Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jikos/hid
[profile/ivi/kernel-x86-ivi.git] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/capability.h>
28 #include <linux/blkdev.h>
29 #include <linux/file.h>
30 #include <linux/quotaops.h>
31 #include <linux/highmem.h>
32 #include <linux/export.h>
33 #include <linux/writeback.h>
34 #include <linux/hash.h>
35 #include <linux/suspend.h>
36 #include <linux/buffer_head.h>
37 #include <linux/task_io_accounting_ops.h>
38 #include <linux/bio.h>
39 #include <linux/notifier.h>
40 #include <linux/cpu.h>
41 #include <linux/bitops.h>
42 #include <linux/mpage.h>
43 #include <linux/bit_spinlock.h>
44 #include <trace/events/block.h>
45
46 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
47
48 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
49
50 void init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
51 {
52         bh->b_end_io = handler;
53         bh->b_private = private;
54 }
55 EXPORT_SYMBOL(init_buffer);
56
57 inline void touch_buffer(struct buffer_head *bh)
58 {
59         trace_block_touch_buffer(bh);
60         mark_page_accessed(bh->b_page);
61 }
62 EXPORT_SYMBOL(touch_buffer);
63
64 static int sleep_on_buffer(void *word)
65 {
66         io_schedule();
67         return 0;
68 }
69
70 void __lock_buffer(struct buffer_head *bh)
71 {
72         wait_on_bit_lock(&bh->b_state, BH_Lock, sleep_on_buffer,
73                                                         TASK_UNINTERRUPTIBLE);
74 }
75 EXPORT_SYMBOL(__lock_buffer);
76
77 void unlock_buffer(struct buffer_head *bh)
78 {
79         clear_bit_unlock(BH_Lock, &bh->b_state);
80         smp_mb__after_clear_bit();
81         wake_up_bit(&bh->b_state, BH_Lock);
82 }
83 EXPORT_SYMBOL(unlock_buffer);
84
85 /*
86  * Returns if the page has dirty or writeback buffers. If all the buffers
87  * are unlocked and clean then the PageDirty information is stale. If
88  * any of the pages are locked, it is assumed they are locked for IO.
89  */
90 void buffer_check_dirty_writeback(struct page *page,
91                                      bool *dirty, bool *writeback)
92 {
93         struct buffer_head *head, *bh;
94         *dirty = false;
95         *writeback = false;
96
97         BUG_ON(!PageLocked(page));
98
99         if (!page_has_buffers(page))
100                 return;
101
102         if (PageWriteback(page))
103                 *writeback = true;
104
105         head = page_buffers(page);
106         bh = head;
107         do {
108                 if (buffer_locked(bh))
109                         *writeback = true;
110
111                 if (buffer_dirty(bh))
112                         *dirty = true;
113
114                 bh = bh->b_this_page;
115         } while (bh != head);
116 }
117 EXPORT_SYMBOL(buffer_check_dirty_writeback);
118
119 /*
120  * Block until a buffer comes unlocked.  This doesn't stop it
121  * from becoming locked again - you have to lock it yourself
122  * if you want to preserve its state.
123  */
124 void __wait_on_buffer(struct buffer_head * bh)
125 {
126         wait_on_bit(&bh->b_state, BH_Lock, sleep_on_buffer, TASK_UNINTERRUPTIBLE);
127 }
128 EXPORT_SYMBOL(__wait_on_buffer);
129
130 static void
131 __clear_page_buffers(struct page *page)
132 {
133         ClearPagePrivate(page);
134         set_page_private(page, 0);
135         page_cache_release(page);
136 }
137
138
139 static int quiet_error(struct buffer_head *bh)
140 {
141         if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
142                 return 0;
143         return 1;
144 }
145
146
147 static void buffer_io_error(struct buffer_head *bh)
148 {
149         char b[BDEVNAME_SIZE];
150         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
151                         bdevname(bh->b_bdev, b),
152                         (unsigned long long)bh->b_blocknr);
153 }
154
155 /*
156  * End-of-IO handler helper function which does not touch the bh after
157  * unlocking it.
158  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
159  * a race there is benign: unlock_buffer() only use the bh's address for
160  * hashing after unlocking the buffer, so it doesn't actually touch the bh
161  * itself.
162  */
163 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
164 {
165         if (uptodate) {
166                 set_buffer_uptodate(bh);
167         } else {
168                 /* This happens, due to failed READA attempts. */
169                 clear_buffer_uptodate(bh);
170         }
171         unlock_buffer(bh);
172 }
173
174 /*
175  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
176  * unlock the buffer. This is what ll_rw_block uses too.
177  */
178 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
179 {
180         __end_buffer_read_notouch(bh, uptodate);
181         put_bh(bh);
182 }
183 EXPORT_SYMBOL(end_buffer_read_sync);
184
185 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
186 {
187         char b[BDEVNAME_SIZE];
188
189         if (uptodate) {
190                 set_buffer_uptodate(bh);
191         } else {
192                 if (!quiet_error(bh)) {
193                         buffer_io_error(bh);
194                         printk(KERN_WARNING "lost page write due to "
195                                         "I/O error on %s\n",
196                                        bdevname(bh->b_bdev, b));
197                 }
198                 set_buffer_write_io_error(bh);
199                 clear_buffer_uptodate(bh);
200         }
201         unlock_buffer(bh);
202         put_bh(bh);
203 }
204 EXPORT_SYMBOL(end_buffer_write_sync);
205
206 /*
207  * Various filesystems appear to want __find_get_block to be non-blocking.
208  * But it's the page lock which protects the buffers.  To get around this,
209  * we get exclusion from try_to_free_buffers with the blockdev mapping's
210  * private_lock.
211  *
212  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
213  * may be quite high.  This code could TryLock the page, and if that
214  * succeeds, there is no need to take private_lock. (But if
215  * private_lock is contended then so is mapping->tree_lock).
216  */
217 static struct buffer_head *
218 __find_get_block_slow(struct block_device *bdev, sector_t block)
219 {
220         struct inode *bd_inode = bdev->bd_inode;
221         struct address_space *bd_mapping = bd_inode->i_mapping;
222         struct buffer_head *ret = NULL;
223         pgoff_t index;
224         struct buffer_head *bh;
225         struct buffer_head *head;
226         struct page *page;
227         int all_mapped = 1;
228
229         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
230         page = find_get_page(bd_mapping, index);
231         if (!page)
232                 goto out;
233
234         spin_lock(&bd_mapping->private_lock);
235         if (!page_has_buffers(page))
236                 goto out_unlock;
237         head = page_buffers(page);
238         bh = head;
239         do {
240                 if (!buffer_mapped(bh))
241                         all_mapped = 0;
242                 else if (bh->b_blocknr == block) {
243                         ret = bh;
244                         get_bh(bh);
245                         goto out_unlock;
246                 }
247                 bh = bh->b_this_page;
248         } while (bh != head);
249
250         /* we might be here because some of the buffers on this page are
251          * not mapped.  This is due to various races between
252          * file io on the block device and getblk.  It gets dealt with
253          * elsewhere, don't buffer_error if we had some unmapped buffers
254          */
255         if (all_mapped) {
256                 char b[BDEVNAME_SIZE];
257
258                 printk("__find_get_block_slow() failed. "
259                         "block=%llu, b_blocknr=%llu\n",
260                         (unsigned long long)block,
261                         (unsigned long long)bh->b_blocknr);
262                 printk("b_state=0x%08lx, b_size=%zu\n",
263                         bh->b_state, bh->b_size);
264                 printk("device %s blocksize: %d\n", bdevname(bdev, b),
265                         1 << bd_inode->i_blkbits);
266         }
267 out_unlock:
268         spin_unlock(&bd_mapping->private_lock);
269         page_cache_release(page);
270 out:
271         return ret;
272 }
273
274 /*
275  * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
276  */
277 static void free_more_memory(void)
278 {
279         struct zone *zone;
280         int nid;
281
282         wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
283         yield();
284
285         for_each_online_node(nid) {
286                 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
287                                                 gfp_zone(GFP_NOFS), NULL,
288                                                 &zone);
289                 if (zone)
290                         try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
291                                                 GFP_NOFS, NULL);
292         }
293 }
294
295 /*
296  * I/O completion handler for block_read_full_page() - pages
297  * which come unlocked at the end of I/O.
298  */
299 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
300 {
301         unsigned long flags;
302         struct buffer_head *first;
303         struct buffer_head *tmp;
304         struct page *page;
305         int page_uptodate = 1;
306
307         BUG_ON(!buffer_async_read(bh));
308
309         page = bh->b_page;
310         if (uptodate) {
311                 set_buffer_uptodate(bh);
312         } else {
313                 clear_buffer_uptodate(bh);
314                 if (!quiet_error(bh))
315                         buffer_io_error(bh);
316                 SetPageError(page);
317         }
318
319         /*
320          * Be _very_ careful from here on. Bad things can happen if
321          * two buffer heads end IO at almost the same time and both
322          * decide that the page is now completely done.
323          */
324         first = page_buffers(page);
325         local_irq_save(flags);
326         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
327         clear_buffer_async_read(bh);
328         unlock_buffer(bh);
329         tmp = bh;
330         do {
331                 if (!buffer_uptodate(tmp))
332                         page_uptodate = 0;
333                 if (buffer_async_read(tmp)) {
334                         BUG_ON(!buffer_locked(tmp));
335                         goto still_busy;
336                 }
337                 tmp = tmp->b_this_page;
338         } while (tmp != bh);
339         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
340         local_irq_restore(flags);
341
342         /*
343          * If none of the buffers had errors and they are all
344          * uptodate then we can set the page uptodate.
345          */
346         if (page_uptodate && !PageError(page))
347                 SetPageUptodate(page);
348         unlock_page(page);
349         return;
350
351 still_busy:
352         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
353         local_irq_restore(flags);
354         return;
355 }
356
357 /*
358  * Completion handler for block_write_full_page() - pages which are unlocked
359  * during I/O, and which have PageWriteback cleared upon I/O completion.
360  */
361 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
362 {
363         char b[BDEVNAME_SIZE];
364         unsigned long flags;
365         struct buffer_head *first;
366         struct buffer_head *tmp;
367         struct page *page;
368
369         BUG_ON(!buffer_async_write(bh));
370
371         page = bh->b_page;
372         if (uptodate) {
373                 set_buffer_uptodate(bh);
374         } else {
375                 if (!quiet_error(bh)) {
376                         buffer_io_error(bh);
377                         printk(KERN_WARNING "lost page write due to "
378                                         "I/O error on %s\n",
379                                bdevname(bh->b_bdev, b));
380                 }
381                 set_bit(AS_EIO, &page->mapping->flags);
382                 set_buffer_write_io_error(bh);
383                 clear_buffer_uptodate(bh);
384                 SetPageError(page);
385         }
386
387         first = page_buffers(page);
388         local_irq_save(flags);
389         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
390
391         clear_buffer_async_write(bh);
392         unlock_buffer(bh);
393         tmp = bh->b_this_page;
394         while (tmp != bh) {
395                 if (buffer_async_write(tmp)) {
396                         BUG_ON(!buffer_locked(tmp));
397                         goto still_busy;
398                 }
399                 tmp = tmp->b_this_page;
400         }
401         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
402         local_irq_restore(flags);
403         end_page_writeback(page);
404         return;
405
406 still_busy:
407         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
408         local_irq_restore(flags);
409         return;
410 }
411 EXPORT_SYMBOL(end_buffer_async_write);
412
413 /*
414  * If a page's buffers are under async readin (end_buffer_async_read
415  * completion) then there is a possibility that another thread of
416  * control could lock one of the buffers after it has completed
417  * but while some of the other buffers have not completed.  This
418  * locked buffer would confuse end_buffer_async_read() into not unlocking
419  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
420  * that this buffer is not under async I/O.
421  *
422  * The page comes unlocked when it has no locked buffer_async buffers
423  * left.
424  *
425  * PageLocked prevents anyone starting new async I/O reads any of
426  * the buffers.
427  *
428  * PageWriteback is used to prevent simultaneous writeout of the same
429  * page.
430  *
431  * PageLocked prevents anyone from starting writeback of a page which is
432  * under read I/O (PageWriteback is only ever set against a locked page).
433  */
434 static void mark_buffer_async_read(struct buffer_head *bh)
435 {
436         bh->b_end_io = end_buffer_async_read;
437         set_buffer_async_read(bh);
438 }
439
440 static void mark_buffer_async_write_endio(struct buffer_head *bh,
441                                           bh_end_io_t *handler)
442 {
443         bh->b_end_io = handler;
444         set_buffer_async_write(bh);
445 }
446
447 void mark_buffer_async_write(struct buffer_head *bh)
448 {
449         mark_buffer_async_write_endio(bh, end_buffer_async_write);
450 }
451 EXPORT_SYMBOL(mark_buffer_async_write);
452
453
454 /*
455  * fs/buffer.c contains helper functions for buffer-backed address space's
456  * fsync functions.  A common requirement for buffer-based filesystems is
457  * that certain data from the backing blockdev needs to be written out for
458  * a successful fsync().  For example, ext2 indirect blocks need to be
459  * written back and waited upon before fsync() returns.
460  *
461  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
462  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
463  * management of a list of dependent buffers at ->i_mapping->private_list.
464  *
465  * Locking is a little subtle: try_to_free_buffers() will remove buffers
466  * from their controlling inode's queue when they are being freed.  But
467  * try_to_free_buffers() will be operating against the *blockdev* mapping
468  * at the time, not against the S_ISREG file which depends on those buffers.
469  * So the locking for private_list is via the private_lock in the address_space
470  * which backs the buffers.  Which is different from the address_space 
471  * against which the buffers are listed.  So for a particular address_space,
472  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
473  * mapping->private_list will always be protected by the backing blockdev's
474  * ->private_lock.
