Merge tag 'master-2014-10-02' of git://git.kernel.org/pub/scm/linux/kernel/git/linvil...
[linux-2.6-microblaze.git] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/capability.h>
28 #include <linux/blkdev.h>
29 #include <linux/file.h>
30 #include <linux/quotaops.h>
31 #include <linux/highmem.h>
32 #include <linux/export.h>
33 #include <linux/writeback.h>
34 #include <linux/hash.h>
35 #include <linux/suspend.h>
36 #include <linux/buffer_head.h>
37 #include <linux/task_io_accounting_ops.h>
38 #include <linux/bio.h>
39 #include <linux/notifier.h>
40 #include <linux/cpu.h>
41 #include <linux/bitops.h>
42 #include <linux/mpage.h>
43 #include <linux/bit_spinlock.h>
44 #include <trace/events/block.h>
45
46 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
47
48 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
49
50 void init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
51 {
52         bh->b_end_io = handler;
53         bh->b_private = private;
54 }
55 EXPORT_SYMBOL(init_buffer);
56
57 inline void touch_buffer(struct buffer_head *bh)
58 {
59         trace_block_touch_buffer(bh);
60         mark_page_accessed(bh->b_page);
61 }
62 EXPORT_SYMBOL(touch_buffer);
63
64 void __lock_buffer(struct buffer_head *bh)
65 {
66         wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
67 }
68 EXPORT_SYMBOL(__lock_buffer);
69
70 void unlock_buffer(struct buffer_head *bh)
71 {
72         clear_bit_unlock(BH_Lock, &bh->b_state);
73         smp_mb__after_atomic();
74         wake_up_bit(&bh->b_state, BH_Lock);
75 }
76 EXPORT_SYMBOL(unlock_buffer);
77
78 /*
79  * Returns if the page has dirty or writeback buffers. If all the buffers
80  * are unlocked and clean then the PageDirty information is stale. If
81  * any of the pages are locked, it is assumed they are locked for IO.
82  */
83 void buffer_check_dirty_writeback(struct page *page,
84                                      bool *dirty, bool *writeback)
85 {
86         struct buffer_head *head, *bh;
87         *dirty = false;
88         *writeback = false;
89
90         BUG_ON(!PageLocked(page));
91
92         if (!page_has_buffers(page))
93                 return;
94
95         if (PageWriteback(page))
96                 *writeback = true;
97
98         head = page_buffers(page);
99         bh = head;
100         do {
101                 if (buffer_locked(bh))
102                         *writeback = true;
103
104                 if (buffer_dirty(bh))
105                         *dirty = true;
106
107                 bh = bh->b_this_page;
108         } while (bh != head);
109 }
110 EXPORT_SYMBOL(buffer_check_dirty_writeback);
111
112 /*
113  * Block until a buffer comes unlocked.  This doesn't stop it
114  * from becoming locked again - you have to lock it yourself
115  * if you want to preserve its state.
116  */
117 void __wait_on_buffer(struct buffer_head * bh)
118 {
119         wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
120 }
121 EXPORT_SYMBOL(__wait_on_buffer);
122
123 static void
124 __clear_page_buffers(struct page *page)
125 {
126         ClearPagePrivate(page);
127         set_page_private(page, 0);
128         page_cache_release(page);
129 }
130
131
132 static int quiet_error(struct buffer_head *bh)
133 {
134         if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
135                 return 0;
136         return 1;
137 }
138
139
140 static void buffer_io_error(struct buffer_head *bh)
141 {
142         char b[BDEVNAME_SIZE];
143         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
144                         bdevname(bh->b_bdev, b),
145                         (unsigned long long)bh->b_blocknr);
146 }
147
148 /*
149  * End-of-IO handler helper function which does not touch the bh after
150  * unlocking it.
151  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
152  * a race there is benign: unlock_buffer() only use the bh's address for
153  * hashing after unlocking the buffer, so it doesn't actually touch the bh
154  * itself.
155  */
156 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
157 {
158         if (uptodate) {
159                 set_buffer_uptodate(bh);
160         } else {
161                 /* This happens, due to failed READA attempts. */
162                 clear_buffer_uptodate(bh);
163         }
164         unlock_buffer(bh);
165 }
166
167 /*
168  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
169  * unlock the buffer. This is what ll_rw_block uses too.
170  */
171 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
172 {
173         __end_buffer_read_notouch(bh, uptodate);
174         put_bh(bh);
175 }
176 EXPORT_SYMBOL(end_buffer_read_sync);
177
178 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
179 {
180         char b[BDEVNAME_SIZE];
181
182         if (uptodate) {
183                 set_buffer_uptodate(bh);
184         } else {
185                 if (!quiet_error(bh)) {
186                         buffer_io_error(bh);
187                         printk(KERN_WARNING "lost page write due to "
188                                         "I/O error on %s\n",
189                                        bdevname(bh->b_bdev, b));
190                 }
191                 set_buffer_write_io_error(bh);
192                 clear_buffer_uptodate(bh);
193         }
194         unlock_buffer(bh);
195         put_bh(bh);
196 }
197 EXPORT_SYMBOL(end_buffer_write_sync);
198
199 /*
200  * Various filesystems appear to want __find_get_block to be non-blocking.
201  * But it's the page lock which protects the buffers.  To get around this,
202  * we get exclusion from try_to_free_buffers with the blockdev mapping's
203  * private_lock.
204  *
205  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
206  * may be quite high.  This code could TryLock the page, and if that
207  * succeeds, there is no need to take private_lock. (But if
208  * private_lock is contended then so is mapping->tree_lock).
209  */
210 static struct buffer_head *
211 __find_get_block_slow(struct block_device *bdev, sector_t block)
212 {
213         struct inode *bd_inode = bdev->bd_inode;
214         struct address_space *bd_mapping = bd_inode->i_mapping;
215         struct buffer_head *ret = NULL;
216         pgoff_t index;
217         struct buffer_head *bh;
218         struct buffer_head *head;
219         struct page *page;
220         int all_mapped = 1;
221
222         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
223         page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
224         if (!page)
225                 goto out;
226
227         spin_lock(&bd_mapping->private_lock);
228         if (!page_has_buffers(page))
229                 goto out_unlock;
230         head = page_buffers(page);
231         bh = head;
232         do {
233                 if (!buffer_mapped(bh))
234                         all_mapped = 0;
235                 else if (bh->b_blocknr == block) {
236                         ret = bh;
237                         get_bh(bh);
238                         goto out_unlock;
239                 }
240                 bh = bh->b_this_page;
241         } while (bh != head);
242
243         /* we might be here because some of the buffers on this page are
244          * not mapped.  This is due to various races between
245          * file io on the block device and getblk.  It gets dealt with
246          * elsewhere, don't buffer_error if we had some unmapped buffers
247          */
248         if (all_mapped) {
249                 char b[BDEVNAME_SIZE];
250
251                 printk("__find_get_block_slow() failed. "
252                         "block=%llu, b_blocknr=%llu\n",
253                         (unsigned long long)block,
254                         (unsigned long long)bh->b_blocknr);
255                 printk("b_state=0x%08lx, b_size=%zu\n",
256                         bh->b_state, bh->b_size);
257                 printk("device %s blocksize: %d\n", bdevname(bdev, b),
258                         1 << bd_inode->i_blkbits);
259         }
260 out_unlock:
261         spin_unlock(&bd_mapping->private_lock);
262         page_cache_release(page);
263 out:
264         return ret;
265 }
266
267 /*
268  * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
269  */
270 static void free_more_memory(void)
271 {
272         struct zone *zone;
273         int nid;
274
275         wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
276         yield();
277
278         for_each_online_node(nid) {
279                 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
280                                                 gfp_zone(GFP_NOFS), NULL,
281                                                 &zone);
282                 if (zone)
283                         try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
284                                                 GFP_NOFS, NULL);
285         }
286 }
287
288 /*
289  * I/O completion handler for block_read_full_page() - pages
290  * which come unlocked at the end of I/O.
291  */
292 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
293 {
294         unsigned long flags;
295         struct buffer_head *first;
296         struct buffer_head *tmp;
297         struct page *page;
298         int page_uptodate = 1;
299
300         BUG_ON(!buffer_async_read(bh));
301
302         page = bh->b_page;
303         if (uptodate) {
304                 set_buffer_uptodate(bh);
305         } else {
306                 clear_buffer_uptodate(bh);
307                 if (!quiet_error(bh))
308                         buffer_io_error(bh);
309                 SetPageError(page);
310         }
311
312         /*
313          * Be _very_ careful from here on. Bad things can happen if
314          * two buffer heads end IO at almost the same time and both
315          * decide that the page is now completely done.
316          */
317         first = page_buffers(page);
318         local_irq_save(flags);
319         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
320         clear_buffer_async_read(bh);
321         unlock_buffer(bh);
322         tmp = bh;
323         do {
324                 if (!buffer_uptodate(tmp))
325                         page_uptodate = 0;
326                 if (buffer_async_read(tmp)) {
327                         BUG_ON(!buffer_locked(tmp));
328                         goto still_busy;
329                 }
330                 tmp = tmp->b_this_page;
331         } while (tmp != bh);
332         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
333         local_irq_restore(flags);
334
335         /*
336          * If none of the buffers had errors and they are all
337          * uptodate then we can set the page uptodate.
338          */
339         if (page_uptodate && !PageError(page))
340                 SetPageUptodate(page);
341         unlock_page(page);
342         return;
343
344 still_busy:
345         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
346         local_irq_restore(flags);
347         return;
348 }
349
350 /*
351  * Completion handler for block_write_full_page() - pages which are unlocked
352  * during I/O, and which have PageWriteback cleared upon I/O completion.
353  */
354 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
355 {
356         char b[BDEVNAME_SIZE];
357         unsigned long flags;
358         struct buffer_head *first;
359         struct buffer_head *tmp;
360         struct page *page;
361
362         BUG_ON(!buffer_async_write(bh));
363
364         page = bh->b_page;
365         if (uptodate) {
366                 set_buffer_uptodate(bh);
367         } else {
368                 if (!quiet_error(bh)) {
369                         buffer_io_error(bh);
370                         printk(KERN_WARNING "lost page write due to "
371                                         "I/O error on %s\n",
372                                bdevname(bh->b_bdev, b));
373                 }
374                 set_bit(AS_EIO, &page->mapping->flags);
375                 set_buffer_write_io_error(bh);
376                 clear_buffer_uptodate(bh);
377                 SetPageError(page);
378         }
379
380         first = page_buffers(page);
381         local_irq_save(flags);
382         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
383
384         clear_buffer_async_write(bh);
385         unlock_buffer(bh);
386         tmp = bh->b_this_page;
387         while (tmp != bh) {
388                 if (buffer_async_write(tmp)) {
389                         BUG_ON(!buffer_locked(tmp));
390                         goto still_busy;
391                 }
392                 tmp = tmp->b_this_page;
393         }
394         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
395         local_irq_restore(flags);
396         end_page_writeback(page);
397         return;
398
399 still_busy:
400         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
401         local_irq_restore(flags);
402         return;
403 }
404 EXPORT_SYMBOL(end_buffer_async_write);
405
406 /*
407  * If a page's buffers are under async readin (end_buffer_async_read
408  * completion) then there is a possibility that another thread of
409  * control could lock one of the buffers after it has completed
410  * but while some of the other buffers have not completed.  This
411  * locked buffer would confuse end_buffer_async_read() into not unlocking
412  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
413  * that this buffer is not under async I/O.
414  *
415  * The page comes unlocked when it has no locked buffer_async buffers
416  * left.
417  *
418  * PageLocked prevents anyone starting new async I/O reads any of
419  * the buffers.
420  *
421  * PageWriteback is used to prevent simultaneous writeout of the same
422  * page.
423  *
424  * PageLocked prevents anyone from starting writeback of a page which is
425  * under read I/O (PageWriteback is only ever set against a locked page).
426  */
427 static void mark_buffer_async_read(struct buffer_head *bh)
428 {
429         bh->b_end_io = end_buffer_async_read;
430         set_buffer_async_read(bh);
431 }
432
433 static void mark_buffer_async_write_endio(struct buffer_head *bh,
434                                           bh_end_io_t *handler)
435 {
436         bh->b_end_io = handler;
437         set_buffer_async_write(bh);
438 }
439
440 void mark_buffer_async_write(struct buffer_head *bh)
441 {
442         mark_buffer_async_write_endio(bh, end_buffer_async_write);
443 }
444 EXPORT_SYMBOL(mark_buffer_async_write);
445
446
447 /*
448  * fs/buffer.c contains helper functions for buffer-backed address space's
449  * fsync functions.  A common requirement for buffer-based filesystems is
450  * that certain data from the backing blockdev needs to be written out for
451  * a successful fsync().  For example, ext2 indirect blocks need to be
452  * written back and waited upon before fsync() returns.
453  *
454  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
455  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
456  * management of a list of dependent buffers at ->i_mapping->private_list.
457  *
458  * Locking is a little subtle: try_to_free_buffers() will remove buffers
459  * from their controlling inode's queue when they are being freed.  But
460  * try_to_free_buffers() will be operating against the *blockdev* mapping
461  * at the time, not against the S_ISREG file which depends on those buffers.
462  * So the locking for private_list is via the private_lock in the address_space
463  * which backs the buffers.  Which is different from the address_space 
464  * against which the buffers are listed.  So for a particular address_space,
465  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
466  * mapping->private_list will always be protected by the backing blockdev's
467  * ->private_lock.
468  *
469  * Which introduces a requirement: all buffers on an address_space's
470  * ->private_list must be from the same address_space: the blockdev's.
471  *
472  * address_spaces which do not place buffers at ->private_list via these
473  * utility functions are free to use private_lock and private_list for
474  * whatever they want.  The only requirement is that list_empty(private_list)
475  * be true at clear_inode() time.
