fs/buffer.c: do not inline exported function
[~shefty/rdma-dev.git] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/capability.h>
28 #include <linux/blkdev.h>
29 #include <linux/file.h>
30 #include <linux/quotaops.h>
31 #include <linux/highmem.h>
32 #include <linux/export.h>
33 #include <linux/writeback.h>
34 #include <linux/hash.h>
35 #include <linux/suspend.h>
36 #include <linux/buffer_head.h>
37 #include <linux/task_io_accounting_ops.h>
38 #include <linux/bio.h>
39 #include <linux/notifier.h>
40 #include <linux/cpu.h>
41 #include <linux/bitops.h>
42 #include <linux/mpage.h>
43 #include <linux/bit_spinlock.h>
44
45 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
46
47 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
48
49 void init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
50 {
51         bh->b_end_io = handler;
52         bh->b_private = private;
53 }
54 EXPORT_SYMBOL(init_buffer);
55
56 static int sleep_on_buffer(void *word)
57 {
58         io_schedule();
59         return 0;
60 }
61
62 void __lock_buffer(struct buffer_head *bh)
63 {
64         wait_on_bit_lock(&bh->b_state, BH_Lock, sleep_on_buffer,
65                                                         TASK_UNINTERRUPTIBLE);
66 }
67 EXPORT_SYMBOL(__lock_buffer);
68
69 void unlock_buffer(struct buffer_head *bh)
70 {
71         clear_bit_unlock(BH_Lock, &bh->b_state);
72         smp_mb__after_clear_bit();
73         wake_up_bit(&bh->b_state, BH_Lock);
74 }
75 EXPORT_SYMBOL(unlock_buffer);
76
77 /*
78  * Block until a buffer comes unlocked.  This doesn't stop it
79  * from becoming locked again - you have to lock it yourself
80  * if you want to preserve its state.
81  */
82 void __wait_on_buffer(struct buffer_head * bh)
83 {
84         wait_on_bit(&bh->b_state, BH_Lock, sleep_on_buffer, TASK_UNINTERRUPTIBLE);
85 }
86 EXPORT_SYMBOL(__wait_on_buffer);
87
88 static void
89 __clear_page_buffers(struct page *page)
90 {
91         ClearPagePrivate(page);
92         set_page_private(page, 0);
93         page_cache_release(page);
94 }
95
96
97 static int quiet_error(struct buffer_head *bh)
98 {
99         if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
100                 return 0;
101         return 1;
102 }
103
104
105 static void buffer_io_error(struct buffer_head *bh)
106 {
107         char b[BDEVNAME_SIZE];
108         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
109                         bdevname(bh->b_bdev, b),
110                         (unsigned long long)bh->b_blocknr);
111 }
112
113 /*
114  * End-of-IO handler helper function which does not touch the bh after
115  * unlocking it.
116  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
117  * a race there is benign: unlock_buffer() only use the bh's address for
118  * hashing after unlocking the buffer, so it doesn't actually touch the bh
119  * itself.
120  */
121 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
122 {
123         if (uptodate) {
124                 set_buffer_uptodate(bh);
125         } else {
126                 /* This happens, due to failed READA attempts. */
127                 clear_buffer_uptodate(bh);
128         }
129         unlock_buffer(bh);
130 }
131
132 /*
133  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
134  * unlock the buffer. This is what ll_rw_block uses too.
135  */
136 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
137 {
138         __end_buffer_read_notouch(bh, uptodate);
139         put_bh(bh);
140 }
141 EXPORT_SYMBOL(end_buffer_read_sync);
142
143 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
144 {
145         char b[BDEVNAME_SIZE];
146
147         if (uptodate) {
148                 set_buffer_uptodate(bh);
149         } else {
150                 if (!quiet_error(bh)) {
151                         buffer_io_error(bh);
152                         printk(KERN_WARNING "lost page write due to "
153                                         "I/O error on %s\n",
154                                        bdevname(bh->b_bdev, b));
155                 }
156                 set_buffer_write_io_error(bh);
157                 clear_buffer_uptodate(bh);
158         }
159         unlock_buffer(bh);
160         put_bh(bh);
161 }
162 EXPORT_SYMBOL(end_buffer_write_sync);
163
164 /*
165  * Various filesystems appear to want __find_get_block to be non-blocking.
166  * But it's the page lock which protects the buffers.  To get around this,
167  * we get exclusion from try_to_free_buffers with the blockdev mapping's
168  * private_lock.
169  *
170  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
171  * may be quite high.  This code could TryLock the page, and if that
172  * succeeds, there is no need to take private_lock. (But if
173  * private_lock is contended then so is mapping->tree_lock).
174  */
175 static struct buffer_head *
176 __find_get_block_slow(struct block_device *bdev, sector_t block)
177 {
178         struct inode *bd_inode = bdev->bd_inode;
179         struct address_space *bd_mapping = bd_inode->i_mapping;
180         struct buffer_head *ret = NULL;
181         pgoff_t index;
182         struct buffer_head *bh;
183         struct buffer_head *head;
184         struct page *page;
185         int all_mapped = 1;
186
187         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
188         page = find_get_page(bd_mapping, index);
189         if (!page)
190                 goto out;
191
192         spin_lock(&bd_mapping->private_lock);
193         if (!page_has_buffers(page))
194                 goto out_unlock;
195         head = page_buffers(page);
196         bh = head;
197         do {
198                 if (!buffer_mapped(bh))
199                         all_mapped = 0;
200                 else if (bh->b_blocknr == block) {
201                         ret = bh;
202                         get_bh(bh);
203                         goto out_unlock;
204                 }
205                 bh = bh->b_this_page;
206         } while (bh != head);
207
208         /* we might be here because some of the buffers on this page are
209          * not mapped.  This is due to various races between
210          * file io on the block device and getblk.  It gets dealt with
211          * elsewhere, don't buffer_error if we had some unmapped buffers
212          */
213         if (all_mapped) {
214                 char b[BDEVNAME_SIZE];
215
216                 printk("__find_get_block_slow() failed. "
217                         "block=%llu, b_blocknr=%llu\n",
218                         (unsigned long long)block,
219                         (unsigned long long)bh->b_blocknr);
220                 printk("b_state=0x%08lx, b_size=%zu\n",
221                         bh->b_state, bh->b_size);
222                 printk("device %s blocksize: %d\n", bdevname(bdev, b),
223                         1 << bd_inode->i_blkbits);
224         }
225 out_unlock:
226         spin_unlock(&bd_mapping->private_lock);
227         page_cache_release(page);
228 out:
229         return ret;
230 }
231
232 /*
233  * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
234  */
235 static void free_more_memory(void)
236 {
237         struct zone *zone;
238         int nid;
239
240         wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
241         yield();
242
243         for_each_online_node(nid) {
244                 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
245                                                 gfp_zone(GFP_NOFS), NULL,
246                                                 &zone);
247                 if (zone)
248                         try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
249                                                 GFP_NOFS, NULL);
250         }
251 }
252
253 /*
254  * I/O completion handler for block_read_full_page() - pages
255  * which come unlocked at the end of I/O.
256  */
257 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
258 {
259         unsigned long flags;
260         struct buffer_head *first;
261         struct buffer_head *tmp;
262         struct page *page;
263         int page_uptodate = 1;
264
265         BUG_ON(!buffer_async_read(bh));
266
267         page = bh->b_page;
268         if (uptodate) {
269                 set_buffer_uptodate(bh);
270         } else {
271                 clear_buffer_uptodate(bh);
272                 if (!quiet_error(bh))
273                         buffer_io_error(bh);
274                 SetPageError(page);
275         }
276
277         /*
278          * Be _very_ careful from here on. Bad things can happen if
279          * two buffer heads end IO at almost the same time and both
280          * decide that the page is now completely done.
281          */
282         first = page_buffers(page);
283         local_irq_save(flags);
284         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
285         clear_buffer_async_read(bh);
286         unlock_buffer(bh);
287         tmp = bh;
288         do {
289                 if (!buffer_uptodate(tmp))
290                         page_uptodate = 0;
291                 if (buffer_async_read(tmp)) {
292                         BUG_ON(!buffer_locked(tmp));
293                         goto still_busy;
294                 }
295                 tmp = tmp->b_this_page;
296         } while (tmp != bh);
297         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
298         local_irq_restore(flags);
299
300         /*
301          * If none of the buffers had errors and they are all
302          * uptodate then we can set the page uptodate.
303          */
304         if (page_uptodate && !PageError(page))
305                 SetPageUptodate(page);
306         unlock_page(page);
307         return;
308
309 still_busy:
310         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
311         local_irq_restore(flags);
312         return;
313 }
314
315 /*
316  * Completion handler for block_write_full_page() - pages which are unlocked
317  * during I/O, and which have PageWriteback cleared upon I/O completion.
318  */
319 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
320 {
321         char b[BDEVNAME_SIZE];
322         unsigned long flags;
323         struct buffer_head *first;
324         struct buffer_head *tmp;
325         struct page *page;
326
327         BUG_ON(!buffer_async_write(bh));
328
329         page = bh->b_page;
330         if (uptodate) {
331                 set_buffer_uptodate(bh);
332         } else {
333                 if (!quiet_error(bh)) {
334                         buffer_io_error(bh);
335                         printk(KERN_WARNING "lost page write due to "
336                                         "I/O error on %s\n",
337                                bdevname(bh->b_bdev, b));
338                 }
339                 set_bit(AS_EIO, &page->mapping->flags);
340                 set_buffer_write_io_error(bh);
341                 clear_buffer_uptodate(bh);
342                 SetPageError(page);
343         }
344
345         first = page_buffers(page);
346         local_irq_save(flags);
347         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
348
349         clear_buffer_async_write(bh);
350         unlock_buffer(bh);
351         tmp = bh->b_this_page;
352         while (tmp != bh) {
353                 if (buffer_async_write(tmp)) {
354                         BUG_ON(!buffer_locked(tmp));
355                         goto still_busy;
356                 }
357                 tmp = tmp->b_this_page;
358         }
359         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
360         local_irq_restore(flags);
361         end_page_writeback(page);
362         return;
363
364 still_busy:
365         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
366         local_irq_restore(flags);
367         return;
368 }
369 EXPORT_SYMBOL(end_buffer_async_write);
370
371 /*
372  * If a page's buffers are under async readin (end_buffer_async_read
373  * completion) then there is a possibility that another thread of
374  * control could lock one of the buffers after it has completed
375  * but while some of the other buffers have not completed.  This
376  * locked buffer would confuse end_buffer_async_read() into not unlocking
377  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
378  * that this buffer is not under async I/O.
379  *
380  * The page comes unlocked when it has no locked buffer_async buffers
381  * left.
382  *
383  * PageLocked prevents anyone starting new async I/O reads any of
384  * the buffers.
385  *
386  * PageWriteback is used to prevent simultaneous writeout of the same
387  * page.
388  *
389  * PageLocked prevents anyone from starting writeback of a page which is
390  * under read I/O (PageWriteback is only ever set against a locked page).
391  */
392 static void mark_buffer_async_read(struct buffer_head *bh)
393 {
394         bh->b_end_io = end_buffer_async_read;
395         set_buffer_async_read(bh);
396 }
397
398 static void mark_buffer_async_write_endio(struct buffer_head *bh,
399                                           bh_end_io_t *handler)
400 {
401         bh->b_end_io = handler;
402         set_buffer_async_write(bh);
403 }
404
405 void mark_buffer_async_write(struct buffer_head *bh)
406 {
407         mark_buffer_async_write_endio(bh, end_buffer_async_write);
408 }
409 EXPORT_SYMBOL(mark_buffer_async_write);
410
411
412 /*
413  * fs/buffer.c contains helper functions for buffer-backed address space's
414  * fsync functions.  A common requirement for buffer-based filesystems is
415  * that certain data from the backing blockdev needs to be written out for
416  * a successful fsync().  For example, ext2 indirect blocks need to be
417  * written back and waited upon before fsync() returns.
418  *
419  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
420  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
421  * management of a list of dependent buffers at ->i_mapping->private_list.
422  *
423  * Locking is a little subtle: try_to_free_buffers() will remove buffers
424  * from their controlling inode's queue when they are being freed.  But
425  * try_to_free_buffers() will be operating against the *blockdev* mapping
426  * at the time, not against the S_ISREG file which depends on those buffers.
427  * So the locking for private_list is via the private_lock in the address_space
428  * which backs the buffers.  Which is different from the address_space 
429  * against which the buffers are listed.  So for a particular address_space,
430  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
431  * mapping->private_list will always be protected by the backing blockdev's
432  * ->private_lock.
433  *
434  * Which introduces a requirement: all buffers on an address_space's
435  * ->private_list must be from the same address_space: the blockdev's.
436  *
437  * address_spaces which do not place buffers at ->private_list via these
438  * utility functions are free to use private_lock and private_list for
439  * whatever they want.  The only requirement is that list_empty(private_list)
440  * be true at clear_inode() time.
441  *
442  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
443  * filesystems should do that.  invalidate_inode_buffers() should just go
444  * BUG_ON(!list_empty).
445  *
446  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
447  * take an address_space, not an inode.  And it should be called
448  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
449  * queued up.
450  *
451  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
452  * list if it is already on a list.  Because if the buffer is on a list,
453  * it *must* already be on the right one.  If not, the filesystem is being
454  * silly.  This will save a ton of locking.  But first we have to ensure
455  * that buffers are taken *off* the old inode's list when they are freed
456  * (presumably in truncate).  That requires careful auditing of all
457  * filesystems (do it inside bforget()).  It could also be done by bringing
458  * b_inode back.
