Merge tag 'jfs-3.11' of git://github.com/kleikamp/linux-shaggy
[platform/adaptation/renesas_rcar/renesas_kernel.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         spin_lock_irq(&mapping->tree_lock);
658         if (page->mapping) {    /* Race with truncate? */
659                 WARN_ON_ONCE(warn && !PageUptodate(page));
660                 account_page_dirtied(page, mapping);
661                 radix_tree_tag_set(&mapping->page_tree,
662                                 page_index(page), PAGECACHE_TAG_DIRTY);
663         }
664         spin_unlock_irq(&mapping->tree_lock);
665         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
666 }
667
668 /*
669  * Add a page to the dirty page list.
670  *
671  * It is a sad fact of life that this function is called from several places
672  * deeply under spinlocking.  It may not sleep.
673  *
674  * If the page has buffers, the uptodate buffers are set dirty, to preserve
675  * dirty-state coherency between the page and the buffers.  It the page does
676  * not have buffers then when they are later attached they will all be set
677  * dirty.
678  *
679  * The buffers are dirtied before the page is dirtied.  There's a small race
680  * window in which a writepage caller may see the page cleanness but not the
681  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
682  * before the buffers, a concurrent writepage caller could clear the page dirty
683  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
684  * page on the dirty page list.
685  *
686  * We use private_lock to lock against try_to_free_buffers while using the
687  * page's buffer list.  Also use this to protect against clean buffers being
688  * added to the page after it was set dirty.
689  *
690  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
691  * address_space though.
692  */
693 int __set_page_dirty_buffers(struct page *page)
694 {
695         int newly_dirty;
696         struct address_space *mapping = page_mapping(page);
697
698         if (unlikely(!mapping))
699                 return !TestSetPageDirty(page);
700
701         spin_lock(&mapping->private_lock);
702         if (page_has_buffers(page)) {
703                 struct buffer_head *head = page_buffers(page);
704                 struct buffer_head *bh = head;
705
706                 do {
707                         set_buffer_dirty(bh);
708                         bh = bh->b_this_page;
709                 } while (bh != head);
710         }
711         newly_dirty = !TestSetPageDirty(page);
712         spin_unlock(&mapping->private_lock);
713
714         if (newly_dirty)
715                 __set_page_dirty(page, mapping, 1);
716         return newly_dirty;
717 }
718 EXPORT_SYMBOL(__set_page_dirty_buffers);
719
720 /*
721  * Write out and wait upon a list of buffers.
722  *
723  * We have conflicting pressures: we want to make sure that all
724  * initially dirty buffers get waited on, but that any subsequently
725  * dirtied buffers don't.  After all, we don't want fsync to last
726  * forever if somebody is actively writing to the file.
727  *
728  * Do this in two main stages: first we copy dirty buffers to a
729  * temporary inode list, queueing the writes as we go.  Then we clean
730  * up, waiting for those writes to complete.
731  * 
732  * During this second stage, any subsequent updates to the file may end
733  * up refiling the buffer on the original inode's dirty list again, so
734  * there is a chance we will end up with a buffer queued for write but
735  * not yet completed on that list.  So, as a final cleanup we go through
736  * the osync code to catch these locked, dirty buffers without requeuing
737  * any newly dirty buffers for write.
738  */
739 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
740 {
741         struct buffer_head *bh;
742         struct list_head tmp;
743         struct address_space *mapping;
744         int err = 0, err2;
745         struct blk_plug plug;
746
747         INIT_LIST_HEAD(&tmp);
748         blk_start_plug(&plug);
749
750         spin_lock(lock);
751         while (!list_empty(list)) {
752                 bh = BH_ENTRY(list->next);
753                 mapping = bh->b_assoc_map;
754                 __remove_assoc_queue(bh);
755                 /* Avoid race with mark_buffer_dirty_inode() which does
756                  * a lockless check and we rely on seeing the dirty bit */
757                 smp_mb();
758                 if (buffer_dirty(bh) || buffer_locked(bh)) {
759                         list_add(&bh->b_assoc_buffers, &tmp);
760                         bh->b_assoc_map = mapping;
761                         if (buffer_dirty(bh)) {
762                                 get_bh(bh);
763                                 spin_unlock(lock);
764                                 /*
765                                  * Ensure any pending I/O completes so that
766                                  * write_dirty_buffer() actually writes the
767                                  * current contents - it is a noop if I/O is
768                                  * still in flight on potentially older
769                                  * contents.
770                                  */
771                                 write_dirty_buffer(bh, WRITE_SYNC);
772
773                                 /*
774                                  * Kick off IO for the previous mapping. Note
775                                  * that we will not run the very last mapping,
776                                  * wait_on_buffer() will do that for us
777                                  * through sync_buffer().
778                                  */
779                                 brelse(bh);
780                                 spin_lock(lock);
781                         }
782                 }
783         }
784
785         spin_unlock(lock);
786         blk_finish_plug(&plug);
787         spin_lock(lock);
788
789         while (!list_empty(&tmp)) {
790                 bh = BH_ENTRY(tmp.prev);
791                 get_bh(bh);
792                 mapping = bh->b_assoc_map;
793                 __remove_assoc_queue(bh);
794                 /* Avoid race with mark_buffer_dirty_inode() which does
795                  * a lockless check and we rely on seeing the dirty bit */
796                 smp_mb();
797                 if (buffer_dirty(bh)) {
798                         list_add(&bh->b_assoc_buffers,
799                                  &mapping->private_list);
800                         bh->b_assoc_map = mapping;
801                 }
802                 spin_unlock(lock);
803                 wait_on_buffer(bh);
804                 if (!buffer_uptodate(bh))
805                         err = -EIO;
806                 brelse(bh);
807                 spin_lock(lock);
808         }
809         
810         spin_unlock(lock);
811         err2 = osync_buffers_list(lock, list);
812         if (err)
813                 return err;
814         else
815                 return err2;
816 }
817
818 /*
819  * Invalidate any and all dirty buffers on a given inode.  We are
820  * probably unmounting the fs, but that doesn't mean we have already
821  * done a sync().  Just drop the buffers from the inode list.
822  *
823  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
824  * assumes that all the buffers are against the blockdev.  Not true
825  * for reiserfs.
826  */
827 void invalidate_inode_buffers(struct inode *inode)
828 {
829         if (inode_has_buffers(inode)) {
830                 struct address_space *mapping = &inode->i_data;
831                 struct list_head *list = &mapping->private_list;
832                 struct address_space *buffer_mapping = mapping->private_data;
833
834                 spin_lock(&buffer_mapping->private_lock);
835                 while (!list_empty(list))
836                         __remove_assoc_queue(BH_ENTRY(list->next));
837                 spin_unlock(&buffer_mapping->private_lock);
838         }
839 }
840 EXPORT_SYMBOL(invalidate_inode_buffers);
841
842 /*
843  * Remove any clean buffers from the inode's buffer list.  This is called
844  * when we're trying to free the inode itself.  Those buffers can pin it.
845  *
846  * Returns true if all buffers were removed.
847  */
848 int remove_inode_buffers(struct inode *inode)
849 {
850         int ret = 1;
851
852         if (inode_has_buffers(inode)) {
853                 struct address_space *mapping = &inode->i_data;
854                 struct list_head *list = &mapping->private_list;
855                 struct address_space *buffer_mapping = mapping->private_data;
856
857                 spin_lock(&buffer_mapping->private_lock);
858                 while (!list_empty(list)) {
859                         struct buffer_head *bh = BH_ENTRY(list->next);
860                         if (buffer_dirty(bh)) {
861                                 ret = 0;
862                                 break;
863                         }
864                         __remove_assoc_queue(bh);
865                 }
866                 spin_unlock(&buffer_mapping->private_lock);
867         }
868         return ret;
869 }
870
871 /*
872  * Create the appropriate buffers when given a page for data area and
873  * the size of each buffer.. Use the bh->b_this_page linked list to
874  * follow the buffers created.  Return NULL if unable to create more
875  * buffers.
876  *
877  * The retry flag is used to differentiate async IO (paging, swapping)
878  * which may not fail from ordinary buffer allocations.
879  */
880 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
881                 int retry)
882 {
883         struct buffer_head *bh, *head;
884         long offset;
885
886 try_again:
887         head = NULL;
888         offset = PAGE_SIZE;
889         while ((offset -= size) >= 0) {
890                 bh = alloc_buffer_head(GFP_NOFS);
891                 if (!bh)
892                         goto no_grow;
893
894                 bh->b_this_page = head;
895                 bh->b_blocknr = -1;
896                 head = bh;
897
898                 bh->b_size = size;
899
900                 /* Link the buffer to its page */
901                 set_bh_page(bh, page, offset);
902         }
903         return head;
904 /*
905  * In case anything failed, we just free everything we got.
906  */
907 no_grow:
908         if (head) {
909                 do {
910                         bh = head;
911                         head = head->b_this_page;
912                         free_buffer_head(bh);
913                 } while (head);
914         }
915
916         /*
917          * Return failure for non-async IO requests.  Async IO requests
918          * are not allowed to fail, so we have to wait until buffer heads
919          * become available.  But we don't want tasks sleeping with 
920          * partially complete buffers, so all were released above.
921          */
922         if (!retry)
923                 return NULL;
924
925         /* We're _really_ low on memory. Now we just
926          * wait for old buffer heads to become free due to
927          * finishing IO.  Since this is an async request and
928          * the reserve list is empty, we're sure there are 
929          * async buffer heads in use.