476  *
477  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
478  * filesystems should do that.  invalidate_inode_buffers() should just go
479  * BUG_ON(!list_empty).
480  *
481  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
482  * take an address_space, not an inode.  And it should be called
483  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
484  * queued up.
485  *
486  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
487  * list if it is already on a list.  Because if the buffer is on a list,
488  * it *must* already be on the right one.  If not, the filesystem is being
489  * silly.  This will save a ton of locking.  But first we have to ensure
490  * that buffers are taken *off* the old inode's list when they are freed
491  * (presumably in truncate).  That requires careful auditing of all
492  * filesystems (do it inside bforget()).  It could also be done by bringing
493  * b_inode back.
494  */
495
496 /*
497  * The buffer's backing address_space's private_lock must be held
498  */
499 static void __remove_assoc_queue(struct buffer_head *bh)
500 {
501         list_del_init(&bh->b_assoc_buffers);
502         WARN_ON(!bh->b_assoc_map);
503         if (buffer_write_io_error(bh))
504                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
505         bh->b_assoc_map = NULL;
506 }
507
508 int inode_has_buffers(struct inode *inode)
509 {
510         return !list_empty(&inode->i_data.private_list);
511 }
512
513 /*
514  * osync is designed to support O_SYNC io.  It waits synchronously for
515  * all already-submitted IO to complete, but does not queue any new
516  * writes to the disk.
517  *
518  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
519  * you dirty the buffers, and then use osync_inode_buffers to wait for
520  * completion.  Any other dirty buffers which are not yet queued for
521  * write will not be flushed to disk by the osync.
522  */
523 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
524 {
525         struct buffer_head *bh;
526         struct list_head *p;
527         int err = 0;
528
529         spin_lock(lock);
530 repeat:
531         list_for_each_prev(p, list) {
532                 bh = BH_ENTRY(p);
533                 if (buffer_locked(bh)) {
534                         get_bh(bh);
535                         spin_unlock(lock);
536                         wait_on_buffer(bh);
537                         if (!buffer_uptodate(bh))
538                                 err = -EIO;
539                         brelse(bh);
540                         spin_lock(lock);
541                         goto repeat;
542                 }
543         }
544         spin_unlock(lock);
545         return err;
546 }
547
548 static void do_thaw_one(struct super_block *sb, void *unused)
549 {
550         char b[BDEVNAME_SIZE];
551         while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
552                 printk(KERN_WARNING "Emergency Thaw on %s\n",
553                        bdevname(sb->s_bdev, b));
554 }
555
556 static void do_thaw_all(struct work_struct *work)
557 {
558         iterate_supers(do_thaw_one, NULL);
559         kfree(work);
560         printk(KERN_WARNING "Emergency Thaw complete\n");
561 }
562
563 /**
564  * emergency_thaw_all -- forcibly thaw every frozen filesystem
565  *
566  * Used for emergency unfreeze of all filesystems via SysRq
567  */
568 void emergency_thaw_all(void)
569 {
570         struct work_struct *work;
571
572         work = kmalloc(sizeof(*work), GFP_ATOMIC);
573         if (work) {
574                 INIT_WORK(work, do_thaw_all);
575                 schedule_work(work);
576         }
577 }
578
579 /**
580  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
581  * @mapping: the mapping which wants those buffers written
582  *
583  * Starts I/O against the buffers at mapping->private_list, and waits upon
584  * that I/O.
585  *
586  * Basically, this is a convenience function for fsync().
587  * @mapping is a file or directory which needs those buffers to be written for
588  * a successful fsync().
589  */
590 int sync_mapping_buffers(struct address_space *mapping)
591 {
592         struct address_space *buffer_mapping = mapping->private_data;
593
594         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
595                 return 0;
596
597         return fsync_buffers_list(&buffer_mapping->private_lock,
598                                         &mapping->private_list);
599 }
600 EXPORT_SYMBOL(sync_mapping_buffers);
601
602 /*
603  * Called when we've recently written block `bblock', and it is known that
604  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
605  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
606  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
607  */
608 void write_boundary_block(struct block_device *bdev,
609                         sector_t bblock, unsigned blocksize)
610 {
611         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
612         if (bh) {
613                 if (buffer_dirty(bh))
614                         ll_rw_block(WRITE, 1, &bh);
615                 put_bh(bh);
616         }
617 }
618
619 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
620 {
621         struct address_space *mapping = inode->i_mapping;
622         struct address_space *buffer_mapping = bh->b_page->mapping;
623
624         mark_buffer_dirty(bh);
625         if (!mapping->private_data) {
626                 mapping->private_data = buffer_mapping;
627         } else {
628                 BUG_ON(mapping->private_data != buffer_mapping);
629         }
630         if (!bh->b_assoc_map) {
631                 spin_lock(&buffer_mapping->private_lock);
632                 list_move_tail(&bh->b_assoc_buffers,
633                                 &mapping->private_list);
634                 bh->b_assoc_map = mapping;
635                 spin_unlock(&buffer_mapping->private_lock);
636         }
637 }
638 EXPORT_SYMBOL(mark_buffer_dirty_inode);
639
640 /*
641  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
642  * dirty.
643  *
644  * If warn is true, then emit a warning if the page is not uptodate and has
645  * not been truncated.
646  */
647 static void __set_page_dirty(struct page *page,
648                 struct address_space *mapping, int warn)
649 {
650         unsigned long flags;
651
652         spin_lock_irqsave(&mapping->tree_lock, flags);
653         if (page->mapping) {    /* Race with truncate? */
654                 WARN_ON_ONCE(warn && !PageUptodate(page));
655                 account_page_dirtied(page, mapping);
656                 radix_tree_tag_set(&mapping->page_tree,
657                                 page_index(page), PAGECACHE_TAG_DIRTY);
658         }
659         spin_unlock_irqrestore(&mapping->tree_lock, flags);
660         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
661 }
662
663 /*
664  * Add a page to the dirty page list.
665  *
666  * It is a sad fact of life that this function is called from several places
667  * deeply under spinlocking.  It may not sleep.
668  *
669  * If the page has buffers, the uptodate buffers are set dirty, to preserve
670  * dirty-state coherency between the page and the buffers.  It the page does
671  * not have buffers then when they are later attached they will all be set
672  * dirty.
673  *
674  * The buffers are dirtied before the page is dirtied.  There's a small race
675  * window in which a writepage caller may see the page cleanness but not the
676  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
677  * before the buffers, a concurrent writepage caller could clear the page dirty
678  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
679  * page on the dirty page list.
680  *
681  * We use private_lock to lock against try_to_free_buffers while using the
682  * page's buffer list.  Also use this to protect against clean buffers being
683  * added to the page after it was set dirty.
684  *
685  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
686  * address_space though.
687  */
688 int __set_page_dirty_buffers(struct page *page)
689 {
690         int newly_dirty;
691         struct address_space *mapping = page_mapping(page);
692
693         if (unlikely(!mapping))
694                 return !TestSetPageDirty(page);
695
696         spin_lock(&mapping->private_lock);
697         if (page_has_buffers(page)) {
698                 struct buffer_head *head = page_buffers(page);
699                 struct buffer_head *bh = head;
700
701                 do {
702                         set_buffer_dirty(bh);
703                         bh = bh->b_this_page;
704                 } while (bh != head);
705         }
706         newly_dirty = !TestSetPageDirty(page);
707         spin_unlock(&mapping->private_lock);
708
709         if (newly_dirty)
710                 __set_page_dirty(page, mapping, 1);
711         return newly_dirty;
712 }
713 EXPORT_SYMBOL(__set_page_dirty_buffers);
714
715 /*
716  * Write out and wait upon a list of buffers.
717  *
718  * We have conflicting pressures: we want to make sure that all
719  * initially dirty buffers get waited on, but that any subsequently
720  * dirtied buffers don't.  After all, we don't want fsync to last
721  * forever if somebody is actively writing to the file.
722  *
723  * Do this in two main stages: first we copy dirty buffers to a
724  * temporary inode list, queueing the writes as we go.  Then we clean
725  * up, waiting for those writes to complete.
726  * 
727  * During this second stage, any subsequent updates to the file may end
728  * up refiling the buffer on the original inode's dirty list again, so
729  * there is a chance we will end up with a buffer queued for write but
730  * not yet completed on that list.  So, as a final cleanup we go through
731  * the osync code to catch these locked, dirty buffers without requeuing
732  * any newly dirty buffers for write.
733  */
734 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
735 {
736         struct buffer_head *bh;
737         struct list_head tmp;
738         struct address_space *mapping;
739         int err = 0, err2;
740         struct blk_plug plug;
741
742         INIT_LIST_HEAD(&tmp);
743         blk_start_plug(&plug);
744
745         spin_lock(lock);
746         while (!list_empty(list)) {
747                 bh = BH_ENTRY(list->next);
748                 mapping = bh->b_assoc_map;
749                 __remove_assoc_queue(bh);
750                 /* Avoid race with mark_buffer_dirty_inode() which does
751                  * a lockless check and we rely on seeing the dirty bit */
752                 smp_mb();
753                 if (buffer_dirty(bh) || buffer_locked(bh)) {
754                         list_add(&bh->b_assoc_buffers, &tmp);
755                         bh->b_assoc_map = mapping;
756                         if (buffer_dirty(bh)) {
757                                 get_bh(bh);
758                                 spin_unlock(lock);
759                                 /*
760                                  * Ensure any pending I/O completes so that
761                                  * write_dirty_buffer() actually writes the
762                                  * current contents - it is a noop if I/O is
763                                  * still in flight on potentially older
764                                  * contents.
765                                  */
766                                 write_dirty_buffer(bh, WRITE_SYNC);
767
768                                 /*
769                                  * Kick off IO for the previous mapping. Note
770                                  * that we will not run the very last mapping,
771                                  * wait_on_buffer() will do that for us
772                                  * through sync_buffer().
773                                  */
774                                 brelse(bh);
775                                 spin_lock(lock);
776                         }
777                 }
778         }
779
780         spin_unlock(lock);
781         blk_finish_plug(&plug);
782         spin_lock(lock);
783
784         while (!list_empty(&tmp)) {
785                 bh = BH_ENTRY(tmp.prev);
786                 get_bh(bh);
787                 mapping = bh->b_assoc_map;
788                 __remove_assoc_queue(bh);
789                 /* Avoid race with mark_buffer_dirty_inode() which does
790                  * a lockless check and we rely on seeing the dirty bit */
791                 smp_mb();
792                 if (buffer_dirty(bh)) {
793                         list_add(&bh->b_assoc_buffers,
794                                  &mapping->private_list);
795                         bh->b_assoc_map = mapping;
796                 }
797                 spin_unlock(lock);
798                 wait_on_buffer(bh);
799                 if (!buffer_uptodate(bh))
800                         err = -EIO;
801                 brelse(bh);
802                 spin_lock(lock);
803         }
804         
805         spin_unlock(lock);
806         err2 = osync_buffers_list(lock, list);
807         if (err)
808                 return err;
809         else
810                 return err2;
811 }
812
813 /*
814  * Invalidate any and all dirty buffers on a given inode.  We are
815  * probably unmounting the fs, but that doesn't mean we have already
816  * done a sync().  Just drop the buffers from the inode list.
817  *
818  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
819  * assumes that all the buffers are against the blockdev.  Not true
820  * for reiserfs.
821  */
822 void invalidate_inode_buffers(struct inode *inode)
823 {
824         if (inode_has_buffers(inode)) {
825                 struct address_space *mapping = &inode->i_data;
826                 struct list_head *list = &mapping->private_list;
827                 struct address_space *buffer_mapping = mapping->private_data;
828
829                 spin_lock(&buffer_mapping->private_lock);
830                 while (!list_empty(list))
831                         __remove_assoc_queue(BH_ENTRY(list->next));
832                 spin_unlock(&buffer_mapping->private_lock);
833         }
834 }
835 EXPORT_SYMBOL(invalidate_inode_buffers);
836
837 /*
838  * Remove any clean buffers from the inode's buffer list.  This is called
839  * when we're trying to free the inode itself.  Those buffers can pin it.
840  *
841  * Returns true if all buffers were removed.
842  */
843 int remove_inode_buffers(struct inode *inode)
844 {
845         int ret = 1;
846
847         if (inode_has_buffers(inode)) {
848                 struct address_space *mapping = &inode->i_data;
849                 struct list_head *list = &mapping->private_list;
850                 struct address_space *buffer_mapping = mapping->private_data;
851
852                 spin_lock(&buffer_mapping->private_lock);
853                 while (!list_empty(list)) {
854                         struct buffer_head *bh = BH_ENTRY(list->next);
855                         if (buffer_dirty(bh)) {
856                                 ret = 0;
857                                 break;
858                         }
859                         __remove_assoc_queue(bh);
860                 }
861                 spin_unlock(&buffer_mapping->private_lock);
862         }
863         return ret;
864 }
865
866 /*
867  * Create the appropriate buffers when given a page for data area and
868  * the size of each buffer.. Use the bh->b_this_page linked list to
869  * follow the buffers created.  Return NULL if unable to create more
870  * buffers.
871  *
872  * The retry flag is used to differentiate async IO (paging, swapping)
873  * which may not fail from ordinary buffer allocations.
874  */
875 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
876                 int retry)
877 {
878         struct buffer_head *bh, *head;
879         long offset;
880
881 try_again:
882         head = NULL;
883         offset = PAGE_SIZE;
884         while ((offset -= size) >= 0) {
885                 bh = alloc_buffer_head(GFP_NOFS);
886                 if (!bh)
887                         goto no_grow;
888
889                 bh->b_this_page = head;
890                 bh->b_blocknr = -1;
891                 head = bh;
892
893                 bh->b_size = size;
894
895                 /* Link the buffer to its page */
896                 set_bh_page(bh, page, offset);
897         }
898         return head;
899 /*
900  * In case anything failed, we just free everything we got.
