do {
2079                 block_end = block_start + blocksize;
2080                 if (block_end > from && block_start < to) {
2081                         if (!buffer_uptodate(bh)) {
2082                                 ret = 0;
2083                                 break;
2084                         }
2085                         if (block_end >= to)
2086                                 break;
2087                 }
2088                 block_start = block_end;
2089                 bh = bh->b_this_page;
2090         } while (bh != head);
2091
2092         return ret;
2093 }
2094 EXPORT_SYMBOL(block_is_partially_uptodate);
2095
2096 /*
2097  * Generic "read page" function for block devices that have the normal
2098  * get_block functionality. This is most of the block device filesystems.
2099  * Reads the page asynchronously --- the unlock_buffer() and
2100  * set/clear_buffer_uptodate() functions propagate buffer state into the
2101  * page struct once IO has completed.
2102  */
2103 int block_read_full_page(struct page *page, get_block_t *get_block)
2104 {
2105         struct inode *inode = page->mapping->host;
2106         sector_t iblock, lblock;
2107         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2108         unsigned int blocksize, bbits;
2109         int nr, i;
2110         int fully_mapped = 1;
2111
2112         head = create_page_buffers(page, inode, 0);
2113         blocksize = head->b_size;
2114         bbits = block_size_bits(blocksize);
2115
2116         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
2117         lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2118         bh = head;
2119         nr = 0;
2120         i = 0;
2121
2122         do {
2123                 if (buffer_uptodate(bh))
2124                         continue;
2125
2126                 if (!buffer_mapped(bh)) {
2127                         int err = 0;
2128
2129                         fully_mapped = 0;
2130                         if (iblock < lblock) {
2131                                 WARN_ON(bh->b_size != blocksize);
2132                                 err = get_block(inode, iblock, bh, 0);
2133                                 if (err)
2134                                         SetPageError(page);
2135                         }
2136                         if (!buffer_mapped(bh)) {
2137                                 zero_user(page, i * blocksize, blocksize);
2138                                 if (!err)
2139                                         set_buffer_uptodate(bh);
2140                                 continue;
2141                         }
2142                         /*
2143                          * get_block() might have updated the buffer
2144                          * synchronously
2145                          */
2146                         if (buffer_uptodate(bh))
2147                                 continue;
2148                 }
2149                 arr[nr++] = bh;
2150         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2151
2152         if (fully_mapped)
2153                 SetPageMappedToDisk(page);
2154
2155         if (!nr) {
2156                 /*
2157                  * All buffers are uptodate - we can set the page uptodate
2158                  * as well. But not if get_block() returned an error.
2159                  */
2160                 if (!PageError(page))
2161                         SetPageUptodate(page);
2162                 unlock_page(page);
2163                 return 0;
2164         }
2165
2166         /* Stage two: lock the buffers */
2167         for (i = 0; i < nr; i++) {
2168                 bh = arr[i];
2169                 lock_buffer(bh);
2170                 mark_buffer_async_read(bh);
2171         }
2172
2173         /*
2174          * Stage 3: start the IO.  Check for uptodateness
2175          * inside the buffer lock in case another process reading
2176          * the underlying blockdev brought it uptodate (the sct fix).
2177          */
2178         for (i = 0; i < nr; i++) {
2179                 bh = arr[i];
2180                 if (buffer_uptodate(bh))
2181                         end_buffer_async_read(bh, 1);
2182                 else
2183                         submit_bh(READ, bh);
2184         }
2185         return 0;
2186 }
2187 EXPORT_SYMBOL(block_read_full_page);
2188
2189 /* utility function for filesystems that need to do work on expanding
2190  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2191  * deal with the hole.  
2192  */
2193 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2194 {
2195         struct address_space *mapping = inode->i_mapping;
2196         struct page *page;
2197         void *fsdata;
2198         int err;
2199
2200         err = inode_newsize_ok(inode, size);
2201         if (err)
2202                 goto out;
2203
2204         err = pagecache_write_begin(NULL, mapping, size, 0,
2205                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2206                                 &page, &fsdata);
2207         if (err)
2208                 goto out;
2209
2210         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2211         BUG_ON(err > 0);
2212
2213 out:
2214         return err;
2215 }
2216 EXPORT_SYMBOL(generic_cont_expand_simple);
2217
2218 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2219                             loff_t pos, loff_t *bytes)
2220 {
2221         struct inode *inode = mapping->host;
2222         unsigned blocksize = 1 << inode->i_blkbits;
2223         struct page *page;
2224         void *fsdata;
2225         pgoff_t index, curidx;
2226         loff_t curpos;
2227         unsigned zerofrom, offset, len;
2228         int err = 0;
2229
2230         index = pos >> PAGE_CACHE_SHIFT;
2231         offset = pos & ~PAGE_CACHE_MASK;
2232
2233         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2234                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2235                 if (zerofrom & (blocksize-1)) {
2236                         *bytes |= (blocksize-1);
2237                         (*bytes)++;
2238                 }
2239                 len = PAGE_CACHE_SIZE - zerofrom;
2240
2241                 err = pagecache_write_begin(file, mapping, curpos, len,
2242                                                 AOP_FLAG_UNINTERRUPTIBLE,
2243                                                 &page, &fsdata);
2244                 if (err)
2245                         goto out;
2246                 zero_user(page, zerofrom, len);
2247                 err = pagecache_write_end(file, mapping, curpos, len, len,
2248                                                 page, fsdata);
2249                 if (err < 0)
2250                         goto out;
2251                 BUG_ON(err != len);
2252                 err = 0;
2253
2254                 balance_dirty_pages_ratelimited(mapping);
2255         }
2256
2257         /* page covers the boundary, find the boundary offset */
2258         if (index == curidx) {
2259                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2260                 /* if we will expand the thing last block will be filled */
2261                 if (offset <= zerofrom) {
2262                         goto out;
2263                 }
2264                 if (zerofrom & (blocksize-1)) {
2265                         *bytes |= (blocksize-1);
2266                         (*bytes)++;
2267                 }
2268                 len = offset - zerofrom;
2269
2270                 err = pagecache_write_begin(file, mapping, curpos, len,
2271                                                 AOP_FLAG_UNINTERRUPTIBLE,
2272                                                 &page, &fsdata);
2273                 if (err)
2274                         goto out;
2275                 zero_user(page, zerofrom, len);
2276                 err = pagecache_write_end(file, mapping, curpos, len, len,
2277                                                 page, fsdata);
2278                 if (err < 0)
2279                         goto out;
2280                 BUG_ON(err != len);
2281                 err = 0;
2282         }
2283 out:
2284         return err;
2285 }
2286
2287 /*
2288  * For moronic filesystems that do not allow holes in file.