930          */
931         free_more_memory();
932         goto try_again;
933 }
934 EXPORT_SYMBOL_GPL(alloc_page_buffers);
935
936 static inline void
937 link_dev_buffers(struct page *page, struct buffer_head *head)
938 {
939         struct buffer_head *bh, *tail;
940
941         bh = head;
942         do {
943                 tail = bh;
944                 bh = bh->b_this_page;
945         } while (bh);
946         tail->b_this_page = head;
947         attach_page_buffers(page, head);
948 }
949
950 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
951 {
952         sector_t retval = ~((sector_t)0);
953         loff_t sz = i_size_read(bdev->bd_inode);
954
955         if (sz) {
956                 unsigned int sizebits = blksize_bits(size);
957                 retval = (sz >> sizebits);
958         }
959         return retval;
960 }
961
962 /*
963  * Initialise the state of a blockdev page's buffers.
964  */ 
965 static sector_t
966 init_page_buffers(struct page *page, struct block_device *bdev,
967                         sector_t block, int size)
968 {
969         struct buffer_head *head = page_buffers(page);
970         struct buffer_head *bh = head;
971         int uptodate = PageUptodate(page);
972         sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
973
974         do {
975                 if (!buffer_mapped(bh)) {
976                         init_buffer(bh, NULL, NULL);
977                         bh->b_bdev = bdev;
978                         bh->b_blocknr = block;
979                         if (uptodate)
980                                 set_buffer_uptodate(bh);
981                         if (block < end_block)
982                                 set_buffer_mapped(bh);
983                 }
984                 block++;
985                 bh = bh->b_this_page;
986         } while (bh != head);
987
988         /*
989          * Caller needs to validate requested block against end of device.
990          */
991         return end_block;
992 }
993
994 /*
995  * Create the page-cache page that contains the requested block.
996  *
997  * This is used purely for blockdev mappings.
998  */
999 static int
1000 grow_dev_page(struct block_device *bdev, sector_t block,
1001                 pgoff_t index, int size, int sizebits)
1002 {
1003         struct inode *inode = bdev->bd_inode;
1004         struct page *page;
1005         struct buffer_head *bh;
1006         sector_t end_block;
1007         int ret = 0;            /* Will call free_more_memory() */
1008
1009         page = find_or_create_page(inode->i_mapping, index,
1010                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
1011         if (!page)
1012                 return ret;
1013
1014         BUG_ON(!PageLocked(page));
1015
1016         if (page_has_buffers(page)) {
1017                 bh = page_buffers(page);
1018                 if (bh->b_size == size) {
1019                         end_block = init_page_buffers(page, bdev,
1020                                                 index << sizebits, size);
1021                         goto done;
1022                 }
1023                 if (!try_to_free_buffers(page))
1024                         goto failed;
1025         }
1026
1027         /*
1028          * Allocate some buffers for this page
1029          */
1030         bh = alloc_page_buffers(page, size, 0);
1031         if (!bh)
1032                 goto failed;
1033
1034         /*
1035          * Link the page to the buffers and initialise them.  Take the
1036          * lock to be atomic wrt __find_get_block(), which does not
1037          * run under the page lock.
1038          */
1039         spin_lock(&inode->i_mapping->private_lock);
1040         link_dev_buffers(page, bh);
1041         end_block = init_page_buffers(page, bdev, index << sizebits, size);
1042         spin_unlock(&inode->i_mapping->private_lock);
1043 done:
1044         ret = (block < end_block) ? 1 : -ENXIO;
1045 failed:
1046         unlock_page(page);
1047         page_cache_release(page);
1048         return ret;
1049 }
1050
1051 /*
1052  * Create buffers for the specified block device block's page.  If
1053  * that page was dirty, the buffers are set dirty also.
1054  */
1055 static int
1056 grow_buffers(struct block_device *bdev, sector_t block, int size)
1057 {
1058         pgoff_t index;
1059         int sizebits;
1060
1061         sizebits = -1;
1062         do {
1063                 sizebits++;
1064         } while ((size << sizebits) < PAGE_SIZE);
1065
1066         index = block >> sizebits;
1067
1068         /*
1069          * Check for a block which wants to lie outside our maximum possible
1070          * pagecache index.  (this comparison is done using sector_t types).
1071          */
1072         if (unlikely(index != block >> sizebits)) {
1073                 char b[BDEVNAME_SIZE];
1074
1075                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1076                         "device %s\n",
1077                         __func__, (unsigned long long)block,
1078                         bdevname(bdev, b));
1079                 return -EIO;
1080         }
1081
1082         /* Create a page with the proper size buffers.. */
1083         return grow_dev_page(bdev, block, index, size, sizebits);
1084 }
1085
1086 static struct buffer_head *
1087 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1088 {
1089         /* Size must be multiple of hard sectorsize */
1090         if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1091                         (size < 512 || size > PAGE_SIZE))) {
1092                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1093                                         size);
1094                 printk(KERN_ERR "logical block size: %d\n",
1095                                         bdev_logical_block_size(bdev));
1096
1097                 dump_stack();
1098                 return NULL;
1099         }
1100
1101         for (;;) {
1102                 struct buffer_head *bh;
1103                 int ret;
1104
1105                 bh = __find_get_block(bdev, block, size);
1106                 if (bh)
1107                         return bh;
1108
1109                 ret = grow_buffers(bdev, block, size);
1110                 if (ret < 0)
1111                         return NULL;
1112                 if (ret == 0)
1113                         free_more_memory();
1114         }
1115 }
1116
1117 /*
1118  * The relationship between dirty buffers and dirty pages:
1119  *
1120  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1121  * the page is tagged dirty in its radix tree.
1122  *
1123  * At all times, the dirtiness of the buffers represents the dirtiness of
1124  * subsections of the page.  If the page has buffers, the page dirty bit is
1125  * merely a hint about the true dirty state.
1126  *
1127  * When a page is set dirty in its entirety, all its buffers are marked dirty
1128  * (if the page has buffers).
1129  *
1130  * When a buffer is marked dirty, its page is dirtied, but the page's other
1131  * buffers are not.
1132  *
1133  * Also.  When blockdev buffers are explicitly read with bread(), they
1134  * individually become uptodate.  But their backing page remains not
1135  * uptodate - even if all of its buffers are uptodate.  A subsequent
1136  * block_read_full_page() against that page will discover all the uptodate
1137  * buffers, will set the page uptodate and will perform no I/O.
1138  */
1139
1140 /**
1141  * mark_buffer_dirty - mark a buffer_head as needing writeout
1142  * @bh: the buffer_head to mark dirty
1143  *
1144  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1145  * backing page dirty, then tag the page as dirty in its address_space's radix
1146  * tree and then attach the address_space's inode to its superblock's dirty
1147  * inode list.
1148  *
1149  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1150  * mapping->tree_lock and mapping->host->i_lock.
1151  */
1152 void mark_buffer_dirty(struct buffer_head *bh)
1153 {
1154         WARN_ON_ONCE(!buffer_uptodate(bh));
1155
1156         trace_block_dirty_buffer(bh);
1157
1158         /*
1159          * Very *carefully* optimize the it-is-already-dirty case.
1160          *
1161          * Don't let the final "is it dirty" escape to before we
1162          * perhaps modified the buffer.
1163          */
1164         if (buffer_dirty(bh)) {
1165                 smp_mb();
1166                 if (buffer_dirty(bh))
1167                         return;
1168         }
1169
1170         if (!test_set_buffer_dirty(bh)) {
1171                 struct page *page = bh->b_page;
1172                 if (!TestSetPageDirty(page)) {
1173                         struct address_space *mapping = page_mapping(page);
1174                         if (mapping)
1175                                 __set_page_dirty(page, mapping, 0);
1176                 }
1177         }
1178 }
1179 EXPORT_SYMBOL(mark_buffer_dirty);
1180
1181 /*
1182  * Decrement a buffer_head's reference count.  If all buffers against a page
1183  * have zero reference count, are clean and unlocked, and if the page is clean
1184  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1185  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1186  * a page but it ends up not being freed, and buffers may later be reattached).
1187  */
1188 void __brelse(struct buffer_head * buf)
1189 {
1190         if (atomic_read(&buf->b_count)) {
1191                 put_bh(buf);
1192                 return;
1193         }
1194         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1195 }
1196 EXPORT_SYMBOL(__brelse);
1197
1198 /*
1199  * bforget() is like brelse(), except it discards any
1200  * potentially dirty data.
1201  */
1202 void __bforget(struct buffer_head *bh)
1203 {
1204         clear_buffer_dirty(bh);
1205         if (bh->b_assoc_map) {
1206                 struct address_space *buffer_mapping = bh->b_page->mapping;
1207
1208                 spin_lock(&buffer_mapping->private_lock);
1209                 list_del_init(&bh->b_assoc_buffers);
1210                 bh->b_assoc_map = NULL;
1211                 spin_unlock(&buffer_mapping->private_lock);
1212         }
1213         __brelse(bh);
1214 }
1215 EXPORT_SYMBOL(__bforget);
1216
1217 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1218 {
1219         lock_buffer(bh);
1220         if (buffer_uptodate(bh)) {
1221                 unlock_buffer(bh);
1222                 return bh;
1223         } else {
1224                 get_bh(bh);
1225                 bh->b_end_io = end_buffer_read_sync;
1226                 submit_bh(READ, bh);
1227                 wait_on_buffer(bh);
1228                 if (buffer_uptodate(bh))
1229                         return bh;
1230         }
1231         brelse(bh);
1232         return NULL;
1233 }
1234
1235 /*
1236  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1237  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1238  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1239  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1240  * CPU's LRUs at the same time.