459  */
460
461 /*
462  * The buffer's backing address_space's private_lock must be held
463  */
464 static void __remove_assoc_queue(struct buffer_head *bh)
465 {
466         list_del_init(&bh->b_assoc_buffers);
467         WARN_ON(!bh->b_assoc_map);
468         if (buffer_write_io_error(bh))
469                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
470         bh->b_assoc_map = NULL;
471 }
472
473 int inode_has_buffers(struct inode *inode)
474 {
475         return !list_empty(&inode->i_data.private_list);
476 }
477
478 /*
479  * osync is designed to support O_SYNC io.  It waits synchronously for
480  * all already-submitted IO to complete, but does not queue any new
481  * writes to the disk.
482  *
483  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
484  * you dirty the buffers, and then use osync_inode_buffers to wait for
485  * completion.  Any other dirty buffers which are not yet queued for
486  * write will not be flushed to disk by the osync.
487  */
488 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
489 {
490         struct buffer_head *bh;
491         struct list_head *p;
492         int err = 0;
493
494         spin_lock(lock);
495 repeat:
496         list_for_each_prev(p, list) {
497                 bh = BH_ENTRY(p);
498                 if (buffer_locked(bh)) {
499                         get_bh(bh);
500                         spin_unlock(lock);
501                         wait_on_buffer(bh);
502                         if (!buffer_uptodate(bh))
503                                 err = -EIO;
504                         brelse(bh);
505                         spin_lock(lock);
506                         goto repeat;
507                 }
508         }
509         spin_unlock(lock);
510         return err;
511 }
512
513 static void do_thaw_one(struct super_block *sb, void *unused)
514 {
515         char b[BDEVNAME_SIZE];
516         while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
517                 printk(KERN_WARNING "Emergency Thaw on %s\n",
518                        bdevname(sb->s_bdev, b));
519 }
520
521 static void do_thaw_all(struct work_struct *work)
522 {
523         iterate_supers(do_thaw_one, NULL);
524         kfree(work);
525         printk(KERN_WARNING "Emergency Thaw complete\n");
526 }
527
528 /**
529  * emergency_thaw_all -- forcibly thaw every frozen filesystem
530  *
531  * Used for emergency unfreeze of all filesystems via SysRq
532  */
533 void emergency_thaw_all(void)
534 {
535         struct work_struct *work;
536
537         work = kmalloc(sizeof(*work), GFP_ATOMIC);
538         if (work) {
539                 INIT_WORK(work, do_thaw_all);
540                 schedule_work(work);
541         }
542 }
543
544 /**
545  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
546  * @mapping: the mapping which wants those buffers written
547  *
548  * Starts I/O against the buffers at mapping->private_list, and waits upon
549  * that I/O.
550  *
551  * Basically, this is a convenience function for fsync().
552  * @mapping is a file or directory which needs those buffers to be written for
553  * a successful fsync().
554  */
555 int sync_mapping_buffers(struct address_space *mapping)
556 {
557         struct address_space *buffer_mapping = mapping->private_data;
558
559         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
560                 return 0;
561
562         return fsync_buffers_list(&buffer_mapping->private_lock,
563                                         &mapping->private_list);
564 }
565 EXPORT_SYMBOL(sync_mapping_buffers);
566
567 /*
568  * Called when we've recently written block `bblock', and it is known that
569  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
570  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
571  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
572  */
573 void write_boundary_block(struct block_device *bdev,
574                         sector_t bblock, unsigned blocksize)
575 {
576         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
577         if (bh) {
578                 if (buffer_dirty(bh))
579                         ll_rw_block(WRITE, 1, &bh);
580                 put_bh(bh);
581         }
582 }
583
584 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
585 {
586         struct address_space *mapping = inode->i_mapping;
587         struct address_space *buffer_mapping = bh->b_page->mapping;
588
589         mark_buffer_dirty(bh);
590         if (!mapping->private_data) {
591                 mapping->private_data = buffer_mapping;
592         } else {
593                 BUG_ON(mapping->private_data != buffer_mapping);
594         }
595         if (!bh->b_assoc_map) {
596                 spin_lock(&buffer_mapping->private_lock);
597                 list_move_tail(&bh->b_assoc_buffers,
598                                 &mapping->private_list);
599                 bh->b_assoc_map = mapping;
600                 spin_unlock(&buffer_mapping->private_lock);
601         }
602 }
603 EXPORT_SYMBOL(mark_buffer_dirty_inode);
604
605 /*
606  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
607  * dirty.
608  *
609  * If warn is true, then emit a warning if the page is not uptodate and has
610  * not been truncated.
611  */
612 static void __set_page_dirty(struct page *page,
613                 struct address_space *mapping, int warn)
614 {
615         spin_lock_irq(&mapping->tree_lock);
616         if (page->mapping) {    /* Race with truncate? */
617                 WARN_ON_ONCE(warn && !PageUptodate(page));
618                 account_page_dirtied(page, mapping);
619                 radix_tree_tag_set(&mapping->page_tree,
620                                 page_index(page), PAGECACHE_TAG_DIRTY);
621         }
622         spin_unlock_irq(&mapping->tree_lock);
623         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
624 }
625
626 /*
627  * Add a page to the dirty page list.
628  *
629  * It is a sad fact of life that this function is called from several places
630  * deeply under spinlocking.  It may not sleep.
631  *
632  * If the page has buffers, the uptodate buffers are set dirty, to preserve
633  * dirty-state coherency between the page and the buffers.  It the page does
634  * not have buffers then when they are later attached they will all be set
635  * dirty.
636  *
637  * The buffers are dirtied before the page is dirtied.  There's a small race
638  * window in which a writepage caller may see the page cleanness but not the
639  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
640  * before the buffers, a concurrent writepage caller could clear the page dirty
641  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
642  * page on the dirty page list.
643  *
644  * We use private_lock to lock against try_to_free_buffers while using the
645  * page's buffer list.  Also use this to protect against clean buffers being
646  * added to the page after it was set dirty.
647  *
648  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
649  * address_space though.
650  */
651 int __set_page_dirty_buffers(struct page *page)
652 {
653         int newly_dirty;
654         struct address_space *mapping = page_mapping(page);
655
656         if (unlikely(!mapping))
657                 return !TestSetPageDirty(page);
658
659         spin_lock(&mapping->private_lock);
660         if (page_has_buffers(page)) {
661                 struct buffer_head *head = page_buffers(page);
662                 struct buffer_head *bh = head;
663
664                 do {
665                         set_buffer_dirty(bh);
666                         bh = bh->b_this_page;
667                 } while (bh != head);
668         }
669         newly_dirty = !TestSetPageDirty(page);
670         spin_unlock(&mapping->private_lock);
671
672         if (newly_dirty)
673                 __set_page_dirty(page, mapping, 1);
674         return newly_dirty;
675 }
676 EXPORT_SYMBOL(__set_page_dirty_buffers);
677
678 /*
679  * Write out and wait upon a list of buffers.
680  *
681  * We have conflicting pressures: we want to make sure that all
682  * initially dirty buffers get waited on, but that any subsequently
683  * dirtied buffers don't.  After all, we don't want fsync to last
684  * forever if somebody is actively writing to the file.
685  *
686  * Do this in two main stages: first we copy dirty buffers to a
687  * temporary inode list, queueing the writes as we go.  Then we clean
688  * up, waiting for those writes to complete.
689  * 
690  * During this second stage, any subsequent updates to the file may end
691  * up refiling the buffer on the original inode's dirty list again, so
692  * there is a chance we will end up with a buffer queued for write but
693  * not yet completed on that list.  So, as a final cleanup we go through
694  * the osync code to catch these locked, dirty buffers without requeuing
695  * any newly dirty buffers for write.
696  */
697 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
698 {
699         struct buffer_head *bh;
700         struct list_head tmp;
701         struct address_space *mapping;
702         int err = 0, err2;
703         struct blk_plug plug;
704
705         INIT_LIST_HEAD(&tmp);
706         blk_start_plug(&plug);
707
708         spin_lock(lock);
709         while (!list_empty(list)) {
710                 bh = BH_ENTRY(list->next);
711                 mapping = bh->b_assoc_map;
712                 __remove_assoc_queue(bh);
713                 /* Avoid race with mark_buffer_dirty_inode() which does
714                  * a lockless check and we rely on seeing the dirty bit */
715                 smp_mb();
716                 if (buffer_dirty(bh) || buffer_locked(bh)) {
717                         list_add(&bh->b_assoc_buffers, &tmp);
718                         bh->b_assoc_map = mapping;
719                         if (buffer_dirty(bh)) {
720                                 get_bh(bh);
721                                 spin_unlock(lock);
722                                 /*
723                                  * Ensure any pending I/O completes so that
724                                  * write_dirty_buffer() actually writes the
725                                  * current contents - it is a noop if I/O is
726                                  * still in flight on potentially older
727                                  * contents.
728                                  */
729                                 write_dirty_buffer(bh, WRITE_SYNC);
730
731                                 /*
732                                  * Kick off IO for the previous mapping. Note
733                                  * that we will not run the very last mapping,
734                                  * wait_on_buffer() will do that for us
735                                  * through sync_buffer().
736                                  */
737                                 brelse(bh);
738                                 spin_lock(lock);
739                         }
740                 }
741         }
742
743         spin_unlock(lock);
744         blk_finish_plug(&plug);
745         spin_lock(lock);
746
747         while (!list_empty(&tmp)) {
748                 bh = BH_ENTRY(tmp.prev);
749                 get_bh(bh);
750                 mapping = bh->b_assoc_map;
751                 __remove_assoc_queue(bh);
752                 /* Avoid race with mark_buffer_dirty_inode() which does
753                  * a lockless check and we rely on seeing the dirty bit */
754                 smp_mb();
755                 if (buffer_dirty(bh)) {
756                         list_add(&bh->b_assoc_buffers,
757                                  &mapping->private_list);
758                         bh->b_assoc_map = mapping;
759                 }
760                 spin_unlock(lock);
761                 wait_on_buffer(bh);
762                 if (!buffer_uptodate(bh))
763                         err = -EIO;
764                 brelse(bh);
765                 spin_lock(lock);
766         }
767         
768         spin_unlock(lock);
769         err2 = osync_buffers_list(lock, list);
770         if (err)
771                 return err;
772         else
773                 return err2;
774 }
775
776 /*
777  * Invalidate any and all dirty buffers on a given inode.  We are
778  * probably unmounting the fs, but that doesn't mean we have already
779  * done a sync().  Just drop the buffers from the inode list.
780  *
781  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
782  * assumes that all the buffers are against the blockdev.  Not true
783  * for reiserfs.
784  */
785 void invalidate_inode_buffers(struct inode *inode)
786 {
787         if (inode_has_buffers(inode)) {
788                 struct address_space *mapping = &inode->i_data;
789                 struct list_head *list = &mapping->private_list;
790                 struct address_space *buffer_mapping = mapping->private_data;
791
792                 spin_lock(&buffer_mapping->private_lock);
793                 while (!list_empty(list))
794                         __remove_assoc_queue(BH_ENTRY(list->next));
795                 spin_unlock(&buffer_mapping->private_lock);
796         }
797 }
798 EXPORT_SYMBOL(invalidate_inode_buffers);
799
800 /*
801  * Remove any clean buffers from the inode's buffer list.  This is called
802  * when we're trying to free the inode itself.  Those buffers can pin it.
803  *
804  * Returns true if all buffers were removed.
805  */
806 int remove_inode_buffers(struct inode *inode)
807 {
808         int ret = 1;
809
810         if (inode_has_buffers(inode)) {
811                 struct address_space *mapping = &inode->i_data;
812                 struct list_head *list = &mapping->private_list;
813                 struct address_space *buffer_mapping = mapping->private_data;
814
815                 spin_lock(&buffer_mapping->private_lock);
816                 while (!list_empty(list)) {
817                         struct buffer_head *bh = BH_ENTRY(list->next);
818                         if (buffer_dirty(bh)) {
819                                 ret = 0;
820                                 break;
821                         }
822                         __remove_assoc_queue(bh);
823                 }
824                 spin_unlock(&buffer_mapping->private_lock);
825         }
826         return ret;
827 }
828
829 /*
830  * Create the appropriate buffers when given a page for data area and
831  * the size of each buffer.. Use the bh->b_this_page linked list to
832  * follow the buffers created.  Return NULL if unable to create more
833  * buffers.
834  *
835  * The retry flag is used to differentiate async IO (paging, swapping)
836  * which may not fail from ordinary buffer allocations.
837  */
838 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
839                 int retry)
840 {
841         struct buffer_head *bh, *head;
842         long offset;
843
844 try_again:
845         head = NULL;
846         offset = PAGE_SIZE;
847         while ((offset -= size) >= 0) {
848                 bh = alloc_buffer_head(GFP_NOFS);
849                 if (!bh)
850                         goto no_grow;
851
852                 bh->b_bdev = NULL;
853                 bh->b_this_page = head;
854                 bh->b_blocknr = -1;
855                 head = bh;
856
857                 bh->b_state = 0;
858                 atomic_set(&bh->b_count, 0);
859                 bh->b_size = size;
860
861                 /* Link the buffer to its page */
862                 set_bh_page(bh, page, offset);
863
864                 init_buffer(bh, NULL, NULL);
865         }
866         return head;
867 /*
868  * In case anything failed, we just free everything we got.