2289  * We may have to extend the file.
2290  */
2291 int cont_write_begin(struct file *file, struct address_space *mapping,
2292                         loff_t pos, unsigned len, unsigned flags,
2293                         struct page **pagep, void **fsdata,
2294                         get_block_t *get_block, loff_t *bytes)
2295 {
2296         struct inode *inode = mapping->host;
2297         unsigned blocksize = 1 << inode->i_blkbits;
2298         unsigned zerofrom;
2299         int err;
2300
2301         err = cont_expand_zero(file, mapping, pos, bytes);
2302         if (err)
2303                 return err;
2304
2305         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2306         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2307                 *bytes |= (blocksize-1);
2308                 (*bytes)++;
2309         }
2310
2311         return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2312 }
2313 EXPORT_SYMBOL(cont_write_begin);
2314
2315 int block_commit_write(struct page *page, unsigned from, unsigned to)
2316 {
2317         struct inode *inode = page->mapping->host;
2318         __block_commit_write(inode,page,from,to);
2319         return 0;
2320 }
2321 EXPORT_SYMBOL(block_commit_write);
2322
2323 /*
2324  * block_page_mkwrite() is not allowed to change the file size as it gets
2325  * called from a page fault handler when a page is first dirtied. Hence we must
2326  * be careful to check for EOF conditions here. We set the page up correctly
2327  * for a written page which means we get ENOSPC checking when writing into
2328  * holes and correct delalloc and unwritten extent mapping on filesystems that
2329  * support these features.
2330  *
2331  * We are not allowed to take the i_mutex here so we have to play games to
2332  * protect against truncate races as the page could now be beyond EOF.  Because
2333  * truncate writes the inode size before removing pages, once we have the
2334  * page lock we can determine safely if the page is beyond EOF. If it is not
2335  * beyond EOF, then the page is guaranteed safe against truncation until we
2336  * unlock the page.
2337  *
2338  * Direct callers of this function should protect against filesystem freezing
2339  * using sb_start_write() - sb_end_write() functions.
2340  */
2341 int __block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2342                          get_block_t get_block)
2343 {
2344         struct page *page = vmf->page;
2345         struct inode *inode = file_inode(vma->vm_file);
2346         unsigned long end;
2347         loff_t size;
2348         int ret;
2349
2350         lock_page(page);
2351         size = i_size_read(inode);
2352         if ((page->mapping != inode->i_mapping) ||
2353             (page_offset(page) > size)) {
2354                 /* We overload EFAULT to mean page got truncated */
2355                 ret = -EFAULT;
2356                 goto out_unlock;
2357         }
2358
2359         /* page is wholly or partially inside EOF */
2360         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2361                 end = size & ~PAGE_CACHE_MASK;
2362         else
2363                 end = PAGE_CACHE_SIZE;
2364
2365         ret = __block_write_begin(page, 0, end, get_block);
2366         if (!ret)
2367                 ret = block_commit_write(page, 0, end);
2368
2369         if (unlikely(ret < 0))
2370                 goto out_unlock;
2371         set_page_dirty(page);
2372         wait_for_stable_page(page);
2373         return 0;
2374 out_unlock:
2375         unlock_page(page);
2376         return ret;
2377 }
2378 EXPORT_SYMBOL(__block_page_mkwrite);
2379
2380 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2381                    get_block_t get_block)
2382 {
2383         int ret;
2384         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
2385
2386         sb_start_pagefault(sb);
2387
2388         /*
2389          * Update file times before taking page lock. We may end up failing the
2390          * fault so this update may be superfluous but who really cares...
2391          */
2392         file_update_time(vma->vm_file);
2393
2394         ret = __block_page_mkwrite(vma, vmf, get_block);
2395         sb_end_pagefault(sb);
2396         return block_page_mkwrite_return(ret);
2397 }
2398 EXPORT_SYMBOL(block_page_mkwrite);
2399
2400 /*
2401  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2402  * immediately, while under the page lock.  So it needs a special end_io
2403  * handler which does not touch the bh after unlocking it.
2404  */
2405 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2406 {
2407         __end_buffer_read_notouch(bh, uptodate);
2408 }
2409
2410 /*
2411  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2412  * the page (converting it to circular linked list and taking care of page
2413  * dirty races).