1241  *
1242  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1243  * sb_find_get_block().
1244  *
1245  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1246  * a local interrupt disable for that.
1247  */
1248
1249 #define BH_LRU_SIZE     8
1250
1251 struct bh_lru {
1252         struct buffer_head *bhs[BH_LRU_SIZE];
1253 };
1254
1255 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1256
1257 #ifdef CONFIG_SMP
1258 #define bh_lru_lock()   local_irq_disable()
1259 #define bh_lru_unlock() local_irq_enable()
1260 #else
1261 #define bh_lru_lock()   preempt_disable()
1262 #define bh_lru_unlock() preempt_enable()
1263 #endif
1264
1265 static inline void check_irqs_on(void)
1266 {
1267 #ifdef irqs_disabled
1268         BUG_ON(irqs_disabled());
1269 #endif
1270 }
1271
1272 /*
1273  * The LRU management algorithm is dopey-but-simple.  Sorry.
1274  */
1275 static void bh_lru_install(struct buffer_head *bh)
1276 {
1277         struct buffer_head *evictee = NULL;
1278
1279         check_irqs_on();
1280         bh_lru_lock();
1281         if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1282                 struct buffer_head *bhs[BH_LRU_SIZE];
1283                 int in;
1284                 int out = 0;
1285
1286                 get_bh(bh);
1287                 bhs[out++] = bh;
1288                 for (in = 0; in < BH_LRU_SIZE; in++) {
1289                         struct buffer_head *bh2 =
1290                                 __this_cpu_read(bh_lrus.bhs[in]);
1291
1292                         if (bh2 == bh) {
1293                                 __brelse(bh2);
1294                         } else {
1295                                 if (out >= BH_LRU_SIZE) {
1296                                         BUG_ON(evictee != NULL);
1297                                         evictee = bh2;
1298                                 } else {
1299                                         bhs[out++] = bh2;
1300                                 }
1301                         }
1302                 }
1303                 while (out < BH_LRU_SIZE)
1304                         bhs[out++] = NULL;
1305                 memcpy(__this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1306         }
1307         bh_lru_unlock();
1308
1309         if (evictee)
1310                 __brelse(evictee);
1311 }
1312
1313 /*
1314  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1315  */
1316 static struct buffer_head *
1317 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1318 {
1319         struct buffer_head *ret = NULL;
1320         unsigned int i;
1321
1322         check_irqs_on();
1323         bh_lru_lock();
1324         for (i = 0; i < BH_LRU_SIZE; i++) {
1325                 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1326
1327                 if (bh && bh->b_bdev == bdev &&
1328                                 bh->b_blocknr == block && bh->b_size == size) {
1329                         if (i) {
1330                                 while (i) {
1331                                         __this_cpu_write(bh_lrus.bhs[i],
1332                                                 __this_cpu_read(bh_lrus.bhs[i - 1]));
1333                                         i--;
1334                                 }
1335                                 __this_cpu_write(bh_lrus.bhs[0], bh);
1336                         }
1337                         get_bh(bh);
1338                         ret = bh;
1339                         break;
1340                 }
1341         }
1342         bh_lru_unlock();
1343         return ret;
1344 }
1345
1346 /*
1347  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1348  * it in the LRU and mark it as accessed.  If it is not present then return
1349  * NULL
1350  */
1351 struct buffer_head *
1352 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1353 {
1354         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1355
1356         if (bh == NULL) {
1357                 bh = __find_get_block_slow(bdev, block);
1358                 if (bh)
1359                         bh_lru_install(bh);
1360         }
1361         if (bh)
1362                 touch_buffer(bh);
1363         return bh;
1364 }
1365 EXPORT_SYMBOL(__find_get_block);
1366
1367 /*
1368  * __getblk will locate (and, if necessary, create) the buffer_head
1369  * which corresponds to the passed block_device, block and size. The
1370  * returned buffer has its reference count incremented.
1371  *
1372  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1373  * attempt is failing.  FIXME, perhaps?
1374  */
1375 struct buffer_head *
1376 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1377 {
1378         struct buffer_head *bh = __find_get_block(bdev, block, size);
1379
1380         might_sleep();
1381         if (bh == NULL)
1382                 bh = __getblk_slow(bdev, block, size);
1383         return bh;
1384 }
1385 EXPORT_SYMBOL(__getblk);
1386
1387 /*
1388  * Do async read-ahead on a buffer..
1389  */
1390 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1391 {
1392         struct buffer_head *bh = __getblk(bdev, block, size);
1393         if (likely(bh)) {
1394                 ll_rw_block(READA, 1, &bh);
1395                 brelse(bh);
1396         }
1397 }
1398 EXPORT_SYMBOL(__breadahead);
1399
1400 /**
1401  *  __bread() - reads a specified block and returns the bh
1402  *  @bdev: the block_device to read from
1403  *  @block: number of block
1404  *  @size: size (in bytes) to read
1405  * 
1406  *  Reads a specified block, and returns buffer head that contains it.
1407  *  It returns NULL if the block was unreadable.
1408  */
1409 struct buffer_head *
1410 __bread(struct block_device *bdev, sector_t block, unsigned size)
1411 {
1412         struct buffer_head *bh = __getblk(bdev, block, size);
1413
1414         if (likely(bh) && !buffer_uptodate(bh))
1415                 bh = __bread_slow(bh);
1416         return bh;
1417 }
1418 EXPORT_SYMBOL(__bread);
1419
1420 /*
1421  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1422  * This doesn't race because it runs in each cpu either in irq
1423  * or with preempt disabled.
1424  */
1425 static void invalidate_bh_lru(void *arg)
1426 {
1427         struct bh_lru *b = &get_cpu_var(bh_lrus);
1428         int i;
1429
1430         for (i = 0; i < BH_LRU_SIZE; i++) {
1431                 brelse(b->bhs[i]);
1432                 b->bhs[i] = NULL;
1433         }
1434         put_cpu_var(bh_lrus);
1435 }
1436
1437 static bool has_bh_in_lru(int cpu, void *dummy)
1438 {
1439         struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1440         int i;
1441         
1442         for (i = 0; i < BH_LRU_SIZE; i++) {
1443                 if (b->bhs[i])
1444                         return 1;
1445         }
1446
1447         return 0;
1448 }
1449
1450 void invalidate_bh_lrus(void)
1451 {
1452         on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1453 }
1454 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1455
1456 void set_bh_page(struct buffer_head *bh,
1457                 struct page *page, unsigned long offset)
1458 {
1459         bh->b_page = page;
1460         BUG_ON(offset >= PAGE_SIZE);
1461         if (PageHighMem(page))
1462                 /*
1463                  * This catches illegal uses and preserves the offset:
1464                  */
1465                 bh->b_data = (char *)(0 + offset);
1466         else
1467                 bh->b_data = page_address(page) + offset;
1468 }
1469 EXPORT_SYMBOL(set_bh_page);
1470
1471 /*
1472  * Called when truncating a buffer on a page completely.
1473  */
1474 static void discard_buffer(struct buffer_head * bh)
1475 {
1476         lock_buffer(bh);
1477         clear_buffer_dirty(bh);
1478         bh->b_bdev = NULL;
1479         clear_buffer_mapped(bh);
1480         clear_buffer_req(bh);
1481         clear_buffer_new(bh);
1482         clear_buffer_delay(bh);
1483         clear_buffer_unwritten(bh);
1484         unlock_buffer(bh);
1485 }
1486
1487 /**
1488  * block_invalidatepage - invalidate part or all of a buffer-backed page
1489  *
1490  * @page: the page which is affected
1491  * @offset: start of the range to invalidate
1492  * @length: length of the range to invalidate
1493  *
1494  * block_invalidatepage() is called when all or part of the page has become
1495  * invalidated by a truncate operation.
1496  *
1497  * block_invalidatepage() does not have to release all buffers, but it must
1498  * ensure that no dirty buffer is left outside @offset and that no I/O
1499  * is underway against any of the blocks which are outside the truncation
1500  * point.  Because the caller is about to free (and possibly reuse) those
1501  * blocks on-disk.
1502  */
1503 void block_invalidatepage(struct page *page, unsigned int offset,
1504                           unsigned int length)
1505 {
1506         struct buffer_head *head, *bh, *next;
1507         unsigned int curr_off = 0;
1508         unsigned int stop = length + offset;
1509
1510         BUG_ON(!PageLocked(page));
1511         if (!page_has_buffers(page))
1512                 goto out;
1513
1514         /*
1515          * Check for overflow
1516          */
1517         BUG_ON(stop > PAGE_CACHE_SIZE || stop < length);
1518
1519         head = page_buffers(page);
1520         bh = head;
1521         do {
1522                 unsigned int next_off = curr_off + bh->b_size;
1523                 next = bh->b_this_page;
1524
1525                 /*
1526                  * Are we still fully in range ?
1527                  */
1528                 if (next_off > stop)
1529                         goto out;
1530
1531                 /*
1532                  * is this block fully invalidated?