869  */
870 no_grow:
871         if (head) {
872                 do {
873                         bh = head;
874                         head = head->b_this_page;
875                         free_buffer_head(bh);
876                 } while (head);
877         }
878
879         /*
880          * Return failure for non-async IO requests.  Async IO requests
881          * are not allowed to fail, so we have to wait until buffer heads
882          * become available.  But we don't want tasks sleeping with 
883          * partially complete buffers, so all were released above.
884          */
885         if (!retry)
886                 return NULL;
887
888         /* We're _really_ low on memory. Now we just
889          * wait for old buffer heads to become free due to
890          * finishing IO.  Since this is an async request and
891          * the reserve list is empty, we're sure there are 
892          * async buffer heads in use.
893          */
894         free_more_memory();
895         goto try_again;
896 }
897 EXPORT_SYMBOL_GPL(alloc_page_buffers);
898
899 static inline void
900 link_dev_buffers(struct page *page, struct buffer_head *head)
901 {
902         struct buffer_head *bh, *tail;
903
904         bh = head;
905         do {
906                 tail = bh;
907                 bh = bh->b_this_page;
908         } while (bh);
909         tail->b_this_page = head;
910         attach_page_buffers(page, head);
911 }
912
913 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
914 {
915         sector_t retval = ~((sector_t)0);
916         loff_t sz = i_size_read(bdev->bd_inode);
917
918         if (sz) {
919                 unsigned int sizebits = blksize_bits(size);
920                 retval = (sz >> sizebits);
921         }
922         return retval;
923 }
924
925 /*
926  * Initialise the state of a blockdev page's buffers.
927  */ 
928 static sector_t
929 init_page_buffers(struct page *page, struct block_device *bdev,
930                         sector_t block, int size)
931 {
932         struct buffer_head *head = page_buffers(page);
933         struct buffer_head *bh = head;
934         int uptodate = PageUptodate(page);
935         sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
936
937         do {
938                 if (!buffer_mapped(bh)) {
939                         init_buffer(bh, NULL, NULL);
940                         bh->b_bdev = bdev;
941                         bh->b_blocknr = block;
942                         if (uptodate)
943                                 set_buffer_uptodate(bh);
944                         if (block < end_block)
945                                 set_buffer_mapped(bh);
946                 }
947                 block++;
948                 bh = bh->b_this_page;
949         } while (bh != head);
950
951         /*
952          * Caller needs to validate requested block against end of device.
953          */
954         return end_block;
955 }
956
957 /*
958  * Create the page-cache page that contains the requested block.
959  *
960  * This is used purely for blockdev mappings.
961  */
962 static int
963 grow_dev_page(struct block_device *bdev, sector_t block,
964                 pgoff_t index, int size, int sizebits)
965 {
966         struct inode *inode = bdev->bd_inode;
967         struct page *page;
968         struct buffer_head *bh;
969         sector_t end_block;
970         int ret = 0;            /* Will call free_more_memory() */
971
972         page = find_or_create_page(inode->i_mapping, index,
973                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
974         if (!page)
975                 return ret;
976
977         BUG_ON(!PageLocked(page));
978
979         if (page_has_buffers(page)) {
980                 bh = page_buffers(page);
981                 if (bh->b_size == size) {
982                         end_block = init_page_buffers(page, bdev,
983                                                 index << sizebits, size);
984                         goto done;
985                 }
986                 if (!try_to_free_buffers(page))
987                         goto failed;
988         }
989
990         /*
991          * Allocate some buffers for this page
992          */
993         bh = alloc_page_buffers(page, size, 0);
994         if (!bh)
995                 goto failed;
996
997         /*
998          * Link the page to the buffers and initialise them.  Take the
999          * lock to be atomic wrt __find_get_block(), which does not
1000          * run under the page lock.
1001          */
1002         spin_lock(&inode->i_mapping->private_lock);
1003         link_dev_buffers(page, bh);
1004         end_block = init_page_buffers(page, bdev, index << sizebits, size);
1005         spin_unlock(&inode->i_mapping->private_lock);
1006 done:
1007         ret = (block < end_block) ? 1 : -ENXIO;
1008 failed:
1009         unlock_page(page);
1010         page_cache_release(page);
1011         return ret;
1012 }
1013
1014 /*
1015  * Create buffers for the specified block device block's page.  If
1016  * that page was dirty, the buffers are set dirty also.
1017  */
1018 static int
1019 grow_buffers(struct block_device *bdev, sector_t block, int size)
1020 {
1021         pgoff_t index;
1022         int sizebits;
1023
1024         sizebits = -1;
1025         do {
1026                 sizebits++;
1027         } while ((size << sizebits) < PAGE_SIZE);
1028
1029         index = block >> sizebits;
1030
1031         /*
1032          * Check for a block which wants to lie outside our maximum possible
1033          * pagecache index.  (this comparison is done using sector_t types).
1034          */
1035         if (unlikely(index != block >> sizebits)) {
1036                 char b[BDEVNAME_SIZE];
1037
1038                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1039                         "device %s\n",
1040                         __func__, (unsigned long long)block,
1041                         bdevname(bdev, b));
1042                 return -EIO;
1043         }
1044
1045         /* Create a page with the proper size buffers.. */
1046         return grow_dev_page(bdev, block, index, size, sizebits);
1047 }
1048
1049 static struct buffer_head *
1050 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1051 {
1052         /* Size must be multiple of hard sectorsize */
1053         if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1054                         (size < 512 || size > PAGE_SIZE))) {
1055                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1056                                         size);
1057                 printk(KERN_ERR "logical block size: %d\n",
1058                                         bdev_logical_block_size(bdev));
1059
1060                 dump_stack();
1061                 return NULL;
1062         }
1063
1064         for (;;) {
1065                 struct buffer_head *bh;
1066                 int ret;
1067
1068                 bh = __find_get_block(bdev, block, size);
1069                 if (bh)
1070                         return bh;
1071
1072                 ret = grow_buffers(bdev, block, size);
1073                 if (ret < 0)
1074                         return NULL;
1075                 if (ret == 0)
1076                         free_more_memory();
1077         }
1078 }
1079
1080 /*
1081  * The relationship between dirty buffers and dirty pages:
1082  *
1083  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1084  * the page is tagged dirty in its radix tree.
1085  *
1086  * At all times, the dirtiness of the buffers represents the dirtiness of
1087  * subsections of the page.  If the page has buffers, the page dirty bit is
1088  * merely a hint about the true dirty state.
1089  *
1090  * When a page is set dirty in its entirety, all its buffers are marked dirty
1091  * (if the page has buffers).
1092  *
1093  * When a buffer is marked dirty, its page is dirtied, but the page's other
1094  * buffers are not.
1095  *
1096  * Also.  When blockdev buffers are explicitly read with bread(), they
1097  * individually become uptodate.  But their backing page remains not
1098  * uptodate - even if all of its buffers are uptodate.  A subsequent
1099  * block_read_full_page() against that page will discover all the uptodate
1100  * buffers, will set the page uptodate and will perform no I/O.
1101  */
1102
1103 /**
1104  * mark_buffer_dirty - mark a buffer_head as needing writeout
1105  * @bh: the buffer_head to mark dirty
1106  *
1107  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1108  * backing page dirty, then tag the page as dirty in its address_space's radix
1109  * tree and then attach the address_space's inode to its superblock's dirty
1110  * inode list.
1111  *
1112  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1113  * mapping->tree_lock and mapping->host->i_lock.
1114  */
1115 void mark_buffer_dirty(struct buffer_head *bh)
1116 {
1117         WARN_ON_ONCE(!buffer_uptodate(bh));
1118
1119         /*
1120          * Very *carefully* optimize the it-is-already-dirty case.
1121          *
1122          * Don't let the final "is it dirty" escape to before we
1123          * perhaps modified the buffer.
1124          */
1125         if (buffer_dirty(bh)) {
1126                 smp_mb();
1127                 if (buffer_dirty(bh))
1128                         return;
1129         }
1130
1131         if (!test_set_buffer_dirty(bh)) {
1132                 struct page *page = bh->b_page;
1133                 if (!TestSetPageDirty(page)) {
1134                         struct address_space *mapping = page_mapping(page);
1135                         if (mapping)
1136                                 __set_page_dirty(page, mapping, 0);
1137                 }
1138         }
1139 }
1140 EXPORT_SYMBOL(mark_buffer_dirty);
1141
1142 /*
1143  * Decrement a buffer_head's reference count.  If all buffers against a page
1144  * have zero reference count, are clean and unlocked, and if the page is clean
1145  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1146  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1147  * a page but it ends up not being freed, and buffers may later be reattached).
1148  */
1149 void __brelse(struct buffer_head * buf)
1150 {
1151         if (atomic_read(&buf->b_count)) {
1152                 put_bh(buf);
1153                 return;
1154         }
1155         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1156 }
1157 EXPORT_SYMBOL(__brelse);
1158
1159 /*
1160  * bforget() is like brelse(), except it discards any
1161  * potentially dirty data.
1162  */
1163 void __bforget(struct buffer_head *bh)
1164 {
1165         clear_buffer_dirty(bh);
1166         if (bh->b_assoc_map) {
1167                 struct address_space *buffer_mapping = bh->b_page->mapping;
1168
1169                 spin_lock(&buffer_mapping->private_lock);
1170                 list_del_init(&bh->b_assoc_buffers);
1171                 bh->b_assoc_map = NULL;
1172                 spin_unlock(&buffer_mapping->private_lock);
1173         }
1174         __brelse(bh);
1175 }
1176 EXPORT_SYMBOL(__bforget);
1177
1178 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1179 {
1180         lock_buffer(bh);
1181         if (buffer_uptodate(bh)) {
1182                 unlock_buffer(bh);
1183                 return bh;
1184         } else {
1185                 get_bh(bh);
1186                 bh->b_end_io = end_buffer_read_sync;
1187                 submit_bh(READ, bh);
1188                 wait_on_buffer(bh);
1189                 if (buffer_uptodate(bh))
1190                         return bh;
1191         }
1192         brelse(bh);
1193         return NULL;
1194 }
1195
1196 /*
1197  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1198  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1199  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1200  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1201  * CPU's LRUs at the same time.
1202  *
1203  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1204  * sb_find_get_block().
1205  *
1206  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1207  * a local interrupt disable for that.
1208  */
1209
1210 #define BH_LRU_SIZE     8
1211
1212 struct bh_lru {
1213         struct buffer_head *bhs[BH_LRU_SIZE];
1214 };
1215
1216 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1217
1218 #ifdef CONFIG_SMP
1219 #define bh_lru_lock()   local_irq_disable()
1220 #define bh_lru_unlock() local_irq_enable()
1221 #else
1222 #define bh_lru_lock()   preempt_disable()
1223 #define bh_lru_unlock() preempt_enable()
1224 #endif
1225
1226 static inline void check_irqs_on(void)
1227 {
1228 #ifdef irqs_disabled
1229         BUG_ON(irqs_disabled());
1230 #endif
1231 }
1232
1233 /*
1234  * The LRU management algorithm is dopey-but-simple.  Sorry.
1235  */
1236 static void bh_lru_install(struct buffer_head *bh)
1237 {
1238         struct buffer_head *evictee = NULL;
1239
1240         check_irqs_on();
1241         bh_lru_lock();
1242         if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1243                 struct buffer_head *bhs[BH_LRU_SIZE];
1244                 int in;
1245                 int out = 0;
1246
1247                 get_bh(bh);
1248                 bhs[out++] = bh;
1249                 for (in = 0; in < BH_LRU_SIZE; in++) {
1250                         struct buffer_head *bh2 =
1251                                 __this_cpu_read(bh_lrus.bhs[in]);
1252
1253                         if (bh2 == bh) {
1254                                 __brelse(bh2);
1255                         } else {
1256                                 if (out >= BH_LRU_SIZE) {
1257                                         BUG_ON(evictee != NULL);
1258                                         evictee = bh2;
1259                                 } else {
1260                                         bhs[out++] = bh2;
1261                                 }
1262                         }
1263                 }
1264                 while (out < BH_LRU_SIZE)
1265                         bhs[out++] = NULL;
1266                 memcpy(__this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1267         }
1268         bh_lru_unlock();
1269
1270         if (evictee)
1271                 __brelse(evictee);
1272 }
1273
1274 /*
1275  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1276  */
1277 static struct buffer_head *
1278 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1279 {
1280         struct buffer_head *ret = NULL;
1281         unsigned int i;
1282
1283         check_irqs_on();
1284         bh_lru_lock();
1285         for (i = 0; i < BH_LRU_SIZE; i++) {
1286                 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1287
1288                 if (bh && bh->b_bdev == bdev &&
1289                                 bh->b_blocknr == block && bh->b_size == size) {
1290                         if (i) {
1291                                 while (i) {
1292                                         __this_cpu_write(bh_lrus.bhs[i],
1293                                                 __this_cpu_read(bh_lrus.bhs[i - 1]));
1294                                         i--;
1295                                 }
1296                                 __this_cpu_write(bh_lrus.bhs[0], bh);
1297                         }
1298                         get_bh(bh);
1299                         ret = bh;
1300                         break;
1301                 }
1302         }
1303         bh_lru_unlock();
1304         return ret;
1305 }
1306
1307 /*
1308  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1309  * it in the LRU and mark it as accessed.  If it is not present then return
1310  * NULL
1311  */
1312 struct buffer_head *
1313 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1314 {
1315         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1316
1317         if (bh == NULL) {
1318                 bh = __find_get_block_slow(bdev, block);
1319                 if (bh)
1320                         bh_lru_install(bh);
1321         }
1322         if (bh)
1323                 touch_buffer(bh);
1324         return bh;
1325 }
1326 EXPORT_SYMBOL(__find_get_block);
1327
1328 /*
1329  * __getblk will locate (and, if necessary, create) the buffer_head
1330  * which corresponds to the passed block_device, block and size. The
1331  * returned buffer has its reference count incremented.