2414  */
2415 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2416 {
2417         struct buffer_head *bh;
2418
2419         BUG_ON(!PageLocked(page));
2420
2421         spin_lock(&page->mapping->private_lock);
2422         bh = head;
2423         do {
2424                 if (PageDirty(page))
2425                         set_buffer_dirty(bh);
2426                 if (!bh->b_this_page)
2427                         bh->b_this_page = head;
2428                 bh = bh->b_this_page;
2429         } while (bh != head);
2430         attach_page_buffers(page, head);
2431         spin_unlock(&page->mapping->private_lock);
2432 }
2433
2434 /*
2435  * On entry, the page is fully not uptodate.
2436  * On exit the page is fully uptodate in the areas outside (from,to)
2437  * The filesystem needs to handle block truncation upon failure.
2438  */
2439 int nobh_write_begin(struct address_space *mapping,
2440                         loff_t pos, unsigned len, unsigned flags,
2441                         struct page **pagep, void **fsdata,
2442                         get_block_t *get_block)
2443 {
2444         struct inode *inode = mapping->host;
2445         const unsigned blkbits = inode->i_blkbits;
2446         const unsigned blocksize = 1 << blkbits;
2447         struct buffer_head *head, *bh;
2448         struct page *page;
2449         pgoff_t index;
2450         unsigned from, to;
2451         unsigned block_in_page;
2452         unsigned block_start, block_end;
2453         sector_t block_in_file;
2454         int nr_reads = 0;
2455         int ret = 0;
2456         int is_mapped_to_disk = 1;
2457
2458         index = pos >> PAGE_CACHE_SHIFT;
2459         from = pos & (PAGE_CACHE_SIZE - 1);
2460         to = from + len;
2461
2462         page = grab_cache_page_write_begin(mapping, index, flags);
2463         if (!page)
2464                 return -ENOMEM;
2465         *pagep = page;
2466         *fsdata = NULL;
2467
2468         if (page_has_buffers(page)) {
2469                 ret = __block_write_begin(page, pos, len, get_block);
2470                 if (unlikely(ret))
2471                         goto out_release;
2472                 return ret;
2473         }
2474
2475         if (PageMappedToDisk(page))
2476                 return 0;
2477
2478         /*
2479          * Allocate buffers so that we can keep track of state, and potentially
2480          * attach them to the page if an error occurs. In the common case of
2481          * no error, they will just be freed again without ever being attached
2482          * to the page (which is all OK, because we're under the page lock).
2483          *
2484          * Be careful: the buffer linked list is a NULL terminated one, rather
2485          * than the circular one we're used to.
2486          */
2487         head = alloc_page_buffers(page, blocksize, 0);
2488         if (!head) {
2489                 ret = -ENOMEM;
2490                 goto out_release;
2491         }
2492
2493         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2494
2495         /*
2496          * We loop across all blocks in the page, whether or not they are
2497          * part of the affected region.  This is so we can discover if the
2498          * page is fully mapped-to-disk.
2499          */
2500         for (block_start = 0, block_in_page = 0, bh = head;
2501                   block_start < PAGE_CACHE_SIZE;
2502                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2503                 int create;
2504
2505                 block_end = block_start + blocksize;
2506                 bh->b_state = 0;
2507                 create = 1;
2508                 if (block_start >= to)
2509                         create = 0;
2510                 ret = get_block(inode, block_in_file + block_in_page,
2511                                         bh, create);
2512                 if (ret)
2513                         goto failed;
2514                 if (!buffer_mapped(bh))
2515                         is_mapped_to_disk = 0;
2516                 if (buffer_new(bh))
2517                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2518                 if (PageUptodate(page)) {
2519                         set_buffer_uptodate(bh);
2520                         continue;
2521                 }
2522                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2523                         zero_user_segments(page, block_start, from,
2524                                                         to, block_end);
2525                         continue;
2526                 }
2527                 if (buffer_uptodate(bh))
2528                         continue;       /* reiserfs does this */
2529                 if (block_start < from || block_end > to) {
2530                         lock_buffer(bh);
2531                         bh->b_end_io = end_buffer_read_nobh;
2532                         submit_bh(READ, bh);
2533                         nr_reads++;
2534                 }
2535         }
2536
2537         if (nr_reads) {
2538                 /*
2539                  * The page is locked, so these buffers are protected from
2540                  * any VM or truncate activity.  Hence we don't need to care
2541                  * for the buffer_head refcounts.
2542                  */
2543                 for (bh = head; bh; bh = bh->b_this_page) {
2544                         wait_on_buffer(bh);
2545                         if (!buffer_uptodate(bh))
2546                                 ret = -EIO;
2547                 }
2548                 if (ret)
2549                         goto failed;
2550         }
2551
2552         if (is_mapped_to_disk)
2553                 SetPageMappedToDisk(page);
2554
2555         *fsdata = head; /* to be released by nobh_write_end */
2556
2557         return 0;
2558
2559 failed:
2560         BUG_ON(!ret);
2561         /*
2562          * Error recovery is a bit difficult. We need to zero out blocks that
2563          * were newly allocated, and dirty them to ensure they get written out.
2564          * Buffers need to be attached to the page at this point, otherwise
2565          * the handling of potential IO errors during writeout would be hard
2566          * (could try doing synchronous writeout, but what if that fails too?)