1533                  */
1534                 if (offset <= curr_off)
1535                         discard_buffer(bh);
1536                 curr_off = next_off;
1537                 bh = next;
1538         } while (bh != head);
1539
1540         /*
1541          * We release buffers only if the entire page is being invalidated.
1542          * The get_block cached value has been unconditionally invalidated,
1543          * so real IO is not possible anymore.
1544          */
1545         if (offset == 0)
1546                 try_to_release_page(page, 0);
1547 out:
1548         return;
1549 }
1550 EXPORT_SYMBOL(block_invalidatepage);
1551
1552
1553 /*
1554  * We attach and possibly dirty the buffers atomically wrt
1555  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1556  * is already excluded via the page lock.
1557  */
1558 void create_empty_buffers(struct page *page,
1559                         unsigned long blocksize, unsigned long b_state)
1560 {
1561         struct buffer_head *bh, *head, *tail;
1562
1563         head = alloc_page_buffers(page, blocksize, 1);
1564         bh = head;
1565         do {
1566                 bh->b_state |= b_state;
1567                 tail = bh;
1568                 bh = bh->b_this_page;
1569         } while (bh);
1570         tail->b_this_page = head;
1571
1572         spin_lock(&page->mapping->private_lock);
1573         if (PageUptodate(page) || PageDirty(page)) {
1574                 bh = head;
1575                 do {
1576                         if (PageDirty(page))
1577                                 set_buffer_dirty(bh);
1578                         if (PageUptodate(page))
1579                                 set_buffer_uptodate(bh);
1580                         bh = bh->b_this_page;
1581                 } while (bh != head);
1582         }
1583         attach_page_buffers(page, head);
1584         spin_unlock(&page->mapping->private_lock);
1585 }
1586 EXPORT_SYMBOL(create_empty_buffers);
1587
1588 /*
1589  * We are taking a block for data and we don't want any output from any
1590  * buffer-cache aliases starting from return from that function and
1591  * until the moment when something will explicitly mark the buffer
1592  * dirty (hopefully that will not happen until we will free that block ;-)
1593  * We don't even need to mark it not-uptodate - nobody can expect
1594  * anything from a newly allocated buffer anyway. We used to used
1595  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1596  * don't want to mark the alias unmapped, for example - it would confuse
1597  * anyone who might pick it with bread() afterwards...
1598  *
1599  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1600  * be writeout I/O going on against recently-freed buffers.  We don't
1601  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1602  * only if we really need to.  That happens here.
1603  */
1604 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1605 {
1606         struct buffer_head *old_bh;
1607
1608         might_sleep();
1609
1610         old_bh = __find_get_block_slow(bdev, block);
1611         if (old_bh) {
1612                 clear_buffer_dirty(old_bh);
1613                 wait_on_buffer(old_bh);
1614                 clear_buffer_req(old_bh);
1615                 __brelse(old_bh);
1616         }
1617 }
1618 EXPORT_SYMBOL(unmap_underlying_metadata);
1619
1620 /*
1621  * Size is a power-of-two in the range 512..PAGE_SIZE,
1622  * and the case we care about most is PAGE_SIZE.
1623  *
1624  * So this *could* possibly be written with those
1625  * constraints in mind (relevant mostly if some
1626  * architecture has a slow bit-scan instruction)
1627  */
1628 static inline int block_size_bits(unsigned int blocksize)
1629 {
1630         return ilog2(blocksize);
1631 }
1632
1633 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1634 {
1635         BUG_ON(!PageLocked(page));
1636
1637         if (!page_has_buffers(page))
1638                 create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1639         return page_buffers(page);
1640 }
1641
1642 /*
1643  * NOTE! All mapped/uptodate combinations are valid:
1644  *
1645  *      Mapped  Uptodate        Meaning
1646  *
1647  *      No      No              "unknown" - must do get_block()
1648  *      No      Yes             "hole" - zero-filled
1649  *      Yes     No              "allocated" - allocated on disk, not read in
1650  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1651  *
1652  * "Dirty" is valid only with the last case (mapped+uptodate).
1653  */
1654
1655 /*
1656  * While block_write_full_page is writing back the dirty buffers under
1657  * the page lock, whoever dirtied the buffers may decide to clean them
1658  * again at any time.  We handle that by only looking at the buffer
1659  * state inside lock_buffer().
1660  *
1661  * If block_write_full_page() is called for regular writeback
1662  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1663  * locked buffer.   This only can happen if someone has written the buffer
1664  * directly, with submit_bh().  At the address_space level PageWriteback
1665  * prevents this contention from occurring.
1666  *
1667  * If block_write_full_page() is called with wbc->sync_mode ==
1668  * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1669  * causes the writes to be flagged as synchronous writes.
1670  */
1671 static int __block_write_full_page(struct inode *inode, struct page *page,
1672                         get_block_t *get_block, struct writeback_control *wbc,
1673                         bh_end_io_t *handler)
1674 {
1675         int err;
1676         sector_t block;
1677         sector_t last_block;
1678         struct buffer_head *bh, *head;
1679         unsigned int blocksize, bbits;
1680         int nr_underway = 0;
1681         int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1682                         WRITE_SYNC : WRITE);
1683
1684         head = create_page_buffers(page, inode,
1685                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1686
1687         /*
1688          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1689          * here, and the (potentially unmapped) buffers may become dirty at
1690          * any time.  If a buffer becomes dirty here after we've inspected it
1691          * then we just miss that fact, and the page stays dirty.
1692          *
1693          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1694          * handle that here by just cleaning them.
1695          */
1696
1697         bh = head;
1698         blocksize = bh->b_size;
1699         bbits = block_size_bits(blocksize);
1700
1701         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1702         last_block = (i_size_read(inode) - 1) >> bbits;
1703
1704         /*
1705          * Get all the dirty buffers mapped to disk addresses and
1706          * handle any aliases from the underlying blockdev's mapping.
1707          */
1708         do {
1709                 if (block > last_block) {
1710                         /*
1711                          * mapped buffers outside i_size will occur, because
1712                          * this page can be outside i_size when there is a
1713                          * truncate in progress.
1714                          */
1715                         /*
1716                          * The buffer was zeroed by block_write_full_page()
1717                          */
1718                         clear_buffer_dirty(bh);
1719                         set_buffer_uptodate(bh);
1720                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1721                            buffer_dirty(bh)) {
1722                         WARN_ON(bh->b_size != blocksize);
1723                         err = get_block(inode, block, bh, 1);
1724                         if (err)
1725                                 goto recover;
1726                         clear_buffer_delay(bh);
1727                         if (buffer_new(bh)) {
1728                                 /* blockdev mappings never come here */
1729                                 clear_buffer_new(bh);
1730                                 unmap_underlying_metadata(bh->b_bdev,
1731                                                         bh->b_blocknr);
1732                         }
1733                 }
1734                 bh = bh->b_this_page;
1735                 block++;
1736         } while (bh != head);
1737
1738         do {
1739                 if (!buffer_mapped(bh))
1740                         continue;
1741                 /*
1742                  * If it's a fully non-blocking write attempt and we cannot
1743                  * lock the buffer then redirty the page.  Note that this can
1744                  * potentially cause a busy-wait loop from writeback threads
1745                  * and kswapd activity, but those code paths have their own
1746                  * higher-level throttling.
1747                  */
1748                 if (wbc->sync_mode != WB_SYNC_NONE) {
1749                         lock_buffer(bh);
1750                 } else if (!trylock_buffer(bh)) {
1751                         redirty_page_for_writepage(wbc, page);
1752                         continue;
1753                 }
1754                 if (test_clear_buffer_dirty(bh)) {
1755                         mark_buffer_async_write_endio(bh, handler);
1756                 } else {
1757                         unlock_buffer(bh);
1758                 }
1759         } while ((bh = bh->b_this_page) != head);
1760
1761         /*
1762          * The page and its buffers are protected by PageWriteback(), so we can
1763          * drop the bh refcounts early.
1764          */
1765         BUG_ON(PageWriteback(page));
1766         set_page_writeback(page);
1767
1768         do {
1769                 struct buffer_head *next = bh->b_this_page;
1770                 if (buffer_async_write(bh)) {
1771                         submit_bh(write_op, bh);
1772                         nr_underway++;
1773                 }
1774                 bh = next;
1775         } while (bh != head);
1776         unlock_page(page);
1777
1778         err = 0;
1779 done:
1780         if (nr_underway == 0) {
1781                 /*
1782                  * The page was marked dirty, but the buffers were
1783                  * clean.  Someone wrote them back by hand with
1784                  * ll_rw_block/submit_bh.  A rare case.
1785                  */
1786                 end_page_writeback(page);
1787
1788                 /*
1789                  * The page and buffer_heads can be released at any time from
1790                  * here on.
1791                  */
1792         }
1793         return err;
1794
1795 recover:
1796         /*
1797          * ENOSPC, or some other error.  We may already have added some
1798          * blocks to the file, so we need to write these out to avoid
1799          * exposing stale data.
1800          * The page is currently locked and not marked for writeback
1801          */
1802         bh = head;
1803         /* Recovery: lock and submit the mapped buffers */
1804         do {
1805                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1806                     !buffer_delay(bh)) {
1807                         lock_buffer(bh);
1808                         mark_buffer_async_write_endio(bh, handler);
1809                 } else {
1810                         /*
1811                          * The buffer may have been set dirty during
1812                          * attachment to a dirty page.