1332  *
1333  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1334  * attempt is failing.  FIXME, perhaps?
1335  */
1336 struct buffer_head *
1337 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1338 {
1339         struct buffer_head *bh = __find_get_block(bdev, block, size);
1340
1341         might_sleep();
1342         if (bh == NULL)
1343                 bh = __getblk_slow(bdev, block, size);
1344         return bh;
1345 }
1346 EXPORT_SYMBOL(__getblk);
1347
1348 /*
1349  * Do async read-ahead on a buffer..
1350  */
1351 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1352 {
1353         struct buffer_head *bh = __getblk(bdev, block, size);
1354         if (likely(bh)) {
1355                 ll_rw_block(READA, 1, &bh);
1356                 brelse(bh);
1357         }
1358 }
1359 EXPORT_SYMBOL(__breadahead);
1360
1361 /**
1362  *  __bread() - reads a specified block and returns the bh
1363  *  @bdev: the block_device to read from
1364  *  @block: number of block
1365  *  @size: size (in bytes) to read
1366  * 
1367  *  Reads a specified block, and returns buffer head that contains it.
1368  *  It returns NULL if the block was unreadable.
1369  */
1370 struct buffer_head *
1371 __bread(struct block_device *bdev, sector_t block, unsigned size)
1372 {
1373         struct buffer_head *bh = __getblk(bdev, block, size);
1374
1375         if (likely(bh) && !buffer_uptodate(bh))
1376                 bh = __bread_slow(bh);
1377         return bh;
1378 }
1379 EXPORT_SYMBOL(__bread);
1380
1381 /*
1382  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1383  * This doesn't race because it runs in each cpu either in irq
1384  * or with preempt disabled.
1385  */
1386 static void invalidate_bh_lru(void *arg)
1387 {
1388         struct bh_lru *b = &get_cpu_var(bh_lrus);
1389         int i;
1390
1391         for (i = 0; i < BH_LRU_SIZE; i++) {
1392                 brelse(b->bhs[i]);
1393                 b->bhs[i] = NULL;
1394         }
1395         put_cpu_var(bh_lrus);
1396 }
1397
1398 static bool has_bh_in_lru(int cpu, void *dummy)
1399 {
1400         struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1401         int i;
1402         
1403         for (i = 0; i < BH_LRU_SIZE; i++) {
1404                 if (b->bhs[i])
1405                         return 1;
1406         }
1407
1408         return 0;
1409 }
1410
1411 void invalidate_bh_lrus(void)
1412 {
1413         on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1414 }
1415 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1416
1417 void set_bh_page(struct buffer_head *bh,
1418                 struct page *page, unsigned long offset)
1419 {
1420         bh->b_page = page;
1421         BUG_ON(offset >= PAGE_SIZE);
1422         if (PageHighMem(page))
1423                 /*
1424                  * This catches illegal uses and preserves the offset:
1425                  */
1426                 bh->b_data = (char *)(0 + offset);
1427         else
1428                 bh->b_data = page_address(page) + offset;
1429 }
1430 EXPORT_SYMBOL(set_bh_page);
1431
1432 /*
1433  * Called when truncating a buffer on a page completely.
1434  */
1435 static void discard_buffer(struct buffer_head * bh)
1436 {
1437         lock_buffer(bh);
1438         clear_buffer_dirty(bh);
1439         bh->b_bdev = NULL;
1440         clear_buffer_mapped(bh);
1441         clear_buffer_req(bh);
1442         clear_buffer_new(bh);
1443         clear_buffer_delay(bh);
1444         clear_buffer_unwritten(bh);
1445         unlock_buffer(bh);
1446 }
1447
1448 /**
1449  * block_invalidatepage - invalidate part or all of a buffer-backed page
1450  *
1451  * @page: the page which is affected
1452  * @offset: the index of the truncation point
1453  *
1454  * block_invalidatepage() is called when all or part of the page has become
1455  * invalidated by a truncate operation.
1456  *
1457  * block_invalidatepage() does not have to release all buffers, but it must
1458  * ensure that no dirty buffer is left outside @offset and that no I/O
1459  * is underway against any of the blocks which are outside the truncation
1460  * point.  Because the caller is about to free (and possibly reuse) those
1461  * blocks on-disk.
1462  */
1463 void block_invalidatepage(struct page *page, unsigned long offset)
1464 {
1465         struct buffer_head *head, *bh, *next;
1466         unsigned int curr_off = 0;
1467
1468         BUG_ON(!PageLocked(page));
1469         if (!page_has_buffers(page))
1470                 goto out;
1471
1472         head = page_buffers(page);
1473         bh = head;
1474         do {
1475                 unsigned int next_off = curr_off + bh->b_size;
1476                 next = bh->b_this_page;
1477
1478                 /*
1479                  * is this block fully invalidated?
1480                  */
1481                 if (offset <= curr_off)
1482                         discard_buffer(bh);
1483                 curr_off = next_off;
1484                 bh = next;
1485         } while (bh != head);
1486
1487         /*
1488          * We release buffers only if the entire page is being invalidated.
1489          * The get_block cached value has been unconditionally invalidated,
1490          * so real IO is not possible anymore.
1491          */
1492         if (offset == 0)
1493                 try_to_release_page(page, 0);
1494 out:
1495         return;
1496 }
1497 EXPORT_SYMBOL(block_invalidatepage);
1498
1499 /*
1500  * We attach and possibly dirty the buffers atomically wrt
1501  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1502  * is already excluded via the page lock.
1503  */
1504 void create_empty_buffers(struct page *page,
1505                         unsigned long blocksize, unsigned long b_state)
1506 {
1507         struct buffer_head *bh, *head, *tail;
1508
1509         head = alloc_page_buffers(page, blocksize, 1);
1510         bh = head;
1511         do {
1512                 bh->b_state |= b_state;
1513                 tail = bh;
1514                 bh = bh->b_this_page;
1515         } while (bh);
1516         tail->b_this_page = head;
1517
1518         spin_lock(&page->mapping->private_lock);
1519         if (PageUptodate(page) || PageDirty(page)) {
1520                 bh = head;
1521                 do {
1522                         if (PageDirty(page))
1523                                 set_buffer_dirty(bh);
1524                         if (PageUptodate(page))
1525                                 set_buffer_uptodate(bh);
1526                         bh = bh->b_this_page;
1527                 } while (bh != head);
1528         }
1529         attach_page_buffers(page, head);
1530         spin_unlock(&page->mapping->private_lock);
1531 }
1532 EXPORT_SYMBOL(create_empty_buffers);
1533
1534 /*
1535  * We are taking a block for data and we don't want any output from any
1536  * buffer-cache aliases starting from return from that function and
1537  * until the moment when something will explicitly mark the buffer
1538  * dirty (hopefully that will not happen until we will free that block ;-)
1539  * We don't even need to mark it not-uptodate - nobody can expect
1540  * anything from a newly allocated buffer anyway. We used to used
1541  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1542  * don't want to mark the alias unmapped, for example - it would confuse
1543  * anyone who might pick it with bread() afterwards...
1544  *
1545  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1546  * be writeout I/O going on against recently-freed buffers.  We don't
1547  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1548  * only if we really need to.  That happens here.
1549  */
1550 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1551 {
1552         struct buffer_head *old_bh;
1553
1554         might_sleep();
1555
1556         old_bh = __find_get_block_slow(bdev, block);
1557         if (old_bh) {
1558                 clear_buffer_dirty(old_bh);
1559                 wait_on_buffer(old_bh);
1560                 clear_buffer_req(old_bh);
1561                 __brelse(old_bh);
1562         }
1563 }
1564 EXPORT_SYMBOL(unmap_underlying_metadata);
1565
1566 /*
1567  * Size is a power-of-two in the range 512..PAGE_SIZE,
1568  * and the case we care about most is PAGE_SIZE.
1569  *
1570  * So this *could* possibly be written with those
1571  * constraints in mind (relevant mostly if some
1572  * architecture has a slow bit-scan instruction)
1573  */
1574 static inline int block_size_bits(unsigned int blocksize)
1575 {
1576         return ilog2(blocksize);
1577 }
1578
1579 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1580 {
1581         BUG_ON(!PageLocked(page));
1582
1583         if (!page_has_buffers(page))
1584                 create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1585         return page_buffers(page);
1586 }
1587
1588 /*
1589  * NOTE! All mapped/uptodate combinations are valid:
1590  *
1591  *      Mapped  Uptodate        Meaning
1592  *
1593  *      No      No              "unknown" - must do get_block()
1594  *      No      Yes             "hole" - zero-filled
1595  *      Yes     No              "allocated" - allocated on disk, not read in
1596  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1597  *
1598  * "Dirty" is valid only with the last case (mapped+uptodate).
1599  */
1600
1601 /*
1602  * While block_write_full_page is writing back the dirty buffers under
1603  * the page lock, whoever dirtied the buffers may decide to clean them
1604  * again at any time.  We handle that by only looking at the buffer
1605  * state inside lock_buffer().
1606  *
1607  * If block_write_full_page() is called for regular writeback
1608  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1609  * locked buffer.   This only can happen if someone has written the buffer
1610  * directly, with submit_bh().  At the address_space level PageWriteback
1611  * prevents this contention from occurring.
1612  *
1613  * If block_write_full_page() is called with wbc->sync_mode ==
1614  * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1615  * causes the writes to be flagged as synchronous writes.
1616  */
1617 static int __block_write_full_page(struct inode *inode, struct page *page,
1618                         get_block_t *get_block, struct writeback_control *wbc,
1619                         bh_end_io_t *handler)
1620 {
1621         int err;
1622         sector_t block;
1623         sector_t last_block;
1624         struct buffer_head *bh, *head;
1625         unsigned int blocksize, bbits;
1626         int nr_underway = 0;
1627         int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1628                         WRITE_SYNC : WRITE);
1629
1630         head = create_page_buffers(page, inode,
1631                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1632
1633         /*
1634          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1635          * here, and the (potentially unmapped) buffers may become dirty at
1636          * any time.  If a buffer becomes dirty here after we've inspected it
1637          * then we just miss that fact, and the page stays dirty.
1638          *
1639          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1640          * handle that here by just cleaning them.
1641          */
1642
1643         bh = head;
1644         blocksize = bh->b_size;
1645         bbits = block_size_bits(blocksize);
1646
1647         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1648         last_block = (i_size_read(inode) - 1) >> bbits;
1649
1650         /*
1651          * Get all the dirty buffers mapped to disk addresses and
1652          * handle any aliases from the underlying blockdev's mapping.
1653          */
1654         do {
1655                 if (block > last_block) {
1656                         /*
1657                          * mapped buffers outside i_size will occur, because
1658                          * this page can be outside i_size when there is a
1659                          * truncate in progress.
1660                          */
1661                         /*
1662                          * The buffer was zeroed by block_write_full_page()
1663                          */
1664                         clear_buffer_dirty(bh);
1665                         set_buffer_uptodate(bh);
1666                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1667                            buffer_dirty(bh)) {
1668                         WARN_ON(bh->b_size != blocksize);
1669                         err = get_block(inode, block, bh, 1);
1670                         if (err)
1671                                 goto recover;
1672                         clear_buffer_delay(bh);
1673                         if (buffer_new(bh)) {
1674                                 /* blockdev mappings never come here */
1675                                 clear_buffer_new(bh);
1676                                 unmap_underlying_metadata(bh->b_bdev,
1677                                                         bh->b_blocknr);
1678                         }
1679                 }
1680                 bh = bh->b_this_page;
1681                 block++;
1682         } while (bh != head);
1683
1684         do {
1685                 if (!buffer_mapped(bh))
1686                         continue;
1687                 /*
1688                  * If it's a fully non-blocking write attempt and we cannot
1689                  * lock the buffer then redirty the page.  Note that this can
1690                  * potentially cause a busy-wait loop from writeback threads
1691                  * and kswapd activity, but those code paths have their own
1692                  * higher-level throttling.
1693                  */
1694                 if (wbc->sync_mode != WB_SYNC_NONE) {
1695                         lock_buffer(bh);
1696                 } else if (!trylock_buffer(bh)) {
1697                         redirty_page_for_writepage(wbc, page);
1698                         continue;
1699                 }
1700                 if (test_clear_buffer_dirty(bh)) {
1701                         mark_buffer_async_write_endio(bh, handler);
1702                 } else {
1703                         unlock_buffer(bh);
1704                 }
1705         } while ((bh = bh->b_this_page) != head);
1706
1707         /*
1708          * The page and its buffers are protected by PageWriteback(), so we can
1709          * drop the bh refcounts early.