475  *
476  * Which introduces a requirement: all buffers on an address_space's
477  * ->private_list must be from the same address_space: the blockdev's.
478  *
479  * address_spaces which do not place buffers at ->private_list via these
480  * utility functions are free to use private_lock and private_list for
481  * whatever they want.  The only requirement is that list_empty(private_list)
482  * be true at clear_inode() time.
483  *
484  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
485  * filesystems should do that.  invalidate_inode_buffers() should just go
486  * BUG_ON(!list_empty).
487  *
488  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
489  * take an address_space, not an inode.  And it should be called
490  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
491  * queued up.
492  *
493  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
494  * list if it is already on a list.  Because if the buffer is on a list,
495  * it *must* already be on the right one.  If not, the filesystem is being
496  * silly.  This will save a ton of locking.  But first we have to ensure
497  * that buffers are taken *off* the old inode's list when they are freed
498  * (presumably in truncate).  That requires careful auditing of all
499  * filesystems (do it inside bforget()).  It could also be done by bringing
500  * b_inode back.
501  */
502
503 /*
504  * The buffer's backing address_space's private_lock must be held
505  */
506 static void __remove_assoc_queue(struct buffer_head *bh)
507 {
508         list_del_init(&bh->b_assoc_buffers);
509         WARN_ON(!bh->b_assoc_map);
510         if (buffer_write_io_error(bh))
511                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
512         bh->b_assoc_map = NULL;
513 }
514
515 int inode_has_buffers(struct inode *inode)
516 {
517         return !list_empty(&inode->i_data.private_list);
518 }
519
520 /*
521  * osync is designed to support O_SYNC io.  It waits synchronously for
522  * all already-submitted IO to complete, but does not queue any new
523  * writes to the disk.
524  *
525  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
526  * you dirty the buffers, and then use osync_inode_buffers to wait for
527  * completion.  Any other dirty buffers which are not yet queued for
528  * write will not be flushed to disk by the osync.
529  */
530 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
531 {
532         struct buffer_head *bh;
533         struct list_head *p;
534         int err = 0;
535
536         spin_lock(lock);
537 repeat:
538         list_for_each_prev(p, list) {
539                 bh = BH_ENTRY(p);
540                 if (buffer_locked(bh)) {
541                         get_bh(bh);
542                         spin_unlock(lock);
543                         wait_on_buffer(bh);
544                         if (!buffer_uptodate(bh))
545                                 err = -EIO;
546                         brelse(bh);
547                         spin_lock(lock);
548                         goto repeat;
549                 }
550         }
551         spin_unlock(lock);
552         return err;
553 }
554
555 static void do_thaw_one(struct super_block *sb, void *unused)
556 {
557         char b[BDEVNAME_SIZE];
558         while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
559                 printk(KERN_WARNING "Emergency Thaw on %s\n",
560                        bdevname(sb->s_bdev, b));
561 }
562
563 static void do_thaw_all(struct work_struct *work)
564 {
565         iterate_supers(do_thaw_one, NULL);
566         kfree(work);
567         printk(KERN_WARNING "Emergency Thaw complete\n");
568 }
569
570 /**
571  * emergency_thaw_all -- forcibly thaw every frozen filesystem
572  *
573  * Used for emergency unfreeze of all filesystems via SysRq
574  */
575 void emergency_thaw_all(void)
576 {
577         struct work_struct *work;
578
579         work = kmalloc(sizeof(*work), GFP_ATOMIC);
580         if (work) {
581                 INIT_WORK(work, do_thaw_all);
582                 schedule_work(work);
583         }
584 }
585
586 /**
587  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
588  * @mapping: the mapping which wants those buffers written
589  *
590  * Starts I/O against the buffers at mapping->private_list, and waits upon
591  * that I/O.
592  *
593  * Basically, this is a convenience function for fsync().
594  * @mapping is a file or directory which needs those buffers to be written for
595  * a successful fsync().
596  */
597 int sync_mapping_buffers(struct address_space *mapping)
598 {
599         struct address_space *buffer_mapping = mapping->private_data;
600
601         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
602                 return 0;
603
604         return fsync_buffers_list(&buffer_mapping->private_lock,
605                                         &mapping->private_list);
606 }
607 EXPORT_SYMBOL(sync_mapping_buffers);
608
609 /*
610  * Called when we've recently written block `bblock', and it is known that
611  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
612  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
613  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
614  */
615 void write_boundary_block(struct block_device *bdev,
616                         sector_t bblock, unsigned blocksize)
617 {
618         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
619         if (bh) {
620                 if (buffer_dirty(bh))
621                         ll_rw_block(WRITE, 1, &bh);
622                 put_bh(bh);
623         }
624 }
625
626 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
627 {
628         struct address_space *mapping = inode->i_mapping;
629         struct address_space *buffer_mapping = bh->b_page->mapping;
630
631         mark_buffer_dirty(bh);
632         if (!mapping->private_data) {
633                 mapping->private_data = buffer_mapping;
634         } else {
635                 BUG_ON(mapping->private_data != buffer_mapping);
636         }
637         if (!bh->b_assoc_map) {
638                 spin_lock(&buffer_mapping->private_lock);
639                 list_move_tail(&bh->b_assoc_buffers,
640                                 &mapping->private_list);
641                 bh->b_assoc_map = mapping;
642                 spin_unlock(&buffer_mapping->private_lock);
643         }
644 }
645 EXPORT_SYMBOL(mark_buffer_dirty_inode);
646
647 /*
648  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
649  * dirty.
650  *
651  * If warn is true, then emit a warning if the page is not uptodate and has
652  * not been truncated.
653  */
654 static void __set_page_dirty(struct page *page,
655                 struct address_space *mapping, int warn)
656 {
657         unsigned long flags;
658
659         spin_lock_irqsave(&mapping->tree_lock, flags);
660         if (page->mapping) {    /* Race with truncate? */
661                 WARN_ON_ONCE(warn && !PageUptodate(page));
662                 account_page_dirtied(page, mapping);
663                 radix_tree_tag_set(&mapping->page_tree,
664                                 page_index(page), PAGECACHE_TAG_DIRTY);
665         }
666         spin_unlock_irqrestore(&mapping->tree_lock, flags);
667         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
668 }
669
670 /*
671  * Add a page to the dirty page list.
672  *
673  * It is a sad fact of life that this function is called from several places
674  * deeply under spinlocking.  It may not sleep.
675  *
676  * If the page has buffers, the uptodate buffers are set dirty, to preserve
677  * dirty-state coherency between the page and the buffers.  It the page does
678  * not have buffers then when they are later attached they will all be set
679  * dirty.
680  *
681  * The buffers are dirtied before the page is dirtied.  There's a small race
682  * window in which a writepage caller may see the page cleanness but not the
683  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
684  * before the buffers, a concurrent writepage caller could clear the page dirty
685  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
686  * page on the dirty page list.
687  *
688  * We use private_lock to lock against try_to_free_buffers while using the
689  * page's buffer list.  Also use this to protect against clean buffers being
690  * added to the page after it was set dirty.
691  *
692  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
693  * address_space though.
694  */
695 int __set_page_dirty_buffers(struct page *page)
696 {
697         int newly_dirty;
698         struct address_space *mapping = page_mapping(page);
699
700         if (unlikely(!mapping))
701                 return !TestSetPageDirty(page);
702
703         spin_lock(&mapping->private_lock);
704         if (page_has_buffers(page)) {
705                 struct buffer_head *head = page_buffers(page);
706                 struct buffer_head *bh = head;
707
708                 do {
709                         set_buffer_dirty(bh);
710                         bh = bh->b_this_page;
711                 } while (bh != head);
712         }
713         newly_dirty = !TestSetPageDirty(page);
714         spin_unlock(&mapping->private_lock);
715
716         if (newly_dirty)
717                 __set_page_dirty(page, mapping, 1);
718         return newly_dirty;
719 }
720 EXPORT_SYMBOL(__set_page_dirty_buffers);
721
722 /*
723  * Write out and wait upon a list of buffers.
724  *
725  * We have conflicting pressures: we want to make sure that all
726  * initially dirty buffers get waited on, but that any subsequently
727  * dirtied buffers don't.  After all, we don't want fsync to last
728  * forever if somebody is actively writing to the file.
729  *
730  * Do this in two main stages: first we copy dirty buffers to a
731  * temporary inode list, queueing the writes as we go.  Then we clean
732  * up, waiting for those writes to complete.
733  * 
734  * During this second stage, any subsequent updates to the file may end
735  * up refiling the buffer on the original inode's dirty list again, so
736  * there is a chance we will end up with a buffer queued for write but
737  * not yet completed on that list.  So, as a final cleanup we go through
738  * the osync code to catch these locked, dirty buffers without requeuing
739  * any newly dirty buffers for write.
740  */
741 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
742 {
743         struct buffer_head *bh;
744         struct list_head tmp;
745         struct address_space *mapping;
746         int err = 0, err2;
747         struct blk_plug plug;
748
749         INIT_LIST_HEAD(&tmp);
750         blk_start_plug(&plug);
751
752         spin_lock(lock);
753         while (!list_empty(list)) {
754                 bh = BH_ENTRY(list->next);
755                 mapping = bh->b_assoc_map;
756                 __remove_assoc_queue(bh);
757                 /* Avoid race with mark_buffer_dirty_inode() which does
758                  * a lockless check and we rely on seeing the dirty bit */
759                 smp_mb();
760                 if (buffer_dirty(bh) || buffer_locked(bh)) {
761                         list_add(&bh->b_assoc_buffers, &tmp);
762                         bh->b_assoc_map = mapping;
763                         if (buffer_dirty(bh)) {
764                                 get_bh(bh);
765                                 spin_unlock(lock);
766                                 /*
767                                  * Ensure any pending I/O completes so that
768                                  * write_dirty_buffer() actually writes the
769                                  * current contents - it is a noop if I/O is
770                                  * still in flight on potentially older
771                                  * contents.
772                                  */
773                                 write_dirty_buffer(bh, WRITE_SYNC);
774
775                                 /*
776                                  * Kick off IO for the previous mapping. Note
777                                  * that we will not run the very last mapping,
778                                  * wait_on_buffer() will do that for us
779                                  * through sync_buffer().
780                                  */
781                                 brelse(bh);
782                                 spin_lock(lock);
783                         }
784                 }
785         }
786
787         spin_unlock(lock);
788         blk_finish_plug(&plug);
789         spin_lock(lock);
790
791         while (!list_empty(&tmp)) {
792                 bh = BH_ENTRY(tmp.prev);
793                 get_bh(bh);
794                 mapping = bh->b_assoc_map;
795                 __remove_assoc_queue(bh);
796                 /* Avoid race with mark_buffer_dirty_inode() which does
797                  * a lockless check and we rely on seeing the dirty bit */
798                 smp_mb();
799                 if (buffer_dirty(bh)) {
800                         list_add(&bh->b_assoc_buffers,
801                                  &mapping->private_list);
802                         bh->b_assoc_map = mapping;
803                 }
804                 spin_unlock(lock);
805                 wait_on_buffer(bh);
806                 if (!buffer_uptodate(bh))
807                         err = -EIO;
808                 brelse(bh);
809                 spin_lock(lock);
810         }
811         
812         spin_unlock(lock);
813         err2 = osync_buffers_list(lock, list);
814         if (err)
815                 return err;
816         else
817                 return err2;
818 }
819
820 /*
821  * Invalidate any and all dirty buffers on a given inode.  We are
822  * probably unmounting the fs, but that doesn't mean we have already
823  * done a sync().  Just drop the buffers from the inode list.
824  *
825  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
826  * assumes that all the buffers are against the blockdev.  Not true
827  * for reiserfs.
828  */
829 void invalidate_inode_buffers(struct inode *inode)
830 {
831         if (inode_has_buffers(inode)) {
832                 struct address_space *mapping = &inode->i_data;
833                 struct list_head *list = &mapping->private_list;
834                 struct address_space *buffer_mapping = mapping->private_data;
835
836                 spin_lock(&buffer_mapping->private_lock);
837                 while (!list_empty(list))
838                         __remove_assoc_queue(BH_ENTRY(list->next));
839                 spin_unlock(&buffer_mapping->private_lock);
840         }
841 }
842 EXPORT_SYMBOL(invalidate_inode_buffers);
843
844 /*
845  * Remove any clean buffers from the inode's buffer list.  This is called
846  * when we're trying to free the inode itself.  Those buffers can pin it.
847  *
848  * Returns true if all buffers were removed.
849  */
850 int remove_inode_buffers(struct inode *inode)
851 {
852         int ret = 1;
853
854         if (inode_has_buffers(inode)) {
855                 struct address_space *mapping = &inode->i_data;
856                 struct list_head *list = &mapping->private_list;
857                 struct address_space *buffer_mapping = mapping->private_data;
858
859                 spin_lock(&buffer_mapping->private_lock);
860                 while (!list_empty(list)) {
861                         struct buffer_head *bh = BH_ENTRY(list->next);
862                         if (buffer_dirty(bh)) {
863                                 ret = 0;
864                                 break;
865                         }
866                         __remove_assoc_queue(bh);
867                 }
868                 spin_unlock(&buffer_mapping->private_lock);
869         }
870         return ret;
871 }
872
873 /*
874  * Create the appropriate buffers when given a page for data area and
875  * the size of each buffer.. Use the bh->b_this_page linked list to
876  * follow the buffers created.  Return NULL if unable to create more
877  * buffers.
878  *
879  * The retry flag is used to differentiate async IO (paging, swapping)
880  * which may not fail from ordinary buffer allocations.