1813                          */
1814                         clear_buffer_dirty(bh);
1815                 }
1816         } while ((bh = bh->b_this_page) != head);
1817         SetPageError(page);
1818         BUG_ON(PageWriteback(page));
1819         mapping_set_error(page->mapping, err);
1820         set_page_writeback(page);
1821         do {
1822                 struct buffer_head *next = bh->b_this_page;
1823                 if (buffer_async_write(bh)) {
1824                         clear_buffer_dirty(bh);
1825                         submit_bh(write_op, bh);
1826                         nr_underway++;
1827                 }
1828                 bh = next;
1829         } while (bh != head);
1830         unlock_page(page);
1831         goto done;
1832 }
1833
1834 /*
1835  * If a page has any new buffers, zero them out here, and mark them uptodate
1836  * and dirty so they'll be written out (in order to prevent uninitialised
1837  * block data from leaking). And clear the new bit.
1838  */
1839 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1840 {
1841         unsigned int block_start, block_end;
1842         struct buffer_head *head, *bh;
1843
1844         BUG_ON(!PageLocked(page));
1845         if (!page_has_buffers(page))
1846                 return;
1847
1848         bh = head = page_buffers(page);
1849         block_start = 0;
1850         do {
1851                 block_end = block_start + bh->b_size;
1852
1853                 if (buffer_new(bh)) {
1854                         if (block_end > from && block_start < to) {
1855                                 if (!PageUptodate(page)) {
1856                                         unsigned start, size;
1857
1858                                         start = max(from, block_start);
1859                                         size = min(to, block_end) - start;
1860
1861                                         zero_user(page, start, size);
1862                                         set_buffer_uptodate(bh);
1863                                 }
1864
1865                                 clear_buffer_new(bh);
1866                                 mark_buffer_dirty(bh);
1867                         }
1868                 }
1869
1870                 block_start = block_end;
1871                 bh = bh->b_this_page;
1872         } while (bh != head);
1873 }
1874 EXPORT_SYMBOL(page_zero_new_buffers);
1875
1876 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1877                 get_block_t *get_block)
1878 {
1879         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1880         unsigned to = from + len;
1881         struct inode *inode = page->mapping->host;
1882         unsigned block_start, block_end;
1883         sector_t block;
1884         int err = 0;
1885         unsigned blocksize, bbits;
1886         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1887
1888         BUG_ON(!PageLocked(page));
1889         BUG_ON(from > PAGE_CACHE_SIZE);
1890         BUG_ON(to > PAGE_CACHE_SIZE);
1891         BUG_ON(from > to);
1892
1893         head = create_page_buffers(page, inode, 0);
1894         blocksize = head->b_size;
1895         bbits = block_size_bits(blocksize);
1896
1897         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1898
1899         for(bh = head, block_start = 0; bh != head || !block_start;
1900             block++, block_start=block_end, bh = bh->b_this_page) {
1901                 block_end = block_start + blocksize;
1902                 if (block_end <= from || block_start >= to) {
1903                         if (PageUptodate(page)) {
1904                                 if (!buffer_uptodate(bh))
1905                                         set_buffer_uptodate(bh);
1906                         }
1907                         continue;
1908                 }
1909                 if (buffer_new(bh))
1910                         clear_buffer_new(bh);
1911                 if (!buffer_mapped(bh)) {
1912                         WARN_ON(bh->b_size != blocksize);
1913                         err = get_block(inode, block, bh, 1);
1914                         if (err)
1915                                 break;
1916                         if (buffer_new(bh)) {
1917                                 unmap_underlying_metadata(bh->b_bdev,
1918                                                         bh->b_blocknr);
1919                                 if (PageUptodate(page)) {
1920                                         clear_buffer_new(bh);
1921                                         set_buffer_uptodate(bh);
1922                                         mark_buffer_dirty(bh);
1923                                         continue;
1924                                 }
1925                                 if (block_end > to || block_start < from)
1926                                         zero_user_segments(page,
1927                                                 to, block_end,
1928                                                 block_start, from);
1929                                 continue;
1930                         }
1931                 }
1932                 if (PageUptodate(page)) {
1933                         if (!buffer_uptodate(bh))
1934                                 set_buffer_uptodate(bh);
1935                         continue; 
1936                 }
1937                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1938                     !buffer_unwritten(bh) &&
1939                      (block_start < from || block_end > to)) {
1940                         ll_rw_block(READ, 1, &bh);
1941                         *wait_bh++=bh;
1942                 }
1943         }
1944         /*
1945          * If we issued read requests - let them complete.
1946          */
1947         while(wait_bh > wait) {
1948                 wait_on_buffer(*--wait_bh);
1949                 if (!buffer_uptodate(*wait_bh))
1950                         err = -EIO;
1951         }
1952         if (unlikely(err))
1953                 page_zero_new_buffers(page, from, to);
1954         return err;
1955 }
1956 EXPORT_SYMBOL(__block_write_begin);
1957
1958 static int __block_commit_write(struct inode *inode, struct page *page,
1959                 unsigned from, unsigned to)
1960 {
1961         unsigned block_start, block_end;
1962         int partial = 0;
1963         unsigned blocksize;
1964         struct buffer_head *bh, *head;
1965
1966         bh = head = page_buffers(page);
1967         blocksize = bh->b_size;
1968
1969         block_start = 0;
1970         do {
1971                 block_end = block_start + blocksize;
1972                 if (block_end <= from || block_start >= to) {
1973                         if (!buffer_uptodate(bh))
1974                                 partial = 1;
1975                 } else {
1976                         set_buffer_uptodate(bh);
1977                         mark_buffer_dirty(bh);
1978                 }
1979                 clear_buffer_new(bh);
1980
1981                 block_start = block_end;
1982                 bh = bh->b_this_page;
1983         } while (bh != head);
1984
1985         /*
1986          * If this is a partial write which happened to make all buffers
1987          * uptodate then we can optimize away a bogus readpage() for
1988          * the next read(). Here we 'discover' whether the page went
1989          * uptodate as a result of this (potentially partial) write.
1990          */
1991         if (!partial)
1992                 SetPageUptodate(page);
1993         return 0;
1994 }
1995
1996 /*
1997  * block_write_begin takes care of the basic task of block allocation and
1998  * bringing partial write blocks uptodate first.
1999  *
2000  * The filesystem needs to handle block truncation upon failure.
2001  */
2002 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2003                 unsigned flags, struct page **pagep, get_block_t *get_block)
2004 {
2005         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
2006         struct page *page;
2007         int status;
2008
2009         page = grab_cache_page_write_begin(mapping, index, flags);
2010         if (!page)
2011                 return -ENOMEM;
2012
2013         status = __block_write_begin(page, pos, len, get_block);
2014         if (unlikely(status)) {
2015                 unlock_page(page);
2016                 page_cache_release(page);
2017                 page = NULL;
2018         }
2019
2020         *pagep = page;
2021         return status;
2022 }
2023 EXPORT_SYMBOL(block_write_begin);
2024
2025 int block_write_end(struct file *file, struct address_space *mapping,
2026                         loff_t pos, unsigned len, unsigned copied,
2027                         struct page *page, void *fsdata)
2028 {
2029         struct inode *inode = mapping->host;
2030         unsigned start;
2031
2032         start = pos & (PAGE_CACHE_SIZE - 1);
2033
2034         if (unlikely(copied < len)) {
2035                 /*
2036                  * The buffers that were written will now be uptodate, so we
2037                  * don't have to worry about a readpage reading them and
2038                  * overwriting a partial write. However if we have encountered
2039                  * a short write and only partially written into a buffer, it
2040                  * will not be marked uptodate, so a readpage might come in and
2041                  * destroy our partial write.
2042                  *
2043                  * Do the simplest thing, and just treat any short write to a
2044                  * non uptodate page as a zero-length write, and force the
2045                  * caller to redo the whole thing.
2046                  */
2047                 if (!PageUptodate(page))
2048                         copied = 0;
2049
2050                 page_zero_new_buffers(page, start+copied, start+len);
2051         }
2052         flush_dcache_page(page);
2053
2054         /* This could be a short (even 0-length) commit */
2055         __block_commit_write(inode, page, start, start+copied);
2056
2057         return copied;
2058 }
2059 EXPORT_SYMBOL(block_write_end);
2060
2061 int generic_write_end(struct file *file, struct address_space *mapping,
2062                         loff_t pos, unsigned len, unsigned copied,
2063                         struct page *page, void *fsdata)
2064 {
2065         struct inode *inode = mapping->host;
2066         int i_size_changed = 0;
2067
2068         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2069
2070         /*
2071          * No need to use i_size_read() here, the i_size
2072          * cannot change under us because we hold i_mutex.
2073          *
2074          * But it's important to update i_size while still holding page lock:
2075          * page writeout could otherwise come in and zero beyond i_size.
2076          */
2077         if (pos+copied > inode->i_size) {
2078                 i_size_write(inode, pos+copied);
2079                 i_size_changed = 1;
2080         }
2081
2082         unlock_page(page);
2083         page_cache_release(page);
2084
2085         /*
2086          * Don't mark the inode dirty under page lock. First, it unnecessarily
2087          * makes the holding time of page lock longer. Second, it forces lock
2088          * ordering of page lock and transaction start for journaling
2089          * filesystems.