901  */
902 no_grow:
903         if (head) {
904                 do {
905                         bh = head;
906                         head = head->b_this_page;
907                         free_buffer_head(bh);
908                 } while (head);
909         }
910
911         /*
912          * Return failure for non-async IO requests.  Async IO requests
913          * are not allowed to fail, so we have to wait until buffer heads
914          * become available.  But we don't want tasks sleeping with 
915          * partially complete buffers, so all were released above.
916          */
917         if (!retry)
918                 return NULL;
919
920         /* We're _really_ low on memory. Now we just
921          * wait for old buffer heads to become free due to
922          * finishing IO.  Since this is an async request and
923          * the reserve list is empty, we're sure there are 
924          * async buffer heads in use.
925          */
926         free_more_memory();
927         goto try_again;
928 }
929 EXPORT_SYMBOL_GPL(alloc_page_buffers);
930
931 static inline void
932 link_dev_buffers(struct page *page, struct buffer_head *head)
933 {
934         struct buffer_head *bh, *tail;
935
936         bh = head;
937         do {
938                 tail = bh;
939                 bh = bh->b_this_page;
940         } while (bh);
941         tail->b_this_page = head;
942         attach_page_buffers(page, head);
943 }
944
945 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
946 {
947         sector_t retval = ~((sector_t)0);
948         loff_t sz = i_size_read(bdev->bd_inode);
949
950         if (sz) {
951                 unsigned int sizebits = blksize_bits(size);
952                 retval = (sz >> sizebits);
953         }
954         return retval;
955 }
956
957 /*
958  * Initialise the state of a blockdev page's buffers.
959  */ 
960 static sector_t
961 init_page_buffers(struct page *page, struct block_device *bdev,
962                         sector_t block, int size)
963 {
964         struct buffer_head *head = page_buffers(page);
965         struct buffer_head *bh = head;
966         int uptodate = PageUptodate(page);
967         sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
968
969         do {
970                 if (!buffer_mapped(bh)) {
971                         init_buffer(bh, NULL, NULL);
972                         bh->b_bdev = bdev;
973                         bh->b_blocknr = block;
974                         if (uptodate)
975                                 set_buffer_uptodate(bh);
976                         if (block < end_block)
977                                 set_buffer_mapped(bh);
978                 }
979                 block++;
980                 bh = bh->b_this_page;
981         } while (bh != head);
982
983         /*
984          * Caller needs to validate requested block against end of device.
985          */
986         return end_block;
987 }
988
989 /*
990  * Create the page-cache page that contains the requested block.
991  *
992  * This is used purely for blockdev mappings.
993  */
994 static int
995 grow_dev_page(struct block_device *bdev, sector_t block,
996                 pgoff_t index, int size, int sizebits)
997 {
998         struct inode *inode = bdev->bd_inode;
999         struct page *page;
1000         struct buffer_head *bh;
1001         sector_t end_block;
1002         int ret = 0;            /* Will call free_more_memory() */
1003         gfp_t gfp_mask;
1004
1005         gfp_mask = mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS;
1006         gfp_mask |= __GFP_MOVABLE;
1007         /*
1008          * XXX: __getblk_slow() can not really deal with failure and
1009          * will endlessly loop on improvised global reclaim.  Prefer
1010          * looping in the allocator rather than here, at least that
1011          * code knows what it's doing.
1012          */
1013         gfp_mask |= __GFP_NOFAIL;
1014
1015         page = find_or_create_page(inode->i_mapping, index, gfp_mask);
1016         if (!page)
1017                 return ret;
1018
1019         BUG_ON(!PageLocked(page));
1020
1021         if (page_has_buffers(page)) {
1022                 bh = page_buffers(page);
1023                 if (bh->b_size == size) {
1024                         end_block = init_page_buffers(page, bdev,
1025                                                 (sector_t)index << sizebits,
1026                                                 size);
1027                         goto done;
1028                 }
1029                 if (!try_to_free_buffers(page))
1030                         goto failed;
1031         }
1032
1033         /*
1034          * Allocate some buffers for this page
1035          */
1036         bh = alloc_page_buffers(page, size, 0);
1037         if (!bh)
1038                 goto failed;
1039
1040         /*
1041          * Link the page to the buffers and initialise them.  Take the
1042          * lock to be atomic wrt __find_get_block(), which does not
1043          * run under the page lock.
1044          */
1045         spin_lock(&inode->i_mapping->private_lock);
1046         link_dev_buffers(page, bh);
1047         end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1048                         size);
1049         spin_unlock(&inode->i_mapping->private_lock);
1050 done:
1051         ret = (block < end_block) ? 1 : -ENXIO;
1052 failed:
1053         unlock_page(page);
1054         page_cache_release(page);
1055         return ret;
1056 }
1057
1058 /*
1059  * Create buffers for the specified block device block's page.  If
1060  * that page was dirty, the buffers are set dirty also.
1061  */
1062 static int
1063 grow_buffers(struct block_device *bdev, sector_t block, int size)
1064 {
1065         pgoff_t index;
1066         int sizebits;
1067
1068         sizebits = -1;
1069         do {
1070                 sizebits++;
1071         } while ((size << sizebits) < PAGE_SIZE);
1072
1073         index = block >> sizebits;
1074
1075         /*
1076          * Check for a block which wants to lie outside our maximum possible
1077          * pagecache index.  (this comparison is done using sector_t types).
1078          */
1079         if (unlikely(index != block >> sizebits)) {
1080                 char b[BDEVNAME_SIZE];
1081
1082                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1083                         "device %s\n",
1084                         __func__, (unsigned long long)block,
1085                         bdevname(bdev, b));
1086                 return -EIO;
1087         }
1088
1089         /* Create a page with the proper size buffers.. */
1090         return grow_dev_page(bdev, block, index, size, sizebits);
1091 }
1092
1093 static struct buffer_head *
1094 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1095 {
1096         /* Size must be multiple of hard sectorsize */
1097         if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1098                         (size < 512 || size > PAGE_SIZE))) {
1099                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1100                                         size);
1101                 printk(KERN_ERR "logical block size: %d\n",
1102                                         bdev_logical_block_size(bdev));
1103
1104                 dump_stack();
1105                 return NULL;
1106         }
1107
1108         for (;;) {
1109                 struct buffer_head *bh;
1110                 int ret;
1111
1112                 bh = __find_get_block(bdev, block, size);
1113                 if (bh)
1114                         return bh;
1115
1116                 ret = grow_buffers(bdev, block, size);
1117                 if (ret < 0)
1118                         return NULL;
1119                 if (ret == 0)
1120                         free_more_memory();
1121         }
1122 }
1123
1124 /*
1125  * The relationship between dirty buffers and dirty pages:
1126  *
1127  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1128  * the page is tagged dirty in its radix tree.
1129  *
1130  * At all times, the dirtiness of the buffers represents the dirtiness of
1131  * subsections of the page.  If the page has buffers, the page dirty bit is
1132  * merely a hint about the true dirty state.
1133  *
1134  * When a page is set dirty in its entirety, all its buffers are marked dirty
1135  * (if the page has buffers).
1136  *
1137  * When a buffer is marked dirty, its page is dirtied, but the page's other
1138  * buffers are not.
1139  *
1140  * Also.  When blockdev buffers are explicitly read with bread(), they
1141  * individually become uptodate.  But their backing page remains not
1142  * uptodate - even if all of its buffers are uptodate.  A subsequent
1143  * block_read_full_page() against that page will discover all the uptodate
1144  * buffers, will set the page uptodate and will perform no I/O.
1145  */
1146
1147 /**
1148  * mark_buffer_dirty - mark a buffer_head as needing writeout
1149  * @bh: the buffer_head to mark dirty
1150  *
1151  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1152  * backing page dirty, then tag the page as dirty in its address_space's radix
1153  * tree and then attach the address_space's inode to its superblock's dirty
1154  * inode list.
1155  *
1156  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1157  * mapping->tree_lock and mapping->host->i_lock.
1158  */
1159 void mark_buffer_dirty(struct buffer_head *bh)
1160 {
1161         WARN_ON_ONCE(!buffer_uptodate(bh));
1162
1163         trace_block_dirty_buffer(bh);
1164
1165         /*
1166          * Very *carefully* optimize the it-is-already-dirty case.
1167          *
1168          * Don't let the final "is it dirty" escape to before we
1169          * perhaps modified the buffer.
1170          */
1171         if (buffer_dirty(bh)) {
1172                 smp_mb();
1173                 if (buffer_dirty(bh))
1174                         return;
1175         }
1176
1177         if (!test_set_buffer_dirty(bh)) {
1178                 struct page *page = bh->b_page;
1179                 if (!TestSetPageDirty(page)) {
1180                         struct address_space *mapping = page_mapping(page);
1181                         if (mapping)
1182                                 __set_page_dirty(page, mapping, 0);
1183                 }
1184         }
1185 }
1186 EXPORT_SYMBOL(mark_buffer_dirty);
1187
1188 /*
1189  * Decrement a buffer_head's reference count.  If all buffers against a page
1190  * have zero reference count, are clean and unlocked, and if the page is clean
1191  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1192  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1193  * a page but it ends up not being freed, and buffers may later be reattached).
1194  */
1195 void __brelse(struct buffer_head * buf)
1196 {
1197         if (atomic_read(&buf->b_count)) {
1198                 put_bh(buf);
1199                 return;
1200         }
1201         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1202 }
1203 EXPORT_SYMBOL(__brelse);
1204
1205 /*
1206  * bforget() is like brelse(), except it discards any
1207  * potentially dirty data.
1208  */
1209 void __bforget(struct buffer_head *bh)
1210 {
1211         clear_buffer_dirty(bh);
1212         if (bh->b_assoc_map) {
1213                 struct address_space *buffer_mapping = bh->b_page->mapping;
1214
1215                 spin_lock(&buffer_mapping->private_lock);
1216                 list_del_init(&bh->b_assoc_buffers);
1217                 bh->b_assoc_map = NULL;
1218                 spin_unlock(&buffer_mapping->private_lock);
1219         }
1220         __brelse(bh);
1221 }
1222 EXPORT_SYMBOL(__bforget);
1223
1224 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1225 {
1226         lock_buffer(bh);
1227         if (buffer_uptodate(bh)) {
1228                 unlock_buffer(bh);
1229                 return bh;
1230         } else {
1231                 get_bh(bh);
1232                 bh->b_end_io = end_buffer_read_sync;
1233                 submit_bh(READ, bh);
1234                 wait_on_buffer(bh);
1235                 if (buffer_uptodate(bh))
1236                         return bh;
1237         }
1238         brelse(bh);
1239         return NULL;
1240 }
1241
1242 /*
1243  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1244  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1245  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1246  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1247  * CPU's LRUs at the same time.
1248  *
1249  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1250  * sb_find_get_block().
1251  *
1252  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1253  * a local interrupt disable for that.
1254  */
1255
1256 #define BH_LRU_SIZE     8
1257
1258 struct bh_lru {
1259         struct buffer_head *bhs[BH_LRU_SIZE];
1260 };
1261
1262 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1263
1264 #ifdef CONFIG_SMP
1265 #define bh_lru_lock()   local_irq_disable()
1266 #define bh_lru_unlock() local_irq_enable()
1267 #else
1268 #define bh_lru_lock()   preempt_disable()
1269 #define bh_lru_unlock() preempt_enable()
1270 #endif
1271
1272 static inline void check_irqs_on(void)
1273 {
1274 #ifdef irqs_disabled
1275         BUG_ON(irqs_disabled());
1276 #endif
1277 }
1278
1279 /*
1280  * The LRU management algorithm is dopey-but-simple.  Sorry.
1281  */
1282 static void bh_lru_install(struct buffer_head *bh)
1283 {
1284         struct buffer_head *evictee = NULL;
1285
1286         check_irqs_on();
1287         bh_lru_lock();
1288         if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1289                 struct buffer_head *bhs[BH_LRU_SIZE];
1290                 int in;
1291                 int out = 0;
1292
1293                 get_bh(bh);
1294                 bhs[out++] = bh;
1295                 for (in = 0; in < BH_LRU_SIZE; in++) {
1296                         struct buffer_head *bh2 =
1297                                 __this_cpu_read(bh_lrus.bhs[in]);
1298
1299                         if (bh2 == bh) {
1300                                 __brelse(bh2);
1301                         } else {
1302                                 if (out >= BH_LRU_SIZE) {
1303                                         BUG_ON(evictee != NULL);
1304                                         evictee = bh2;
1305                                 } else {
1306                                         bhs[out++] = bh2;
1307                                 }
1308                         }
1309                 }
1310                 while (out < BH_LRU_SIZE)
1311                         bhs[out++] = NULL;
1312                 memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1313         }
1314         bh_lru_unlock();
1315
1316         if (evictee)
1317                 __brelse(evictee);
1318 }
1319
1320 /*
1321  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1322  */
1323 static struct buffer_head *
1324 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1325 {
1326         struct buffer_head *ret = NULL;
1327         unsigned int i;
1328
1329         check_irqs_on();
1330         bh_lru_lock();
1331         for (i = 0; i < BH_LRU_SIZE; i++) {
1332                 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1333
1334                 if (bh && bh->b_bdev == bdev &&
1335                                 bh->b_blocknr == block && bh->b_size == size) {
1336                         if (i) {
1337                                 while (i) {
1338                                         __this_cpu_write(bh_lrus.bhs[i],
1339                                                 __this_cpu_read(bh_lrus.bhs[i - 1]));
1340                                         i--;
1341                                 }
1342                                 __this_cpu_write(bh_lrus.bhs[0], bh);
1343                         }
1344                         get_bh(bh);
1345                         ret = bh;
1346                         break;
1347                 }
1348         }
1349         bh_lru_unlock();
1350         return ret;
1351 }
1352
1353 /*
1354  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1355  * it in the LRU and mark it as accessed.  If it is not present then return
1356  * NULL
1357  */
1358 struct buffer_head *
1359 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1360 {
1361         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1362
1363         if (bh == NULL) {
1364                 /* __find_get_block_slow will mark the page accessed */
1365                 bh = __find_get_block_slow(bdev, block);
1366                 if (bh)
1367                         bh_lru_install(bh);
1368         } else
1369                 touch_buffer(bh);
1370
1371         return bh;
1372 }
1373 EXPORT_SYMBOL(__find_get_block);
1374
1375 /*
1376  * __getblk will locate (and, if necessary, create) the buffer_head
1377  * which corresponds to the passed block_device, block and size. The
1378  * returned buffer has its reference count incremented.