2567          */
2568         attach_nobh_buffers(page, head);
2569         page_zero_new_buffers(page, from, to);
2570
2571 out_release:
2572         unlock_page(page);
2573         page_cache_release(page);
2574         *pagep = NULL;
2575
2576         return ret;
2577 }
2578 EXPORT_SYMBOL(nobh_write_begin);
2579
2580 int nobh_write_end(struct file *file, struct address_space *mapping,
2581                         loff_t pos, unsigned len, unsigned copied,
2582                         struct page *page, void *fsdata)
2583 {
2584         struct inode *inode = page->mapping->host;
2585         struct buffer_head *head = fsdata;
2586         struct buffer_head *bh;
2587         BUG_ON(fsdata != NULL && page_has_buffers(page));
2588
2589         if (unlikely(copied < len) && head)
2590                 attach_nobh_buffers(page, head);
2591         if (page_has_buffers(page))
2592                 return generic_write_end(file, mapping, pos, len,
2593                                         copied, page, fsdata);
2594
2595         SetPageUptodate(page);
2596         set_page_dirty(page);
2597         if (pos+copied > inode->i_size) {
2598                 i_size_write(inode, pos+copied);
2599                 mark_inode_dirty(inode);
2600         }
2601
2602         unlock_page(page);
2603         page_cache_release(page);
2604
2605         while (head) {
2606                 bh = head;
2607                 head = head->b_this_page;
2608                 free_buffer_head(bh);
2609         }
2610
2611         return copied;
2612 }
2613 EXPORT_SYMBOL(nobh_write_end);
2614
2615 /*
2616  * nobh_writepage() - based on block_full_write_page() except
2617  * that it tries to operate without attaching bufferheads to
2618  * the page.
2619  */
2620 int nobh_writepage(struct page *page, get_block_t *get_block,
2621                         struct writeback_control *wbc)
2622 {
2623         struct inode * const inode = page->mapping->host;
2624         loff_t i_size = i_size_read(inode);
2625         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2626         unsigned offset;
2627         int ret;
2628
2629         /* Is the page fully inside i_size? */
2630         if (page->index < end_index)
2631                 goto out;
2632
2633         /* Is the page fully outside i_size? (truncate in progress) */
2634         offset = i_size & (PAGE_CACHE_SIZE-1);
2635         if (page->index >= end_index+1 || !offset) {
2636                 /*
2637                  * The page may have dirty, unmapped buffers.  For example,
2638                  * they may have been added in ext3_writepage().  Make them
2639                  * freeable here, so the page does not leak.
2640                  */
2641 #if 0
2642                 /* Not really sure about this  - do we need this ? */
2643                 if (page->mapping->a_ops->invalidatepage)
2644                         page->mapping->a_ops->invalidatepage(page, offset);
2645 #endif
2646                 unlock_page(page);
2647                 return 0; /* don't care */
2648         }
2649
2650         /*
2651          * The page straddles i_size.  It must be zeroed out on each and every
2652          * writepage invocation because it may be mmapped.  "A file is mapped
2653          * in multiples of the page size.  For a file that is not a multiple of
2654          * the  page size, the remaining memory is zeroed when mapped, and
2655          * writes to that region are not written out to the file."
2656          */
2657         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2658 out:
2659         ret = mpage_writepage(page, get_block, wbc);
2660         if (ret == -EAGAIN)
2661                 ret = __block_write_full_page(inode, page, get_block, wbc,
2662                                               end_buffer_async_write);
2663         return ret;
2664 }
2665 EXPORT_SYMBOL(nobh_writepage);
2666
2667 int nobh_truncate_page(struct address_space *mapping,
2668                         loff_t from, get_block_t *get_block)
2669 {
2670         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2671         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2672         unsigned blocksize;
2673         sector_t iblock;
2674         unsigned length, pos;
2675         struct inode *inode = mapping->host;
2676         struct page *page;
2677         struct buffer_head map_bh;
2678         int err;
2679
2680         blocksize = 1 << inode->i_blkbits;
2681         length = offset & (blocksize - 1);
2682
2683         /* Block boundary? Nothing to do */
2684         if (!length)
2685                 return 0;
2686
2687         length = blocksize - length;
2688         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2689
2690         page = grab_cache_page(mapping, index);
2691         err = -ENOMEM;
2692         if (!page)
2693                 goto out;
2694
2695         if (page_has_buffers(page)) {
2696 has_buffers:
2697                 unlock_page(page);
2698                 page_cache_release(page);
2699                 return block_truncate_page(mapping, from, get_block);
2700         }
2701
2702         /* Find the buffer that contains "offset" */
2703         pos = blocksize;
2704         while (offset >= pos) {
2705                 iblock++;
2706                 pos += blocksize;
2707         }
2708
2709         map_bh.b_size = blocksize;
2710         map_bh.b_state = 0;
2711         err = get_block(inode, iblock, &map_bh, 0);
2712         if (err)
2713                 goto unlock;
2714         /* unmapped? It's a hole - nothing to do */
2715         if (!buffer_mapped(&map_bh))
2716                 goto unlock;
2717
2718         /* Ok, it's mapped. Make sure it's up-to-date */
2719         if (!PageUptodate(page)) {
2720                 err = mapping->a_ops->readpage(NULL, page);
2721                 if (err) {
2722                         page_cache_release(page);
2723                         goto out;
2724                 }
2725                 lock_page(page);
2726                 if (!PageUptodate(page)) {
2727                         err = -EIO;
2728                         goto unlock;
2729                 }
2730                 if (page_has_buffers(page))
2731                         goto has_buffers;
2732         }
2733         zero_user(page, offset, length);
2734         set_page_dirty(page);
2735         err = 0;