2090          */
2091         if (i_size_changed)
2092                 mark_inode_dirty(inode);
2093
2094         return copied;
2095 }
2096 EXPORT_SYMBOL(generic_write_end);
2097
2098 /*
2099  * block_is_partially_uptodate checks whether buffers within a page are
2100  * uptodate or not.
2101  *
2102  * Returns true if all buffers which correspond to a file portion
2103  * we want to read are uptodate.
2104  */
2105 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2106                                         unsigned long from)
2107 {
2108         unsigned block_start, block_end, blocksize;
2109         unsigned to;
2110         struct buffer_head *bh, *head;
2111         int ret = 1;
2112
2113         if (!page_has_buffers(page))
2114                 return 0;
2115
2116         head = page_buffers(page);
2117         blocksize = head->b_size;
2118         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2119         to = from + to;
2120         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2121                 return 0;
2122
2123         bh = head;
2124         block_start = 0;
2125         do {
2126                 block_end = block_start + blocksize;
2127                 if (block_end > from && block_start < to) {
2128                         if (!buffer_uptodate(bh)) {
2129                                 ret = 0;
2130                                 break;
2131                         }
2132                         if (block_end >= to)
2133                                 break;
2134                 }
2135                 block_start = block_end;
2136                 bh = bh->b_this_page;
2137         } while (bh != head);
2138
2139         return ret;
2140 }
2141 EXPORT_SYMBOL(block_is_partially_uptodate);
2142
2143 /*
2144  * Generic "read page" function for block devices that have the normal
2145  * get_block functionality. This is most of the block device filesystems.
2146  * Reads the page asynchronously --- the unlock_buffer() and
2147  * set/clear_buffer_uptodate() functions propagate buffer state into the
2148  * page struct once IO has completed.
2149  */
2150 int block_read_full_page(struct page *page, get_block_t *get_block)
2151 {
2152         struct inode *inode = page->mapping->host;
2153         sector_t iblock, lblock;
2154         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2155         unsigned int blocksize, bbits;
2156         int nr, i;
2157         int fully_mapped = 1;
2158
2159         head = create_page_buffers(page, inode, 0);
2160         blocksize = head->b_size;
2161         bbits = block_size_bits(blocksize);
2162
2163         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
2164         lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2165         bh = head;
2166         nr = 0;
2167         i = 0;
2168
2169         do {
2170                 if (buffer_uptodate(bh))
2171                         continue;
2172
2173                 if (!buffer_mapped(bh)) {
2174                         int err = 0;
2175
2176                         fully_mapped = 0;
2177                         if (iblock < lblock) {
2178                                 WARN_ON(bh->b_size != blocksize);
2179                                 err = get_block(inode, iblock, bh, 0);
2180                                 if (err)
2181                                         SetPageError(page);
2182                         }
2183                         if (!buffer_mapped(bh)) {
2184                                 zero_user(page, i * blocksize, blocksize);
2185                                 if (!err)
2186                                         set_buffer_uptodate(bh);
2187                                 continue;
2188                         }
2189                         /*
2190                          * get_block() might have updated the buffer
2191                          * synchronously
2192                          */
2193                         if (buffer_uptodate(bh))
2194                                 continue;
2195                 }
2196                 arr[nr++] = bh;
2197         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2198
2199         if (fully_mapped)
2200                 SetPageMappedToDisk(page);
2201
2202         if (!nr) {
2203                 /*
2204                  * All buffers are uptodate - we can set the page uptodate
2205                  * as well. But not if get_block() returned an error.
2206                  */
2207                 if (!PageError(page))
2208                         SetPageUptodate(page);
2209                 unlock_page(page);
2210                 return 0;
2211         }
2212
2213         /* Stage two: lock the buffers */
2214         for (i = 0; i < nr; i++) {
2215                 bh = arr[i];
2216                 lock_buffer(bh);
2217                 mark_buffer_async_read(bh);
2218         }
2219
2220         /*
2221          * Stage 3: start the IO.  Check for uptodateness
2222          * inside the buffer lock in case another process reading
2223          * the underlying blockdev brought it uptodate (the sct fix).
2224          */
2225         for (i = 0; i < nr; i++) {
2226                 bh = arr[i];
2227                 if (buffer_uptodate(bh))
2228                         end_buffer_async_read(bh, 1);
2229                 else
2230                         submit_bh(READ, bh);
2231         }
2232         return 0;
2233 }
2234 EXPORT_SYMBOL(block_read_full_page);
2235
2236 /* utility function for filesystems that need to do work on expanding
2237  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2238  * deal with the hole.  
2239  */
2240 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2241 {
2242         struct address_space *mapping = inode->i_mapping;
2243         struct page *page;
2244         void *fsdata;
2245         int err;
2246
2247         err = inode_newsize_ok(inode, size);
2248         if (err)
2249                 goto out;
2250
2251         err = pagecache_write_begin(NULL, mapping, size, 0,
2252                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2253                                 &page, &fsdata);
2254         if (err)
2255                 goto out;
2256
2257         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2258         BUG_ON(err > 0);
2259
2260 out:
2261         return err;
2262 }
2263 EXPORT_SYMBOL(generic_cont_expand_simple);
2264
2265 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2266                             loff_t pos, loff_t *bytes)
2267 {
2268         struct inode *inode = mapping->host;
2269         unsigned blocksize = 1 << inode->i_blkbits;
2270         struct page *page;
2271         void *fsdata;
2272         pgoff_t index, curidx;
2273         loff_t curpos;
2274         unsigned zerofrom, offset, len;
2275         int err = 0;
2276
2277         index = pos >> PAGE_CACHE_SHIFT;
2278         offset = pos & ~PAGE_CACHE_MASK;
2279
2280         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2281                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2282                 if (zerofrom & (blocksize-1)) {
2283                         *bytes |= (blocksize-1);
2284                         (*bytes)++;
2285                 }
2286                 len = PAGE_CACHE_SIZE - zerofrom;
2287
2288                 err = pagecache_write_begin(file, mapping, curpos, len,
2289                                                 AOP_FLAG_UNINTERRUPTIBLE,
2290                                                 &page, &fsdata);
2291                 if (err)
2292                         goto out;
2293                 zero_user(page, zerofrom, len);
2294                 err = pagecache_write_end(file, mapping, curpos, len, len,
2295                                                 page, fsdata);
2296                 if (err < 0)
2297                         goto out;
2298                 BUG_ON(err != len);
2299                 err = 0;
2300
2301                 balance_dirty_pages_ratelimited(mapping);
2302         }
2303
2304         /* page covers the boundary, find the boundary offset */
2305         if (index == curidx) {
2306                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2307                 /* if we will expand the thing last block will be filled */
2308                 if (offset <= zerofrom) {
2309                         goto out;
2310                 }
2311                 if (zerofrom & (blocksize-1)) {
2312                         *bytes |= (blocksize-1);
2313                         (*bytes)++;
2314                 }
2315                 len = offset - zerofrom;
2316
2317                 err = pagecache_write_begin(file, mapping, curpos, len,
2318                                                 AOP_FLAG_UNINTERRUPTIBLE,
2319                                                 &page, &fsdata);
2320                 if (err)
2321                         goto out;
2322                 zero_user(page, zerofrom, len);
2323                 err = pagecache_write_end(file, mapping, curpos, len, len,
2324                                                 page, fsdata);
2325                 if (err < 0)
2326                         goto out;
2327                 BUG_ON(err != len);
2328                 err = 0;
2329         }
2330 out:
2331         return err;
2332 }
2333
2334 /*
2335  * For moronic filesystems that do not allow holes in file.
2336  * We may have to extend the file.
2337  */
2338 int cont_write_begin(struct file *file, struct address_space *mapping,
2339                         loff_t pos, unsigned len, unsigned flags,
2340                         struct page **pagep, void **fsdata,
2341                         get_block_t *get_block, loff_t *bytes)
2342 {
2343         struct inode *inode = mapping->host;
2344         unsigned blocksize = 1 << inode->i_blkbits;
2345         unsigned zerofrom;
2346         int err;
2347
2348         err = cont_expand_zero(file, mapping, pos, bytes);
2349         if (err)
2350                 return err;
2351
2352         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2353         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2354                 *bytes |= (blocksize-1);
2355                 (*bytes)++;
2356         }
2357
2358         return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2359 }
2360 EXPORT_SYMBOL(cont_write_begin);
2361
2362 int block_commit_write(struct page *page, unsigned from, unsigned to)
2363 {
2364         struct inode *inode = page->mapping->host;
2365         __block_commit_write(inode,page,from,to);
2366         return 0;
2367 }
2368 EXPORT_SYMBOL(block_commit_write);
2369
2370 /*
2371  * block_page_mkwrite() is not allowed to change the file size as it gets
2372  * called from a page fault handler when a page is first dirtied. Hence we must
2373  * be careful to check for EOF conditions here. We set the page up correctly
2374  * for a written page which means we get ENOSPC checking when writing into
2375  * holes and correct delalloc and unwritten extent mapping on filesystems that
2376  * support these features.
2377  *
2378  * We are not allowed to take the i_mutex here so we have to play games to
2379  * protect against truncate races as the page could now be beyond EOF.  Because
2380  * truncate writes the inode size before removing pages, once we have the
2381  * page lock we can determine safely if the page is beyond EOF. If it is not
2382  * beyond EOF, then the page is guaranteed safe against truncation until we
2383  * unlock the page.
2384  *
2385  * Direct callers of this function should protect against filesystem freezing
2386  * using sb_start_write() - sb_end_write() functions.