881  */
882 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
883                 int retry)
884 {
885         struct buffer_head *bh, *head;
886         long offset;
887
888 try_again:
889         head = NULL;
890         offset = PAGE_SIZE;
891         while ((offset -= size) >= 0) {
892                 bh = alloc_buffer_head(GFP_NOFS);
893                 if (!bh)
894                         goto no_grow;
895
896                 bh->b_this_page = head;
897                 bh->b_blocknr = -1;
898                 head = bh;
899
900                 bh->b_size = size;
901
902                 /* Link the buffer to its page */
903                 set_bh_page(bh, page, offset);
904         }
905         return head;
906 /*
907  * In case anything failed, we just free everything we got.
908  */
909 no_grow:
910         if (head) {
911                 do {
912                         bh = head;
913                         head = head->b_this_page;
914                         free_buffer_head(bh);
915                 } while (head);
916         }
917
918         /*
919          * Return failure for non-async IO requests.  Async IO requests
920          * are not allowed to fail, so we have to wait until buffer heads
921          * become available.  But we don't want tasks sleeping with 
922          * partially complete buffers, so all were released above.
923          */
924         if (!retry)
925                 return NULL;
926
927         /* We're _really_ low on memory. Now we just
928          * wait for old buffer heads to become free due to
929          * finishing IO.  Since this is an async request and
930          * the reserve list is empty, we're sure there are 
931          * async buffer heads in use.
932          */
933         free_more_memory();
934         goto try_again;
935 }
936 EXPORT_SYMBOL_GPL(alloc_page_buffers);
937
938 static inline void
939 link_dev_buffers(struct page *page, struct buffer_head *head)
940 {
941         struct buffer_head *bh, *tail;
942
943         bh = head;
944         do {
945                 tail = bh;
946                 bh = bh->b_this_page;
947         } while (bh);
948         tail->b_this_page = head;
949         attach_page_buffers(page, head);
950 }
951
952 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
953 {
954         sector_t retval = ~((sector_t)0);
955         loff_t sz = i_size_read(bdev->bd_inode);
956
957         if (sz) {
958                 unsigned int sizebits = blksize_bits(size);
959                 retval = (sz >> sizebits);
960         }
961         return retval;
962 }
963
964 /*
965  * Initialise the state of a blockdev page's buffers.
966  */ 
967 static sector_t
968 init_page_buffers(struct page *page, struct block_device *bdev,
969                         sector_t block, int size)
970 {
971         struct buffer_head *head = page_buffers(page);
972         struct buffer_head *bh = head;
973         int uptodate = PageUptodate(page);
974         sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
975
976         do {
977                 if (!buffer_mapped(bh)) {
978                         init_buffer(bh, NULL, NULL);
979                         bh->b_bdev = bdev;
980                         bh->b_blocknr = block;
981                         if (uptodate)
982                                 set_buffer_uptodate(bh);
983                         if (block < end_block)
984                                 set_buffer_mapped(bh);
985                 }
986                 block++;
987                 bh = bh->b_this_page;
988         } while (bh != head);
989
990         /*
991          * Caller needs to validate requested block against end of device.
992          */
993         return end_block;
994 }
995
996 /*
997  * Create the page-cache page that contains the requested block.
998  *
999  * This is used purely for blockdev mappings.
1000  */
1001 static int
1002 grow_dev_page(struct block_device *bdev, sector_t block,
1003                 pgoff_t index, int size, int sizebits)
1004 {
1005         struct inode *inode = bdev->bd_inode;
1006         struct page *page;
1007         struct buffer_head *bh;
1008         sector_t end_block;
1009         int ret = 0;            /* Will call free_more_memory() */
1010         gfp_t gfp_mask;
1011
1012         gfp_mask = mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS;
1013         gfp_mask |= __GFP_MOVABLE;
1014         /*
1015          * XXX: __getblk_slow() can not really deal with failure and
1016          * will endlessly loop on improvised global reclaim.  Prefer
1017          * looping in the allocator rather than here, at least that
1018          * code knows what it's doing.
1019          */
1020         gfp_mask |= __GFP_NOFAIL;
1021
1022         page = find_or_create_page(inode->i_mapping, index, gfp_mask);
1023         if (!page)
1024                 return ret;
1025
1026         BUG_ON(!PageLocked(page));
1027
1028         if (page_has_buffers(page)) {
1029                 bh = page_buffers(page);
1030                 if (bh->b_size == size) {
1031                         end_block = init_page_buffers(page, bdev,
1032                                                 index << sizebits, size);
1033                         goto done;
1034                 }
1035                 if (!try_to_free_buffers(page))
1036                         goto failed;
1037         }
1038
1039         /*
1040          * Allocate some buffers for this page
1041          */
1042         bh = alloc_page_buffers(page, size, 0);
1043         if (!bh)
1044                 goto failed;
1045
1046         /*
1047          * Link the page to the buffers and initialise them.  Take the
1048          * lock to be atomic wrt __find_get_block(), which does not
1049          * run under the page lock.
1050          */
1051         spin_lock(&inode->i_mapping->private_lock);
1052         link_dev_buffers(page, bh);
1053         end_block = init_page_buffers(page, bdev, index << sizebits, size);
1054         spin_unlock(&inode->i_mapping->private_lock);
1055 done:
1056         ret = (block < end_block) ? 1 : -ENXIO;
1057 failed:
1058         unlock_page(page);
1059         page_cache_release(page);
1060         return ret;
1061 }
1062
1063 /*
1064  * Create buffers for the specified block device block's page.  If
1065  * that page was dirty, the buffers are set dirty also.
1066  */
1067 static int
1068 grow_buffers(struct block_device *bdev, sector_t block, int size)
1069 {
1070         pgoff_t index;
1071         int sizebits;
1072
1073         sizebits = -1;
1074         do {
1075                 sizebits++;
1076         } while ((size << sizebits) < PAGE_SIZE);
1077
1078         index = block >> sizebits;
1079
1080         /*
1081          * Check for a block which wants to lie outside our maximum possible
1082          * pagecache index.  (this comparison is done using sector_t types).
1083          */
1084         if (unlikely(index != block >> sizebits)) {
1085                 char b[BDEVNAME_SIZE];
1086
1087                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1088                         "device %s\n",
1089                         __func__, (unsigned long long)block,
1090                         bdevname(bdev, b));
1091                 return -EIO;
1092         }
1093
1094         /* Create a page with the proper size buffers.. */
1095         return grow_dev_page(bdev, block, index, size, sizebits);
1096 }
1097
1098 static struct buffer_head *
1099 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1100 {
1101         /* Size must be multiple of hard sectorsize */
1102         if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1103                         (size < 512 || size > PAGE_SIZE))) {
1104                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1105                                         size);
1106                 printk(KERN_ERR "logical block size: %d\n",
1107                                         bdev_logical_block_size(bdev));
1108
1109                 dump_stack();
1110                 return NULL;
1111         }
1112
1113         for (;;) {
1114                 struct buffer_head *bh;
1115                 int ret;
1116
1117                 bh = __find_get_block(bdev, block, size);
1118                 if (bh)
1119                         return bh;
1120
1121                 ret = grow_buffers(bdev, block, size);
1122                 if (ret < 0)
1123                         return NULL;
1124                 if (ret == 0)
1125                         free_more_memory();
1126         }
1127 }
1128
1129 /*
1130  * The relationship between dirty buffers and dirty pages:
1131  *
1132  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1133  * the page is tagged dirty in its radix tree.
1134  *
1135  * At all times, the dirtiness of the buffers represents the dirtiness of
1136  * subsections of the page.  If the page has buffers, the page dirty bit is
1137  * merely a hint about the true dirty state.
1138  *
1139  * When a page is set dirty in its entirety, all its buffers are marked dirty
1140  * (if the page has buffers).
1141  *
1142  * When a buffer is marked dirty, its page is dirtied, but the page's other
1143  * buffers are not.
1144  *
1145  * Also.  When blockdev buffers are explicitly read with bread(), they
1146  * individually become uptodate.  But their backing page remains not
1147  * uptodate - even if all of its buffers are uptodate.  A subsequent
1148  * block_read_full_page() against that page will discover all the uptodate
1149  * buffers, will set the page uptodate and will perform no I/O.
1150  */
1151
1152 /**
1153  * mark_buffer_dirty - mark a buffer_head as needing writeout
1154  * @bh: the buffer_head to mark dirty
1155  *
1156  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1157  * backing page dirty, then tag the page as dirty in its address_space's radix
1158  * tree and then attach the address_space's inode to its superblock's dirty
1159  * inode list.
1160  *
1161  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1162  * mapping->tree_lock and mapping->host->i_lock.
1163  */
1164 void mark_buffer_dirty(struct buffer_head *bh)
1165 {
1166         WARN_ON_ONCE(!buffer_uptodate(bh));
1167
1168         trace_block_dirty_buffer(bh);
1169
1170         /*
1171          * Very *carefully* optimize the it-is-already-dirty case.
1172          *
1173          * Don't let the final "is it dirty" escape to before we
1174          * perhaps modified the buffer.
1175          */
1176         if (buffer_dirty(bh)) {
1177                 smp_mb();
1178                 if (buffer_dirty(bh))
1179                         return;
1180         }
1181
1182         if (!test_set_buffer_dirty(bh)) {
1183                 struct page *page = bh->b_page;
1184                 if (!TestSetPageDirty(page)) {
1185                         struct address_space *mapping = page_mapping(page);
1186                         if (mapping)
1187                                 __set_page_dirty(page, mapping, 0);
1188                 }
1189         }
1190 }
1191 EXPORT_SYMBOL(mark_buffer_dirty);
1192
1193 /*
1194  * Decrement a buffer_head's reference count.  If all buffers against a page
1195  * have zero reference count, are clean and unlocked, and if the page is clean
1196  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1197  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1198  * a page but it ends up not being freed, and buffers may later be reattached).
1199  */
1200 void __brelse(struct buffer_head * buf)
1201 {
1202         if (atomic_read(&buf->b_count)) {
1203                 put_bh(buf);
1204                 return;
1205         }
1206         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1207 }
1208 EXPORT_SYMBOL(__brelse);
1209
1210 /*
1211  * bforget() is like brelse(), except it discards any
1212  * potentially dirty data.
1213  */
1214 void __bforget(struct buffer_head *bh)
1215 {
1216         clear_buffer_dirty(bh);
1217         if (bh->b_assoc_map) {
1218                 struct address_space *buffer_mapping = bh->b_page->mapping;
1219
1220                 spin_lock(&buffer_mapping->private_lock);
1221                 list_del_init(&bh->b_assoc_buffers);
1222                 bh->b_assoc_map = NULL;
1223                 spin_unlock(&buffer_mapping->private_lock);
1224         }
1225         __brelse(bh);
1226 }
1227 EXPORT_SYMBOL(__bforget);
1228
1229 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1230 {
1231         lock_buffer(bh);
1232         if (buffer_uptodate(bh)) {
1233                 unlock_buffer(bh);
1234                 return bh;
1235         } else {
1236                 get_bh(bh);
1237                 bh->b_end_io = end_buffer_read_sync;
1238                 submit_bh(READ, bh);
1239                 wait_on_buffer(bh);
1240                 if (buffer_uptodate(bh))
1241                         return bh;
1242         }
1243         brelse(bh);
1244         return NULL;
1245 }
1246
1247 /*
1248  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1249  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1250  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1251  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1252  * CPU's LRUs at the same time.
1253  *
1254  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1255  * sb_find_get_block().
1256  *
1257  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1258  * a local interrupt disable for that.
1259  */
1260
1261 #define BH_LRU_SIZE     8
1262
1263 struct bh_lru {
1264         struct buffer_head *bhs[BH_LRU_SIZE];
1265 };
1266
1267 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1268
1269 #ifdef CONFIG_SMP
1270 #define bh_lru_lock()   local_irq_disable()
1271 #define bh_lru_unlock() local_irq_enable()
1272 #else
1273 #define bh_lru_lock()   preempt_disable()
1274 #define bh_lru_unlock() preempt_enable()
1275 #endif
1276
1277 static inline void check_irqs_on(void)
1278 {
1279 #ifdef irqs_disabled
1280         BUG_ON(irqs_disabled());
1281 #endif
1282 }
1283
1284 /*
1285  * The LRU management algorithm is dopey-but-simple.  Sorry.
1286  */
1287 static void bh_lru_install(struct buffer_head *bh)
1288 {
1289         struct buffer_head *evictee = NULL;
1290
1291         check_irqs_on();
1292         bh_lru_lock();
1293         if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1294                 struct buffer_head *bhs[BH_LRU_SIZE];
1295                 int in;
1296                 int out = 0;
1297
1298                 get_bh(bh);
1299                 bhs[out++] = bh;
1300                 for (in = 0; in < BH_LRU_SIZE; in++) {
1301                         struct buffer_head *bh2 =
1302                                 __this_cpu_read(bh_lrus.bhs[in]);
1303
1304                         if (bh2 == bh) {
1305                                 __brelse(bh2);
1306                         } else {
1307                                 if (out >= BH_LRU_SIZE) {
1308                                         BUG_ON(evictee != NULL);
1309                                         evictee = bh2;
1310                                 } else {
1311                                         bhs[out++] = bh2;
1312                                 }
1313                         }
1314                 }
1315                 while (out < BH_LRU_SIZE)
1316                         bhs[out++] = NULL;
1317                 memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1318         }
1319         bh_lru_unlock();
1320
1321         if (evictee)
1322                 __brelse(evictee);
1323 }
1324
1325 /*
1326  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1327  */
1328 static struct buffer_head *
1329 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1330 {
1331         struct buffer_head *ret = NULL;
1332         unsigned int i;
1333
1334         check_irqs_on();
1335         bh_lru_lock();
1336         for (i = 0; i < BH_LRU_SIZE; i++) {
1337                 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1338
1339                 if (bh && bh->b_bdev == bdev &&
1340                                 bh->b_blocknr == block && bh->b_size == size) {
1341                         if (i) {
1342                                 while (i) {
1343                                         __this_cpu_write(bh_lrus.bhs[i],
1344                                                 __this_cpu_read(bh_lrus.bhs[i - 1]));
1345                                         i--;
1346                                 }
1347                                 __this_cpu_write(bh_lrus.bhs[0], bh);
1348                         }
1349                         get_bh(bh);
1350                         ret = bh;
1351                         break;
1352                 }
1353         }
1354         bh_lru_unlock();
1355         return ret;
1356 }
1357
1358 /*
1359  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1360  * it in the LRU and mark it as accessed.  If it is not present then return
1361  * NULL
1362  */
1363 struct buffer_head *
1364 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1365 {
1366         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1367
1368         if (bh == NULL) {
1369                 bh = __find_get_block_slow(bdev, block);
1370                 if (bh)
1371                         bh_lru_install(bh);
1372         }
1373         if (bh)
1374                 touch_buffer(bh);
1375         return bh;
1376 }
1377 EXPORT_SYMBOL(__find_get_block);
1378
1379 /*
1380  * __getblk will locate (and, if necessary, create) the buffer_head
1381  * which corresponds to the passed block_device, block and size. The
1382  * returned buffer has its reference count incremented.