2736
2737 unlock:
2738         unlock_page(page);
2739         page_cache_release(page);
2740 out:
2741         return err;
2742 }
2743 EXPORT_SYMBOL(nobh_truncate_page);
2744
2745 int block_truncate_page(struct address_space *mapping,
2746                         loff_t from, get_block_t *get_block)
2747 {
2748         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2749         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2750         unsigned blocksize;
2751         sector_t iblock;
2752         unsigned length, pos;
2753         struct inode *inode = mapping->host;
2754         struct page *page;
2755         struct buffer_head *bh;
2756         int err;
2757
2758         blocksize = 1 << inode->i_blkbits;
2759         length = offset & (blocksize - 1);
2760
2761         /* Block boundary? Nothing to do */
2762         if (!length)
2763                 return 0;
2764
2765         length = blocksize - length;
2766         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2767         
2768         page = grab_cache_page(mapping, index);
2769         err = -ENOMEM;
2770         if (!page)
2771                 goto out;
2772
2773         if (!page_has_buffers(page))
2774                 create_empty_buffers(page, blocksize, 0);
2775
2776         /* Find the buffer that contains "offset" */
2777         bh = page_buffers(page);
2778         pos = blocksize;
2779         while (offset >= pos) {
2780                 bh = bh->b_this_page;
2781                 iblock++;
2782                 pos += blocksize;
2783         }
2784
2785         err = 0;
2786         if (!buffer_mapped(bh)) {
2787                 WARN_ON(bh->b_size != blocksize);
2788                 err = get_block(inode, iblock, bh, 0);
2789                 if (err)
2790                         goto unlock;
2791                 /* unmapped? It's a hole - nothing to do */
2792                 if (!buffer_mapped(bh))
2793                         goto unlock;
2794         }
2795
2796         /* Ok, it's mapped. Make sure it's up-to-date */
2797         if (PageUptodate(page))
2798                 set_buffer_uptodate(bh);
2799
2800         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2801                 err = -EIO;
2802                 ll_rw_block(READ, 1, &bh);
2803                 wait_on_buffer(bh);
2804                 /* Uhhuh. Read error. Complain and punt. */
2805                 if (!buffer_uptodate(bh))
2806                         goto unlock;
2807         }
2808
2809         zero_user(page, offset, length);
2810         mark_buffer_dirty(bh);
2811         err = 0;
2812
2813 unlock:
2814         unlock_page(page);
2815         page_cache_release(page);
2816 out:
2817         return err;
2818 }
2819 EXPORT_SYMBOL(block_truncate_page);
2820
2821 /*
2822  * The generic ->writepage function for buffer-backed address_spaces
2823  * this form passes in the end_io handler used to finish the IO.
2824  */
2825 int block_write_full_page_endio(struct page *page, get_block_t *get_block,
2826                         struct writeback_control *wbc, bh_end_io_t *handler)
2827 {
2828         struct inode * const inode = page->mapping->host;
2829         loff_t i_size = i_size_read(inode);
2830         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2831         unsigned offset;
2832
2833         /* Is the page fully inside i_size? */
2834         if (page->index < end_index)
2835                 return __block_write_full_page(inode, page, get_block, wbc,
2836                                                handler);
2837
2838         /* Is the page fully outside i_size? (truncate in progress) */
2839         offset = i_size & (PAGE_CACHE_SIZE-1);
2840         if (page->index >= end_index+1 || !offset) {
2841                 /*
2842                  * The page may have dirty, unmapped buffers.  For example,
2843                  * they may have been added in ext3_writepage().  Make them
2844                  * freeable here, so the page does not leak.
2845                  */
2846                 do_invalidatepage(page, 0);
2847                 unlock_page(page);
2848                 return 0; /* don't care */
2849         }
2850
2851         /*
2852          * The page straddles i_size.  It must be zeroed out on each and every
2853          * writepage invocation because it may be mmapped.  "A file is mapped
2854          * in multiples of the page size.  For a file that is not a multiple of
2855          * the  page size, the remaining memory is zeroed when mapped, and
2856          * writes to that region are not written out to the file."
2857          */
2858         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2859         return __block_write_full_page(inode, page, get_block, wbc, handler);
2860 }
2861 EXPORT_SYMBOL(block_write_full_page_endio);
2862
2863 /*
2864  * The generic ->writepage function for buffer-backed address_spaces
2865  */
2866 int block_write_full_page(struct page *page, get_block_t *get_block,
2867                         struct writeback_control *wbc)
2868 {
2869         return block_write_full_page_endio(page, get_block, wbc,
2870                                            end_buffer_async_write);
2871 }
2872 EXPORT_SYMBOL(block_write_full_page);
2873
2874 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2875                             get_block_t *get_block)
2876 {
2877         struct buffer_head tmp;
2878         struct inode *inode = mapping->host;
2879         tmp.b_state = 0;
2880         tmp.b_blocknr = 0;
2881         tmp.b_size = 1 << inode->i_blkbits;
2882         get_block(inode, block, &tmp, 0);
2883         return tmp.b_blocknr;
2884 }
2885 EXPORT_SYMBOL(generic_block_bmap);
2886
2887 static void end_bio_bh_io_sync(struct bio *bio, int err)
2888 {
2889         struct buffer_head *bh = bio->bi_private;
2890
2891         if (err == -EOPNOTSUPP) {
2892                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2893         }
2894
2895         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2896                 set_bit(BH_Quiet, &bh->b_state);
2897
2898         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2899         bio_put(bio);
2900 }
2901
2902 /*
2903  * This allows us to do IO even on the odd last sectors
2904  * of a device, even if the bh block size is some multiple
2905  * of the physical sector size.