1710          */
1711         BUG_ON(PageWriteback(page));
1712         set_page_writeback(page);
1713
1714         do {
1715                 struct buffer_head *next = bh->b_this_page;
1716                 if (buffer_async_write(bh)) {
1717                         submit_bh(write_op, bh);
1718                         nr_underway++;
1719                 }
1720                 bh = next;
1721         } while (bh != head);
1722         unlock_page(page);
1723
1724         err = 0;
1725 done:
1726         if (nr_underway == 0) {
1727                 /*
1728                  * The page was marked dirty, but the buffers were
1729                  * clean.  Someone wrote them back by hand with
1730                  * ll_rw_block/submit_bh.  A rare case.
1731                  */
1732                 end_page_writeback(page);
1733
1734                 /*
1735                  * The page and buffer_heads can be released at any time from
1736                  * here on.
1737                  */
1738         }
1739         return err;
1740
1741 recover:
1742         /*
1743          * ENOSPC, or some other error.  We may already have added some
1744          * blocks to the file, so we need to write these out to avoid
1745          * exposing stale data.
1746          * The page is currently locked and not marked for writeback
1747          */
1748         bh = head;
1749         /* Recovery: lock and submit the mapped buffers */
1750         do {
1751                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1752                     !buffer_delay(bh)) {
1753                         lock_buffer(bh);
1754                         mark_buffer_async_write_endio(bh, handler);
1755                 } else {
1756                         /*
1757                          * The buffer may have been set dirty during
1758                          * attachment to a dirty page.
1759                          */
1760                         clear_buffer_dirty(bh);
1761                 }
1762         } while ((bh = bh->b_this_page) != head);
1763         SetPageError(page);
1764         BUG_ON(PageWriteback(page));
1765         mapping_set_error(page->mapping, err);
1766         set_page_writeback(page);
1767         do {
1768                 struct buffer_head *next = bh->b_this_page;
1769                 if (buffer_async_write(bh)) {
1770                         clear_buffer_dirty(bh);
1771                         submit_bh(write_op, bh);
1772                         nr_underway++;
1773                 }
1774                 bh = next;
1775         } while (bh != head);
1776         unlock_page(page);
1777         goto done;
1778 }
1779
1780 /*
1781  * If a page has any new buffers, zero them out here, and mark them uptodate
1782  * and dirty so they'll be written out (in order to prevent uninitialised
1783  * block data from leaking). And clear the new bit.
1784  */
1785 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1786 {
1787         unsigned int block_start, block_end;
1788         struct buffer_head *head, *bh;
1789
1790         BUG_ON(!PageLocked(page));
1791         if (!page_has_buffers(page))
1792                 return;
1793
1794         bh = head = page_buffers(page);
1795         block_start = 0;
1796         do {
1797                 block_end = block_start + bh->b_size;
1798
1799                 if (buffer_new(bh)) {
1800                         if (block_end > from && block_start < to) {
1801                                 if (!PageUptodate(page)) {
1802                                         unsigned start, size;
1803
1804                                         start = max(from, block_start);
1805                                         size = min(to, block_end) - start;
1806
1807                                         zero_user(page, start, size);
1808                                         set_buffer_uptodate(bh);
1809                                 }
1810
1811                                 clear_buffer_new(bh);
1812                                 mark_buffer_dirty(bh);
1813                         }
1814                 }
1815
1816                 block_start = block_end;
1817                 bh = bh->b_this_page;
1818         } while (bh != head);
1819 }
1820 EXPORT_SYMBOL(page_zero_new_buffers);
1821
1822 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1823                 get_block_t *get_block)
1824 {
1825         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1826         unsigned to = from + len;
1827         struct inode *inode = page->mapping->host;
1828         unsigned block_start, block_end;
1829         sector_t block;
1830         int err = 0;
1831         unsigned blocksize, bbits;
1832         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1833
1834         BUG_ON(!PageLocked(page));
1835         BUG_ON(from > PAGE_CACHE_SIZE);
1836         BUG_ON(to > PAGE_CACHE_SIZE);
1837         BUG_ON(from > to);
1838
1839         head = create_page_buffers(page, inode, 0);
1840         blocksize = head->b_size;
1841         bbits = block_size_bits(blocksize);
1842
1843         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1844
1845         for(bh = head, block_start = 0; bh != head || !block_start;
1846             block++, block_start=block_end, bh = bh->b_this_page) {
1847                 block_end = block_start + blocksize;
1848                 if (block_end <= from || block_start >= to) {
1849                         if (PageUptodate(page)) {
1850                                 if (!buffer_uptodate(bh))
1851                                         set_buffer_uptodate(bh);
1852                         }
1853                         continue;
1854                 }
1855                 if (buffer_new(bh))
1856                         clear_buffer_new(bh);
1857                 if (!buffer_mapped(bh)) {
1858                         WARN_ON(bh->b_size != blocksize);
1859                         err = get_block(inode, block, bh, 1);
1860                         if (err)
1861                                 break;
1862                         if (buffer_new(bh)) {
1863                                 unmap_underlying_metadata(bh->b_bdev,
1864                                                         bh->b_blocknr);
1865                                 if (PageUptodate(page)) {
1866                                         clear_buffer_new(bh);
1867                                         set_buffer_uptodate(bh);
1868                                         mark_buffer_dirty(bh);
1869                                         continue;
1870                                 }
1871                                 if (block_end > to || block_start < from)
1872                                         zero_user_segments(page,
1873                                                 to, block_end,
1874                                                 block_start, from);
1875                                 continue;
1876                         }
1877                 }
1878                 if (PageUptodate(page)) {
1879                         if (!buffer_uptodate(bh))
1880                                 set_buffer_uptodate(bh);
1881                         continue; 
1882                 }
1883                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1884                     !buffer_unwritten(bh) &&
1885                      (block_start < from || block_end > to)) {
1886                         ll_rw_block(READ, 1, &bh);
1887                         *wait_bh++=bh;
1888                 }
1889         }
1890         /*
1891          * If we issued read requests - let them complete.
1892          */
1893         while(wait_bh > wait) {
1894                 wait_on_buffer(*--wait_bh);
1895                 if (!buffer_uptodate(*wait_bh))
1896                         err = -EIO;
1897         }
1898         if (unlikely(err))
1899                 page_zero_new_buffers(page, from, to);
1900         return err;
1901 }
1902 EXPORT_SYMBOL(__block_write_begin);
1903
1904 static int __block_commit_write(struct inode *inode, struct page *page,
1905                 unsigned from, unsigned to)
1906 {
1907         unsigned block_start, block_end;
1908         int partial = 0;
1909         unsigned blocksize;
1910         struct buffer_head *bh, *head;
1911
1912         bh = head = page_buffers(page);
1913         blocksize = bh->b_size;
1914
1915         block_start = 0;
1916         do {
1917                 block_end = block_start + blocksize;
1918                 if (block_end <= from || block_start >= to) {
1919                         if (!buffer_uptodate(bh))
1920                                 partial = 1;
1921                 } else {
1922                         set_buffer_uptodate(bh);
1923                         mark_buffer_dirty(bh);
1924                 }
1925                 clear_buffer_new(bh);
1926
1927                 block_start = block_end;
1928                 bh = bh->b_this_page;
1929         } while (bh != head);
1930
1931         /*
1932          * If this is a partial write which happened to make all buffers
1933          * uptodate then we can optimize away a bogus readpage() for
1934          * the next read(). Here we 'discover' whether the page went
1935          * uptodate as a result of this (potentially partial) write.
1936          */
1937         if (!partial)
1938                 SetPageUptodate(page);
1939         return 0;
1940 }
1941
1942 /*
1943  * block_write_begin takes care of the basic task of block allocation and
1944  * bringing partial write blocks uptodate first.
1945  *
1946  * The filesystem needs to handle block truncation upon failure.
1947  */
1948 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
1949                 unsigned flags, struct page **pagep, get_block_t *get_block)
1950 {
1951         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1952         struct page *page;
1953         int status;
1954
1955         page = grab_cache_page_write_begin(mapping, index, flags);
1956         if (!page)
1957                 return -ENOMEM;
1958
1959         status = __block_write_begin(page, pos, len, get_block);
1960         if (unlikely(status)) {
1961                 unlock_page(page);
1962                 page_cache_release(page);
1963                 page = NULL;
1964         }
1965
1966         *pagep = page;
1967         return status;
1968 }
1969 EXPORT_SYMBOL(block_write_begin);
1970
1971 int block_write_end(struct file *file, struct address_space *mapping,
1972                         loff_t pos, unsigned len, unsigned copied,
1973                         struct page *page, void *fsdata)
1974 {
1975         struct inode *inode = mapping->host;
1976         unsigned start;
1977
1978         start = pos & (PAGE_CACHE_SIZE - 1);
1979
1980         if (unlikely(copied < len)) {
1981                 /*
1982                  * The buffers that were written will now be uptodate, so we
1983                  * don't have to worry about a readpage reading them and
1984                  * overwriting a partial write. However if we have encountered
1985                  * a short write and only partially written into a buffer, it
1986                  * will not be marked uptodate, so a readpage might come in and
1987                  * destroy our partial write.
1988                  *
1989                  * Do the simplest thing, and just treat any short write to a
1990                  * non uptodate page as a zero-length write, and force the
1991                  * caller to redo the whole thing.
1992                  */
1993                 if (!PageUptodate(page))
1994                         copied = 0;
1995
1996                 page_zero_new_buffers(page, start+copied, start+len);
1997         }
1998         flush_dcache_page(page);
1999
2000         /* This could be a short (even 0-length) commit */
2001         __block_commit_write(inode, page, start, start+copied);
2002
2003         return copied;
2004 }
2005 EXPORT_SYMBOL(block_write_end);
2006
2007 int generic_write_end(struct file *file, struct address_space *mapping,
2008                         loff_t pos, unsigned len, unsigned copied,
2009                         struct page *page, void *fsdata)
2010 {
2011         struct inode *inode = mapping->host;
2012         int i_size_changed = 0;
2013
2014         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2015
2016         /*
2017          * No need to use i_size_read() here, the i_size
2018          * cannot change under us because we hold i_mutex.
2019          *
2020          * But it's important to update i_size while still holding page lock:
2021          * page writeout could otherwise come in and zero beyond i_size.
2022          */
2023         if (pos+copied > inode->i_size) {
2024                 i_size_write(inode, pos+copied);
2025                 i_size_changed = 1;
2026         }
2027
2028         unlock_page(page);
2029         page_cache_release(page);
2030
2031         /*
2032          * Don't mark the inode dirty under page lock. First, it unnecessarily
2033          * makes the holding time of page lock longer. Second, it forces lock
2034          * ordering of page lock and transaction start for journaling
2035          * filesystems.
2036          */
2037         if (i_size_changed)
2038                 mark_inode_dirty(inode);
2039
2040         return copied;
2041 }
2042 EXPORT_SYMBOL(generic_write_end);
2043
2044 /*
2045  * block_is_partially_uptodate checks whether buffers within a page are
2046  * uptodate or not.
2047  *
2048  * Returns true if all buffers which correspond to a file portion
2049  * we want to read are uptodate.
2050  */
2051 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2052                                         unsigned long from)
2053 {
2054         unsigned block_start, block_end, blocksize;
2055         unsigned to;
2056         struct buffer_head *bh, *head;
2057         int ret = 1;
2058
2059         if (!page_has_buffers(page))
2060                 return 0;
2061
2062         head = page_buffers(page);
2063         blocksize = head->b_size;
2064         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2065         to = from + to;
2066         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2067                 return 0;
2068
2069         bh = head;
2070         block_start = 0;
2071         do {
2072                 block_end = block_start + blocksize;
2073                 if (block_end > from && block_start < to) {
2074                         if (!buffer_uptodate(bh)) {
2075                                 ret = 0;
2076                                 break;
2077                         }
2078                         if (block_end >= to)
2079                                 break;
2080                 }
2081                 block_start = block_end;
2082                 bh = bh->b_this_page;
2083         } while (bh != head);
2084
2085         return ret;
2086 }
2087 EXPORT_SYMBOL(block_is_partially_uptodate);
2088
2089 /*
2090  * Generic "read page" function for block devices that have the normal
2091  * get_block functionality. This is most of the block device filesystems.
2092  * Reads the page asynchronously --- the unlock_buffer() and
2093  * set/clear_buffer_uptodate() functions propagate buffer state into the
2094  * page struct once IO has completed.
2095  */
2096 int block_read_full_page(struct page *page, get_block_t *get_block)
2097 {
2098         struct inode *inode = page->mapping->host;
2099         sector_t iblock, lblock;
2100         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2101         unsigned int blocksize, bbits;
2102         int nr, i;
2103         int fully_mapped = 1;
2104
2105         head = create_page_buffers(page, inode, 0);
2106         blocksize = head->b_size;
2107         bbits = block_size_bits(blocksize);
2108
2109         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
2110         lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2111         bh = head;
2112         nr = 0;
2113         i = 0;
2114
2115         do {
2116                 if (buffer_uptodate(bh))
2117                         continue;
2118
2119                 if (!buffer_mapped(bh)) {
2120                         int err = 0;
2121
2122                         fully_mapped = 0;
2123                         if (iblock < lblock) {
2124                                 WARN_ON(bh->b_size != blocksize);
2125                                 err = get_block(inode, iblock, bh, 0);
2126                                 if (err)
2127                                         SetPageError(page);
2128                         }
2129                         if (!buffer_mapped(bh)) {
2130                                 zero_user(page, i * blocksize, blocksize);
2131                                 if (!err)
2132                                         set_buffer_uptodate(bh);
2133                                 continue;
2134                         }
2135                         /*
2136                          * get_block() might have updated the buffer
2137                          * synchronously
2138                          */
2139                         if (buffer_uptodate(bh))
2140                                 continue;
2141                 }
2142                 arr[nr++] = bh;
2143         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2144
2145         if (fully_mapped)
2146                 SetPageMappedToDisk(page);
2147
2148         if (!nr) {
2149                 /*
2150                  * All buffers are uptodate - we can set the page uptodate
2151                  * as well. But not if get_block() returned an error.