1383  *
1384  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1385  * attempt is failing.  FIXME, perhaps?
1386  */
1387 struct buffer_head *
1388 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1389 {
1390         struct buffer_head *bh = __find_get_block(bdev, block, size);
1391
1392         might_sleep();
1393         if (bh == NULL)
1394                 bh = __getblk_slow(bdev, block, size);
1395         return bh;
1396 }
1397 EXPORT_SYMBOL(__getblk);
1398
1399 /*
1400  * Do async read-ahead on a buffer..
1401  */
1402 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1403 {
1404         struct buffer_head *bh = __getblk(bdev, block, size);
1405         if (likely(bh)) {
1406                 ll_rw_block(READA, 1, &bh);
1407                 brelse(bh);
1408         }
1409 }
1410 EXPORT_SYMBOL(__breadahead);
1411
1412 /**
1413  *  __bread() - reads a specified block and returns the bh
1414  *  @bdev: the block_device to read from
1415  *  @block: number of block
1416  *  @size: size (in bytes) to read
1417  * 
1418  *  Reads a specified block, and returns buffer head that contains it.
1419  *  It returns NULL if the block was unreadable.
1420  */
1421 struct buffer_head *
1422 __bread(struct block_device *bdev, sector_t block, unsigned size)
1423 {
1424         struct buffer_head *bh = __getblk(bdev, block, size);
1425
1426         if (likely(bh) && !buffer_uptodate(bh))
1427                 bh = __bread_slow(bh);
1428         return bh;
1429 }
1430 EXPORT_SYMBOL(__bread);
1431
1432 /*
1433  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1434  * This doesn't race because it runs in each cpu either in irq
1435  * or with preempt disabled.
1436  */
1437 static void invalidate_bh_lru(void *arg)
1438 {
1439         struct bh_lru *b = &get_cpu_var(bh_lrus);
1440         int i;
1441
1442         for (i = 0; i < BH_LRU_SIZE; i++) {
1443                 brelse(b->bhs[i]);
1444                 b->bhs[i] = NULL;
1445         }
1446         put_cpu_var(bh_lrus);
1447 }
1448
1449 static bool has_bh_in_lru(int cpu, void *dummy)
1450 {
1451         struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1452         int i;
1453         
1454         for (i = 0; i < BH_LRU_SIZE; i++) {
1455                 if (b->bhs[i])
1456                         return 1;
1457         }
1458
1459         return 0;
1460 }
1461
1462 void invalidate_bh_lrus(void)
1463 {
1464         on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1465 }
1466 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1467
1468 void set_bh_page(struct buffer_head *bh,
1469                 struct page *page, unsigned long offset)
1470 {
1471         bh->b_page = page;
1472         BUG_ON(offset >= PAGE_SIZE);
1473         if (PageHighMem(page))
1474                 /*
1475                  * This catches illegal uses and preserves the offset:
1476                  */
1477                 bh->b_data = (char *)(0 + offset);
1478         else
1479                 bh->b_data = page_address(page) + offset;
1480 }
1481 EXPORT_SYMBOL(set_bh_page);
1482
1483 /*
1484  * Called when truncating a buffer on a page completely.
1485  */
1486 static void discard_buffer(struct buffer_head * bh)
1487 {
1488         lock_buffer(bh);
1489         clear_buffer_dirty(bh);
1490         bh->b_bdev = NULL;
1491         clear_buffer_mapped(bh);
1492         clear_buffer_req(bh);
1493         clear_buffer_new(bh);
1494         clear_buffer_delay(bh);
1495         clear_buffer_unwritten(bh);
1496         unlock_buffer(bh);
1497 }
1498
1499 /**
1500  * block_invalidatepage - invalidate part or all of a buffer-backed page
1501  *
1502  * @page: the page which is affected
1503  * @offset: start of the range to invalidate
1504  * @length: length of the range to invalidate
1505  *
1506  * block_invalidatepage() is called when all or part of the page has become
1507  * invalidated by a truncate operation.
1508  *
1509  * block_invalidatepage() does not have to release all buffers, but it must
1510  * ensure that no dirty buffer is left outside @offset and that no I/O
1511  * is underway against any of the blocks which are outside the truncation
1512  * point.  Because the caller is about to free (and possibly reuse) those
1513  * blocks on-disk.
1514  */
1515 void block_invalidatepage(struct page *page, unsigned int offset,
1516                           unsigned int length)
1517 {
1518         struct buffer_head *head, *bh, *next;
1519         unsigned int curr_off = 0;
1520         unsigned int stop = length + offset;
1521
1522         BUG_ON(!PageLocked(page));
1523         if (!page_has_buffers(page))
1524                 goto out;
1525
1526         /*
1527          * Check for overflow
1528          */
1529         BUG_ON(stop > PAGE_CACHE_SIZE || stop < length);
1530
1531         head = page_buffers(page);
1532         bh = head;
1533         do {
1534                 unsigned int next_off = curr_off + bh->b_size;
1535                 next = bh->b_this_page;
1536
1537                 /*
1538                  * Are we still fully in range ?
1539                  */
1540                 if (next_off > stop)
1541                         goto out;
1542
1543                 /*
1544                  * is this block fully invalidated?
1545                  */
1546                 if (offset <= curr_off)
1547                         discard_buffer(bh);
1548                 curr_off = next_off;
1549                 bh = next;
1550         } while (bh != head);
1551
1552         /*
1553          * We release buffers only if the entire page is being invalidated.
1554          * The get_block cached value has been unconditionally invalidated,
1555          * so real IO is not possible anymore.
1556          */
1557         if (offset == 0)
1558                 try_to_release_page(page, 0);
1559 out:
1560         return;
1561 }
1562 EXPORT_SYMBOL(block_invalidatepage);
1563
1564
1565 /*
1566  * We attach and possibly dirty the buffers atomically wrt
1567  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1568  * is already excluded via the page lock.
1569  */
1570 void create_empty_buffers(struct page *page,
1571                         unsigned long blocksize, unsigned long b_state)
1572 {
1573         struct buffer_head *bh, *head, *tail;
1574
1575         head = alloc_page_buffers(page, blocksize, 1);
1576         bh = head;
1577         do {
1578                 bh->b_state |= b_state;
1579                 tail = bh;
1580                 bh = bh->b_this_page;
1581         } while (bh);
1582         tail->b_this_page = head;
1583
1584         spin_lock(&page->mapping->private_lock);
1585         if (PageUptodate(page) || PageDirty(page)) {
1586                 bh = head;
1587                 do {
1588                         if (PageDirty(page))
1589                                 set_buffer_dirty(bh);
1590                         if (PageUptodate(page))
1591                                 set_buffer_uptodate(bh);
1592                         bh = bh->b_this_page;
1593                 } while (bh != head);
1594         }
1595         attach_page_buffers(page, head);
1596         spin_unlock(&page->mapping->private_lock);
1597 }
1598 EXPORT_SYMBOL(create_empty_buffers);
1599
1600 /*
1601  * We are taking a block for data and we don't want any output from any
1602  * buffer-cache aliases starting from return from that function and
1603  * until the moment when something will explicitly mark the buffer
1604  * dirty (hopefully that will not happen until we will free that block ;-)
1605  * We don't even need to mark it not-uptodate - nobody can expect
1606  * anything from a newly allocated buffer anyway. We used to used
1607  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1608  * don't want to mark the alias unmapped, for example - it would confuse
1609  * anyone who might pick it with bread() afterwards...
1610  *
1611  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1612  * be writeout I/O going on against recently-freed buffers.  We don't
1613  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1614  * only if we really need to.  That happens here.
1615  */
1616 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1617 {
1618         struct buffer_head *old_bh;
1619
1620         might_sleep();
1621
1622         old_bh = __find_get_block_slow(bdev, block);
1623         if (old_bh) {
1624                 clear_buffer_dirty(old_bh);
1625                 wait_on_buffer(old_bh);
1626                 clear_buffer_req(old_bh);
1627                 __brelse(old_bh);
1628         }
1629 }
1630 EXPORT_SYMBOL(unmap_underlying_metadata);
1631
1632 /*
1633  * Size is a power-of-two in the range 512..PAGE_SIZE,
1634  * and the case we care about most is PAGE_SIZE.
1635  *
1636  * So this *could* possibly be written with those
1637  * constraints in mind (relevant mostly if some
1638  * architecture has a slow bit-scan instruction)
1639  */
1640 static inline int block_size_bits(unsigned int blocksize)
1641 {
1642         return ilog2(blocksize);
1643 }
1644
1645 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1646 {
1647         BUG_ON(!PageLocked(page));
1648
1649         if (!page_has_buffers(page))
1650                 create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1651         return page_buffers(page);
1652 }
1653
1654 /*
1655  * NOTE! All mapped/uptodate combinations are valid:
1656  *
1657  *      Mapped  Uptodate        Meaning
1658  *
1659  *      No      No              "unknown" - must do get_block()
1660  *      No      Yes             "hole" - zero-filled
1661  *      Yes     No              "allocated" - allocated on disk, not read in
1662  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1663  *
1664  * "Dirty" is valid only with the last case (mapped+uptodate).
1665  */
1666
1667 /*
1668  * While block_write_full_page is writing back the dirty buffers under
1669  * the page lock, whoever dirtied the buffers may decide to clean them
1670  * again at any time.  We handle that by only looking at the buffer
1671  * state inside lock_buffer().
1672  *
1673  * If block_write_full_page() is called for regular writeback
1674  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1675  * locked buffer.   This only can happen if someone has written the buffer
1676  * directly, with submit_bh().  At the address_space level PageWriteback
1677  * prevents this contention from occurring.
1678  *
1679  * If block_write_full_page() is called with wbc->sync_mode ==
1680  * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1681  * causes the writes to be flagged as synchronous writes.
1682  */
1683 static int __block_write_full_page(struct inode *inode, struct page *page,
1684                         get_block_t *get_block, struct writeback_control *wbc,
1685                         bh_end_io_t *handler)
1686 {
1687         int err;
1688         sector_t block;
1689         sector_t last_block;
1690         struct buffer_head *bh, *head;
1691         unsigned int blocksize, bbits;
1692         int nr_underway = 0;
1693         int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1694                         WRITE_SYNC : WRITE);
1695
1696         head = create_page_buffers(page, inode,
1697                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1698
1699         /*
1700          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1701          * here, and the (potentially unmapped) buffers may become dirty at
1702          * any time.  If a buffer becomes dirty here after we've inspected it
1703          * then we just miss that fact, and the page stays dirty.
1704          *
1705          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1706          * handle that here by just cleaning them.
1707          */
1708
1709         bh = head;
1710         blocksize = bh->b_size;
1711         bbits = block_size_bits(blocksize);
1712
1713         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1714         last_block = (i_size_read(inode) - 1) >> bbits;
1715
1716         /*
1717          * Get all the dirty buffers mapped to disk addresses and
1718          * handle any aliases from the underlying blockdev's mapping.
1719          */
1720         do {
1721                 if (block > last_block) {
1722                         /*
1723                          * mapped buffers outside i_size will occur, because
1724                          * this page can be outside i_size when there is a
1725                          * truncate in progress.
1726                          */
1727                         /*
1728                          * The buffer was zeroed by block_write_full_page()
1729                          */
1730                         clear_buffer_dirty(bh);
1731                         set_buffer_uptodate(bh);
1732                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1733                            buffer_dirty(bh)) {
1734                         WARN_ON(bh->b_size != blocksize);
1735                         err = get_block(inode, block, bh, 1);
1736                         if (err)
1737                                 goto recover;
1738                         clear_buffer_delay(bh);
1739                         if (buffer_new(bh)) {
1740                                 /* blockdev mappings never come here */
1741                                 clear_buffer_new(bh);
1742                                 unmap_underlying_metadata(bh->b_bdev,
1743                                                         bh->b_blocknr);
1744                         }
1745                 }
1746                 bh = bh->b_this_page;
1747                 block++;
1748         } while (bh != head);
1749
1750         do {
1751                 if (!buffer_mapped(bh))
1752                         continue;
1753                 /*
1754                  * If it's a fully non-blocking write attempt and we cannot
1755                  * lock the buffer then redirty the page.  Note that this can
1756                  * potentially cause a busy-wait loop from writeback threads
1757                  * and kswapd activity, but those code paths have their own
1758                  * higher-level throttling.