1379  *
1380  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1381  * attempt is failing.  FIXME, perhaps?
1382  */
1383 struct buffer_head *
1384 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1385 {
1386         struct buffer_head *bh = __find_get_block(bdev, block, size);
1387
1388         might_sleep();
1389         if (bh == NULL)
1390                 bh = __getblk_slow(bdev, block, size);
1391         return bh;
1392 }
1393 EXPORT_SYMBOL(__getblk);
1394
1395 /*
1396  * Do async read-ahead on a buffer..
1397  */
1398 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1399 {
1400         struct buffer_head *bh = __getblk(bdev, block, size);
1401         if (likely(bh)) {
1402                 ll_rw_block(READA, 1, &bh);
1403                 brelse(bh);
1404         }
1405 }
1406 EXPORT_SYMBOL(__breadahead);
1407
1408 /**
1409  *  __bread() - reads a specified block and returns the bh
1410  *  @bdev: the block_device to read from
1411  *  @block: number of block
1412  *  @size: size (in bytes) to read
1413  * 
1414  *  Reads a specified block, and returns buffer head that contains it.
1415  *  It returns NULL if the block was unreadable.
1416  */
1417 struct buffer_head *
1418 __bread(struct block_device *bdev, sector_t block, unsigned size)
1419 {
1420         struct buffer_head *bh = __getblk(bdev, block, size);
1421
1422         if (likely(bh) && !buffer_uptodate(bh))
1423                 bh = __bread_slow(bh);
1424         return bh;
1425 }
1426 EXPORT_SYMBOL(__bread);
1427
1428 /*
1429  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1430  * This doesn't race because it runs in each cpu either in irq
1431  * or with preempt disabled.
1432  */
1433 static void invalidate_bh_lru(void *arg)
1434 {
1435         struct bh_lru *b = &get_cpu_var(bh_lrus);
1436         int i;
1437
1438         for (i = 0; i < BH_LRU_SIZE; i++) {
1439                 brelse(b->bhs[i]);
1440                 b->bhs[i] = NULL;
1441         }
1442         put_cpu_var(bh_lrus);
1443 }
1444
1445 static bool has_bh_in_lru(int cpu, void *dummy)
1446 {
1447         struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1448         int i;
1449         
1450         for (i = 0; i < BH_LRU_SIZE; i++) {
1451                 if (b->bhs[i])
1452                         return 1;
1453         }
1454
1455         return 0;
1456 }
1457
1458 void invalidate_bh_lrus(void)
1459 {
1460         on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1461 }
1462 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1463
1464 void set_bh_page(struct buffer_head *bh,
1465                 struct page *page, unsigned long offset)
1466 {
1467         bh->b_page = page;
1468         BUG_ON(offset >= PAGE_SIZE);
1469         if (PageHighMem(page))
1470                 /*
1471                  * This catches illegal uses and preserves the offset:
1472                  */
1473                 bh->b_data = (char *)(0 + offset);
1474         else
1475                 bh->b_data = page_address(page) + offset;
1476 }
1477 EXPORT_SYMBOL(set_bh_page);
1478
1479 /*
1480  * Called when truncating a buffer on a page completely.
1481  */
1482
1483 /* Bits that are cleared during an invalidate */
1484 #define BUFFER_FLAGS_DISCARD \
1485         (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1486          1 << BH_Delay | 1 << BH_Unwritten)
1487
1488 static void discard_buffer(struct buffer_head * bh)
1489 {
1490         unsigned long b_state, b_state_old;
1491
1492         lock_buffer(bh);
1493         clear_buffer_dirty(bh);
1494         bh->b_bdev = NULL;
1495         b_state = bh->b_state;
1496         for (;;) {
1497                 b_state_old = cmpxchg(&bh->b_state, b_state,
1498                                       (b_state & ~BUFFER_FLAGS_DISCARD));
1499                 if (b_state_old == b_state)
1500                         break;
1501                 b_state = b_state_old;
1502         }
1503         unlock_buffer(bh);
1504 }
1505
1506 /**
1507  * block_invalidatepage - invalidate part or all of a buffer-backed page
1508  *
1509  * @page: the page which is affected
1510  * @offset: start of the range to invalidate
1511  * @length: length of the range to invalidate
1512  *
1513  * block_invalidatepage() is called when all or part of the page has become
1514  * invalidated by a truncate operation.
1515  *
1516  * block_invalidatepage() does not have to release all buffers, but it must
1517  * ensure that no dirty buffer is left outside @offset and that no I/O
1518  * is underway against any of the blocks which are outside the truncation
1519  * point.  Because the caller is about to free (and possibly reuse) those
1520  * blocks on-disk.
1521  */
1522 void block_invalidatepage(struct page *page, unsigned int offset,
1523                           unsigned int length)
1524 {
1525         struct buffer_head *head, *bh, *next;
1526         unsigned int curr_off = 0;
1527         unsigned int stop = length + offset;
1528
1529         BUG_ON(!PageLocked(page));
1530         if (!page_has_buffers(page))
1531                 goto out;
1532
1533         /*
1534          * Check for overflow
1535          */
1536         BUG_ON(stop > PAGE_CACHE_SIZE || stop < length);
1537
1538         head = page_buffers(page);
1539         bh = head;
1540         do {
1541                 unsigned int next_off = curr_off + bh->b_size;
1542                 next = bh->b_this_page;
1543
1544                 /*
1545                  * Are we still fully in range ?
1546                  */
1547                 if (next_off > stop)
1548                         goto out;
1549
1550                 /*
1551                  * is this block fully invalidated?
1552                  */
1553                 if (offset <= curr_off)
1554                         discard_buffer(bh);
1555                 curr_off = next_off;
1556                 bh = next;
1557         } while (bh != head);
1558
1559         /*
1560          * We release buffers only if the entire page is being invalidated.
1561          * The get_block cached value has been unconditionally invalidated,
1562          * so real IO is not possible anymore.
1563          */
1564         if (offset == 0)
1565                 try_to_release_page(page, 0);
1566 out:
1567         return;
1568 }
1569 EXPORT_SYMBOL(block_invalidatepage);
1570
1571
1572 /*
1573  * We attach and possibly dirty the buffers atomically wrt
1574  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1575  * is already excluded via the page lock.
1576  */
1577 void create_empty_buffers(struct page *page,
1578                         unsigned long blocksize, unsigned long b_state)
1579 {
1580         struct buffer_head *bh, *head, *tail;
1581
1582         head = alloc_page_buffers(page, blocksize, 1);
1583         bh = head;
1584         do {
1585                 bh->b_state |= b_state;
1586                 tail = bh;
1587                 bh = bh->b_this_page;
1588         } while (bh);
1589         tail->b_this_page = head;
1590
1591         spin_lock(&page->mapping->private_lock);
1592         if (PageUptodate(page) || PageDirty(page)) {
1593                 bh = head;
1594                 do {
1595                         if (PageDirty(page))
1596                                 set_buffer_dirty(bh);
1597                         if (PageUptodate(page))
1598                                 set_buffer_uptodate(bh);
1599                         bh = bh->b_this_page;
1600                 } while (bh != head);
1601         }
1602         attach_page_buffers(page, head);
1603         spin_unlock(&page->mapping->private_lock);
1604 }
1605 EXPORT_SYMBOL(create_empty_buffers);
1606
1607 /*
1608  * We are taking a block for data and we don't want any output from any
1609  * buffer-cache aliases starting from return from that function and
1610  * until the moment when something will explicitly mark the buffer
1611  * dirty (hopefully that will not happen until we will free that block ;-)
1612  * We don't even need to mark it not-uptodate - nobody can expect
1613  * anything from a newly allocated buffer anyway. We used to used
1614  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1615  * don't want to mark the alias unmapped, for example - it would confuse
1616  * anyone who might pick it with bread() afterwards...
1617  *
1618  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1619  * be writeout I/O going on against recently-freed buffers.  We don't
1620  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1621  * only if we really need to.  That happens here.
1622  */
1623 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1624 {
1625         struct buffer_head *old_bh;
1626
1627         might_sleep();
1628
1629         old_bh = __find_get_block_slow(bdev, block);
1630         if (old_bh) {
1631                 clear_buffer_dirty(old_bh);
1632                 wait_on_buffer(old_bh);
1633                 clear_buffer_req(old_bh);
1634                 __brelse(old_bh);
1635         }
1636 }
1637 EXPORT_SYMBOL(unmap_underlying_metadata);
1638
1639 /*
1640  * Size is a power-of-two in the range 512..PAGE_SIZE,
1641  * and the case we care about most is PAGE_SIZE.
1642  *
1643  * So this *could* possibly be written with those
1644  * constraints in mind (relevant mostly if some
1645  * architecture has a slow bit-scan instruction)
1646  */
1647 static inline int block_size_bits(unsigned int blocksize)
1648 {
1649         return ilog2(blocksize);
1650 }
1651
1652 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1653 {
1654         BUG_ON(!PageLocked(page));
1655
1656         if (!page_has_buffers(page))
1657                 create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1658         return page_buffers(page);
1659 }
1660
1661 /*
1662  * NOTE! All mapped/uptodate combinations are valid:
1663  *
1664  *      Mapped  Uptodate        Meaning
1665  *
1666  *      No      No              "unknown" - must do get_block()
1667  *      No      Yes             "hole" - zero-filled
1668  *      Yes     No              "allocated" - allocated on disk, not read in
1669  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1670  *
1671  * "Dirty" is valid only with the last case (mapped+uptodate).
1672  */
1673
1674 /*
1675  * While block_write_full_page is writing back the dirty buffers under
1676  * the page lock, whoever dirtied the buffers may decide to clean them
1677  * again at any time.  We handle that by only looking at the buffer
1678  * state inside lock_buffer().
1679  *
1680  * If block_write_full_page() is called for regular writeback
1681  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1682  * locked buffer.   This only can happen if someone has written the buffer
1683  * directly, with submit_bh().  At the address_space level PageWriteback
1684  * prevents this contention from occurring.
1685  *
1686  * If block_write_full_page() is called with wbc->sync_mode ==
1687  * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1688  * causes the writes to be flagged as synchronous writes.
1689  */
1690 static int __block_write_full_page(struct inode *inode, struct page *page,
1691                         get_block_t *get_block, struct writeback_control *wbc,
1692                         bh_end_io_t *handler)
1693 {
1694         int err;
1695         sector_t block;
1696         sector_t last_block;
1697         struct buffer_head *bh, *head;
1698         unsigned int blocksize, bbits;
1699         int nr_underway = 0;
1700         int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1701                         WRITE_SYNC : WRITE);
1702
1703         head = create_page_buffers(page, inode,
1704                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1705
1706         /*
1707          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1708          * here, and the (potentially unmapped) buffers may become dirty at
1709          * any time.  If a buffer becomes dirty here after we've inspected it
1710          * then we just miss that fact, and the page stays dirty.
1711          *
1712          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1713          * handle that here by just cleaning them.
1714          */
1715
1716         bh = head;
1717         blocksize = bh->b_size;
1718         bbits = block_size_bits(blocksize);
1719
1720         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1721         last_block = (i_size_read(inode) - 1) >> bbits;
1722
1723         /*
1724          * Get all the dirty buffers mapped to disk addresses and
1725          * handle any aliases from the underlying blockdev's mapping.
1726          */
1727         do {
1728                 if (block > last_block) {
1729                         /*
1730                          * mapped buffers outside i_size will occur, because
1731                          * this page can be outside i_size when there is a
1732                          * truncate in progress.
1733                          */
1734                         /*
1735                          * The buffer was zeroed by block_write_full_page()
1736                          */
1737                         clear_buffer_dirty(bh);
1738                         set_buffer_uptodate(bh);
1739                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1740                            buffer_dirty(bh)) {
1741                         WARN_ON(bh->b_size != blocksize);
1742                         err = get_block(inode, block, bh, 1);
1743                         if (err)
1744                                 goto recover;
1745                         clear_buffer_delay(bh);
1746                         if (buffer_new(bh)) {
1747                                 /* blockdev mappings never come here */
1748                                 clear_buffer_new(bh);
1749                                 unmap_underlying_metadata(bh->b_bdev,
1750                                                         bh->b_blocknr);
1751                         }
1752                 }
1753                 bh = bh->b_this_page;
1754                 block++;
1755         } while (bh != head);
1756
1757         do {
1758                 if (!buffer_mapped(bh))
1759                         continue;
1760                 /*
1761                  * If it's a fully non-blocking write attempt and we cannot
1762                  * lock the buffer then redirty the page.  Note that this can
1763                  * potentially cause a busy-wait loop from writeback threads
1764                  * and kswapd activity, but those code paths have their own
1765                  * higher-level throttling.