2152                  */
2153                 if (!PageError(page))
2154                         SetPageUptodate(page);
2155                 unlock_page(page);
2156                 return 0;
2157         }
2158
2159         /* Stage two: lock the buffers */
2160         for (i = 0; i < nr; i++) {
2161                 bh = arr[i];
2162                 lock_buffer(bh);
2163                 mark_buffer_async_read(bh);
2164         }
2165
2166         /*
2167          * Stage 3: start the IO.  Check for uptodateness
2168          * inside the buffer lock in case another process reading
2169          * the underlying blockdev brought it uptodate (the sct fix).
2170          */
2171         for (i = 0; i < nr; i++) {
2172                 bh = arr[i];
2173                 if (buffer_uptodate(bh))
2174                         end_buffer_async_read(bh, 1);
2175                 else
2176                         submit_bh(READ, bh);
2177         }
2178         return 0;
2179 }
2180 EXPORT_SYMBOL(block_read_full_page);
2181
2182 /* utility function for filesystems that need to do work on expanding
2183  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2184  * deal with the hole.  
2185  */
2186 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2187 {
2188         struct address_space *mapping = inode->i_mapping;
2189         struct page *page;
2190         void *fsdata;
2191         int err;
2192
2193         err = inode_newsize_ok(inode, size);
2194         if (err)
2195                 goto out;
2196
2197         err = pagecache_write_begin(NULL, mapping, size, 0,
2198                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2199                                 &page, &fsdata);
2200         if (err)
2201                 goto out;
2202
2203         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2204         BUG_ON(err > 0);
2205
2206 out:
2207         return err;
2208 }
2209 EXPORT_SYMBOL(generic_cont_expand_simple);
2210
2211 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2212                             loff_t pos, loff_t *bytes)
2213 {
2214         struct inode *inode = mapping->host;
2215         unsigned blocksize = 1 << inode->i_blkbits;
2216         struct page *page;
2217         void *fsdata;
2218         pgoff_t index, curidx;
2219         loff_t curpos;
2220         unsigned zerofrom, offset, len;
2221         int err = 0;
2222
2223         index = pos >> PAGE_CACHE_SHIFT;
2224         offset = pos & ~PAGE_CACHE_MASK;
2225
2226         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2227                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2228                 if (zerofrom & (blocksize-1)) {
2229                         *bytes |= (blocksize-1);
2230                         (*bytes)++;
2231                 }
2232                 len = PAGE_CACHE_SIZE - zerofrom;
2233
2234                 err = pagecache_write_begin(file, mapping, curpos, len,
2235                                                 AOP_FLAG_UNINTERRUPTIBLE,
2236                                                 &page, &fsdata);
2237                 if (err)
2238                         goto out;
2239                 zero_user(page, zerofrom, len);
2240                 err = pagecache_write_end(file, mapping, curpos, len, len,
2241                                                 page, fsdata);
2242                 if (err < 0)
2243                         goto out;
2244                 BUG_ON(err != len);
2245                 err = 0;
2246
2247                 balance_dirty_pages_ratelimited(mapping);
2248         }
2249
2250         /* page covers the boundary, find the boundary offset */
2251         if (index == curidx) {
2252                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2253                 /* if we will expand the thing last block will be filled */
2254                 if (offset <= zerofrom) {
2255                         goto out;
2256                 }
2257                 if (zerofrom & (blocksize-1)) {
2258                         *bytes |= (blocksize-1);
2259                         (*bytes)++;
2260                 }
2261                 len = offset - zerofrom;
2262
2263                 err = pagecache_write_begin(file, mapping, curpos, len,
2264                                                 AOP_FLAG_UNINTERRUPTIBLE,
2265                                                 &page, &fsdata);
2266                 if (err)
2267                         goto out;
2268                 zero_user(page, zerofrom, len);
2269                 err = pagecache_write_end(file, mapping, curpos, len, len,
2270                                                 page, fsdata);
2271                 if (err < 0)
2272                         goto out;
2273                 BUG_ON(err != len);
2274                 err = 0;
2275         }
2276 out:
2277         return err;
2278 }
2279
2280 /*
2281  * For moronic filesystems that do not allow holes in file.
2282  * We may have to extend the file.
2283  */
2284 int cont_write_begin(struct file *file, struct address_space *mapping,
2285                         loff_t pos, unsigned len, unsigned flags,
2286                         struct page **pagep, void **fsdata,
2287                         get_block_t *get_block, loff_t *bytes)
2288 {
2289         struct inode *inode = mapping->host;
2290         unsigned blocksize = 1 << inode->i_blkbits;
2291         unsigned zerofrom;
2292         int err;
2293
2294         err = cont_expand_zero(file, mapping, pos, bytes);
2295         if (err)
2296                 return err;
2297
2298         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2299         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2300                 *bytes |= (blocksize-1);
2301                 (*bytes)++;
2302         }
2303
2304         return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2305 }
2306 EXPORT_SYMBOL(cont_write_begin);
2307
2308 int block_commit_write(struct page *page, unsigned from, unsigned to)
2309 {
2310         struct inode *inode = page->mapping->host;
2311         __block_commit_write(inode,page,from,to);
2312         return 0;
2313 }
2314 EXPORT_SYMBOL(block_commit_write);
2315
2316 /*
2317  * block_page_mkwrite() is not allowed to change the file size as it gets
2318  * called from a page fault handler when a page is first dirtied. Hence we must
2319  * be careful to check for EOF conditions here. We set the page up correctly
2320  * for a written page which means we get ENOSPC checking when writing into
2321  * holes and correct delalloc and unwritten extent mapping on filesystems that
2322  * support these features.
2323  *
2324  * We are not allowed to take the i_mutex here so we have to play games to
2325  * protect against truncate races as the page could now be beyond EOF.  Because
2326  * truncate writes the inode size before removing pages, once we have the
2327  * page lock we can determine safely if the page is beyond EOF. If it is not
2328  * beyond EOF, then the page is guaranteed safe against truncation until we
2329  * unlock the page.
2330  *
2331  * Direct callers of this function should protect against filesystem freezing
2332  * using sb_start_write() - sb_end_write() functions.
2333  */
2334 int __block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2335                          get_block_t get_block)
2336 {
2337         struct page *page = vmf->page;
2338         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2339         unsigned long end;
2340         loff_t size;
2341         int ret;
2342
2343         lock_page(page);
2344         size = i_size_read(inode);
2345         if ((page->mapping != inode->i_mapping) ||
2346             (page_offset(page) > size)) {
2347                 /* We overload EFAULT to mean page got truncated */
2348                 ret = -EFAULT;
2349                 goto out_unlock;
2350         }
2351
2352         /* page is wholly or partially inside EOF */
2353         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2354                 end = size & ~PAGE_CACHE_MASK;
2355         else
2356                 end = PAGE_CACHE_SIZE;
2357
2358         ret = __block_write_begin(page, 0, end, get_block);
2359         if (!ret)
2360                 ret = block_commit_write(page, 0, end);
2361
2362         if (unlikely(ret < 0))
2363                 goto out_unlock;
2364         set_page_dirty(page);
2365         wait_on_page_writeback(page);
2366         return 0;
2367 out_unlock:
2368         unlock_page(page);
2369         return ret;
2370 }
2371 EXPORT_SYMBOL(__block_page_mkwrite);
2372
2373 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2374                    get_block_t get_block)
2375 {
2376         int ret;
2377         struct super_block *sb = vma->vm_file->f_path.dentry->d_inode->i_sb;
2378
2379         sb_start_pagefault(sb);
2380
2381         /*
2382          * Update file times before taking page lock. We may end up failing the
2383          * fault so this update may be superfluous but who really cares...
2384          */
2385         file_update_time(vma->vm_file);
2386
2387         ret = __block_page_mkwrite(vma, vmf, get_block);
2388         sb_end_pagefault(sb);
2389         return block_page_mkwrite_return(ret);
2390 }
2391 EXPORT_SYMBOL(block_page_mkwrite);
2392
2393 /*
2394  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2395  * immediately, while under the page lock.  So it needs a special end_io
2396  * handler which does not touch the bh after unlocking it.
2397  */
2398 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2399 {
2400         __end_buffer_read_notouch(bh, uptodate);
2401 }
2402
2403 /*
2404  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2405  * the page (converting it to circular linked list and taking care of page
2406  * dirty races).
2407  */
2408 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2409 {
2410         struct buffer_head *bh;
2411
2412         BUG_ON(!PageLocked(page));
2413
2414         spin_lock(&page->mapping->private_lock);
2415         bh = head;
2416         do {
2417                 if (PageDirty(page))
2418                         set_buffer_dirty(bh);
2419                 if (!bh->b_this_page)
2420                         bh->b_this_page = head;
2421                 bh = bh->b_this_page;
2422         } while (bh != head);
2423         attach_page_buffers(page, head);
2424         spin_unlock(&page->mapping->private_lock);
2425 }
2426
2427 /*
2428  * On entry, the page is fully not uptodate.
2429  * On exit the page is fully uptodate in the areas outside (from,to)
2430  * The filesystem needs to handle block truncation upon failure.
2431  */
2432 int nobh_write_begin(struct address_space *mapping,
2433                         loff_t pos, unsigned len, unsigned flags,
2434                         struct page **pagep, void **fsdata,
2435                         get_block_t *get_block)
2436 {
2437         struct inode *inode = mapping->host;
2438         const unsigned blkbits = inode->i_blkbits;
2439         const unsigned blocksize = 1 << blkbits;
2440         struct buffer_head *head, *bh;
2441         struct page *page;
2442         pgoff_t index;
2443         unsigned from, to;
2444         unsigned block_in_page;
2445         unsigned block_start, block_end;
2446         sector_t block_in_file;
2447         int nr_reads = 0;
2448         int ret = 0;
2449         int is_mapped_to_disk = 1;
2450
2451         index = pos >> PAGE_CACHE_SHIFT;
2452         from = pos & (PAGE_CACHE_SIZE - 1);
2453         to = from + len;
2454
2455         page = grab_cache_page_write_begin(mapping, index, flags);
2456         if (!page)
2457                 return -ENOMEM;
2458         *pagep = page;
2459         *fsdata = NULL;
2460
2461         if (page_has_buffers(page)) {
2462                 ret = __block_write_begin(page, pos, len, get_block);
2463                 if (unlikely(ret))
2464                         goto out_release;
2465                 return ret;
2466         }
2467
2468         if (PageMappedToDisk(page))
2469                 return 0;
2470
2471         /*
2472          * Allocate buffers so that we can keep track of state, and potentially
2473          * attach them to the page if an error occurs. In the common case of
2474          * no error, they will just be freed again without ever being attached
2475          * to the page (which is all OK, because we're under the page lock).
2476          *
2477          * Be careful: the buffer linked list is a NULL terminated one, rather
2478          * than the circular one we're used to.
2479          */
2480         head = alloc_page_buffers(page, blocksize, 0);
2481         if (!head) {
2482                 ret = -ENOMEM;
2483                 goto out_release;
2484         }
2485
2486         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2487
2488         /*
2489          * We loop across all blocks in the page, whether or not they are
2490          * part of the affected region.  This is so we can discover if the
2491          * page is fully mapped-to-disk.
2492          */
2493         for (block_start = 0, block_in_page = 0, bh = head;
2494                   block_start < PAGE_CACHE_SIZE;
2495                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2496                 int create;
2497
2498                 block_end = block_start + blocksize;
2499                 bh->b_state = 0;
2500                 create = 1;
2501                 if (block_start >= to)
2502                         create = 0;
2503                 ret = get_block(inode, block_in_file + block_in_page,
2504                                         bh, create);
2505                 if (ret)
2506                         goto failed;
2507                 if (!buffer_mapped(bh))
2508                         is_mapped_to_disk = 0;
2509                 if (buffer_new(bh))
2510                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2511                 if (PageUptodate(page)) {
2512                         set_buffer_uptodate(bh);
2513                         continue;
2514                 }
2515                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2516                         zero_user_segments(page, block_start, from,
2517                                                         to, block_end);
2518                         continue;
2519                 }
2520                 if (buffer_uptodate(bh))
2521                         continue;       /* reiserfs does this */
2522                 if (block_start < from || block_end > to) {
2523                         lock_buffer(bh);
2524                         bh->b_end_io = end_buffer_read_nobh;
2525                         submit_bh(READ, bh);
2526                         nr_reads++;
2527                 }
2528         }
2529
2530         if (nr_reads) {
2531                 /*
2532                  * The page is locked, so these buffers are protected from
2533                  * any VM or truncate activity.  Hence we don't need to care
2534                  * for the buffer_head refcounts.