1759                  */
1760                 if (wbc->sync_mode != WB_SYNC_NONE) {
1761                         lock_buffer(bh);
1762                 } else if (!trylock_buffer(bh)) {
1763                         redirty_page_for_writepage(wbc, page);
1764                         continue;
1765                 }
1766                 if (test_clear_buffer_dirty(bh)) {
1767                         mark_buffer_async_write_endio(bh, handler);
1768                 } else {
1769                         unlock_buffer(bh);
1770                 }
1771         } while ((bh = bh->b_this_page) != head);
1772
1773         /*
1774          * The page and its buffers are protected by PageWriteback(), so we can
1775          * drop the bh refcounts early.
1776          */
1777         BUG_ON(PageWriteback(page));
1778         set_page_writeback(page);
1779
1780         do {
1781                 struct buffer_head *next = bh->b_this_page;
1782                 if (buffer_async_write(bh)) {
1783                         submit_bh(write_op, bh);
1784                         nr_underway++;
1785                 }
1786                 bh = next;
1787         } while (bh != head);
1788         unlock_page(page);
1789
1790         err = 0;
1791 done:
1792         if (nr_underway == 0) {
1793                 /*
1794                  * The page was marked dirty, but the buffers were
1795                  * clean.  Someone wrote them back by hand with
1796                  * ll_rw_block/submit_bh.  A rare case.
1797                  */
1798                 end_page_writeback(page);
1799
1800                 /*
1801                  * The page and buffer_heads can be released at any time from
1802                  * here on.
1803                  */
1804         }
1805         return err;
1806
1807 recover:
1808         /*
1809          * ENOSPC, or some other error.  We may already have added some
1810          * blocks to the file, so we need to write these out to avoid
1811          * exposing stale data.
1812          * The page is currently locked and not marked for writeback
1813          */
1814         bh = head;
1815         /* Recovery: lock and submit the mapped buffers */
1816         do {
1817                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1818                     !buffer_delay(bh)) {
1819                         lock_buffer(bh);
1820                         mark_buffer_async_write_endio(bh, handler);
1821                 } else {
1822                         /*
1823                          * The buffer may have been set dirty during
1824                          * attachment to a dirty page.
1825                          */
1826                         clear_buffer_dirty(bh);
1827                 }
1828         } while ((bh = bh->b_this_page) != head);
1829         SetPageError(page);
1830         BUG_ON(PageWriteback(page));
1831         mapping_set_error(page->mapping, err);
1832         set_page_writeback(page);
1833         do {
1834                 struct buffer_head *next = bh->b_this_page;
1835                 if (buffer_async_write(bh)) {
1836                         clear_buffer_dirty(bh);
1837                         submit_bh(write_op, bh);
1838                         nr_underway++;
1839                 }
1840                 bh = next;
1841         } while (bh != head);
1842         unlock_page(page);
1843         goto done;
1844 }
1845
1846 /*
1847  * If a page has any new buffers, zero them out here, and mark them uptodate
1848  * and dirty so they'll be written out (in order to prevent uninitialised
1849  * block data from leaking). And clear the new bit.
1850  */
1851 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1852 {
1853         unsigned int block_start, block_end;
1854         struct buffer_head *head, *bh;
1855
1856         BUG_ON(!PageLocked(page));
1857         if (!page_has_buffers(page))
1858                 return;
1859
1860         bh = head = page_buffers(page);
1861         block_start = 0;
1862         do {
1863                 block_end = block_start + bh->b_size;
1864
1865                 if (buffer_new(bh)) {
1866                         if (block_end > from && block_start < to) {
1867                                 if (!PageUptodate(page)) {
1868                                         unsigned start, size;
1869
1870                                         start = max(from, block_start);
1871                                         size = min(to, block_end) - start;
1872
1873                                         zero_user(page, start, size);
1874                                         set_buffer_uptodate(bh);
1875                                 }
1876
1877                                 clear_buffer_new(bh);
1878                                 mark_buffer_dirty(bh);
1879                         }
1880                 }
1881
1882                 block_start = block_end;
1883                 bh = bh->b_this_page;
1884         } while (bh != head);
1885 }
1886 EXPORT_SYMBOL(page_zero_new_buffers);
1887
1888 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1889                 get_block_t *get_block)
1890 {
1891         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1892         unsigned to = from + len;
1893         struct inode *inode = page->mapping->host;
1894         unsigned block_start, block_end;
1895         sector_t block;
1896         int err = 0;
1897         unsigned blocksize, bbits;
1898         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1899
1900         BUG_ON(!PageLocked(page));
1901         BUG_ON(from > PAGE_CACHE_SIZE);
1902         BUG_ON(to > PAGE_CACHE_SIZE);
1903         BUG_ON(from > to);
1904
1905         head = create_page_buffers(page, inode, 0);
1906         blocksize = head->b_size;
1907         bbits = block_size_bits(blocksize);
1908
1909         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1910
1911         for(bh = head, block_start = 0; bh != head || !block_start;
1912             block++, block_start=block_end, bh = bh->b_this_page) {
1913                 block_end = block_start + blocksize;
1914                 if (block_end <= from || block_start >= to) {
1915                         if (PageUptodate(page)) {
1916                                 if (!buffer_uptodate(bh))
1917                                         set_buffer_uptodate(bh);
1918                         }
1919                         continue;
1920                 }
1921                 if (buffer_new(bh))
1922                         clear_buffer_new(bh);
1923                 if (!buffer_mapped(bh)) {
1924                         WARN_ON(bh->b_size != blocksize);
1925                         err = get_block(inode, block, bh, 1);
1926                         if (err)
1927                                 break;
1928                         if (buffer_new(bh)) {
1929                                 unmap_underlying_metadata(bh->b_bdev,
1930                                                         bh->b_blocknr);
1931                                 if (PageUptodate(page)) {
1932                                         clear_buffer_new(bh);
1933                                         set_buffer_uptodate(bh);
1934                                         mark_buffer_dirty(bh);
1935                                         continue;
1936                                 }
1937                                 if (block_end > to || block_start < from)
1938                                         zero_user_segments(page,
1939                                                 to, block_end,
1940                                                 block_start, from);
1941                                 continue;
1942                         }
1943                 }
1944                 if (PageUptodate(page)) {
1945                         if (!buffer_uptodate(bh))
1946                                 set_buffer_uptodate(bh);
1947                         continue; 
1948                 }
1949                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1950                     !buffer_unwritten(bh) &&
1951                      (block_start < from || block_end > to)) {
1952                         ll_rw_block(READ, 1, &bh);
1953                         *wait_bh++=bh;
1954                 }
1955         }
1956         /*
1957          * If we issued read requests - let them complete.
1958          */
1959         while(wait_bh > wait) {
1960                 wait_on_buffer(*--wait_bh);
1961                 if (!buffer_uptodate(*wait_bh))
1962                         err = -EIO;
1963         }
1964         if (unlikely(err))
1965                 page_zero_new_buffers(page, from, to);
1966         return err;
1967 }
1968 EXPORT_SYMBOL(__block_write_begin);
1969
1970 static int __block_commit_write(struct inode *inode, struct page *page,
1971                 unsigned from, unsigned to)
1972 {
1973         unsigned block_start, block_end;
1974         int partial = 0;
1975         unsigned blocksize;
1976         struct buffer_head *bh, *head;
1977
1978         bh = head = page_buffers(page);
1979         blocksize = bh->b_size;
1980
1981         block_start = 0;
1982         do {
1983                 block_end = block_start + blocksize;
1984                 if (block_end <= from || block_start >= to) {
1985                         if (!buffer_uptodate(bh))
1986                                 partial = 1;
1987                 } else {
1988                         set_buffer_uptodate(bh);
1989                         mark_buffer_dirty(bh);
1990                 }
1991                 clear_buffer_new(bh);
1992
1993                 block_start = block_end;
1994                 bh = bh->b_this_page;
1995         } while (bh != head);
1996
1997         /*
1998          * If this is a partial write which happened to make all buffers
1999          * uptodate then we can optimize away a bogus readpage() for
2000          * the next read(). Here we 'discover' whether the page went
2001          * uptodate as a result of this (potentially partial) write.
2002          */
2003         if (!partial)
2004                 SetPageUptodate(page);
2005         return 0;
2006 }
2007
2008 /*
2009  * block_write_begin takes care of the basic task of block allocation and
2010  * bringing partial write blocks uptodate first.
2011  *
2012  * The filesystem needs to handle block truncation upon failure.
2013  */
2014 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2015                 unsigned flags, struct page **pagep, get_block_t *get_block)
2016 {
2017         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
2018         struct page *page;
2019         int status;
2020
2021         page = grab_cache_page_write_begin(mapping, index, flags);
2022         if (!page)
2023                 return -ENOMEM;
2024
2025         status = __block_write_begin(page, pos, len, get_block);
2026         if (unlikely(status)) {
2027                 unlock_page(page);
2028                 page_cache_release(page);
2029                 page = NULL;
2030         }
2031
2032         *pagep = page;
2033         return status;
2034 }
2035 EXPORT_SYMBOL(block_write_begin);
2036
2037 int block_write_end(struct file *file, struct address_space *mapping,
2038                         loff_t pos, unsigned len, unsigned copied,
2039                         struct page *page, void *fsdata)
2040 {
2041         struct inode *inode = mapping->host;
2042         unsigned start;
2043
2044         start = pos & (PAGE_CACHE_SIZE - 1);
2045
2046         if (unlikely(copied < len)) {
2047                 /*
2048                  * The buffers that were written will now be uptodate, so we
2049                  * don't have to worry about a readpage reading them and
2050                  * overwriting a partial write. However if we have encountered
2051                  * a short write and only partially written into a buffer, it
2052                  * will not be marked uptodate, so a readpage might come in and
2053                  * destroy our partial write.
2054                  *
2055                  * Do the simplest thing, and just treat any short write to a
2056                  * non uptodate page as a zero-length write, and force the
2057                  * caller to redo the whole thing.
2058                  */
2059                 if (!PageUptodate(page))
2060                         copied = 0;
2061
2062                 page_zero_new_buffers(page, start+copied, start+len);
2063         }
2064         flush_dcache_page(page);
2065
2066         /* This could be a short (even 0-length) commit */
2067         __block_commit_write(inode, page, start, start+copied);
2068
2069         return copied;
2070 }
2071 EXPORT_SYMBOL(block_write_end);
2072
2073 int generic_write_end(struct file *file, struct address_space *mapping,
2074                         loff_t pos, unsigned len, unsigned copied,
2075                         struct page *page, void *fsdata)
2076 {
2077         struct inode *inode = mapping->host;
2078         int i_size_changed = 0;
2079
2080         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2081
2082         /*
2083          * No need to use i_size_read() here, the i_size
2084          * cannot change under us because we hold i_mutex.
2085          *
2086          * But it's important to update i_size while still holding page lock:
2087          * page writeout could otherwise come in and zero beyond i_size.
2088          */
2089         if (pos+copied > inode->i_size) {
2090                 i_size_write(inode, pos+copied);
2091                 i_size_changed = 1;
2092         }
2093
2094         unlock_page(page);
2095         page_cache_release(page);
2096
2097         /*
2098          * Don't mark the inode dirty under page lock. First, it unnecessarily
2099          * makes the holding time of page lock longer. Second, it forces lock
2100          * ordering of page lock and transaction start for journaling
2101          * filesystems.
2102          */
2103         if (i_size_changed)
2104                 mark_inode_dirty(inode);
2105
2106         return copied;
2107 }
2108 EXPORT_SYMBOL(generic_write_end);
2109
2110 /*
2111  * block_is_partially_uptodate checks whether buffers within a page are
2112  * uptodate or not.
2113  *
2114  * Returns true if all buffers which correspond to a file portion
2115  * we want to read are uptodate.
2116  */
2117 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2118                                         unsigned long from)
2119 {
2120         unsigned block_start, block_end, blocksize;
2121         unsigned to;
2122         struct buffer_head *bh, *head;
2123         int ret = 1;
2124
2125         if (!page_has_buffers(page))
2126                 return 0;
2127
2128         head = page_buffers(page);
2129         blocksize = head->b_size;
2130         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2131         to = from + to;
2132         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2133                 return 0;
2134
2135         bh = head;
2136         block_start = 0;
2137         do {
2138                 block_end = block_start + blocksize;
2139                 if (block_end > from && block_start < to) {
2140                         if (!buffer_uptodate(bh)) {
2141                                 ret = 0;
2142                                 break;
2143                         }
2144                         if (block_end >= to)
2145                                 break;
2146                 }
2147                 block_start = block_end;
2148                 bh = bh->b_this_page;
2149         } while (bh != head);
2150
2151         return ret;
2152 }
2153 EXPORT_SYMBOL(block_is_partially_uptodate);
2154
2155 /*
2156  * Generic "read page" function for block devices that have the normal
2157  * get_block functionality. This is most of the block device filesystems.
2158  * Reads the page asynchronously --- the unlock_buffer() and
2159  * set/clear_buffer_uptodate() functions propagate buffer state into the
2160  * page struct once IO has completed.