2387  */
2388 int __block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2389                          get_block_t get_block)
2390 {
2391         struct page *page = vmf->page;
2392         struct inode *inode = file_inode(vma->vm_file);
2393         unsigned long end;
2394         loff_t size;
2395         int ret;
2396
2397         lock_page(page);
2398         size = i_size_read(inode);
2399         if ((page->mapping != inode->i_mapping) ||
2400             (page_offset(page) > size)) {
2401                 /* We overload EFAULT to mean page got truncated */
2402                 ret = -EFAULT;
2403                 goto out_unlock;
2404         }
2405
2406         /* page is wholly or partially inside EOF */
2407         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2408                 end = size & ~PAGE_CACHE_MASK;
2409         else
2410                 end = PAGE_CACHE_SIZE;
2411
2412         ret = __block_write_begin(page, 0, end, get_block);
2413         if (!ret)
2414                 ret = block_commit_write(page, 0, end);
2415
2416         if (unlikely(ret < 0))
2417                 goto out_unlock;
2418         set_page_dirty(page);
2419         wait_for_stable_page(page);
2420         return 0;
2421 out_unlock:
2422         unlock_page(page);
2423         return ret;
2424 }
2425 EXPORT_SYMBOL(__block_page_mkwrite);
2426
2427 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2428                    get_block_t get_block)
2429 {
2430         int ret;
2431         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
2432
2433         sb_start_pagefault(sb);
2434
2435         /*
2436          * Update file times before taking page lock. We may end up failing the
2437          * fault so this update may be superfluous but who really cares...
2438          */
2439         file_update_time(vma->vm_file);
2440
2441         ret = __block_page_mkwrite(vma, vmf, get_block);
2442         sb_end_pagefault(sb);
2443         return block_page_mkwrite_return(ret);
2444 }
2445 EXPORT_SYMBOL(block_page_mkwrite);
2446
2447 /*
2448  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2449  * immediately, while under the page lock.  So it needs a special end_io
2450  * handler which does not touch the bh after unlocking it.
2451  */
2452 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2453 {
2454         __end_buffer_read_notouch(bh, uptodate);
2455 }
2456
2457 /*
2458  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2459  * the page (converting it to circular linked list and taking care of page
2460  * dirty races).
2461  */
2462 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2463 {
2464         struct buffer_head *bh;
2465
2466         BUG_ON(!PageLocked(page));
2467
2468         spin_lock(&page->mapping->private_lock);
2469         bh = head;
2470         do {
2471                 if (PageDirty(page))
2472                         set_buffer_dirty(bh);
2473                 if (!bh->b_this_page)
2474                         bh->b_this_page = head;
2475                 bh = bh->b_this_page;
2476         } while (bh != head);
2477         attach_page_buffers(page, head);
2478         spin_unlock(&page->mapping->private_lock);
2479 }
2480
2481 /*
2482  * On entry, the page is fully not uptodate.
2483  * On exit the page is fully uptodate in the areas outside (from,to)
2484  * The filesystem needs to handle block truncation upon failure.
2485  */
2486 int nobh_write_begin(struct address_space *mapping,
2487                         loff_t pos, unsigned len, unsigned flags,
2488                         struct page **pagep, void **fsdata,
2489                         get_block_t *get_block)
2490 {
2491         struct inode *inode = mapping->host;
2492         const unsigned blkbits = inode->i_blkbits;
2493         const unsigned blocksize = 1 << blkbits;
2494         struct buffer_head *head, *bh;
2495         struct page *page;
2496         pgoff_t index;
2497         unsigned from, to;
2498         unsigned block_in_page;
2499         unsigned block_start, block_end;
2500         sector_t block_in_file;
2501         int nr_reads = 0;
2502         int ret = 0;
2503         int is_mapped_to_disk = 1;
2504
2505         index = pos >> PAGE_CACHE_SHIFT;
2506         from = pos & (PAGE_CACHE_SIZE - 1);
2507         to = from + len;
2508
2509         page = grab_cache_page_write_begin(mapping, index, flags);
2510         if (!page)
2511                 return -ENOMEM;
2512         *pagep = page;
2513         *fsdata = NULL;
2514
2515         if (page_has_buffers(page)) {
2516                 ret = __block_write_begin(page, pos, len, get_block);
2517                 if (unlikely(ret))
2518                         goto out_release;
2519                 return ret;
2520         }
2521
2522         if (PageMappedToDisk(page))
2523                 return 0;
2524
2525         /*
2526          * Allocate buffers so that we can keep track of state, and potentially
2527          * attach them to the page if an error occurs. In the common case of
2528          * no error, they will just be freed again without ever being attached
2529          * to the page (which is all OK, because we're under the page lock).
2530          *
2531          * Be careful: the buffer linked list is a NULL terminated one, rather
2532          * than the circular one we're used to.
2533          */
2534         head = alloc_page_buffers(page, blocksize, 0);
2535         if (!head) {
2536                 ret = -ENOMEM;
2537                 goto out_release;
2538         }
2539
2540         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2541
2542         /*
2543          * We loop across all blocks in the page, whether or not they are
2544          * part of the affected region.  This is so we can discover if the
2545          * page is fully mapped-to-disk.
2546          */
2547         for (block_start = 0, block_in_page = 0, bh = head;
2548                   block_start < PAGE_CACHE_SIZE;
2549                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2550                 int create;
2551
2552                 block_end = block_start + blocksize;
2553                 bh->b_state = 0;
2554                 create = 1;
2555                 if (block_start >= to)
2556                         create = 0;
2557                 ret = get_block(inode, block_in_file + block_in_page,
2558                                         bh, create);
2559                 if (ret)
2560                         goto failed;
2561                 if (!buffer_mapped(bh))
2562                         is_mapped_to_disk = 0;
2563                 if (buffer_new(bh))
2564                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2565                 if (PageUptodate(page)) {
2566                         set_buffer_uptodate(bh);
2567                         continue;
2568                 }
2569                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2570                         zero_user_segments(page, block_start, from,
2571                                                         to, block_end);
2572                         continue;
2573                 }
2574                 if (buffer_uptodate(bh))
2575                         continue;       /* reiserfs does this */
2576                 if (block_start < from || block_end > to) {
2577                         lock_buffer(bh);
2578                         bh->b_end_io = end_buffer_read_nobh;
2579                         submit_bh(READ, bh);
2580                         nr_reads++;
2581                 }
2582         }
2583
2584         if (nr_reads) {
2585                 /*
2586                  * The page is locked, so these buffers are protected from
2587                  * any VM or truncate activity.  Hence we don't need to care
2588                  * for the buffer_head refcounts.
2589                  */
2590                 for (bh = head; bh; bh = bh->b_this_page) {
2591                         wait_on_buffer(bh);
2592                         if (!buffer_uptodate(bh))
2593                                 ret = -EIO;
2594                 }
2595                 if (ret)
2596                         goto failed;
2597         }
2598
2599         if (is_mapped_to_disk)
2600                 SetPageMappedToDisk(page);
2601
2602         *fsdata = head; /* to be released by nobh_write_end */
2603
2604         return 0;
2605
2606 failed:
2607         BUG_ON(!ret);
2608         /*
2609          * Error recovery is a bit difficult. We need to zero out blocks that
2610          * were newly allocated, and dirty them to ensure they get written out.
2611          * Buffers need to be attached to the page at this point, otherwise
2612          * the handling of potential IO errors during writeout would be hard
2613          * (could try doing synchronous writeout, but what if that fails too?)
2614          */
2615         attach_nobh_buffers(page, head);
2616         page_zero_new_buffers(page, from, to);
2617
2618 out_release:
2619         unlock_page(page);
2620         page_cache_release(page);
2621         *pagep = NULL;
2622
2623         return ret;
2624 }
2625 EXPORT_SYMBOL(nobh_write_begin);
2626
2627 int nobh_write_end(struct file *file, struct address_space *mapping,
2628                         loff_t pos, unsigned len, unsigned copied,
2629                         struct page *page, void *fsdata)
2630 {
2631         struct inode *inode = page->mapping->host;
2632         struct buffer_head *head = fsdata;
2633         struct buffer_head *bh;
2634         BUG_ON(fsdata != NULL && page_has_buffers(page));
2635
2636         if (unlikely(copied < len) && head)
2637                 attach_nobh_buffers(page, head);
2638         if (page_has_buffers(page))
2639                 return generic_write_end(file, mapping, pos, len,
2640                                         copied, page, fsdata);
2641
2642         SetPageUptodate(page);
2643         set_page_dirty(page);
2644         if (pos+copied > inode->i_size) {
2645                 i_size_write(inode, pos+copied);
2646                 mark_inode_dirty(inode);
2647         }
2648
2649         unlock_page(page);
2650         page_cache_release(page);
2651
2652         while (head) {
2653                 bh = head;
2654                 head = head->b_this_page;
2655                 free_buffer_head(bh);
2656         }
2657
2658         return copied;
2659 }
2660 EXPORT_SYMBOL(nobh_write_end);
2661
2662 /*
2663  * nobh_writepage() - based on block_full_write_page() except
2664  * that it tries to operate without attaching bufferheads to
2665  * the page.