1766                  */
1767                 if (wbc->sync_mode != WB_SYNC_NONE) {
1768                         lock_buffer(bh);
1769                 } else if (!trylock_buffer(bh)) {
1770                         redirty_page_for_writepage(wbc, page);
1771                         continue;
1772                 }
1773                 if (test_clear_buffer_dirty(bh)) {
1774                         mark_buffer_async_write_endio(bh, handler);
1775                 } else {
1776                         unlock_buffer(bh);
1777                 }
1778         } while ((bh = bh->b_this_page) != head);
1779
1780         /*
1781          * The page and its buffers are protected by PageWriteback(), so we can
1782          * drop the bh refcounts early.
1783          */
1784         BUG_ON(PageWriteback(page));
1785         set_page_writeback(page);
1786
1787         do {
1788                 struct buffer_head *next = bh->b_this_page;
1789                 if (buffer_async_write(bh)) {
1790                         submit_bh(write_op, bh);
1791                         nr_underway++;
1792                 }
1793                 bh = next;
1794         } while (bh != head);
1795         unlock_page(page);
1796
1797         err = 0;
1798 done:
1799         if (nr_underway == 0) {
1800                 /*
1801                  * The page was marked dirty, but the buffers were
1802                  * clean.  Someone wrote them back by hand with
1803                  * ll_rw_block/submit_bh.  A rare case.
1804                  */
1805                 end_page_writeback(page);
1806
1807                 /*
1808                  * The page and buffer_heads can be released at any time from
1809                  * here on.
1810                  */
1811         }
1812         return err;
1813
1814 recover:
1815         /*
1816          * ENOSPC, or some other error.  We may already have added some
1817          * blocks to the file, so we need to write these out to avoid
1818          * exposing stale data.
1819          * The page is currently locked and not marked for writeback
1820          */
1821         bh = head;
1822         /* Recovery: lock and submit the mapped buffers */
1823         do {
1824                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1825                     !buffer_delay(bh)) {
1826                         lock_buffer(bh);
1827                         mark_buffer_async_write_endio(bh, handler);
1828                 } else {
1829                         /*
1830                          * The buffer may have been set dirty during
1831                          * attachment to a dirty page.
1832                          */
1833                         clear_buffer_dirty(bh);
1834                 }
1835         } while ((bh = bh->b_this_page) != head);
1836         SetPageError(page);
1837         BUG_ON(PageWriteback(page));
1838         mapping_set_error(page->mapping, err);
1839         set_page_writeback(page);
1840         do {
1841                 struct buffer_head *next = bh->b_this_page;
1842                 if (buffer_async_write(bh)) {
1843                         clear_buffer_dirty(bh);
1844                         submit_bh(write_op, bh);
1845                         nr_underway++;
1846                 }
1847                 bh = next;
1848         } while (bh != head);
1849         unlock_page(page);
1850         goto done;
1851 }
1852
1853 /*
1854  * If a page has any new buffers, zero them out here, and mark them uptodate
1855  * and dirty so they'll be written out (in order to prevent uninitialised
1856  * block data from leaking). And clear the new bit.
1857  */
1858 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1859 {
1860         unsigned int block_start, block_end;
1861         struct buffer_head *head, *bh;
1862
1863         BUG_ON(!PageLocked(page));
1864         if (!page_has_buffers(page))
1865                 return;
1866
1867         bh = head = page_buffers(page);
1868         block_start = 0;
1869         do {
1870                 block_end = block_start + bh->b_size;
1871
1872                 if (buffer_new(bh)) {
1873                         if (block_end > from && block_start < to) {
1874                                 if (!PageUptodate(page)) {
1875                                         unsigned start, size;
1876
1877                                         start = max(from, block_start);
1878                                         size = min(to, block_end) - start;
1879
1880                                         zero_user(page, start, size);
1881                                         set_buffer_uptodate(bh);
1882                                 }
1883
1884                                 clear_buffer_new(bh);
1885                                 mark_buffer_dirty(bh);
1886                         }
1887                 }
1888
1889                 block_start = block_end;
1890                 bh = bh->b_this_page;
1891         } while (bh != head);
1892 }
1893 EXPORT_SYMBOL(page_zero_new_buffers);
1894
1895 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1896                 get_block_t *get_block)
1897 {
1898         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1899         unsigned to = from + len;
1900         struct inode *inode = page->mapping->host;
1901         unsigned block_start, block_end;
1902         sector_t block;
1903         int err = 0;
1904         unsigned blocksize, bbits;
1905         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1906
1907         BUG_ON(!PageLocked(page));
1908         BUG_ON(from > PAGE_CACHE_SIZE);
1909         BUG_ON(to > PAGE_CACHE_SIZE);
1910         BUG_ON(from > to);
1911
1912         head = create_page_buffers(page, inode, 0);
1913         blocksize = head->b_size;
1914         bbits = block_size_bits(blocksize);
1915
1916         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1917
1918         for(bh = head, block_start = 0; bh != head || !block_start;
1919             block++, block_start=block_end, bh = bh->b_this_page) {
1920                 block_end = block_start + blocksize;
1921                 if (block_end <= from || block_start >= to) {
1922                         if (PageUptodate(page)) {
1923                                 if (!buffer_uptodate(bh))
1924                                         set_buffer_uptodate(bh);
1925                         }
1926                         continue;
1927                 }
1928                 if (buffer_new(bh))
1929                         clear_buffer_new(bh);
1930                 if (!buffer_mapped(bh)) {
1931                         WARN_ON(bh->b_size != blocksize);
1932                         err = get_block(inode, block, bh, 1);
1933                         if (err)
1934                                 break;
1935                         if (buffer_new(bh)) {
1936                                 unmap_underlying_metadata(bh->b_bdev,
1937                                                         bh->b_blocknr);
1938                                 if (PageUptodate(page)) {
1939                                         clear_buffer_new(bh);
1940                                         set_buffer_uptodate(bh);
1941                                         mark_buffer_dirty(bh);
1942                                         continue;
1943                                 }
1944                                 if (block_end > to || block_start < from)
1945                                         zero_user_segments(page,
1946                                                 to, block_end,
1947                                                 block_start, from);
1948                                 continue;
1949                         }
1950                 }
1951                 if (PageUptodate(page)) {
1952                         if (!buffer_uptodate(bh))
1953                                 set_buffer_uptodate(bh);
1954                         continue; 
1955                 }
1956                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1957                     !buffer_unwritten(bh) &&
1958                      (block_start < from || block_end > to)) {
1959                         ll_rw_block(READ, 1, &bh);
1960                         *wait_bh++=bh;
1961                 }
1962         }
1963         /*
1964          * If we issued read requests - let them complete.
1965          */
1966         while(wait_bh > wait) {
1967                 wait_on_buffer(*--wait_bh);
1968                 if (!buffer_uptodate(*wait_bh))
1969                         err = -EIO;
1970         }
1971         if (unlikely(err))
1972                 page_zero_new_buffers(page, from, to);
1973         return err;
1974 }
1975 EXPORT_SYMBOL(__block_write_begin);
1976
1977 static int __block_commit_write(struct inode *inode, struct page *page,
1978                 unsigned from, unsigned to)
1979 {
1980         unsigned block_start, block_end;
1981         int partial = 0;
1982         unsigned blocksize;
1983         struct buffer_head *bh, *head;
1984
1985         bh = head = page_buffers(page);
1986         blocksize = bh->b_size;
1987
1988         block_start = 0;
1989         do {
1990                 block_end = block_start + blocksize;
1991                 if (block_end <= from || block_start >= to) {
1992                         if (!buffer_uptodate(bh))
1993                                 partial = 1;
1994                 } else {
1995                         set_buffer_uptodate(bh);
1996                         mark_buffer_dirty(bh);
1997                 }
1998                 clear_buffer_new(bh);
1999
2000                 block_start = block_end;
2001                 bh = bh->b_this_page;
2002         } while (bh != head);
2003
2004         /*
2005          * If this is a partial write which happened to make all buffers
2006          * uptodate then we can optimize away a bogus readpage() for
2007          * the next read(). Here we 'discover' whether the page went
2008          * uptodate as a result of this (potentially partial) write.
2009          */
2010         if (!partial)
2011                 SetPageUptodate(page);
2012         return 0;
2013 }
2014
2015 /*
2016  * block_write_begin takes care of the basic task of block allocation and
2017  * bringing partial write blocks uptodate first.
2018  *
2019  * The filesystem needs to handle block truncation upon failure.
2020  */
2021 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2022                 unsigned flags, struct page **pagep, get_block_t *get_block)
2023 {
2024         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
2025         struct page *page;
2026         int status;
2027
2028         page = grab_cache_page_write_begin(mapping, index, flags);
2029         if (!page)
2030                 return -ENOMEM;
2031
2032         status = __block_write_begin(page, pos, len, get_block);
2033         if (unlikely(status)) {
2034                 unlock_page(page);
2035                 page_cache_release(page);
2036                 page = NULL;
2037         }
2038
2039         *pagep = page;
2040         return status;
2041 }
2042 EXPORT_SYMBOL(block_write_begin);
2043
2044 int block_write_end(struct file *file, struct address_space *mapping,
2045                         loff_t pos, unsigned len, unsigned copied,
2046                         struct page *page, void *fsdata)
2047 {
2048         struct inode *inode = mapping->host;
2049         unsigned start;
2050
2051         start = pos & (PAGE_CACHE_SIZE - 1);
2052
2053         if (unlikely(copied < len)) {
2054                 /*
2055                  * The buffers that were written will now be uptodate, so we
2056                  * don't have to worry about a readpage reading them and
2057                  * overwriting a partial write. However if we have encountered
2058                  * a short write and only partially written into a buffer, it
2059                  * will not be marked uptodate, so a readpage might come in and
2060                  * destroy our partial write.
2061                  *
2062                  * Do the simplest thing, and just treat any short write to a
2063                  * non uptodate page as a zero-length write, and force the
2064                  * caller to redo the whole thing.
2065                  */
2066                 if (!PageUptodate(page))
2067                         copied = 0;
2068
2069                 page_zero_new_buffers(page, start+copied, start+len);
2070         }
2071         flush_dcache_page(page);
2072
2073         /* This could be a short (even 0-length) commit */
2074         __block_commit_write(inode, page, start, start+copied);
2075
2076         return copied;
2077 }
2078 EXPORT_SYMBOL(block_write_end);
2079
2080 int generic_write_end(struct file *file, struct address_space *mapping,
2081                         loff_t pos, unsigned len, unsigned copied,
2082                         struct page *page, void *fsdata)
2083 {
2084         struct inode *inode = mapping->host;
2085         int i_size_changed = 0;
2086
2087         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2088
2089         /*
2090          * No need to use i_size_read() here, the i_size
2091          * cannot change under us because we hold i_mutex.
2092          *
2093          * But it's important to update i_size while still holding page lock:
2094          * page writeout could otherwise come in and zero beyond i_size.
2095          */
2096         if (pos+copied > inode->i_size) {
2097                 i_size_write(inode, pos+copied);
2098                 i_size_changed = 1;
2099         }
2100
2101         unlock_page(page);
2102         page_cache_release(page);
2103
2104         /*
2105          * Don't mark the inode dirty under page lock. First, it unnecessarily
2106          * makes the holding time of page lock longer. Second, it forces lock
2107          * ordering of page lock and transaction start for journaling
2108          * filesystems.
2109          */
2110         if (i_size_changed)
2111                 mark_inode_dirty(inode);
2112
2113         return copied;
2114 }
2115 EXPORT_SYMBOL(generic_write_end);
2116
2117 /*
2118  * block_is_partially_uptodate checks whether buffers within a page are
2119  * uptodate or not.
2120  *
2121  * Returns true if all buffers which correspond to a file portion
2122  * we want to read are uptodate.
2123  */
2124 int block_is_partially_uptodate(struct page *page, unsigned long from,
2125                                         unsigned long count)
2126 {
2127         unsigned block_start, block_end, blocksize;
2128         unsigned to;
2129         struct buffer_head *bh, *head;
2130         int ret = 1;
2131
2132         if (!page_has_buffers(page))
2133                 return 0;
2134
2135         head = page_buffers(page);
2136         blocksize = head->b_size;
2137         to = min_t(unsigned, PAGE_CACHE_SIZE - from, count);
2138         to = from + to;
2139         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2140                 return 0;
2141
2142         bh = head;
2143         block_start = 0;
2144         do {
2145                 block_end = block_start + blocksize;
2146                 if (block_end > from && block_start < to) {
2147                         if (!buffer_uptodate(bh)) {
2148                                 ret = 0;
2149                                 break;
2150                         }
2151                         if (block_end >= to)
2152                                 break;
2153                 }
2154                 block_start = block_end;
2155                 bh = bh->b_this_page;
2156         } while (bh != head);
2157
2158         return ret;
2159 }
2160 EXPORT_SYMBOL(block_is_partially_uptodate);
2161
2162 /*
2163  * Generic "read page" function for block devices that have the normal
2164  * get_block functionality. This is most of the block device filesystems.
2165  * Reads the page asynchronously --- the unlock_buffer() and
2166  * set/clear_buffer_uptodate() functions propagate buffer state into the
2167  * page struct once IO has completed.