2535                  */
2536                 for (bh = head; bh; bh = bh->b_this_page) {
2537                         wait_on_buffer(bh);
2538                         if (!buffer_uptodate(bh))
2539                                 ret = -EIO;
2540                 }
2541                 if (ret)
2542                         goto failed;
2543         }
2544
2545         if (is_mapped_to_disk)
2546                 SetPageMappedToDisk(page);
2547
2548         *fsdata = head; /* to be released by nobh_write_end */
2549
2550         return 0;
2551
2552 failed:
2553         BUG_ON(!ret);
2554         /*
2555          * Error recovery is a bit difficult. We need to zero out blocks that
2556          * were newly allocated, and dirty them to ensure they get written out.
2557          * Buffers need to be attached to the page at this point, otherwise
2558          * the handling of potential IO errors during writeout would be hard
2559          * (could try doing synchronous writeout, but what if that fails too?)
2560          */
2561         attach_nobh_buffers(page, head);
2562         page_zero_new_buffers(page, from, to);
2563
2564 out_release:
2565         unlock_page(page);
2566         page_cache_release(page);
2567         *pagep = NULL;
2568
2569         return ret;
2570 }
2571 EXPORT_SYMBOL(nobh_write_begin);
2572
2573 int nobh_write_end(struct file *file, struct address_space *mapping,
2574                         loff_t pos, unsigned len, unsigned copied,
2575                         struct page *page, void *fsdata)
2576 {
2577         struct inode *inode = page->mapping->host;
2578         struct buffer_head *head = fsdata;
2579         struct buffer_head *bh;
2580         BUG_ON(fsdata != NULL && page_has_buffers(page));
2581
2582         if (unlikely(copied < len) && head)
2583                 attach_nobh_buffers(page, head);
2584         if (page_has_buffers(page))
2585                 return generic_write_end(file, mapping, pos, len,
2586                                         copied, page, fsdata);
2587
2588         SetPageUptodate(page);
2589         set_page_dirty(page);
2590         if (pos+copied > inode->i_size) {
2591                 i_size_write(inode, pos+copied);
2592                 mark_inode_dirty(inode);
2593         }
2594
2595         unlock_page(page);
2596         page_cache_release(page);
2597
2598         while (head) {
2599                 bh = head;
2600                 head = head->b_this_page;
2601                 free_buffer_head(bh);
2602         }
2603
2604         return copied;
2605 }
2606 EXPORT_SYMBOL(nobh_write_end);
2607
2608 /*
2609  * nobh_writepage() - based on block_full_write_page() except
2610  * that it tries to operate without attaching bufferheads to
2611  * the page.
2612  */
2613 int nobh_writepage(struct page *page, get_block_t *get_block,
2614                         struct writeback_control *wbc)
2615 {
2616         struct inode * const inode = page->mapping->host;
2617         loff_t i_size = i_size_read(inode);
2618         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2619         unsigned offset;
2620         int ret;
2621
2622         /* Is the page fully inside i_size? */
2623         if (page->index < end_index)
2624                 goto out;
2625
2626         /* Is the page fully outside i_size? (truncate in progress) */
2627         offset = i_size & (PAGE_CACHE_SIZE-1);
2628         if (page->index >= end_index+1 || !offset) {
2629                 /*
2630                  * The page may have dirty, unmapped buffers.  For example,
2631                  * they may have been added in ext3_writepage().  Make them
2632                  * freeable here, so the page does not leak.
2633                  */
2634 #if 0
2635                 /* Not really sure about this  - do we need this ? */
2636                 if (page->mapping->a_ops->invalidatepage)
2637                         page->mapping->a_ops->invalidatepage(page, offset);
2638 #endif
2639                 unlock_page(page);
2640                 return 0; /* don't care */
2641         }
2642
2643         /*
2644          * The page straddles i_size.  It must be zeroed out on each and every
2645          * writepage invocation because it may be mmapped.  "A file is mapped
2646          * in multiples of the page size.  For a file that is not a multiple of
2647          * the  page size, the remaining memory is zeroed when mapped, and
2648          * writes to that region are not written out to the file."
2649          */
2650         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2651 out:
2652         ret = mpage_writepage(page, get_block, wbc);
2653         if (ret == -EAGAIN)
2654                 ret = __block_write_full_page(inode, page, get_block, wbc,
2655                                               end_buffer_async_write);
2656         return ret;
2657 }
2658 EXPORT_SYMBOL(nobh_writepage);
2659
2660 int nobh_truncate_page(struct address_space *mapping,
2661                         loff_t from, get_block_t *get_block)
2662 {
2663         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2664         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2665         unsigned blocksize;
2666         sector_t iblock;
2667         unsigned length, pos;
2668         struct inode *inode = mapping->host;
2669         struct page *page;
2670         struct buffer_head map_bh;
2671         int err;
2672
2673         blocksize = 1 << inode->i_blkbits;
2674         length = offset & (blocksize - 1);
2675
2676         /* Block boundary? Nothing to do */
2677         if (!length)
2678                 return 0;
2679
2680         length = blocksize - length;
2681         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2682
2683         page = grab_cache_page(mapping, index);
2684         err = -ENOMEM;
2685         if (!page)
2686                 goto out;
2687
2688         if (page_has_buffers(page)) {
2689 has_buffers:
2690                 unlock_page(page);
2691                 page_cache_release(page);
2692                 return block_truncate_page(mapping, from, get_block);
2693         }
2694
2695         /* Find the buffer that contains "offset" */
2696         pos = blocksize;
2697         while (offset >= pos) {
2698                 iblock++;
2699                 pos += blocksize;
2700         }
2701
2702         map_bh.b_size = blocksize;
2703         map_bh.b_state = 0;
2704         err = get_block(inode, iblock, &map_bh, 0);
2705         if (err)
2706                 goto unlock;
2707         /* unmapped? It's a hole - nothing to do */
2708         if (!buffer_mapped(&map_bh))
2709                 goto unlock;
2710
2711         /* Ok, it's mapped. Make sure it's up-to-date */
2712         if (!PageUptodate(page)) {
2713                 err = mapping->a_ops->readpage(NULL, page);
2714                 if (err) {
2715                         page_cache_release(page);
2716                         goto out;
2717                 }
2718                 lock_page(page);
2719                 if (!PageUptodate(page)) {
2720                         err = -EIO;
2721                         goto unlock;
2722                 }
2723                 if (page_has_buffers(page))
2724                         goto has_buffers;
2725         }
2726         zero_user(page, offset, length);
2727         set_page_dirty(page);
2728         err = 0;
2729
2730 unlock:
2731         unlock_page(page);
2732         page_cache_release(page);
2733 out:
2734         return err;
2735 }
2736 EXPORT_SYMBOL(nobh_truncate_page);
2737
2738 int block_truncate_page(struct address_space *mapping,
2739                         loff_t from, get_block_t *get_block)
2740 {
2741         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2742         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2743         unsigned blocksize;
2744         sector_t iblock;
2745         unsigned length, pos;
2746         struct inode *inode = mapping->host;
2747         struct page *page;
2748         struct buffer_head *bh;
2749         int err;
2750
2751         blocksize = 1 << inode->i_blkbits;
2752         length = offset & (blocksize - 1);
2753
2754         /* Block boundary? Nothing to do */
2755         if (!length)
2756                 return 0;
2757
2758         length = blocksize - length;
2759         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2760         
2761         page = grab_cache_page(mapping, index);
2762         err = -ENOMEM;
2763         if (!page)
2764                 goto out;
2765
2766         if (!page_has_buffers(page))
2767                 create_empty_buffers(page, blocksize, 0);
2768
2769         /* Find the buffer that contains "offset" */
2770         bh = page_buffers(page);
2771         pos = blocksize;
2772         while (offset >= pos) {
2773                 bh = bh->b_this_page;
2774                 iblock++;
2775                 pos += blocksize;
2776         }
2777
2778         err = 0;
2779         if (!buffer_mapped(bh)) {
2780                 WARN_ON(bh->b_size != blocksize);
2781                 err = get_block(inode, iblock, bh, 0);
2782                 if (err)
2783                         goto unlock;
2784                 /* unmapped? It's a hole - nothing to do */
2785                 if (!buffer_mapped(bh))
2786                         goto unlock;
2787         }
2788
2789         /* Ok, it's mapped. Make sure it's up-to-date */
2790         if (PageUptodate(page))
2791                 set_buffer_uptodate(bh);
2792
2793         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2794                 err = -EIO;
2795                 ll_rw_block(READ, 1, &bh);
2796                 wait_on_buffer(bh);
2797                 /* Uhhuh. Read error. Complain and punt. */
2798                 if (!buffer_uptodate(bh))
2799                         goto unlock;
2800         }
2801
2802         zero_user(page, offset, length);
2803         mark_buffer_dirty(bh);
2804         err = 0;
2805
2806 unlock:
2807         unlock_page(page);
2808         page_cache_release(page);
2809 out:
2810         return err;
2811 }
2812 EXPORT_SYMBOL(block_truncate_page);
2813
2814 /*
2815  * The generic ->writepage function for buffer-backed address_spaces
2816  * this form passes in the end_io handler used to finish the IO.
2817  */
2818 int block_write_full_page_endio(struct page *page, get_block_t *get_block,
2819                         struct writeback_control *wbc, bh_end_io_t *handler)
2820 {
2821         struct inode * const inode = page->mapping->host;
2822         loff_t i_size = i_size_read(inode);
2823         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2824         unsigned offset;
2825
2826         /* Is the page fully inside i_size? */
2827         if (page->index < end_index)
2828                 return __block_write_full_page(inode, page, get_block, wbc,
2829                                                handler);
2830
2831         /* Is the page fully outside i_size? (truncate in progress) */
2832         offset = i_size & (PAGE_CACHE_SIZE-1);
2833         if (page->index >= end_index+1 || !offset) {
2834                 /*
2835                  * The page may have dirty, unmapped buffers.  For example,
2836                  * they may have been added in ext3_writepage().  Make them
2837                  * freeable here, so the page does not leak.
2838                  */
2839                 do_invalidatepage(page, 0);
2840                 unlock_page(page);
2841                 return 0; /* don't care */
2842         }
2843
2844         /*
2845          * The page straddles i_size.  It must be zeroed out on each and every
2846          * writepage invocation because it may be mmapped.  "A file is mapped
2847          * in multiples of the page size.  For a file that is not a multiple of
2848          * the  page size, the remaining memory is zeroed when mapped, and
2849          * writes to that region are not written out to the file."
2850          */
2851         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2852         return __block_write_full_page(inode, page, get_block, wbc, handler);
2853 }
2854 EXPORT_SYMBOL(block_write_full_page_endio);
2855
2856 /*
2857  * The generic ->writepage function for buffer-backed address_spaces
2858  */
2859 int block_write_full_page(struct page *page, get_block_t *get_block,
2860                         struct writeback_control *wbc)
2861 {
2862         return block_write_full_page_endio(page, get_block, wbc,
2863                                            end_buffer_async_write);
2864 }
2865 EXPORT_SYMBOL(block_write_full_page);
2866
2867 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2868                             get_block_t *get_block)
2869 {
2870         struct buffer_head tmp;
2871         struct inode *inode = mapping->host;
2872         tmp.b_state = 0;
2873         tmp.b_blocknr = 0;
2874         tmp.b_size = 1 << inode->i_blkbits;
2875         get_block(inode, block, &tmp, 0);
2876         return tmp.b_blocknr;
2877 }
2878 EXPORT_SYMBOL(generic_block_bmap);
2879
2880 static void end_bio_bh_io_sync(struct bio *bio, int err)
2881 {
2882         struct buffer_head *bh = bio->bi_private;
2883
2884         if (err == -EOPNOTSUPP) {
2885                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2886         }
2887
2888         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2889                 set_bit(BH_Quiet, &bh->b_state);
2890
2891         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2892         bio_put(bio);
2893 }
2894
2895 /*
2896  * This allows us to do IO even on the odd last sectors
2897  * of a device, even if the bh block size is some multiple
2898  * of the physical sector size.
2899  *
2900  * We'll just truncate the bio to the size of the device,
2901  * and clear the end of the buffer head manually.
2902  *
2903  * Truly out-of-range accesses will turn into actual IO
2904  * errors, this only handles the "we need to be able to
2905  * do IO at the final sector" case.
2906  */
2907 static void guard_bh_eod(int rw, struct bio *bio, struct buffer_head *bh)
2908 {
2909         sector_t maxsector;
2910         unsigned bytes;
2911
2912         maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
2913         if (!maxsector)
2914                 return;
2915
2916         /*
2917          * If the *whole* IO is past the end of the device,
2918          * let it through, and the IO layer will turn it into
2919          * an EIO.