2161  */
2162 int block_read_full_page(struct page *page, get_block_t *get_block)
2163 {
2164         struct inode *inode = page->mapping->host;
2165         sector_t iblock, lblock;
2166         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2167         unsigned int blocksize, bbits;
2168         int nr, i;
2169         int fully_mapped = 1;
2170
2171         head = create_page_buffers(page, inode, 0);
2172         blocksize = head->b_size;
2173         bbits = block_size_bits(blocksize);
2174
2175         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
2176         lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2177         bh = head;
2178         nr = 0;
2179         i = 0;
2180
2181         do {
2182                 if (buffer_uptodate(bh))
2183                         continue;
2184
2185                 if (!buffer_mapped(bh)) {
2186                         int err = 0;
2187
2188                         fully_mapped = 0;
2189                         if (iblock < lblock) {
2190                                 WARN_ON(bh->b_size != blocksize);
2191                                 err = get_block(inode, iblock, bh, 0);
2192                                 if (err)
2193                                         SetPageError(page);
2194                         }
2195                         if (!buffer_mapped(bh)) {
2196                                 zero_user(page, i * blocksize, blocksize);
2197                                 if (!err)
2198                                         set_buffer_uptodate(bh);
2199                                 continue;
2200                         }
2201                         /*
2202                          * get_block() might have updated the buffer
2203                          * synchronously
2204                          */
2205                         if (buffer_uptodate(bh))
2206                                 continue;
2207                 }
2208                 arr[nr++] = bh;
2209         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2210
2211         if (fully_mapped)
2212                 SetPageMappedToDisk(page);
2213
2214         if (!nr) {
2215                 /*
2216                  * All buffers are uptodate - we can set the page uptodate
2217                  * as well. But not if get_block() returned an error.
2218                  */
2219                 if (!PageError(page))
2220                         SetPageUptodate(page);
2221                 unlock_page(page);
2222                 return 0;
2223         }
2224
2225         /* Stage two: lock the buffers */
2226         for (i = 0; i < nr; i++) {
2227                 bh = arr[i];
2228                 lock_buffer(bh);
2229                 mark_buffer_async_read(bh);
2230         }
2231
2232         /*
2233          * Stage 3: start the IO.  Check for uptodateness
2234          * inside the buffer lock in case another process reading
2235          * the underlying blockdev brought it uptodate (the sct fix).
2236          */
2237         for (i = 0; i < nr; i++) {
2238                 bh = arr[i];
2239                 if (buffer_uptodate(bh))
2240                         end_buffer_async_read(bh, 1);
2241                 else
2242                         submit_bh(READ, bh);
2243         }
2244         return 0;
2245 }
2246 EXPORT_SYMBOL(block_read_full_page);
2247
2248 /* utility function for filesystems that need to do work on expanding
2249  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2250  * deal with the hole.  
2251  */
2252 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2253 {
2254         struct address_space *mapping = inode->i_mapping;
2255         struct page *page;
2256         void *fsdata;
2257         int err;
2258
2259         err = inode_newsize_ok(inode, size);
2260         if (err)
2261                 goto out;
2262
2263         err = pagecache_write_begin(NULL, mapping, size, 0,
2264                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2265                                 &page, &fsdata);
2266         if (err)
2267                 goto out;
2268
2269         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2270         BUG_ON(err > 0);
2271
2272 out:
2273         return err;
2274 }
2275 EXPORT_SYMBOL(generic_cont_expand_simple);
2276
2277 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2278                             loff_t pos, loff_t *bytes)
2279 {
2280         struct inode *inode = mapping->host;
2281         unsigned blocksize = 1 << inode->i_blkbits;
2282         struct page *page;
2283         void *fsdata;
2284         pgoff_t index, curidx;
2285         loff_t curpos;
2286         unsigned zerofrom, offset, len;
2287         int err = 0;
2288
2289         index = pos >> PAGE_CACHE_SHIFT;
2290         offset = pos & ~PAGE_CACHE_MASK;
2291
2292         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2293                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2294                 if (zerofrom & (blocksize-1)) {
2295                         *bytes |= (blocksize-1);
2296                         (*bytes)++;
2297                 }
2298                 len = PAGE_CACHE_SIZE - zerofrom;
2299
2300                 err = pagecache_write_begin(file, mapping, curpos, len,
2301                                                 AOP_FLAG_UNINTERRUPTIBLE,
2302                                                 &page, &fsdata);
2303                 if (err)
2304                         goto out;
2305                 zero_user(page, zerofrom, len);
2306                 err = pagecache_write_end(file, mapping, curpos, len, len,
2307                                                 page, fsdata);
2308                 if (err < 0)
2309                         goto out;
2310                 BUG_ON(err != len);
2311                 err = 0;
2312
2313                 balance_dirty_pages_ratelimited(mapping);
2314         }
2315
2316         /* page covers the boundary, find the boundary offset */
2317         if (index == curidx) {
2318                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2319                 /* if we will expand the thing last block will be filled */
2320                 if (offset <= zerofrom) {
2321                         goto out;
2322                 }
2323                 if (zerofrom & (blocksize-1)) {
2324                         *bytes |= (blocksize-1);
2325                         (*bytes)++;
2326                 }
2327                 len = offset - zerofrom;
2328
2329                 err = pagecache_write_begin(file, mapping, curpos, len,
2330                                                 AOP_FLAG_UNINTERRUPTIBLE,
2331                                                 &page, &fsdata);
2332                 if (err)
2333                         goto out;
2334                 zero_user(page, zerofrom, len);
2335                 err = pagecache_write_end(file, mapping, curpos, len, len,
2336                                                 page, fsdata);
2337                 if (err < 0)
2338                         goto out;
2339                 BUG_ON(err != len);
2340                 err = 0;
2341         }
2342 out:
2343         return err;
2344 }
2345
2346 /*
2347  * For moronic filesystems that do not allow holes in file.
2348  * We may have to extend the file.
2349  */
2350 int cont_write_begin(struct file *file, struct address_space *mapping,
2351                         loff_t pos, unsigned len, unsigned flags,
2352                         struct page **pagep, void **fsdata,
2353                         get_block_t *get_block, loff_t *bytes)
2354 {
2355         struct inode *inode = mapping->host;
2356         unsigned blocksize = 1 << inode->i_blkbits;
2357         unsigned zerofrom;
2358         int err;
2359
2360         err = cont_expand_zero(file, mapping, pos, bytes);
2361         if (err)
2362                 return err;
2363
2364         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2365         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2366                 *bytes |= (blocksize-1);
2367                 (*bytes)++;
2368         }
2369
2370         return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2371 }
2372 EXPORT_SYMBOL(cont_write_begin);
2373
2374 int block_commit_write(struct page *page, unsigned from, unsigned to)
2375 {
2376         struct inode *inode = page->mapping->host;
2377         __block_commit_write(inode,page,from,to);
2378         return 0;
2379 }
2380 EXPORT_SYMBOL(block_commit_write);
2381
2382 /*
2383  * block_page_mkwrite() is not allowed to change the file size as it gets
2384  * called from a page fault handler when a page is first dirtied. Hence we must
2385  * be careful to check for EOF conditions here. We set the page up correctly
2386  * for a written page which means we get ENOSPC checking when writing into
2387  * holes and correct delalloc and unwritten extent mapping on filesystems that
2388  * support these features.
2389  *
2390  * We are not allowed to take the i_mutex here so we have to play games to
2391  * protect against truncate races as the page could now be beyond EOF.  Because
2392  * truncate writes the inode size before removing pages, once we have the
2393  * page lock we can determine safely if the page is beyond EOF. If it is not
2394  * beyond EOF, then the page is guaranteed safe against truncation until we
2395  * unlock the page.
2396  *
2397  * Direct callers of this function should protect against filesystem freezing
2398  * using sb_start_write() - sb_end_write() functions.
2399  */
2400 int __block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2401                          get_block_t get_block)
2402 {
2403         struct page *page = vmf->page;
2404         struct inode *inode = file_inode(vma->vm_file);
2405         unsigned long end;
2406         loff_t size;
2407         int ret;
2408
2409         lock_page(page);
2410         size = i_size_read(inode);
2411         if ((page->mapping != inode->i_mapping) ||
2412             (page_offset(page) > size)) {
2413                 /* We overload EFAULT to mean page got truncated */
2414                 ret = -EFAULT;
2415                 goto out_unlock;
2416         }
2417
2418         /* page is wholly or partially inside EOF */
2419         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2420                 end = size & ~PAGE_CACHE_MASK;
2421         else
2422                 end = PAGE_CACHE_SIZE;
2423
2424         ret = __block_write_begin(page, 0, end, get_block);
2425         if (!ret)
2426                 ret = block_commit_write(page, 0, end);
2427
2428         if (unlikely(ret < 0))
2429                 goto out_unlock;
2430         set_page_dirty(page);
2431         wait_for_stable_page(page);
2432         return 0;
2433 out_unlock:
2434         unlock_page(page);
2435         return ret;
2436 }
2437 EXPORT_SYMBOL(__block_page_mkwrite);
2438
2439 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2440                    get_block_t get_block)
2441 {
2442         int ret;
2443         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
2444
2445         sb_start_pagefault(sb);
2446
2447         /*
2448          * Update file times before taking page lock. We may end up failing the
2449          * fault so this update may be superfluous but who really cares...
2450          */
2451         file_update_time(vma->vm_file);
2452
2453         ret = __block_page_mkwrite(vma, vmf, get_block);
2454         sb_end_pagefault(sb);
2455         return block_page_mkwrite_return(ret);
2456 }
2457 EXPORT_SYMBOL(block_page_mkwrite);
2458
2459 /*
2460  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2461  * immediately, while under the page lock.  So it needs a special end_io
2462  * handler which does not touch the bh after unlocking it.
2463  */
2464 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2465 {
2466         __end_buffer_read_notouch(bh, uptodate);
2467 }
2468
2469 /*
2470  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2471  * the page (converting it to circular linked list and taking care of page
2472  * dirty races).
2473  */
2474 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2475 {
2476         struct buffer_head *bh;
2477
2478         BUG_ON(!PageLocked(page));
2479
2480         spin_lock(&page->mapping->private_lock);
2481         bh = head;
2482         do {
2483                 if (PageDirty(page))
2484                         set_buffer_dirty(bh);
2485                 if (!bh->b_this_page)
2486                         bh->b_this_page = head;
2487                 bh = bh->b_this_page;
2488         } while (bh != head);
2489         attach_page_buffers(page, head);
2490         spin_unlock(&page->mapping->private_lock);
2491 }
2492
2493 /*
2494  * On entry, the page is fully not uptodate.
2495  * On exit the page is fully uptodate in the areas outside (from,to)
2496  * The filesystem needs to handle block truncation upon failure.
2497  */
2498 int nobh_write_begin(struct address_space *mapping,
2499                         loff_t pos, unsigned len, unsigned flags,
2500                         struct page **pagep, void **fsdata,
2501                         get_block_t *get_block)
2502 {
2503         struct inode *inode = mapping->host;
2504         const unsigned blkbits = inode->i_blkbits;
2505         const unsigned blocksize = 1 << blkbits;
2506         struct buffer_head *head, *bh;
2507         struct page *page;
2508         pgoff_t index;
2509         unsigned from, to;
2510         unsigned block_in_page;
2511         unsigned block_start, block_end;
2512         sector_t block_in_file;
2513         int nr_reads = 0;
2514         int ret = 0;
2515         int is_mapped_to_disk = 1;
2516
2517         index = pos >> PAGE_CACHE_SHIFT;
2518         from = pos & (PAGE_CACHE_SIZE - 1);
2519         to = from + len;
2520
2521         page = grab_cache_page_write_begin(mapping, index, flags);
2522         if (!page)
2523                 return -ENOMEM;
2524         *pagep = page;
2525         *fsdata = NULL;
2526
2527         if (page_has_buffers(page)) {
2528                 ret = __block_write_begin(page, pos, len, get_block);
2529                 if (unlikely(ret))
2530                         goto out_release;
2531                 return ret;
2532         }
2533
2534         if (PageMappedToDisk(page))
2535                 return 0;
2536
2537         /*
2538          * Allocate buffers so that we can keep track of state, and potentially
2539          * attach them to the page if an error occurs. In the common case of
2540          * no error, they will just be freed again without ever being attached
2541          * to the page (which is all OK, because we're under the page lock).
2542          *
2543          * Be careful: the buffer linked list is a NULL terminated one, rather
2544          * than the circular one we're used to.
2545          */
2546         head = alloc_page_buffers(page, blocksize, 0);
2547         if (!head) {
2548                 ret = -ENOMEM;
2549                 goto out_release;
2550         }
2551
2552         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2553
2554         /*
2555          * We loop across all blocks in the page, whether or not they are
2556          * part of the affected region.  This is so we can discover if the
2557          * page is fully mapped-to-disk.
2558          */
2559         for (block_start = 0, block_in_page = 0, bh = head;
2560                   block_start < PAGE_CACHE_SIZE;
2561                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2562                 int create;
2563
2564                 block_end = block_start + blocksize;
2565                 bh->b_state = 0;
2566                 create = 1;
2567                 if (block_start >= to)
2568                         create = 0;
2569                 ret = get_block(inode, block_in_file + block_in_page,
2570                                         bh, create);
2571                 if (ret)
2572                         goto failed;
2573                 if (!buffer_mapped(bh))
2574                         is_mapped_to_disk = 0;
2575                 if (buffer_new(bh))
2576                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2577                 if (PageUptodate(page)) {
2578                         set_buffer_uptodate(bh);
2579                         continue;
2580                 }
2581                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2582                         zero_user_segments(page, block_start, from,
2583                                                         to, block_end);
2584                         continue;
2585                 }
2586                 if (buffer_uptodate(bh))
2587                         continue;       /* reiserfs does this */
2588                 if (block_start < from || block_end > to) {
2589                         lock_buffer(bh);
2590                         bh->b_end_io = end_buffer_read_nobh;
2591                         submit_bh(READ, bh);
2592                         nr_reads++;
2593                 }
2594         }
2595
2596         if (nr_reads) {
2597                 /*
2598                  * The page is locked, so these buffers are protected from
2599                  * any VM or truncate activity.  Hence we don't need to care
2600                  * for the buffer_head refcounts.