2168  */
2169 int block_read_full_page(struct page *page, get_block_t *get_block)
2170 {
2171         struct inode *inode = page->mapping->host;
2172         sector_t iblock, lblock;
2173         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2174         unsigned int blocksize, bbits;
2175         int nr, i;
2176         int fully_mapped = 1;
2177
2178         head = create_page_buffers(page, inode, 0);
2179         blocksize = head->b_size;
2180         bbits = block_size_bits(blocksize);
2181
2182         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
2183         lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2184         bh = head;
2185         nr = 0;
2186         i = 0;
2187
2188         do {
2189                 if (buffer_uptodate(bh))
2190                         continue;
2191
2192                 if (!buffer_mapped(bh)) {
2193                         int err = 0;
2194
2195                         fully_mapped = 0;
2196                         if (iblock < lblock) {
2197                                 WARN_ON(bh->b_size != blocksize);
2198                                 err = get_block(inode, iblock, bh, 0);
2199                                 if (err)
2200                                         SetPageError(page);
2201                         }
2202                         if (!buffer_mapped(bh)) {
2203                                 zero_user(page, i * blocksize, blocksize);
2204                                 if (!err)
2205                                         set_buffer_uptodate(bh);
2206                                 continue;
2207                         }
2208                         /*
2209                          * get_block() might have updated the buffer
2210                          * synchronously
2211                          */
2212                         if (buffer_uptodate(bh))
2213                                 continue;
2214                 }
2215                 arr[nr++] = bh;
2216         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2217
2218         if (fully_mapped)
2219                 SetPageMappedToDisk(page);
2220
2221         if (!nr) {
2222                 /*
2223                  * All buffers are uptodate - we can set the page uptodate
2224                  * as well. But not if get_block() returned an error.
2225                  */
2226                 if (!PageError(page))
2227                         SetPageUptodate(page);
2228                 unlock_page(page);
2229                 return 0;
2230         }
2231
2232         /* Stage two: lock the buffers */
2233         for (i = 0; i < nr; i++) {
2234                 bh = arr[i];
2235                 lock_buffer(bh);
2236                 mark_buffer_async_read(bh);
2237         }
2238
2239         /*
2240          * Stage 3: start the IO.  Check for uptodateness
2241          * inside the buffer lock in case another process reading
2242          * the underlying blockdev brought it uptodate (the sct fix).
2243          */
2244         for (i = 0; i < nr; i++) {
2245                 bh = arr[i];
2246                 if (buffer_uptodate(bh))
2247                         end_buffer_async_read(bh, 1);
2248                 else
2249                         submit_bh(READ, bh);
2250         }
2251         return 0;
2252 }
2253 EXPORT_SYMBOL(block_read_full_page);
2254
2255 /* utility function for filesystems that need to do work on expanding
2256  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2257  * deal with the hole.  
2258  */
2259 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2260 {
2261         struct address_space *mapping = inode->i_mapping;
2262         struct page *page;
2263         void *fsdata;
2264         int err;
2265
2266         err = inode_newsize_ok(inode, size);
2267         if (err)
2268                 goto out;
2269
2270         err = pagecache_write_begin(NULL, mapping, size, 0,
2271                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2272                                 &page, &fsdata);
2273         if (err)
2274                 goto out;
2275
2276         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2277         BUG_ON(err > 0);
2278
2279 out:
2280         return err;
2281 }
2282 EXPORT_SYMBOL(generic_cont_expand_simple);
2283
2284 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2285                             loff_t pos, loff_t *bytes)
2286 {
2287         struct inode *inode = mapping->host;
2288         unsigned blocksize = 1 << inode->i_blkbits;
2289         struct page *page;
2290         void *fsdata;
2291         pgoff_t index, curidx;
2292         loff_t curpos;
2293         unsigned zerofrom, offset, len;
2294         int err = 0;
2295
2296         index = pos >> PAGE_CACHE_SHIFT;
2297         offset = pos & ~PAGE_CACHE_MASK;
2298
2299         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2300                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2301                 if (zerofrom & (blocksize-1)) {
2302                         *bytes |= (blocksize-1);
2303                         (*bytes)++;
2304                 }
2305                 len = PAGE_CACHE_SIZE - zerofrom;
2306
2307                 err = pagecache_write_begin(file, mapping, curpos, len,
2308                                                 AOP_FLAG_UNINTERRUPTIBLE,
2309                                                 &page, &fsdata);
2310                 if (err)
2311                         goto out;
2312                 zero_user(page, zerofrom, len);
2313                 err = pagecache_write_end(file, mapping, curpos, len, len,
2314                                                 page, fsdata);
2315                 if (err < 0)
2316                         goto out;
2317                 BUG_ON(err != len);
2318                 err = 0;
2319
2320                 balance_dirty_pages_ratelimited(mapping);
2321         }
2322
2323         /* page covers the boundary, find the boundary offset */
2324         if (index == curidx) {
2325                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2326                 /* if we will expand the thing last block will be filled */
2327                 if (offset <= zerofrom) {
2328                         goto out;
2329                 }
2330                 if (zerofrom & (blocksize-1)) {
2331                         *bytes |= (blocksize-1);
2332                         (*bytes)++;
2333                 }
2334                 len = offset - zerofrom;
2335
2336                 err = pagecache_write_begin(file, mapping, curpos, len,
2337                                                 AOP_FLAG_UNINTERRUPTIBLE,
2338                                                 &page, &fsdata);
2339                 if (err)
2340                         goto out;
2341                 zero_user(page, zerofrom, len);
2342                 err = pagecache_write_end(file, mapping, curpos, len, len,
2343                                                 page, fsdata);
2344                 if (err < 0)
2345                         goto out;
2346                 BUG_ON(err != len);
2347                 err = 0;
2348         }
2349 out:
2350         return err;
2351 }
2352
2353 /*
2354  * For moronic filesystems that do not allow holes in file.
2355  * We may have to extend the file.
2356  */
2357 int cont_write_begin(struct file *file, struct address_space *mapping,
2358                         loff_t pos, unsigned len, unsigned flags,
2359                         struct page **pagep, void **fsdata,
2360                         get_block_t *get_block, loff_t *bytes)
2361 {
2362         struct inode *inode = mapping->host;
2363         unsigned blocksize = 1 << inode->i_blkbits;
2364         unsigned zerofrom;
2365         int err;
2366
2367         err = cont_expand_zero(file, mapping, pos, bytes);
2368         if (err)
2369                 return err;
2370
2371         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2372         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2373                 *bytes |= (blocksize-1);
2374                 (*bytes)++;
2375         }
2376
2377         return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2378 }
2379 EXPORT_SYMBOL(cont_write_begin);
2380
2381 int block_commit_write(struct page *page, unsigned from, unsigned to)
2382 {
2383         struct inode *inode = page->mapping->host;
2384         __block_commit_write(inode,page,from,to);
2385         return 0;
2386 }
2387 EXPORT_SYMBOL(block_commit_write);
2388
2389 /*
2390  * block_page_mkwrite() is not allowed to change the file size as it gets
2391  * called from a page fault handler when a page is first dirtied. Hence we must
2392  * be careful to check for EOF conditions here. We set the page up correctly
2393  * for a written page which means we get ENOSPC checking when writing into
2394  * holes and correct delalloc and unwritten extent mapping on filesystems that
2395  * support these features.
2396  *
2397  * We are not allowed to take the i_mutex here so we have to play games to
2398  * protect against truncate races as the page could now be beyond EOF.  Because
2399  * truncate writes the inode size before removing pages, once we have the
2400  * page lock we can determine safely if the page is beyond EOF. If it is not
2401  * beyond EOF, then the page is guaranteed safe against truncation until we
2402  * unlock the page.
2403  *
2404  * Direct callers of this function should protect against filesystem freezing
2405  * using sb_start_write() - sb_end_write() functions.
2406  */
2407 int __block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2408                          get_block_t get_block)
2409 {
2410         struct page *page = vmf->page;
2411         struct inode *inode = file_inode(vma->vm_file);
2412         unsigned long end;
2413         loff_t size;
2414         int ret;
2415
2416         lock_page(page);
2417         size = i_size_read(inode);
2418         if ((page->mapping != inode->i_mapping) ||
2419             (page_offset(page) > size)) {
2420                 /* We overload EFAULT to mean page got truncated */
2421                 ret = -EFAULT;
2422                 goto out_unlock;
2423         }
2424
2425         /* page is wholly or partially inside EOF */
2426         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2427                 end = size & ~PAGE_CACHE_MASK;
2428         else
2429                 end = PAGE_CACHE_SIZE;
2430
2431         ret = __block_write_begin(page, 0, end, get_block);
2432         if (!ret)
2433                 ret = block_commit_write(page, 0, end);
2434
2435         if (unlikely(ret < 0))
2436                 goto out_unlock;
2437         set_page_dirty(page);
2438         wait_for_stable_page(page);
2439         return 0;
2440 out_unlock:
2441         unlock_page(page);
2442         return ret;
2443 }
2444 EXPORT_SYMBOL(__block_page_mkwrite);
2445
2446 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2447                    get_block_t get_block)
2448 {
2449         int ret;
2450         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
2451
2452         sb_start_pagefault(sb);
2453
2454         /*
2455          * Update file times before taking page lock. We may end up failing the
2456          * fault so this update may be superfluous but who really cares...
2457          */
2458         file_update_time(vma->vm_file);
2459
2460         ret = __block_page_mkwrite(vma, vmf, get_block);
2461         sb_end_pagefault(sb);
2462         return block_page_mkwrite_return(ret);
2463 }
2464 EXPORT_SYMBOL(block_page_mkwrite);
2465
2466 /*
2467  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2468  * immediately, while under the page lock.  So it needs a special end_io
2469  * handler which does not touch the bh after unlocking it.
2470  */
2471 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2472 {
2473         __end_buffer_read_notouch(bh, uptodate);
2474 }
2475
2476 /*
2477  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2478  * the page (converting it to circular linked list and taking care of page
2479  * dirty races).
2480  */
2481 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2482 {
2483         struct buffer_head *bh;
2484
2485         BUG_ON(!PageLocked(page));
2486
2487         spin_lock(&page->mapping->private_lock);
2488         bh = head;
2489         do {
2490                 if (PageDirty(page))
2491                         set_buffer_dirty(bh);
2492                 if (!bh->b_this_page)
2493                         bh->b_this_page = head;
2494                 bh = bh->b_this_page;
2495         } while (bh != head);
2496         attach_page_buffers(page, head);
2497         spin_unlock(&page->mapping->private_lock);
2498 }
2499
2500 /*
2501  * On entry, the page is fully not uptodate.
2502  * On exit the page is fully uptodate in the areas outside (from,to)
2503  * The filesystem needs to handle block truncation upon failure.
2504  */
2505 int nobh_write_begin(struct address_space *mapping,
2506                         loff_t pos, unsigned len, unsigned flags,
2507                         struct page **pagep, void **fsdata,
2508                         get_block_t *get_block)
2509 {
2510         struct inode *inode = mapping->host;
2511         const unsigned blkbits = inode->i_blkbits;
2512         const unsigned blocksize = 1 << blkbits;
2513         struct buffer_head *head, *bh;
2514         struct page *page;
2515         pgoff_t index;
2516         unsigned from, to;
2517         unsigned block_in_page;
2518         unsigned block_start, block_end;
2519         sector_t block_in_file;
2520         int nr_reads = 0;
2521         int ret = 0;
2522         int is_mapped_to_disk = 1;
2523
2524         index = pos >> PAGE_CACHE_SHIFT;
2525         from = pos & (PAGE_CACHE_SIZE - 1);
2526         to = from + len;
2527
2528         page = grab_cache_page_write_begin(mapping, index, flags);
2529         if (!page)
2530                 return -ENOMEM;
2531         *pagep = page;
2532         *fsdata = NULL;
2533
2534         if (page_has_buffers(page)) {
2535                 ret = __block_write_begin(page, pos, len, get_block);
2536                 if (unlikely(ret))
2537                         goto out_release;
2538                 return ret;
2539         }
2540
2541         if (PageMappedToDisk(page))
2542                 return 0;
2543
2544         /*
2545          * Allocate buffers so that we can keep track of state, and potentially
2546          * attach them to the page if an error occurs. In the common case of
2547          * no error, they will just be freed again without ever being attached
2548          * to the page (which is all OK, because we're under the page lock).
2549          *
2550          * Be careful: the buffer linked list is a NULL terminated one, rather
2551          * than the circular one we're used to.
2552          */
2553         head = alloc_page_buffers(page, blocksize, 0);
2554         if (!head) {
2555                 ret = -ENOMEM;
2556                 goto out_release;
2557         }
2558
2559         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2560
2561         /*
2562          * We loop across all blocks in the page, whether or not they are
2563          * part of the affected region.  This is so we can discover if the
2564          * page is fully mapped-to-disk.
2565          */
2566         for (block_start = 0, block_in_page = 0, bh = head;
2567                   block_start < PAGE_CACHE_SIZE;
2568                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2569                 int create;
2570
2571                 block_end = block_start + blocksize;
2572                 bh->b_state = 0;
2573                 create = 1;
2574                 if (block_start >= to)
2575                         create = 0;
2576                 ret = get_block(inode, block_in_file + block_in_page,
2577                                         bh, create);
2578                 if (ret)
2579                         goto failed;
2580                 if (!buffer_mapped(bh))
2581                         is_mapped_to_disk = 0;
2582                 if (buffer_new(bh))
2583                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2584                 if (PageUptodate(page)) {
2585                         set_buffer_uptodate(bh);
2586                         continue;
2587                 }
2588                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2589                         zero_user_segments(page, block_start, from,
2590                                                         to, block_end);
2591                         continue;
2592                 }
2593                 if (buffer_uptodate(bh))
2594                         continue;       /* reiserfs does this */
2595                 if (block_start < from || block_end > to) {
2596                         lock_buffer(bh);
2597                         bh->b_end_io = end_buffer_read_nobh;
2598                         submit_bh(READ, bh);
2599                         nr_reads++;
2600                 }
2601         }
2602
2603         if (nr_reads) {
2604                 /*
2605                  * The page is locked, so these buffers are protected from
2606                  * any VM or truncate activity.  Hence we don't need to care
2607                  * for the buffer_head refcounts.