2920          */
2921         if (unlikely(bio->bi_sector >= maxsector))
2922                 return;
2923
2924         maxsector -= bio->bi_sector;
2925         bytes = bio->bi_size;
2926         if (likely((bytes >> 9) <= maxsector))
2927                 return;
2928
2929         /* Uhhuh. We've got a bh that straddles the device size! */
2930         bytes = maxsector << 9;
2931
2932         /* Truncate the bio.. */
2933         bio->bi_size = bytes;
2934         bio->bi_io_vec[0].bv_len = bytes;
2935
2936         /* ..and clear the end of the buffer for reads */
2937         if ((rw & RW_MASK) == READ) {
2938                 void *kaddr = kmap_atomic(bh->b_page);
2939                 memset(kaddr + bh_offset(bh) + bytes, 0, bh->b_size - bytes);
2940                 kunmap_atomic(kaddr);
2941         }
2942 }
2943
2944 int submit_bh(int rw, struct buffer_head * bh)
2945 {
2946         struct bio *bio;
2947         int ret = 0;
2948
2949         BUG_ON(!buffer_locked(bh));
2950         BUG_ON(!buffer_mapped(bh));
2951         BUG_ON(!bh->b_end_io);
2952         BUG_ON(buffer_delay(bh));
2953         BUG_ON(buffer_unwritten(bh));
2954
2955         /*
2956          * Only clear out a write error when rewriting
2957          */
2958         if (test_set_buffer_req(bh) && (rw & WRITE))
2959                 clear_buffer_write_io_error(bh);
2960
2961         /*
2962          * from here on down, it's all bio -- do the initial mapping,
2963          * submit_bio -> generic_make_request may further map this bio around
2964          */
2965         bio = bio_alloc(GFP_NOIO, 1);
2966
2967         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2968         bio->bi_bdev = bh->b_bdev;
2969         bio->bi_io_vec[0].bv_page = bh->b_page;
2970         bio->bi_io_vec[0].bv_len = bh->b_size;
2971         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2972
2973         bio->bi_vcnt = 1;
2974         bio->bi_idx = 0;
2975         bio->bi_size = bh->b_size;
2976
2977         bio->bi_end_io = end_bio_bh_io_sync;
2978         bio->bi_private = bh;
2979
2980         /* Take care of bh's that straddle the end of the device */
2981         guard_bh_eod(rw, bio, bh);
2982
2983         bio_get(bio);
2984         submit_bio(rw, bio);
2985
2986         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2987                 ret = -EOPNOTSUPP;
2988
2989         bio_put(bio);
2990         return ret;
2991 }
2992 EXPORT_SYMBOL(submit_bh);
2993
2994 /**
2995  * ll_rw_block: low-level access to block devices (DEPRECATED)
2996  * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
2997  * @nr: number of &struct buffer_heads in the array
2998  * @bhs: array of pointers to &struct buffer_head
2999  *
3000  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3001  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
3002  * %READA option is described in the documentation for generic_make_request()
3003  * which ll_rw_block() calls.
3004  *
3005  * This function drops any buffer that it cannot get a lock on (with the
3006  * BH_Lock state bit), any buffer that appears to be clean when doing a write
3007  * request, and any buffer that appears to be up-to-date when doing read
3008  * request.  Further it marks as clean buffers that are processed for
3009  * writing (the buffer cache won't assume that they are actually clean
3010  * until the buffer gets unlocked).
3011  *
3012  * ll_rw_block sets b_end_io to simple completion handler that marks
3013  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
3014  * any waiters. 
3015  *
3016  * All of the buffers must be for the same device, and must also be a
3017  * multiple of the current approved size for the device.
3018  */
3019 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
3020 {
3021         int i;
3022
3023         for (i = 0; i < nr; i++) {
3024                 struct buffer_head *bh = bhs[i];
3025
3026                 if (!trylock_buffer(bh))
3027                         continue;
3028                 if (rw == WRITE) {
3029                         if (test_clear_buffer_dirty(bh)) {
3030                                 bh->b_end_io = end_buffer_write_sync;
3031                                 get_bh(bh);
3032                                 submit_bh(WRITE, bh);
3033                                 continue;
3034                         }
3035                 } else {
3036                         if (!buffer_uptodate(bh)) {
3037                                 bh->b_end_io = end_buffer_read_sync;
3038                                 get_bh(bh);
3039                                 submit_bh(rw, bh);
3040                                 continue;
3041                         }
3042                 }
3043                 unlock_buffer(bh);
3044         }
3045 }
3046 EXPORT_SYMBOL(ll_rw_block);
3047
3048 void write_dirty_buffer(struct buffer_head *bh, int rw)
3049 {
3050         lock_buffer(bh);
3051         if (!test_clear_buffer_dirty(bh)) {
3052                 unlock_buffer(bh);
3053                 return;
3054         }
3055         bh->b_end_io = end_buffer_write_sync;
3056         get_bh(bh);
3057         submit_bh(rw, bh);
3058 }
3059 EXPORT_SYMBOL(write_dirty_buffer);
3060
3061 /*
3062  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3063  * and then start new I/O and then wait upon it.  The caller must have a ref on
3064  * the buffer_head.
3065  */
3066 int __sync_dirty_buffer(struct buffer_head *bh, int rw)
3067 {
3068         int ret = 0;
3069
3070         WARN_ON(atomic_read(&bh->b_count) < 1);
3071         lock_buffer(bh);
3072         if (test_clear_buffer_dirty(bh)) {
3073                 get_bh(bh);
3074                 bh->b_end_io = end_buffer_write_sync;
3075                 ret = submit_bh(rw, bh);
3076                 wait_on_buffer(bh);
3077                 if (!ret && !buffer_uptodate(bh))
3078                         ret = -EIO;
3079         } else {
3080                 unlock_buffer(bh);
3081         }
3082         return ret;
3083 }
3084 EXPORT_SYMBOL(__sync_dirty_buffer);
3085
3086 int sync_dirty_buffer(struct buffer_head *bh)
3087 {
3088         return __sync_dirty_buffer(bh, WRITE_SYNC);
3089 }
3090 EXPORT_SYMBOL(sync_dirty_buffer);
3091
3092 /*
3093  * try_to_free_buffers() checks if all the buffers on this particular page
3094  * are unused, and releases them if so.
3095  *
3096  * Exclusion against try_to_free_buffers may be obtained by either
3097  * locking the page or by holding its mapping's private_lock.
3098  *
3099  * If the page is dirty but all the buffers are clean then we need to
3100  * be sure to mark the page clean as well.  This is because the page
3101  * may be against a block device, and a later reattachment of buffers
3102  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3103  * filesystem data on the same device.
3104  *
3105  * The same applies to regular filesystem pages: if all the buffers are
3106  * clean then we set the page clean and proceed.  To do that, we require
3107  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3108  * private_lock.
3109  *
3110  * try_to_free_buffers() is non-blocking.
3111  */
3112 static inline int buffer_busy(struct buffer_head *bh)
3113 {
3114         return atomic_read(&bh->b_count) |
3115                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3116 }
3117
3118 static int
3119 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3120 {
3121         struct buffer_head *head = page_buffers(page);
3122         struct buffer_head *bh;
3123
3124         bh = head;
3125         do {
3126                 if (buffer_write_io_error(bh) && page->mapping)
3127                         set_bit(AS_EIO, &page->mapping->flags);
3128                 if (buffer_busy(bh))
3129                         goto failed;
3130                 bh = bh->b_this_page;
3131         } while (bh != head);
3132
3133         do {
3134                 struct buffer_head *next = bh->b_this_page;
3135
3136                 if (bh->b_assoc_map)
3137                         __remove_assoc_queue(bh);
3138                 bh = next;
3139         } while (bh != head);
3140         *buffers_to_free = head;
3141         __clear_page_buffers(page);
3142         return 1;
3143 failed:
3144         return 0;
3145 }
3146
3147 int try_to_free_buffers(struct page *page)
3148 {
3149         struct address_space * const mapping = page->mapping;
3150         struct buffer_head *buffers_to_free = NULL;
3151         int ret = 0;
3152
3153         BUG_ON(!PageLocked(page));
3154         if (PageWriteback(page))
3155                 return 0;
3156
3157         if (mapping == NULL) {          /* can this still happen? */
3158                 ret = drop_buffers(page, &buffers_to_free);
3159                 goto out;
3160         }
3161
3162         spin_lock(&mapping->private_lock);
3163         ret = drop_buffers(page, &buffers_to_free);
3164
3165         /*
3166          * If the filesystem writes its buffers by hand (eg ext3)
3167          * then we can have clean buffers against a dirty page.  We
3168          * clean the page here; otherwise the VM will never notice
3169          * that the filesystem did any IO at all.
3170          *
3171          * Also, during truncate, discard_buffer will have marked all
3172          * the page's buffers clean.  We discover that here and clean
3173          * the page also.
3174          *
3175          * private_lock must be held over this entire operation in order
3176          * to synchronise against __set_page_dirty_buffers and prevent the
3177          * dirty bit from being lost.
3178          */
3179         if (ret)
3180                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3181         spin_unlock(&mapping->private_lock);
3182 out:
3183         if (buffers_to_free) {
3184                 struct buffer_head *bh = buffers_to_free;
3185
3186                 do {
3187                         struct buffer_head *next = bh->b_this_page;
3188                         free_buffer_head(bh);
3189                         bh = next;
3190                 } while (bh != buffers_to_free);
3191         }
3192         return ret;
3193 }
3194 EXPORT_SYMBOL(try_to_free_buffers);
3195
3196 /*
3197  * There are no bdflush tunables left.  But distributions are
3198  * still running obsolete flush daemons, so we terminate them here.
3199  *
3200  * Use of bdflush() is deprecated and will be removed in a future kernel.
3201  * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3202  */
3203 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3204 {
3205         static int msg_count;
3206
3207         if (!capable(CAP_SYS_ADMIN))
3208                 return -EPERM;
3209
3210         if (msg_count < 5) {
3211                 msg_count++;
3212                 printk(KERN_INFO
3213                         "warning: process `%s' used the obsolete bdflush"
3214                         " system call\n", current->comm);
3215                 printk(KERN_INFO "Fix your initscripts?\n");
3216         }
3217
3218         if (func == 1)
3219                 do_exit(0);
3220         return 0;
3221 }
3222
3223 /*
3224  * Buffer-head allocation
3225  */
3226 static struct kmem_cache *bh_cachep __read_mostly;
3227
3228 /*
3229  * Once the number of bh's in the machine exceeds this level, we start
3230  * stripping them in writeback.
3231  */
3232 static int max_buffer_heads;
3233
3234 int buffer_heads_over_limit;
3235
3236 struct bh_accounting {
3237         int nr;                 /* Number of live bh's */
3238         int ratelimit;          /* Limit cacheline bouncing */
3239 };
3240
3241 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3242
3243 static void recalc_bh_state(void)
3244 {
3245         int i;
3246         int tot = 0;
3247
3248         if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3249                 return;
3250         __this_cpu_write(bh_accounting.ratelimit, 0);
3251         for_each_online_cpu(i)
3252                 tot += per_cpu(bh_accounting, i).nr;
3253         buffer_heads_over_limit = (tot > max_buffer_heads);
3254 }
3255
3256 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3257 {
3258         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3259         if (ret) {
3260                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3261                 preempt_disable();
3262                 __this_cpu_inc(bh_accounting.nr);
3263                 recalc_bh_state();
3264                 preempt_enable();
3265         }
3266         return ret;
3267 }
3268 EXPORT_SYMBOL(alloc_buffer_head);
3269
3270 void free_buffer_head(struct buffer_head *bh)
3271 {
3272         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3273         kmem_cache_free(bh_cachep, bh);
3274         preempt_disable();
3275         __this_cpu_dec(bh_accounting.nr);
3276         recalc_bh_state();
3277         preempt_enable();
3278 }
3279 EXPORT_SYMBOL(free_buffer_head);
3280
3281 static void buffer_exit_cpu(int cpu)
3282 {
3283         int i;
3284         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3285
3286         for (i = 0; i < BH_LRU_SIZE; i++) {
3287                 brelse(b->bhs[i]);
3288                 b->bhs[i] = NULL;
3289         }
3290         this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3291         per_cpu(bh_accounting, cpu).nr = 0;
3292 }
3293
3294 static int buffer_cpu_notify(struct notifier_block *self,
3295                               unsigned long action, void *hcpu)
3296 {
3297         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3298                 buffer_exit_cpu((unsigned long)hcpu);
3299         return NOTIFY_OK;
3300 }
3301
3302 /**
3303  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3304  * @bh: struct buffer_head
3305  *
3306  * Return true if the buffer is up-to-date and false,
3307  * with the buffer locked, if not.
3308  */
3309 int bh_uptodate_or_lock(struct buffer_head *bh)
3310 {
3311         if (!buffer_uptodate(bh)) {
3312                 lock_buffer(bh);
3313                 if (!buffer_uptodate(bh))
3314                         return 0;
3315                 unlock_buffer(bh);
3316         }
3317         return 1;
3318 }
3319 EXPORT_SYMBOL(bh_uptodate_or_lock);
3320
3321 /**
3322  * bh_submit_read - Submit a locked buffer for reading
3323  * @bh: struct buffer_head
3324  *
3325  * Returns zero on success and -EIO on error.
3326  */
3327 int bh_submit_read(struct buffer_head *bh)
3328 {
3329         BUG_ON(!buffer_locked(bh));
3330
3331         if (buffer_uptodate(bh)) {
3332                 unlock_buffer(bh);
3333                 return 0;
3334         }
3335
3336         get_bh(bh);
3337         bh->b_end_io = end_buffer_read_sync;
3338         submit_bh(READ, bh);
3339         wait_on_buffer(bh);
3340         if (buffer_uptodate(bh))
3341                 return 0;
3342         return -EIO;
3343 }
3344 EXPORT_SYMBOL(bh_submit_read);
3345
3346 void __init buffer_init(void)
3347 {
3348         int nrpages;
3349
3350         bh_cachep = kmem_cache_create("buffer_head",
3351                         sizeof(struct buffer_head), 0,
3352                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3353                                 SLAB_MEM_SPREAD),
3354                                 NULL);
3355
3356         /*
3357          * Limit the bh occupancy to 10% of ZONE_NORMAL
3358          */
3359         nrpages = (nr_free_buffer_pages() * 10) / 100;
3360         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3361         hotcpu_notifier(buffer_cpu_notify, 0);
3362 }