2601                  */
2602                 for (bh = head; bh; bh = bh->b_this_page) {
2603                         wait_on_buffer(bh);
2604                         if (!buffer_uptodate(bh))
2605                                 ret = -EIO;
2606                 }
2607                 if (ret)
2608                         goto failed;
2609         }
2610
2611         if (is_mapped_to_disk)
2612                 SetPageMappedToDisk(page);
2613
2614         *fsdata = head; /* to be released by nobh_write_end */
2615
2616         return 0;
2617
2618 failed:
2619         BUG_ON(!ret);
2620         /*
2621          * Error recovery is a bit difficult. We need to zero out blocks that
2622          * were newly allocated, and dirty them to ensure they get written out.
2623          * Buffers need to be attached to the page at this point, otherwise
2624          * the handling of potential IO errors during writeout would be hard
2625          * (could try doing synchronous writeout, but what if that fails too?)
2626          */
2627         attach_nobh_buffers(page, head);
2628         page_zero_new_buffers(page, from, to);
2629
2630 out_release:
2631         unlock_page(page);
2632         page_cache_release(page);
2633         *pagep = NULL;
2634
2635         return ret;
2636 }
2637 EXPORT_SYMBOL(nobh_write_begin);
2638
2639 int nobh_write_end(struct file *file, struct address_space *mapping,
2640                         loff_t pos, unsigned len, unsigned copied,
2641                         struct page *page, void *fsdata)
2642 {
2643         struct inode *inode = page->mapping->host;
2644         struct buffer_head *head = fsdata;
2645         struct buffer_head *bh;
2646         BUG_ON(fsdata != NULL && page_has_buffers(page));
2647
2648         if (unlikely(copied < len) && head)
2649                 attach_nobh_buffers(page, head);
2650         if (page_has_buffers(page))
2651                 return generic_write_end(file, mapping, pos, len,
2652                                         copied, page, fsdata);
2653
2654         SetPageUptodate(page);
2655         set_page_dirty(page);
2656         if (pos+copied > inode->i_size) {
2657                 i_size_write(inode, pos+copied);
2658                 mark_inode_dirty(inode);
2659         }
2660
2661         unlock_page(page);
2662         page_cache_release(page);
2663
2664         while (head) {
2665                 bh = head;
2666                 head = head->b_this_page;
2667                 free_buffer_head(bh);
2668         }
2669
2670         return copied;
2671 }
2672 EXPORT_SYMBOL(nobh_write_end);
2673
2674 /*
2675  * nobh_writepage() - based on block_full_write_page() except
2676  * that it tries to operate without attaching bufferheads to
2677  * the page.
2678  */
2679 int nobh_writepage(struct page *page, get_block_t *get_block,
2680                         struct writeback_control *wbc)
2681 {
2682         struct inode * const inode = page->mapping->host;
2683         loff_t i_size = i_size_read(inode);
2684         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2685         unsigned offset;
2686         int ret;
2687
2688         /* Is the page fully inside i_size? */
2689         if (page->index < end_index)
2690                 goto out;
2691
2692         /* Is the page fully outside i_size? (truncate in progress) */
2693         offset = i_size & (PAGE_CACHE_SIZE-1);
2694         if (page->index >= end_index+1 || !offset) {
2695                 /*
2696                  * The page may have dirty, unmapped buffers.  For example,
2697                  * they may have been added in ext3_writepage().  Make them
2698                  * freeable here, so the page does not leak.
2699                  */
2700 #if 0
2701                 /* Not really sure about this  - do we need this ? */
2702                 if (page->mapping->a_ops->invalidatepage)
2703                         page->mapping->a_ops->invalidatepage(page, offset);
2704 #endif
2705                 unlock_page(page);
2706                 return 0; /* don't care */
2707         }
2708
2709         /*
2710          * The page straddles i_size.  It must be zeroed out on each and every
2711          * writepage invocation because it may be mmapped.  "A file is mapped
2712          * in multiples of the page size.  For a file that is not a multiple of
2713          * the  page size, the remaining memory is zeroed when mapped, and
2714          * writes to that region are not written out to the file."
2715          */
2716         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2717 out:
2718         ret = mpage_writepage(page, get_block, wbc);
2719         if (ret == -EAGAIN)
2720                 ret = __block_write_full_page(inode, page, get_block, wbc,
2721                                               end_buffer_async_write);
2722         return ret;
2723 }
2724 EXPORT_SYMBOL(nobh_writepage);
2725
2726 int nobh_truncate_page(struct address_space *mapping,
2727                         loff_t from, get_block_t *get_block)
2728 {
2729         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2730         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2731         unsigned blocksize;
2732         sector_t iblock;
2733         unsigned length, pos;
2734         struct inode *inode = mapping->host;
2735         struct page *page;
2736         struct buffer_head map_bh;
2737         int err;
2738
2739         blocksize = 1 << inode->i_blkbits;
2740         length = offset & (blocksize - 1);
2741
2742         /* Block boundary? Nothing to do */
2743         if (!length)
2744                 return 0;
2745
2746         length = blocksize - length;
2747         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2748
2749         page = grab_cache_page(mapping, index);
2750         err = -ENOMEM;
2751         if (!page)
2752                 goto out;
2753
2754         if (page_has_buffers(page)) {
2755 has_buffers:
2756                 unlock_page(page);
2757                 page_cache_release(page);
2758                 return block_truncate_page(mapping, from, get_block);
2759         }
2760
2761         /* Find the buffer that contains "offset" */
2762         pos = blocksize;
2763         while (offset >= pos) {
2764                 iblock++;
2765                 pos += blocksize;
2766         }
2767
2768         map_bh.b_size = blocksize;
2769         map_bh.b_state = 0;
2770         err = get_block(inode, iblock, &map_bh, 0);
2771         if (err)
2772                 goto unlock;
2773         /* unmapped? It's a hole - nothing to do */
2774         if (!buffer_mapped(&map_bh))
2775                 goto unlock;
2776
2777         /* Ok, it's mapped. Make sure it's up-to-date */
2778         if (!PageUptodate(page)) {
2779                 err = mapping->a_ops->readpage(NULL, page);
2780                 if (err) {
2781                         page_cache_release(page);
2782                         goto out;
2783                 }
2784                 lock_page(page);
2785                 if (!PageUptodate(page)) {
2786                         err = -EIO;
2787                         goto unlock;
2788                 }
2789                 if (page_has_buffers(page))
2790                         goto has_buffers;
2791         }
2792         zero_user(page, offset, length);
2793         set_page_dirty(page);
2794         err = 0;
2795
2796 unlock:
2797         unlock_page(page);
2798         page_cache_release(page);
2799 out:
2800         return err;
2801 }
2802 EXPORT_SYMBOL(nobh_truncate_page);
2803
2804 int block_truncate_page(struct address_space *mapping,
2805                         loff_t from, get_block_t *get_block)
2806 {
2807         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2808         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2809         unsigned blocksize;
2810         sector_t iblock;
2811         unsigned length, pos;
2812         struct inode *inode = mapping->host;
2813         struct page *page;
2814         struct buffer_head *bh;
2815         int err;
2816
2817         blocksize = 1 << inode->i_blkbits;
2818         length = offset & (blocksize - 1);
2819
2820         /* Block boundary? Nothing to do */
2821         if (!length)
2822                 return 0;
2823
2824         length = blocksize - length;
2825         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2826         
2827         page = grab_cache_page(mapping, index);
2828         err = -ENOMEM;
2829         if (!page)
2830                 goto out;
2831
2832         if (!page_has_buffers(page))
2833                 create_empty_buffers(page, blocksize, 0);
2834
2835         /* Find the buffer that contains "offset" */
2836         bh = page_buffers(page);
2837         pos = blocksize;
2838         while (offset >= pos) {
2839                 bh = bh->b_this_page;
2840                 iblock++;
2841                 pos += blocksize;
2842         }
2843
2844         err = 0;
2845         if (!buffer_mapped(bh)) {
2846                 WARN_ON(bh->b_size != blocksize);
2847                 err = get_block(inode, iblock, bh, 0);
2848                 if (err)
2849                         goto unlock;
2850                 /* unmapped? It's a hole - nothing to do */
2851                 if (!buffer_mapped(bh))
2852                         goto unlock;
2853         }
2854
2855         /* Ok, it's mapped. Make sure it's up-to-date */
2856         if (PageUptodate(page))
2857                 set_buffer_uptodate(bh);
2858
2859         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2860                 err = -EIO;
2861                 ll_rw_block(READ, 1, &bh);
2862                 wait_on_buffer(bh);
2863                 /* Uhhuh. Read error. Complain and punt. */
2864                 if (!buffer_uptodate(bh))
2865                         goto unlock;
2866         }
2867
2868         zero_user(page, offset, length);
2869         mark_buffer_dirty(bh);
2870         err = 0;
2871
2872 unlock:
2873         unlock_page(page);
2874         page_cache_release(page);
2875 out:
2876         return err;
2877 }
2878 EXPORT_SYMBOL(block_truncate_page);
2879
2880 /*
2881  * The generic ->writepage function for buffer-backed address_spaces
2882  * this form passes in the end_io handler used to finish the IO.
2883  */
2884 int block_write_full_page_endio(struct page *page, get_block_t *get_block,
2885                         struct writeback_control *wbc, bh_end_io_t *handler)
2886 {
2887         struct inode * const inode = page->mapping->host;
2888         loff_t i_size = i_size_read(inode);
2889         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2890         unsigned offset;
2891
2892         /* Is the page fully inside i_size? */
2893         if (page->index < end_index)
2894                 return __block_write_full_page(inode, page, get_block, wbc,
2895                                                handler);
2896
2897         /* Is the page fully outside i_size? (truncate in progress) */
2898         offset = i_size & (PAGE_CACHE_SIZE-1);
2899         if (page->index >= end_index+1 || !offset) {
2900                 /*
2901                  * The page may have dirty, unmapped buffers.  For example,
2902                  * they may have been added in ext3_writepage().  Make them
2903                  * freeable here, so the page does not leak.
2904                  */
2905                 do_invalidatepage(page, 0, PAGE_CACHE_SIZE);
2906                 unlock_page(page);
2907                 return 0; /* don't care */
2908         }
2909
2910         /*
2911          * The page straddles i_size.  It must be zeroed out on each and every
2912          * writepage invocation because it may be mmapped.  "A file is mapped
2913          * in multiples of the page size.  For a file that is not a multiple of
2914          * the  page size, the remaining memory is zeroed when mapped, and
2915          * writes to that region are not written out to the file."
2916          */
2917         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2918         return __block_write_full_page(inode, page, get_block, wbc, handler);
2919 }
2920 EXPORT_SYMBOL(block_write_full_page_endio);
2921
2922 /*
2923  * The generic ->writepage function for buffer-backed address_spaces
2924  */
2925 int block_write_full_page(struct page *page, get_block_t *get_block,
2926                         struct writeback_control *wbc)
2927 {
2928         return block_write_full_page_endio(page, get_block, wbc,
2929                                            end_buffer_async_write);
2930 }
2931 EXPORT_SYMBOL(block_write_full_page);
2932
2933 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2934                             get_block_t *get_block)
2935 {
2936         struct buffer_head tmp;
2937         struct inode *inode = mapping->host;
2938         tmp.b_state = 0;
2939         tmp.b_blocknr = 0;
2940         tmp.b_size = 1 << inode->i_blkbits;
2941         get_block(inode, block, &tmp, 0);
2942         return tmp.b_blocknr;
2943 }
2944 EXPORT_SYMBOL(generic_block_bmap);
2945
2946 static void end_bio_bh_io_sync(struct bio *bio, int err)
2947 {
2948         struct buffer_head *bh = bio->bi_private;
2949
2950         if (err == -EOPNOTSUPP) {
2951                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2952         }
2953
2954         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2955                 set_bit(BH_Quiet, &bh->b_state);
2956
2957         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2958         bio_put(bio);
2959 }
2960
2961 /*
2962  * This allows us to do IO even on the odd last sectors
2963  * of a device, even if the bh block size is some multiple
2964  * of the physical sector size.
2965  *
2966  * We'll just truncate the bio to the size of the device,
2967  * and clear the end of the buffer head manually.
2968  *
2969  * Truly out-of-range accesses will turn into actual IO
2970  * errors, this only handles the "we need to be able to
2971  * do IO at the final sector" case.
2972  */
2973 static void guard_bh_eod(int rw, struct bio *bio, struct buffer_head *bh)
2974 {
2975         sector_t maxsector;
2976         unsigned bytes;
2977
2978         maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
2979         if (!maxsector)
2980                 return;
2981
2982         /*
2983          * If the *whole* IO is past the end of the device,
2984          * let it through, and the IO layer will turn it into
2985          * an EIO.