2906  *
2907  * We'll just truncate the bio to the size of the device,
2908  * and clear the end of the buffer head manually.
2909  *
2910  * Truly out-of-range accesses will turn into actual IO
2911  * errors, this only handles the "we need to be able to
2912  * do IO at the final sector" case.
2913  */
2914 static void guard_bh_eod(int rw, struct bio *bio, struct buffer_head *bh)
2915 {
2916         sector_t maxsector;
2917         unsigned bytes;
2918
2919         maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
2920         if (!maxsector)
2921                 return;
2922
2923         /*
2924          * If the *whole* IO is past the end of the device,
2925          * let it through, and the IO layer will turn it into
2926          * an EIO.
2927          */
2928         if (unlikely(bio->bi_sector >= maxsector))
2929                 return;
2930
2931         maxsector -= bio->bi_sector;
2932         bytes = bio->bi_size;
2933         if (likely((bytes >> 9) <= maxsector))
2934                 return;
2935
2936         /* Uhhuh. We've got a bh that straddles the device size! */
2937         bytes = maxsector << 9;
2938
2939         /* Truncate the bio.. */
2940         bio->bi_size = bytes;
2941         bio->bi_io_vec[0].bv_len = bytes;
2942
2943         /* ..and clear the end of the buffer for reads */
2944         if ((rw & RW_MASK) == READ) {
2945                 void *kaddr = kmap_atomic(bh->b_page);
2946                 memset(kaddr + bh_offset(bh) + bytes, 0, bh->b_size - bytes);
2947                 kunmap_atomic(kaddr);
2948                 flush_dcache_page(bh->b_page);
2949         }
2950 }
2951
2952 int _submit_bh(int rw, struct buffer_head *bh, unsigned long bio_flags)
2953 {
2954         struct bio *bio;
2955         int ret = 0;
2956
2957         BUG_ON(!buffer_locked(bh));
2958         BUG_ON(!buffer_mapped(bh));
2959         BUG_ON(!bh->b_end_io);
2960         BUG_ON(buffer_delay(bh));
2961         BUG_ON(buffer_unwritten(bh));
2962
2963         /*
2964          * Only clear out a write error when rewriting
2965          */
2966         if (test_set_buffer_req(bh) && (rw & WRITE))
2967                 clear_buffer_write_io_error(bh);
2968
2969         /*
2970          * from here on down, it's all bio -- do the initial mapping,
2971          * submit_bio -> generic_make_request may further map this bio around
2972          */
2973         bio = bio_alloc(GFP_NOIO, 1);
2974
2975         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2976         bio->bi_bdev = bh->b_bdev;
2977         bio->bi_io_vec[0].bv_page = bh->b_page;
2978         bio->bi_io_vec[0].bv_len = bh->b_size;
2979         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2980
2981         bio->bi_vcnt = 1;
2982         bio->bi_size = bh->b_size;
2983
2984         bio->bi_end_io = end_bio_bh_io_sync;
2985         bio->bi_private = bh;
2986         bio->bi_flags |= bio_flags;
2987
2988         /* Take care of bh's that straddle the end of the device */
2989         guard_bh_eod(rw, bio, bh);
2990
2991         if (buffer_meta(bh))
2992                 rw |= REQ_META;
2993         if (buffer_prio(bh))
2994                 rw |= REQ_PRIO;
2995
2996         bio_get(bio);
2997         submit_bio(rw, bio);
2998
2999         if (bio_flagged(bio, BIO_EOPNOTSUPP))
3000                 ret = -EOPNOTSUPP;
3001
3002         bio_put(bio);
3003         return ret;
3004 }
3005 EXPORT_SYMBOL_GPL(_submit_bh);
3006
3007 int submit_bh(int rw, struct buffer_head *bh)
3008 {
3009         return _submit_bh(rw, bh, 0);
3010 }
3011 EXPORT_SYMBOL(submit_bh);
3012
3013 /**
3014  * ll_rw_block: low-level access to block devices (DEPRECATED)
3015  * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
3016  * @nr: number of &struct buffer_heads in the array
3017  * @bhs: array of pointers to &struct buffer_head
3018  *
3019  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3020  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
3021  * %READA option is described in the documentation for generic_make_request()
3022  * which ll_rw_block() calls.
3023  *
3024  * This function drops any buffer that it cannot get a lock on (with the
3025  * BH_Lock state bit), any buffer that appears to be clean when doing a write
3026  * request, and any buffer that appears to be up-to-date when doing read
3027  * request.  Further it marks as clean buffers that are processed for
3028  * writing (the buffer cache won't assume that they are actually clean
3029  * until the buffer gets unlocked).
3030  *
3031  * ll_rw_block sets b_end_io to simple completion handler that marks
3032  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
3033  * any waiters. 
3034  *
3035  * All of the buffers must be for the same device, and must also be a
3036  * multiple of the current approved size for the device.