2608                  */
2609                 for (bh = head; bh; bh = bh->b_this_page) {
2610                         wait_on_buffer(bh);
2611                         if (!buffer_uptodate(bh))
2612                                 ret = -EIO;
2613                 }
2614                 if (ret)
2615                         goto failed;
2616         }
2617
2618         if (is_mapped_to_disk)
2619                 SetPageMappedToDisk(page);
2620
2621         *fsdata = head; /* to be released by nobh_write_end */
2622
2623         return 0;
2624
2625 failed:
2626         BUG_ON(!ret);
2627         /*
2628          * Error recovery is a bit difficult. We need to zero out blocks that
2629          * were newly allocated, and dirty them to ensure they get written out.
2630          * Buffers need to be attached to the page at this point, otherwise
2631          * the handling of potential IO errors during writeout would be hard
2632          * (could try doing synchronous writeout, but what if that fails too?)
2633          */
2634         attach_nobh_buffers(page, head);
2635         page_zero_new_buffers(page, from, to);
2636
2637 out_release:
2638         unlock_page(page);
2639         page_cache_release(page);
2640         *pagep = NULL;
2641
2642         return ret;
2643 }
2644 EXPORT_SYMBOL(nobh_write_begin);
2645
2646 int nobh_write_end(struct file *file, struct address_space *mapping,
2647                         loff_t pos, unsigned len, unsigned copied,
2648                         struct page *page, void *fsdata)
2649 {
2650         struct inode *inode = page->mapping->host;
2651         struct buffer_head *head = fsdata;
2652         struct buffer_head *bh;
2653         BUG_ON(fsdata != NULL && page_has_buffers(page));
2654
2655         if (unlikely(copied < len) && head)
2656                 attach_nobh_buffers(page, head);
2657         if (page_has_buffers(page))
2658                 return generic_write_end(file, mapping, pos, len,
2659                                         copied, page, fsdata);
2660
2661         SetPageUptodate(page);
2662         set_page_dirty(page);
2663         if (pos+copied > inode->i_size) {
2664                 i_size_write(inode, pos+copied);
2665                 mark_inode_dirty(inode);
2666         }
2667
2668         unlock_page(page);
2669         page_cache_release(page);
2670
2671         while (head) {
2672                 bh = head;
2673                 head = head->b_this_page;
2674                 free_buffer_head(bh);
2675         }
2676
2677         return copied;
2678 }
2679 EXPORT_SYMBOL(nobh_write_end);
2680
2681 /*
2682  * nobh_writepage() - based on block_full_write_page() except
2683  * that it tries to operate without attaching bufferheads to
2684  * the page.
2685  */
2686 int nobh_writepage(struct page *page, get_block_t *get_block,
2687                         struct writeback_control *wbc)
2688 {
2689         struct inode * const inode = page->mapping->host;
2690         loff_t i_size = i_size_read(inode);
2691         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2692         unsigned offset;
2693         int ret;
2694
2695         /* Is the page fully inside i_size? */
2696         if (page->index < end_index)
2697                 goto out;
2698
2699         /* Is the page fully outside i_size? (truncate in progress) */
2700         offset = i_size & (PAGE_CACHE_SIZE-1);
2701         if (page->index >= end_index+1 || !offset) {
2702                 /*
2703                  * The page may have dirty, unmapped buffers.  For example,
2704                  * they may have been added in ext3_writepage().  Make them
2705                  * freeable here, so the page does not leak.
2706                  */
2707 #if 0
2708                 /* Not really sure about this  - do we need this ? */
2709                 if (page->mapping->a_ops->invalidatepage)
2710                         page->mapping->a_ops->invalidatepage(page, offset);
2711 #endif
2712                 unlock_page(page);
2713                 return 0; /* don't care */
2714         }
2715
2716         /*
2717          * The page straddles i_size.  It must be zeroed out on each and every
2718          * writepage invocation because it may be mmapped.  "A file is mapped
2719          * in multiples of the page size.  For a file that is not a multiple of
2720          * the  page size, the remaining memory is zeroed when mapped, and
2721          * writes to that region are not written out to the file."
2722          */
2723         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2724 out:
2725         ret = mpage_writepage(page, get_block, wbc);
2726         if (ret == -EAGAIN)
2727                 ret = __block_write_full_page(inode, page, get_block, wbc,
2728                                               end_buffer_async_write);
2729         return ret;
2730 }
2731 EXPORT_SYMBOL(nobh_writepage);
2732
2733 int nobh_truncate_page(struct address_space *mapping,
2734                         loff_t from, get_block_t *get_block)
2735 {
2736         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2737         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2738         unsigned blocksize;
2739         sector_t iblock;
2740         unsigned length, pos;
2741         struct inode *inode = mapping->host;
2742         struct page *page;
2743         struct buffer_head map_bh;
2744         int err;
2745
2746         blocksize = 1 << inode->i_blkbits;
2747         length = offset & (blocksize - 1);
2748
2749         /* Block boundary? Nothing to do */
2750         if (!length)
2751                 return 0;
2752
2753         length = blocksize - length;
2754         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2755
2756         page = grab_cache_page(mapping, index);
2757         err = -ENOMEM;
2758         if (!page)
2759                 goto out;
2760
2761         if (page_has_buffers(page)) {
2762 has_buffers:
2763                 unlock_page(page);
2764                 page_cache_release(page);
2765                 return block_truncate_page(mapping, from, get_block);
2766         }
2767
2768         /* Find the buffer that contains "offset" */
2769         pos = blocksize;
2770         while (offset >= pos) {
2771                 iblock++;
2772                 pos += blocksize;
2773         }
2774
2775         map_bh.b_size = blocksize;
2776         map_bh.b_state = 0;
2777         err = get_block(inode, iblock, &map_bh, 0);
2778         if (err)
2779                 goto unlock;
2780         /* unmapped? It's a hole - nothing to do */
2781         if (!buffer_mapped(&map_bh))
2782                 goto unlock;
2783
2784         /* Ok, it's mapped. Make sure it's up-to-date */
2785         if (!PageUptodate(page)) {
2786                 err = mapping->a_ops->readpage(NULL, page);
2787                 if (err) {
2788                         page_cache_release(page);
2789                         goto out;
2790                 }
2791                 lock_page(page);
2792                 if (!PageUptodate(page)) {
2793                         err = -EIO;
2794                         goto unlock;
2795                 }
2796                 if (page_has_buffers(page))
2797                         goto has_buffers;
2798         }
2799         zero_user(page, offset, length);
2800         set_page_dirty(page);
2801         err = 0;
2802
2803 unlock:
2804         unlock_page(page);
2805         page_cache_release(page);
2806 out:
2807         return err;
2808 }
2809 EXPORT_SYMBOL(nobh_truncate_page);
2810
2811 int block_truncate_page(struct address_space *mapping,
2812                         loff_t from, get_block_t *get_block)
2813 {
2814         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2815         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2816         unsigned blocksize;
2817         sector_t iblock;
2818         unsigned length, pos;
2819         struct inode *inode = mapping->host;
2820         struct page *page;
2821         struct buffer_head *bh;
2822         int err;
2823
2824         blocksize = 1 << inode->i_blkbits;
2825         length = offset & (blocksize - 1);
2826
2827         /* Block boundary? Nothing to do */
2828         if (!length)
2829                 return 0;
2830
2831         length = blocksize - length;
2832         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2833         
2834         page = grab_cache_page(mapping, index);
2835         err = -ENOMEM;
2836         if (!page)
2837                 goto out;
2838
2839         if (!page_has_buffers(page))
2840                 create_empty_buffers(page, blocksize, 0);
2841
2842         /* Find the buffer that contains "offset" */
2843         bh = page_buffers(page);
2844         pos = blocksize;
2845         while (offset >= pos) {
2846                 bh = bh->b_this_page;
2847                 iblock++;
2848                 pos += blocksize;
2849         }
2850
2851         err = 0;
2852         if (!buffer_mapped(bh)) {
2853                 WARN_ON(bh->b_size != blocksize);
2854                 err = get_block(inode, iblock, bh, 0);
2855                 if (err)
2856                         goto unlock;
2857                 /* unmapped? It's a hole - nothing to do */
2858                 if (!buffer_mapped(bh))
2859                         goto unlock;
2860         }
2861
2862         /* Ok, it's mapped. Make sure it's up-to-date */
2863         if (PageUptodate(page))
2864                 set_buffer_uptodate(bh);
2865
2866         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2867                 err = -EIO;
2868                 ll_rw_block(READ, 1, &bh);
2869                 wait_on_buffer(bh);
2870                 /* Uhhuh. Read error. Complain and punt. */
2871                 if (!buffer_uptodate(bh))
2872                         goto unlock;
2873         }
2874
2875         zero_user(page, offset, length);
2876         mark_buffer_dirty(bh);
2877         err = 0;
2878
2879 unlock:
2880         unlock_page(page);
2881         page_cache_release(page);
2882 out:
2883         return err;
2884 }
2885 EXPORT_SYMBOL(block_truncate_page);
2886
2887 /*
2888  * The generic ->writepage function for buffer-backed address_spaces
2889  */
2890 int block_write_full_page(struct page *page, get_block_t *get_block,
2891                         struct writeback_control *wbc)
2892 {
2893         struct inode * const inode = page->mapping->host;
2894         loff_t i_size = i_size_read(inode);
2895         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2896         unsigned offset;
2897
2898         /* Is the page fully inside i_size? */
2899         if (page->index < end_index)
2900                 return __block_write_full_page(inode, page, get_block, wbc,
2901                                                end_buffer_async_write);
2902
2903         /* Is the page fully outside i_size? (truncate in progress) */
2904         offset = i_size & (PAGE_CACHE_SIZE-1);
2905         if (page->index >= end_index+1 || !offset) {
2906                 /*
2907                  * The page may have dirty, unmapped buffers.  For example,
2908                  * they may have been added in ext3_writepage().  Make them
2909                  * freeable here, so the page does not leak.
2910                  */
2911                 do_invalidatepage(page, 0, PAGE_CACHE_SIZE);
2912                 unlock_page(page);
2913                 return 0; /* don't care */
2914         }
2915
2916         /*
2917          * The page straddles i_size.  It must be zeroed out on each and every
2918          * writepage invocation because it may be mmapped.  "A file is mapped
2919          * in multiples of the page size.  For a file that is not a multiple of
2920          * the  page size, the remaining memory is zeroed when mapped, and
2921          * writes to that region are not written out to the file."
2922          */
2923         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2924         return __block_write_full_page(inode, page, get_block, wbc,
2925                                                         end_buffer_async_write);
2926 }
2927 EXPORT_SYMBOL(block_write_full_page);
2928
2929 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2930                             get_block_t *get_block)
2931 {
2932         struct buffer_head tmp;
2933         struct inode *inode = mapping->host;
2934         tmp.b_state = 0;
2935         tmp.b_blocknr = 0;
2936         tmp.b_size = 1 << inode->i_blkbits;
2937         get_block(inode, block, &tmp, 0);
2938         return tmp.b_blocknr;
2939 }
2940 EXPORT_SYMBOL(generic_block_bmap);
2941
2942 static void end_bio_bh_io_sync(struct bio *bio, int err)
2943 {
2944         struct buffer_head *bh = bio->bi_private;
2945
2946         if (err == -EOPNOTSUPP) {
2947                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2948         }
2949
2950         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2951                 set_bit(BH_Quiet, &bh->b_state);
2952
2953         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2954         bio_put(bio);
2955 }
2956
2957 /*
2958  * This allows us to do IO even on the odd last sectors
2959  * of a device, even if the bh block size is some multiple
2960  * of the physical sector size.
2961  *
2962  * We'll just truncate the bio to the size of the device,
2963  * and clear the end of the buffer head manually.
2964  *
2965  * Truly out-of-range accesses will turn into actual IO
2966  * errors, this only handles the "we need to be able to
2967  * do IO at the final sector" case.
2968  */
2969 static void guard_bh_eod(int rw, struct bio *bio, struct buffer_head *bh)
2970 {
2971         sector_t maxsector;
2972         unsigned bytes;
2973
2974         maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
2975         if (!maxsector)
2976                 return;
2977
2978         /*
2979          * If the *whole* IO is past the end of the device,
2980          * let it through, and the IO layer will turn it into
2981          * an EIO.