2986          */
2987         if (unlikely(bio->bi_iter.bi_sector >= maxsector))
2988                 return;
2989
2990         maxsector -= bio->bi_iter.bi_sector;
2991         bytes = bio->bi_iter.bi_size;
2992         if (likely((bytes >> 9) <= maxsector))
2993                 return;
2994
2995         /* Uhhuh. We've got a bh that straddles the device size! */
2996         bytes = maxsector << 9;
2997
2998         /* Truncate the bio.. */
2999         bio->bi_iter.bi_size = bytes;
3000         bio->bi_io_vec[0].bv_len = bytes;
3001
3002         /* ..and clear the end of the buffer for reads */
3003         if ((rw & RW_MASK) == READ) {
3004                 void *kaddr = kmap_atomic(bh->b_page);
3005                 memset(kaddr + bh_offset(bh) + bytes, 0, bh->b_size - bytes);
3006                 kunmap_atomic(kaddr);
3007                 flush_dcache_page(bh->b_page);
3008         }
3009 }
3010
3011 int _submit_bh(int rw, struct buffer_head *bh, unsigned long bio_flags)
3012 {
3013         struct bio *bio;
3014         int ret = 0;
3015
3016         BUG_ON(!buffer_locked(bh));
3017         BUG_ON(!buffer_mapped(bh));
3018         BUG_ON(!bh->b_end_io);
3019         BUG_ON(buffer_delay(bh));
3020         BUG_ON(buffer_unwritten(bh));
3021
3022         /*
3023          * Only clear out a write error when rewriting
3024          */
3025         if (test_set_buffer_req(bh) && (rw & WRITE))
3026                 clear_buffer_write_io_error(bh);
3027
3028         /*
3029          * from here on down, it's all bio -- do the initial mapping,
3030          * submit_bio -> generic_make_request may further map this bio around
3031          */
3032         bio = bio_alloc(GFP_NOIO, 1);
3033
3034         bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3035         bio->bi_bdev = bh->b_bdev;
3036         bio->bi_io_vec[0].bv_page = bh->b_page;
3037         bio->bi_io_vec[0].bv_len = bh->b_size;
3038         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
3039
3040         bio->bi_vcnt = 1;
3041         bio->bi_iter.bi_size = bh->b_size;
3042
3043         bio->bi_end_io = end_bio_bh_io_sync;
3044         bio->bi_private = bh;
3045         bio->bi_flags |= bio_flags;
3046
3047         /* Take care of bh's that straddle the end of the device */
3048         guard_bh_eod(rw, bio, bh);
3049
3050         if (buffer_meta(bh))
3051                 rw |= REQ_META;
3052         if (buffer_prio(bh))
3053                 rw |= REQ_PRIO;
3054
3055         bio_get(bio);
3056         submit_bio(rw, bio);
3057
3058         if (bio_flagged(bio, BIO_EOPNOTSUPP))
3059                 ret = -EOPNOTSUPP;
3060
3061         bio_put(bio);
3062         return ret;
3063 }
3064 EXPORT_SYMBOL_GPL(_submit_bh);
3065
3066 int submit_bh(int rw, struct buffer_head *bh)
3067 {
3068         return _submit_bh(rw, bh, 0);
3069 }
3070 EXPORT_SYMBOL(submit_bh);
3071
3072 /**
3073  * ll_rw_block: low-level access to block devices (DEPRECATED)
3074  * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
3075  * @nr: number of &struct buffer_heads in the array
3076  * @bhs: array of pointers to &struct buffer_head
3077  *
3078  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3079  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
3080  * %READA option is described in the documentation for generic_make_request()
3081  * which ll_rw_block() calls.
3082  *
3083  * This function drops any buffer that it cannot get a lock on (with the
3084  * BH_Lock state bit), any buffer that appears to be clean when doing a write
3085  * request, and any buffer that appears to be up-to-date when doing read
3086  * request.  Further it marks as clean buffers that are processed for
3087  * writing (the buffer cache won't assume that they are actually clean
3088  * until the buffer gets unlocked).
3089  *
3090  * ll_rw_block sets b_end_io to simple completion handler that marks
3091  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
3092  * any waiters. 
3093  *
3094  * All of the buffers must be for the same device, and must also be a
3095  * multiple of the current approved size for the device.
3096  */
3097 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
3098 {
3099         int i;
3100
3101         for (i = 0; i < nr; i++) {
3102                 struct buffer_head *bh = bhs[i];
3103
3104                 if (!trylock_buffer(bh))
3105                         continue;
3106                 if (rw == WRITE) {
3107                         if (test_clear_buffer_dirty(bh)) {
3108                                 bh->b_end_io = end_buffer_write_sync;
3109                                 get_bh(bh);
3110                                 submit_bh(WRITE, bh);
3111                                 continue;
3112                         }
3113                 } else {
3114                         if (!buffer_uptodate(bh)) {
3115                                 bh->b_end_io = end_buffer_read_sync;
3116                                 get_bh(bh);
3117                                 submit_bh(rw, bh);
3118                                 continue;
3119                         }
3120                 }
3121                 unlock_buffer(bh);
3122         }
3123 }
3124 EXPORT_SYMBOL(ll_rw_block);
3125
3126 void write_dirty_buffer(struct buffer_head *bh, int rw)
3127 {
3128         lock_buffer(bh);
3129         if (!test_clear_buffer_dirty(bh)) {
3130                 unlock_buffer(bh);
3131                 return;
3132         }
3133         bh->b_end_io = end_buffer_write_sync;
3134         get_bh(bh);
3135         submit_bh(rw, bh);
3136 }
3137 EXPORT_SYMBOL(write_dirty_buffer);
3138
3139 /*
3140  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3141  * and then start new I/O and then wait upon it.  The caller must have a ref on
3142  * the buffer_head.
3143  */
3144 int __sync_dirty_buffer(struct buffer_head *bh, int rw)
3145 {
3146         int ret = 0;
3147
3148         WARN_ON(atomic_read(&bh->b_count) < 1);
3149         lock_buffer(bh);
3150         if (test_clear_buffer_dirty(bh)) {
3151                 get_bh(bh);
3152                 bh->b_end_io = end_buffer_write_sync;
3153                 ret = submit_bh(rw, bh);
3154                 wait_on_buffer(bh);
3155                 if (!ret && !buffer_uptodate(bh))
3156                         ret = -EIO;
3157         } else {
3158                 unlock_buffer(bh);
3159         }
3160         return ret;
3161 }
3162 EXPORT_SYMBOL(__sync_dirty_buffer);
3163
3164 int sync_dirty_buffer(struct buffer_head *bh)
3165 {
3166         return __sync_dirty_buffer(bh, WRITE_SYNC);
3167 }
3168 EXPORT_SYMBOL(sync_dirty_buffer);
3169
3170 /*
3171  * try_to_free_buffers() checks if all the buffers on this particular page
3172  * are unused, and releases them if so.
3173  *
3174  * Exclusion against try_to_free_buffers may be obtained by either
3175  * locking the page or by holding its mapping's private_lock.
3176  *
3177  * If the page is dirty but all the buffers are clean then we need to
3178  * be sure to mark the page clean as well.  This is because the page
3179  * may be against a block device, and a later reattachment of buffers
3180  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3181  * filesystem data on the same device.
3182  *
3183  * The same applies to regular filesystem pages: if all the buffers are
3184  * clean then we set the page clean and proceed.  To do that, we require
3185  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3186  * private_lock.
3187  *
3188  * try_to_free_buffers() is non-blocking.
3189  */
3190 static inline int buffer_busy(struct buffer_head *bh)
3191 {
3192         return atomic_read(&bh->b_count) |
3193                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3194 }
3195
3196 static int
3197 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3198 {
3199         struct buffer_head *head = page_buffers(page);
3200         struct buffer_head *bh;
3201
3202         bh = head;
3203         do {
3204                 if (buffer_write_io_error(bh) && page->mapping)
3205                         set_bit(AS_EIO, &page->mapping->flags);
3206                 if (buffer_busy(bh))
3207                         goto failed;
3208                 bh = bh->b_this_page;
3209         } while (bh != head);
3210
3211         do {
3212                 struct buffer_head *next = bh->b_this_page;
3213
3214                 if (bh->b_assoc_map)
3215                         __remove_assoc_queue(bh);
3216                 bh = next;
3217         } while (bh != head);
3218         *buffers_to_free = head;
3219         __clear_page_buffers(page);
3220         return 1;
3221 failed:
3222         return 0;
3223 }
3224
3225 int try_to_free_buffers(struct page *page)
3226 {
3227         struct address_space * const mapping = page->mapping;
3228         struct buffer_head *buffers_to_free = NULL;
3229         int ret = 0;
3230
3231         BUG_ON(!PageLocked(page));
3232         if (PageWriteback(page))
3233                 return 0;
3234
3235         if (mapping == NULL) {          /* can this still happen? */
3236                 ret = drop_buffers(page, &buffers_to_free);
3237                 goto out;
3238         }
3239
3240         spin_lock(&mapping->private_lock);
3241         ret = drop_buffers(page, &buffers_to_free);
3242
3243         /*
3244          * If the filesystem writes its buffers by hand (eg ext3)
3245          * then we can have clean buffers against a dirty page.  We
3246          * clean the page here; otherwise the VM will never notice
3247          * that the filesystem did any IO at all.
3248          *
3249          * Also, during truncate, discard_buffer will have marked all
3250          * the page's buffers clean.  We discover that here and clean
3251          * the page also.
3252          *
3253          * private_lock must be held over this entire operation in order
3254          * to synchronise against __set_page_dirty_buffers and prevent the
3255          * dirty bit from being lost.
3256          */
3257         if (ret)
3258                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3259         spin_unlock(&mapping->private_lock);
3260 out:
3261         if (buffers_to_free) {
3262                 struct buffer_head *bh = buffers_to_free;
3263
3264                 do {
3265                         struct buffer_head *next = bh->b_this_page;
3266                         free_buffer_head(bh);
3267                         bh = next;
3268                 } while (bh != buffers_to_free);
3269         }
3270         return ret;
3271 }
3272 EXPORT_SYMBOL(try_to_free_buffers);
3273
3274 /*
3275  * There are no bdflush tunables left.  But distributions are
3276  * still running obsolete flush daemons, so we terminate them here.
3277  *
3278  * Use of bdflush() is deprecated and will be removed in a future kernel.
3279  * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3280  */
3281 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3282 {
3283         static int msg_count;
3284
3285         if (!capable(CAP_SYS_ADMIN))
3286                 return -EPERM;
3287
3288         if (msg_count < 5) {
3289                 msg_count++;
3290                 printk(KERN_INFO
3291                         "warning: process `%s' used the obsolete bdflush"
3292                         " system call\n", current->comm);
3293                 printk(KERN_INFO "Fix your initscripts?\n");
3294         }
3295
3296         if (func == 1)
3297                 do_exit(0);
3298         return 0;
3299 }
3300
3301 /*
3302  * Buffer-head allocation
3303  */
3304 static struct kmem_cache *bh_cachep __read_mostly;
3305
3306 /*
3307  * Once the number of bh's in the machine exceeds this level, we start
3308  * stripping them in writeback.
3309  */
3310 static unsigned long max_buffer_heads;
3311
3312 int buffer_heads_over_limit;
3313
3314 struct bh_accounting {
3315         int nr;                 /* Number of live bh's */
3316         int ratelimit;          /* Limit cacheline bouncing */
3317 };
3318
3319 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3320
3321 static void recalc_bh_state(void)
3322 {
3323         int i;
3324         int tot = 0;
3325
3326         if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3327                 return;
3328         __this_cpu_write(bh_accounting.ratelimit, 0);
3329         for_each_online_cpu(i)
3330                 tot += per_cpu(bh_accounting, i).nr;
3331         buffer_heads_over_limit = (tot > max_buffer_heads);
3332 }
3333
3334 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3335 {
3336         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3337         if (ret) {
3338                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3339                 preempt_disable();
3340                 __this_cpu_inc(bh_accounting.nr);
3341                 recalc_bh_state();
3342                 preempt_enable();
3343         }
3344         return ret;
3345 }
3346 EXPORT_SYMBOL(alloc_buffer_head);
3347
3348 void free_buffer_head(struct buffer_head *bh)
3349 {
3350         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3351         kmem_cache_free(bh_cachep, bh);
3352         preempt_disable();
3353         __this_cpu_dec(bh_accounting.nr);
3354         recalc_bh_state();
3355         preempt_enable();
3356 }
3357 EXPORT_SYMBOL(free_buffer_head);
3358
3359 static void buffer_exit_cpu(int cpu)
3360 {
3361         int i;
3362         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3363
3364         for (i = 0; i < BH_LRU_SIZE; i++) {
3365                 brelse(b->bhs[i]);
3366                 b->bhs[i] = NULL;
3367         }
3368         this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3369         per_cpu(bh_accounting, cpu).nr = 0;
3370 }
3371
3372 static int buffer_cpu_notify(struct notifier_block *self,
3373                               unsigned long action, void *hcpu)
3374 {
3375         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3376                 buffer_exit_cpu((unsigned long)hcpu);
3377         return NOTIFY_OK;
3378 }
3379
3380 /**
3381  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3382  * @bh: struct buffer_head
3383  *
3384  * Return true if the buffer is up-to-date and false,
3385  * with the buffer locked, if not.
3386  */
3387 int bh_uptodate_or_lock(struct buffer_head *bh)
3388 {
3389         if (!buffer_uptodate(bh)) {
3390                 lock_buffer(bh);
3391                 if (!buffer_uptodate(bh))
3392                         return 0;
3393                 unlock_buffer(bh);
3394         }
3395         return 1;
3396 }
3397 EXPORT_SYMBOL(bh_uptodate_or_lock);
3398
3399 /**
3400  * bh_submit_read - Submit a locked buffer for reading
3401  * @bh: struct buffer_head
3402  *
3403  * Returns zero on success and -EIO on error.
3404  */
3405 int bh_submit_read(struct buffer_head *bh)
3406 {
3407         BUG_ON(!buffer_locked(bh));
3408
3409         if (buffer_uptodate(bh)) {
3410                 unlock_buffer(bh);
3411                 return 0;
3412         }
3413
3414         get_bh(bh);
3415         bh->b_end_io = end_buffer_read_sync;
3416         submit_bh(READ, bh);
3417         wait_on_buffer(bh);
3418         if (buffer_uptodate(bh))
3419                 return 0;
3420         return -EIO;
3421 }
3422 EXPORT_SYMBOL(bh_submit_read);
3423
3424 void __init buffer_init(void)
3425 {
3426         unsigned long nrpages;
3427
3428         bh_cachep = kmem_cache_create("buffer_head",
3429                         sizeof(struct buffer_head), 0,
3430                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3431                                 SLAB_MEM_SPREAD),
3432                                 NULL);
3433
3434         /*
3435          * Limit the bh occupancy to 10% of ZONE_NORMAL
3436          */
3437         nrpages = (nr_free_buffer_pages() * 10) / 100;
3438         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3439         hotcpu_notifier(buffer_cpu_notify, 0);
3440 }