3037  */
3038 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
3039 {
3040         int i;
3041
3042         for (i = 0; i < nr; i++) {
3043                 struct buffer_head *bh = bhs[i];
3044
3045                 if (!trylock_buffer(bh))
3046                         continue;
3047                 if (rw == WRITE) {
3048                         if (test_clear_buffer_dirty(bh)) {
3049                                 bh->b_end_io = end_buffer_write_sync;
3050                                 get_bh(bh);
3051                                 submit_bh(WRITE, bh);
3052                                 continue;
3053                         }
3054                 } else {
3055                         if (!buffer_uptodate(bh)) {
3056                                 bh->b_end_io = end_buffer_read_sync;
3057                                 get_bh(bh);
3058                                 submit_bh(rw, bh);
3059                                 continue;
3060                         }
3061                 }
3062                 unlock_buffer(bh);
3063         }
3064 }
3065 EXPORT_SYMBOL(ll_rw_block);
3066
3067 void write_dirty_buffer(struct buffer_head *bh, int rw)
3068 {
3069         lock_buffer(bh);
3070         if (!test_clear_buffer_dirty(bh)) {
3071                 unlock_buffer(bh);
3072                 return;
3073         }
3074         bh->b_end_io = end_buffer_write_sync;
3075         get_bh(bh);
3076         submit_bh(rw, bh);
3077 }
3078 EXPORT_SYMBOL(write_dirty_buffer);
3079
3080 /*
3081  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3082  * and then start new I/O and then wait upon it.  The caller must have a ref on
3083  * the buffer_head.
3084  */
3085 int __sync_dirty_buffer(struct buffer_head *bh, int rw)
3086 {
3087         int ret = 0;
3088
3089         WARN_ON(atomic_read(&bh->b_count) < 1);
3090         lock_buffer(bh);
3091         if (test_clear_buffer_dirty(bh)) {
3092                 get_bh(bh);
3093                 bh->b_end_io = end_buffer_write_sync;
3094                 ret = submit_bh(rw, bh);
3095                 wait_on_buffer(bh);
3096                 if (!ret && !buffer_uptodate(bh))
3097                         ret = -EIO;
3098         } else {
3099                 unlock_buffer(bh);
3100         }
3101         return ret;
3102 }
3103 EXPORT_SYMBOL(__sync_dirty_buffer);
3104
3105 int sync_dirty_buffer(struct buffer_head *bh)
3106 {
3107         return __sync_dirty_buffer(bh, WRITE_SYNC);
3108 }
3109 EXPORT_SYMBOL(sync_dirty_buffer);
3110
3111 /*
3112  * try_to_free_buffers() checks if all the buffers on this particular page
3113  * are unused, and releases them if so.
3114  *
3115  * Exclusion against try_to_free_buffers may be obtained by either
3116  * locking the page or by holding its mapping's private_lock.
3117  *
3118  * If the page is dirty but all the buffers are clean then we need to
3119  * be sure to mark the page clean as well.  This is because the page
3120  * may be against a block device, and a later reattachment of buffers
3121  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3122  * filesystem data on the same device.
3123  *
3124  * The same applies to regular filesystem pages: if all the buffers are
3125  * clean then we set the page clean and proceed.  To do that, we require
3126  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3127  * private_lock.
3128  *
3129  * try_to_free_buffers() is non-blocking.
3130  */
3131 static inline int buffer_busy(struct buffer_head *bh)
3132 {
3133         return atomic_read(&bh->b_count) |
3134                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3135 }
3136
3137 static int
3138 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3139 {
3140         struct buffer_head *head = page_buffers(page);
3141         struct buffer_head *bh;
3142
3143         bh = head;
3144         do {
3145                 if (buffer_write_io_error(bh) && page->mapping)
3146                         set_bit(AS_EIO, &page->mapping->flags);
3147                 if (buffer_busy(bh))
3148                         goto failed;
3149                 bh = bh->b_this_page;
3150         } while (bh != head);
3151
3152         do {
3153                 struct buffer_head *next = bh->b_this_page;
3154
3155                 if (bh->b_assoc_map)
3156                         __remove_assoc_queue(bh);
3157                 bh = next;
3158         } while (bh != head);
3159         *buffers_to_free = head;
3160         __clear_page_buffers(page);
3161         return 1;
3162 failed:
3163         return 0;
3164 }
3165
3166 int try_to_free_buffers(struct page *page)
3167 {
3168         struct address_space * const mapping = page->mapping;
3169         struct buffer_head *buffers_to_free = NULL;
3170         int ret = 0;
3171
3172         BUG_ON(!PageLocked(page));
3173         if (PageWriteback(page))
3174                 return 0;
3175
3176         if (mapping == NULL) {          /* can this still happen? */
3177                 ret = drop_buffers(page, &buffers_to_free);
3178                 goto out;
3179         }
3180
3181         spin_lock(&mapping->private_lock);
3182         ret = drop_buffers(page, &buffers_to_free);
3183
3184         /*
3185          * If the filesystem writes its buffers by hand (eg ext3)
3186          * then we can have clean buffers against a dirty page.  We
3187          * clean the page here; otherwise the VM will never notice
3188          * that the filesystem did any IO at all.
3189          *
3190          * Also, during truncate, discard_buffer will have marked all
3191          * the page's buffers clean.  We discover that here and clean
3192          * the page also.
3193          *
3194          * private_lock must be held over this entire operation in order
3195          * to synchronise against __set_page_dirty_buffers and prevent the
3196          * dirty bit from being lost.