2982          */
2983         if (unlikely(bio->bi_iter.bi_sector >= maxsector))
2984                 return;
2985
2986         maxsector -= bio->bi_iter.bi_sector;
2987         bytes = bio->bi_iter.bi_size;
2988         if (likely((bytes >> 9) <= maxsector))
2989                 return;
2990
2991         /* Uhhuh. We've got a bh that straddles the device size! */
2992         bytes = maxsector << 9;
2993
2994         /* Truncate the bio.. */
2995         bio->bi_iter.bi_size = bytes;
2996         bio->bi_io_vec[0].bv_len = bytes;
2997
2998         /* ..and clear the end of the buffer for reads */
2999         if ((rw & RW_MASK) == READ) {
3000                 void *kaddr = kmap_atomic(bh->b_page);
3001                 memset(kaddr + bh_offset(bh) + bytes, 0, bh->b_size - bytes);
3002                 kunmap_atomic(kaddr);
3003                 flush_dcache_page(bh->b_page);
3004         }
3005 }
3006
3007 int _submit_bh(int rw, struct buffer_head *bh, unsigned long bio_flags)
3008 {
3009         struct bio *bio;
3010         int ret = 0;
3011
3012         BUG_ON(!buffer_locked(bh));
3013         BUG_ON(!buffer_mapped(bh));
3014         BUG_ON(!bh->b_end_io);
3015         BUG_ON(buffer_delay(bh));
3016         BUG_ON(buffer_unwritten(bh));
3017
3018         /*
3019          * Only clear out a write error when rewriting
3020          */
3021         if (test_set_buffer_req(bh) && (rw & WRITE))
3022                 clear_buffer_write_io_error(bh);
3023
3024         /*
3025          * from here on down, it's all bio -- do the initial mapping,
3026          * submit_bio -> generic_make_request may further map this bio around
3027          */
3028         bio = bio_alloc(GFP_NOIO, 1);
3029
3030         bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3031         bio->bi_bdev = bh->b_bdev;
3032         bio->bi_io_vec[0].bv_page = bh->b_page;
3033         bio->bi_io_vec[0].bv_len = bh->b_size;
3034         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
3035
3036         bio->bi_vcnt = 1;
3037         bio->bi_iter.bi_size = bh->b_size;
3038
3039         bio->bi_end_io = end_bio_bh_io_sync;
3040         bio->bi_private = bh;
3041         bio->bi_flags |= bio_flags;
3042
3043         /* Take care of bh's that straddle the end of the device */
3044         guard_bh_eod(rw, bio, bh);
3045
3046         if (buffer_meta(bh))
3047                 rw |= REQ_META;
3048         if (buffer_prio(bh))
3049                 rw |= REQ_PRIO;
3050
3051         bio_get(bio);
3052         submit_bio(rw, bio);
3053
3054         if (bio_flagged(bio, BIO_EOPNOTSUPP))
3055                 ret = -EOPNOTSUPP;
3056
3057         bio_put(bio);
3058         return ret;
3059 }
3060 EXPORT_SYMBOL_GPL(_submit_bh);
3061
3062 int submit_bh(int rw, struct buffer_head *bh)
3063 {
3064         return _submit_bh(rw, bh, 0);
3065 }
3066 EXPORT_SYMBOL(submit_bh);
3067
3068 /**
3069  * ll_rw_block: low-level access to block devices (DEPRECATED)
3070  * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
3071  * @nr: number of &struct buffer_heads in the array
3072  * @bhs: array of pointers to &struct buffer_head
3073  *
3074  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3075  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
3076  * %READA option is described in the documentation for generic_make_request()
3077  * which ll_rw_block() calls.
3078  *
3079  * This function drops any buffer that it cannot get a lock on (with the
3080  * BH_Lock state bit), any buffer that appears to be clean when doing a write
3081  * request, and any buffer that appears to be up-to-date when doing read
3082  * request.  Further it marks as clean buffers that are processed for
3083  * writing (the buffer cache won't assume that they are actually clean
3084  * until the buffer gets unlocked).
3085  *
3086  * ll_rw_block sets b_end_io to simple completion handler that marks
3087  * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3088  * any waiters. 
3089  *
3090  * All of the buffers must be for the same device, and must also be a
3091  * multiple of the current approved size for the device.
3092  */
3093 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
3094 {
3095         int i;
3096
3097         for (i = 0; i < nr; i++) {
3098                 struct buffer_head *bh = bhs[i];
3099
3100                 if (!trylock_buffer(bh))
3101                         continue;
3102                 if (rw == WRITE) {
3103                         if (test_clear_buffer_dirty(bh)) {
3104                                 bh->b_end_io = end_buffer_write_sync;
3105                                 get_bh(bh);
3106                                 submit_bh(WRITE, bh);
3107                                 continue;
3108                         }
3109                 } else {
3110                         if (!buffer_uptodate(bh)) {
3111                                 bh->b_end_io = end_buffer_read_sync;
3112                                 get_bh(bh);
3113                                 submit_bh(rw, bh);
3114                                 continue;
3115                         }
3116                 }
3117                 unlock_buffer(bh);
3118         }
3119 }
3120 EXPORT_SYMBOL(ll_rw_block);
3121
3122 void write_dirty_buffer(struct buffer_head *bh, int rw)
3123 {
3124         lock_buffer(bh);
3125         if (!test_clear_buffer_dirty(bh)) {
3126                 unlock_buffer(bh);
3127                 return;
3128         }
3129         bh->b_end_io = end_buffer_write_sync;
3130         get_bh(bh);
3131         submit_bh(rw, bh);
3132 }
3133 EXPORT_SYMBOL(write_dirty_buffer);
3134
3135 /*
3136  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3137  * and then start new I/O and then wait upon it.  The caller must have a ref on
3138  * the buffer_head.
3139  */
3140 int __sync_dirty_buffer(struct buffer_head *bh, int rw)
3141 {
3142         int ret = 0;
3143
3144         WARN_ON(atomic_read(&bh->b_count) < 1);
3145         lock_buffer(bh);
3146         if (test_clear_buffer_dirty(bh)) {
3147                 get_bh(bh);
3148                 bh->b_end_io = end_buffer_write_sync;
3149                 ret = submit_bh(rw, bh);
3150                 wait_on_buffer(bh);
3151                 if (!ret && !buffer_uptodate(bh))
3152                         ret = -EIO;
3153         } else {
3154                 unlock_buffer(bh);
3155         }
3156         return ret;
3157 }
3158 EXPORT_SYMBOL(__sync_dirty_buffer);
3159
3160 int sync_dirty_buffer(struct buffer_head *bh)
3161 {
3162         return __sync_dirty_buffer(bh, WRITE_SYNC);
3163 }
3164 EXPORT_SYMBOL(sync_dirty_buffer);
3165
3166 /*
3167  * try_to_free_buffers() checks if all the buffers on this particular page
3168  * are unused, and releases them if so.
3169  *
3170  * Exclusion against try_to_free_buffers may be obtained by either
3171  * locking the page or by holding its mapping's private_lock.
3172  *
3173  * If the page is dirty but all the buffers are clean then we need to
3174  * be sure to mark the page clean as well.  This is because the page
3175  * may be against a block device, and a later reattachment of buffers
3176  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3177  * filesystem data on the same device.
3178  *
3179  * The same applies to regular filesystem pages: if all the buffers are
3180  * clean then we set the page clean and proceed.  To do that, we require
3181  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3182  * private_lock.
3183  *
3184  * try_to_free_buffers() is non-blocking.
3185  */
3186 static inline int buffer_busy(struct buffer_head *bh)
3187 {
3188         return atomic_read(&bh->b_count) |
3189                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3190 }
3191
3192 static int
3193 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3194 {
3195         struct buffer_head *head = page_buffers(page);
3196         struct buffer_head *bh;
3197
3198         bh = head;
3199         do {
3200                 if (buffer_write_io_error(bh) && page->mapping)
3201                         set_bit(AS_EIO, &page->mapping->flags);
3202                 if (buffer_busy(bh))
3203                         goto failed;
3204                 bh = bh->b_this_page;
3205         } while (bh != head);
3206
3207         do {
3208                 struct buffer_head *next = bh->b_this_page;
3209
3210                 if (bh->b_assoc_map)
3211                         __remove_assoc_queue(bh);
3212                 bh = next;
3213         } while (bh != head);
3214         *buffers_to_free = head;
3215         __clear_page_buffers(page);
3216         return 1;
3217 failed:
3218         return 0;
3219 }
3220
3221 int try_to_free_buffers(struct page *page)
3222 {
3223         struct address_space * const mapping = page->mapping;
3224         struct buffer_head *buffers_to_free = NULL;
3225         int ret = 0;
3226
3227         BUG_ON(!PageLocked(page));
3228         if (PageWriteback(page))
3229                 return 0;
3230
3231         if (mapping == NULL) {          /* can this still happen? */
3232                 ret = drop_buffers(page, &buffers_to_free);
3233                 goto out;
3234         }
3235
3236         spin_lock(&mapping->private_lock);
3237         ret = drop_buffers(page, &buffers_to_free);
3238
3239         /*
3240          * If the filesystem writes its buffers by hand (eg ext3)
3241          * then we can have clean buffers against a dirty page.  We
3242          * clean the page here; otherwise the VM will never notice
3243          * that the filesystem did any IO at all.
3244          *
3245          * Also, during truncate, discard_buffer will have marked all
3246          * the page's buffers clean.  We discover that here and clean
3247          * the page also.
3248          *
3249          * private_lock must be held over this entire operation in order
3250          * to synchronise against __set_page_dirty_buffers and prevent the
3251          * dirty bit from being lost.
3252          */
3253         if (ret)
3254                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3255         spin_unlock(&mapping->private_lock);
3256 out:
3257         if (buffers_to_free) {
3258                 struct buffer_head *bh = buffers_to_free;
3259
3260                 do {
3261                         struct buffer_head *next = bh->b_this_page;
3262                         free_buffer_head(bh);
3263                         bh = next;
3264                 } while (bh != buffers_to_free);
3265         }
3266         return ret;
3267 }
3268 EXPORT_SYMBOL(try_to_free_buffers);
3269
3270 /*
3271  * There are no bdflush tunables left.  But distributions are
3272  * still running obsolete flush daemons, so we terminate them here.
3273  *
3274  * Use of bdflush() is deprecated and will be removed in a future kernel.
3275  * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3276  */
3277 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3278 {
3279         static int msg_count;
3280
3281         if (!capable(CAP_SYS_ADMIN))
3282                 return -EPERM;
3283
3284         if (msg_count < 5) {
3285                 msg_count++;
3286                 printk(KERN_INFO
3287                         "warning: process `%s' used the obsolete bdflush"
3288                         " system call\n", current->comm);
3289                 printk(KERN_INFO "Fix your initscripts?\n");
3290         }
3291
3292         if (func == 1)
3293                 do_exit(0);
3294         return 0;
3295 }
3296
3297 /*
3298  * Buffer-head allocation
3299  */
3300 static struct kmem_cache *bh_cachep __read_mostly;
3301
3302 /*
3303  * Once the number of bh's in the machine exceeds this level, we start
3304  * stripping them in writeback.
3305  */
3306 static unsigned long max_buffer_heads;
3307
3308 int buffer_heads_over_limit;
3309
3310 struct bh_accounting {
3311         int nr;                 /* Number of live bh's */
3312         int ratelimit;          /* Limit cacheline bouncing */
3313 };
3314
3315 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3316
3317 static void recalc_bh_state(void)
3318 {
3319         int i;
3320         int tot = 0;
3321
3322         if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3323                 return;
3324         __this_cpu_write(bh_accounting.ratelimit, 0);
3325         for_each_online_cpu(i)
3326                 tot += per_cpu(bh_accounting, i).nr;
3327         buffer_heads_over_limit = (tot > max_buffer_heads);
3328 }
3329
3330 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3331 {
3332         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3333         if (ret) {
3334                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3335                 preempt_disable();
3336                 __this_cpu_inc(bh_accounting.nr);
3337                 recalc_bh_state();
3338                 preempt_enable();
3339         }
3340         return ret;
3341 }
3342 EXPORT_SYMBOL(alloc_buffer_head);
3343
3344 void free_buffer_head(struct buffer_head *bh)
3345 {
3346         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3347         kmem_cache_free(bh_cachep, bh);
3348         preempt_disable();
3349         __this_cpu_dec(bh_accounting.nr);
3350         recalc_bh_state();
3351         preempt_enable();
3352 }
3353 EXPORT_SYMBOL(free_buffer_head);
3354
3355 static void buffer_exit_cpu(int cpu)
3356 {
3357         int i;
3358         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3359
3360         for (i = 0; i < BH_LRU_SIZE; i++) {
3361                 brelse(b->bhs[i]);
3362                 b->bhs[i] = NULL;
3363         }
3364         this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3365         per_cpu(bh_accounting, cpu).nr = 0;
3366 }
3367
3368 static int buffer_cpu_notify(struct notifier_block *self,
3369                               unsigned long action, void *hcpu)
3370 {
3371         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3372                 buffer_exit_cpu((unsigned long)hcpu);
3373         return NOTIFY_OK;
3374 }
3375
3376 /**
3377  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3378  * @bh: struct buffer_head
3379  *
3380  * Return true if the buffer is up-to-date and false,
3381  * with the buffer locked, if not.
3382  */
3383 int bh_uptodate_or_lock(struct buffer_head *bh)
3384 {
3385         if (!buffer_uptodate(bh)) {
3386                 lock_buffer(bh);
3387                 if (!buffer_uptodate(bh))
3388                         return 0;
3389                 unlock_buffer(bh);
3390         }
3391         return 1;
3392 }
3393 EXPORT_SYMBOL(bh_uptodate_or_lock);
3394
3395 /**
3396  * bh_submit_read - Submit a locked buffer for reading
3397  * @bh: struct buffer_head
3398  *
3399  * Returns zero on success and -EIO on error.
3400  */
3401 int bh_submit_read(struct buffer_head *bh)
3402 {
3403         BUG_ON(!buffer_locked(bh));
3404
3405         if (buffer_uptodate(bh)) {
3406                 unlock_buffer(bh);
3407                 return 0;
3408         }
3409
3410         get_bh(bh);
3411         bh->b_end_io = end_buffer_read_sync;
3412         submit_bh(READ, bh);
3413         wait_on_buffer(bh);
3414         if (buffer_uptodate(bh))
3415                 return 0;
3416         return -EIO;
3417 }
3418 EXPORT_SYMBOL(bh_submit_read);
3419
3420 void __init buffer_init(void)
3421 {
3422         unsigned long nrpages;
3423
3424         bh_cachep = kmem_cache_create("buffer_head",
3425                         sizeof(struct buffer_head), 0,
3426                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3427                                 SLAB_MEM_SPREAD),
3428                                 NULL);
3429
3430         /*
3431          * Limit the bh occupancy to 10% of ZONE_NORMAL
3432          */
3433         nrpages = (nr_free_buffer_pages() * 10) / 100;
3434         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3435         hotcpu_notifier(buffer_cpu_notify, 0);
3436 }