3197          */
3198         if (ret)
3199                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3200         spin_unlock(&mapping->private_lock);
3201 out:
3202         if (buffers_to_free) {
3203                 struct buffer_head *bh = buffers_to_free;
3204
3205                 do {
3206                         struct buffer_head *next = bh->b_this_page;
3207                         free_buffer_head(bh);
3208                         bh = next;
3209                 } while (bh != buffers_to_free);
3210         }
3211         return ret;
3212 }
3213 EXPORT_SYMBOL(try_to_free_buffers);
3214
3215 /*
3216  * There are no bdflush tunables left.  But distributions are
3217  * still running obsolete flush daemons, so we terminate them here.
3218  *
3219  * Use of bdflush() is deprecated and will be removed in a future kernel.
3220  * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3221  */
3222 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3223 {
3224         static int msg_count;
3225
3226         if (!capable(CAP_SYS_ADMIN))
3227                 return -EPERM;
3228
3229         if (msg_count < 5) {
3230                 msg_count++;
3231                 printk(KERN_INFO
3232                         "warning: process `%s' used the obsolete bdflush"
3233                         " system call\n", current->comm);
3234                 printk(KERN_INFO "Fix your initscripts?\n");
3235         }
3236
3237         if (func == 1)
3238                 do_exit(0);
3239         return 0;
3240 }
3241
3242 /*
3243  * Buffer-head allocation
3244  */
3245 static struct kmem_cache *bh_cachep __read_mostly;
3246
3247 /*
3248  * Once the number of bh's in the machine exceeds this level, we start
3249  * stripping them in writeback.
3250  */
3251 static unsigned long max_buffer_heads;
3252
3253 int buffer_heads_over_limit;
3254
3255 struct bh_accounting {
3256         int nr;                 /* Number of live bh's */
3257         int ratelimit;          /* Limit cacheline bouncing */
3258 };
3259
3260 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3261
3262 static void recalc_bh_state(void)
3263 {
3264         int i;
3265         int tot = 0;
3266
3267         if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3268                 return;
3269         __this_cpu_write(bh_accounting.ratelimit, 0);
3270         for_each_online_cpu(i)
3271                 tot += per_cpu(bh_accounting, i).nr;
3272         buffer_heads_over_limit = (tot > max_buffer_heads);
3273 }
3274
3275 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3276 {
3277         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3278         if (ret) {
3279                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3280                 preempt_disable();
3281                 __this_cpu_inc(bh_accounting.nr);
3282                 recalc_bh_state();
3283                 preempt_enable();
3284         }
3285         return ret;
3286 }
3287 EXPORT_SYMBOL(alloc_buffer_head);
3288
3289 void free_buffer_head(struct buffer_head *bh)
3290 {
3291         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3292         kmem_cache_free(bh_cachep, bh);
3293         preempt_disable();
3294         __this_cpu_dec(bh_accounting.nr);
3295         recalc_bh_state();
3296         preempt_enable();
3297 }
3298 EXPORT_SYMBOL(free_buffer_head);
3299
3300 static void buffer_exit_cpu(int cpu)
3301 {
3302         int i;
3303         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3304
3305         for (i = 0; i < BH_LRU_SIZE; i++) {
3306                 brelse(b->bhs[i]);
3307                 b->bhs[i] = NULL;
3308         }
3309         this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3310         per_cpu(bh_accounting, cpu).nr = 0;
3311 }
3312
3313 static int buffer_cpu_notify(struct notifier_block *self,
3314                               unsigned long action, void *hcpu)
3315 {
3316         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3317                 buffer_exit_cpu((unsigned long)hcpu);
3318         return NOTIFY_OK;
3319 }
3320
3321 /**
3322  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3323  * @bh: struct buffer_head
3324  *
3325  * Return true if the buffer is up-to-date and false,
3326  * with the buffer locked, if not.
3327  */
3328 int bh_uptodate_or_lock(struct buffer_head *bh)
3329 {
3330         if (!buffer_uptodate(bh)) {
3331                 lock_buffer(bh);
3332                 if (!buffer_uptodate(bh))
3333                         return 0;
3334                 unlock_buffer(bh);
3335         }
3336         return 1;
3337 }
3338 EXPORT_SYMBOL(bh_uptodate_or_lock);
3339
3340 /**
3341  * bh_submit_read - Submit a locked buffer for reading
3342  * @bh: struct buffer_head
3343  *
3344  * Returns zero on success and -EIO on error.
3345  */
3346 int bh_submit_read(struct buffer_head *bh)
3347 {
3348         BUG_ON(!buffer_locked(bh));
3349
3350         if (buffer_uptodate(bh)) {
3351                 unlock_buffer(bh);
3352                 return 0;
3353         }
3354
3355         get_bh(bh);
3356         bh->b_end_io = end_buffer_read_sync;
3357         submit_bh(READ, bh);
3358         wait_on_buffer(bh);
3359         if (buffer_uptodate(bh))
3360                 return 0;
3361         return -EIO;
3362 }
3363 EXPORT_SYMBOL(bh_submit_read);
3364
3365 void __init buffer_init(void)
3366 {
3367         unsigned long nrpages;
3368
3369         bh_cachep = kmem_cache_create("buffer_head",
3370                         sizeof(struct buffer_head), 0,
3371                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3372                                 SLAB_MEM_SPREAD),
3373                                 NULL);
3374
3375         /*
3376          * Limit the bh occupancy to 10% of ZONE_NORMAL
3377          */
3378         nrpages = (nr_free_buffer_pages() * 10) / 100;
3379         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3380         hotcpu_notifier(buffer_cpu_notify, 0);
3381 }