2666  */
2667 int nobh_writepage(struct page *page, get_block_t *get_block,
2668                         struct writeback_control *wbc)
2669 {
2670         struct inode * const inode = page->mapping->host;
2671         loff_t i_size = i_size_read(inode);
2672         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2673         unsigned offset;
2674         int ret;
2675
2676         /* Is the page fully inside i_size? */
2677         if (page->index < end_index)
2678                 goto out;
2679
2680         /* Is the page fully outside i_size? (truncate in progress) */
2681         offset = i_size & (PAGE_CACHE_SIZE-1);
2682         if (page->index >= end_index+1 || !offset) {
2683                 /*
2684                  * The page may have dirty, unmapped buffers.  For example,
2685                  * they may have been added in ext3_writepage().  Make them
2686                  * freeable here, so the page does not leak.
2687                  */
2688 #if 0
2689                 /* Not really sure about this  - do we need this ? */
2690                 if (page->mapping->a_ops->invalidatepage)
2691                         page->mapping->a_ops->invalidatepage(page, offset);
2692 #endif
2693                 unlock_page(page);
2694                 return 0; /* don't care */
2695         }
2696
2697         /*
2698          * The page straddles i_size.  It must be zeroed out on each and every
2699          * writepage invocation because it may be mmapped.  "A file is mapped
2700          * in multiples of the page size.  For a file that is not a multiple of
2701          * the  page size, the remaining memory is zeroed when mapped, and
2702          * writes to that region are not written out to the file."
2703          */
2704         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2705 out:
2706         ret = mpage_writepage(page, get_block, wbc);
2707         if (ret == -EAGAIN)
2708                 ret = __block_write_full_page(inode, page, get_block, wbc,
2709                                               end_buffer_async_write);
2710         return ret;
2711 }
2712 EXPORT_SYMBOL(nobh_writepage);
2713
2714 int nobh_truncate_page(struct address_space *mapping,
2715                         loff_t from, get_block_t *get_block)
2716 {
2717         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2718         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2719         unsigned blocksize;
2720         sector_t iblock;
2721         unsigned length, pos;
2722         struct inode *inode = mapping->host;
2723         struct page *page;
2724         struct buffer_head map_bh;
2725         int err;
2726
2727         blocksize = 1 << inode->i_blkbits;
2728         length = offset & (blocksize - 1);
2729
2730         /* Block boundary? Nothing to do */
2731         if (!length)
2732                 return 0;
2733
2734         length = blocksize - length;
2735         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2736
2737         page = grab_cache_page(mapping, index);
2738         err = -ENOMEM;
2739         if (!page)
2740                 goto out;
2741
2742         if (page_has_buffers(page)) {
2743 has_buffers:
2744                 unlock_page(page);
2745                 page_cache_release(page);
2746                 return block_truncate_page(mapping, from, get_block);
2747         }
2748
2749         /* Find the buffer that contains "offset" */
2750         pos = blocksize;
2751         while (offset >= pos) {
2752                 iblock++;
2753                 pos += blocksize;
2754         }
2755
2756         map_bh.b_size = blocksize;
2757         map_bh.b_state = 0;
2758         err = get_block(inode, iblock, &map_bh, 0);
2759         if (err)
2760                 goto unlock;
2761         /* unmapped? It's a hole - nothing to do */
2762         if (!buffer_mapped(&map_bh))
2763                 goto unlock;
2764
2765         /* Ok, it's mapped. Make sure it's up-to-date */
2766         if (!PageUptodate(page)) {
2767                 err = mapping->a_ops->readpage(NULL, page);
2768                 if (err) {
2769                         page_cache_release(page);
2770                         goto out;
2771                 }
2772                 lock_page(page);
2773                 if (!PageUptodate(page)) {
2774                         err = -EIO;
2775                         goto unlock;
2776                 }
2777                 if (page_has_buffers(page))
2778                         goto has_buffers;
2779         }
2780         zero_user(page, offset, length);
2781         set_page_dirty(page);
2782         err = 0;
2783
2784 unlock:
2785         unlock_page(page);
2786         page_cache_release(page);
2787 out:
2788         return err;
2789 }
2790 EXPORT_SYMBOL(nobh_truncate_page);
2791
2792 int block_truncate_page(struct address_space *mapping,
2793                         loff_t from, get_block_t *get_block)
2794 {
2795         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2796         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2797         unsigned blocksize;
2798         sector_t iblock;
2799         unsigned length, pos;
2800         struct inode *inode = mapping->host;
2801         struct page *page;
2802         struct buffer_head *bh;
2803         int err;
2804
2805         blocksize = 1 << inode->i_blkbits;
2806         length = offset & (blocksize - 1);
2807
2808         /* Block boundary? Nothing to do */
2809         if (!length)
2810                 return 0;
2811
2812         length = blocksize - length;
2813         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2814         
2815         page = grab_cache_page(mapping, index);
2816         err = -ENOMEM;
2817         if (!page)
2818                 goto out;
2819
2820         if (!page_has_buffers(page))
2821                 create_empty_buffers(page, blocksize, 0);
2822
2823         /* Find the buffer that contains "offset" */
2824         bh = page_buffers(page);
2825         pos = blocksize;
2826         while (offset >= pos) {
2827                 bh = bh->b_this_page;
2828                 iblock++;
2829                 pos += blocksize;
2830         }
2831
2832         err = 0;
2833         if (!buffer_mapped(bh)) {
2834                 WARN_ON(bh->b_size != blocksize);
2835                 err = get_block(inode, iblock, bh, 0);
2836                 if (err)
2837                         goto unlock;
2838                 /* unmapped? It's a hole - nothing to do */
2839                 if (!buffer_mapped(bh))
2840                         goto unlock;
2841         }
2842
2843         /* Ok, it's mapped. Make sure it's up-to-date */
2844         if (PageUptodate(page))
2845                 set_buffer_uptodate(bh);
2846
2847         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2848                 err = -EIO;
2849                 ll_rw_block(READ, 1, &bh);
2850                 wait_on_buffer(bh);
2851                 /* Uhhuh. Read error. Complain and punt. */
2852                 if (!buffer_uptodate(bh))
2853                         goto unlock;
2854         }
2855
2856         zero_user(page, offset, length);
2857         mark_buffer_dirty(bh);
2858         err = 0;
2859
2860 unlock:
2861         unlock_page(page);
2862         page_cache_release(page);
2863 out:
2864         return err;
2865 }
2866 EXPORT_SYMBOL(block_truncate_page);
2867
2868 /*
2869  * The generic ->writepage function for buffer-backed address_spaces
2870  * this form passes in the end_io handler used to finish the IO.
2871  */
2872 int block_write_full_page_endio(struct page *page, get_block_t *get_block,
2873                         struct writeback_control *wbc, bh_end_io_t *handler)
2874 {
2875         struct inode * const inode = page->mapping->host;
2876         loff_t i_size = i_size_read(inode);
2877         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2878         unsigned offset;
2879
2880         /* Is the page fully inside i_size? */
2881         if (page->index < end_index)
2882                 return __block_write_full_page(inode, page, get_block, wbc,
2883                                                handler);
2884
2885         /* Is the page fully outside i_size? (truncate in progress) */
2886         offset = i_size & (PAGE_CACHE_SIZE-1);
2887         if (page->index >= end_index+1 || !offset) {
2888                 /*
2889                  * The page may have dirty, unmapped buffers.  For example,
2890                  * they may have been added in ext3_writepage().  Make them
2891                  * freeable here, so the page does not leak.
2892                  */
2893                 do_invalidatepage(page, 0, PAGE_CACHE_SIZE);
2894                 unlock_page(page);
2895                 return 0; /* don't care */
2896         }
2897
2898         /*
2899          * The page straddles i_size.  It must be zeroed out on each and every
2900          * writepage invocation because it may be mmapped.  "A file is mapped
2901          * in multiples of the page size.  For a file that is not a multiple of
2902          * the  page size, the remaining memory is zeroed when mapped, and
2903          * writes to that region are not written out to the file."
2904          */
2905         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2906         return __block_write_full_page(inode, page, get_block, wbc, handler);
2907 }
2908 EXPORT_SYMBOL(block_write_full_page_endio);
2909
2910 /*
2911  * The generic ->writepage function for buffer-backed address_spaces
2912  */
2913 int block_write_full_page(struct page *page, get_block_t *get_block,
2914                         struct writeback_control *wbc)
2915 {
2916         return block_write_full_page_endio(page, get_block, wbc,
2917                                            end_buffer_async_write);
2918 }
2919 EXPORT_SYMBOL(block_write_full_page);
2920
2921 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2922                             get_block_t *get_block)
2923 {
2924         struct buffer_head tmp;
2925         struct inode *inode = mapping->host;
2926         tmp.b_state = 0;
2927         tmp.b_blocknr = 0;
2928         tmp.b_size = 1 << inode->i_blkbits;
2929         get_block(inode, block, &tmp, 0);
2930         return tmp.b_blocknr;
2931 }
2932 EXPORT_SYMBOL(generic_block_bmap);
2933
2934 static void end_bio_bh_io_sync(struct bio *bio, int err)
2935 {
2936         struct buffer_head *bh = bio->bi_private;
2937
2938         if (err == -EOPNOTSUPP) {
2939                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2940         }
2941
2942         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2943                 set_bit(BH_Quiet, &bh->b_state);
2944
2945         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2946         bio_put(bio);
2947 }
2948
2949 /*
2950  * This allows us to do IO even on the odd last sectors
2951  * of a device, even if the bh block size is some multiple
2952  * of the physical sector size.
2953  *
2954  * We'll just truncate the bio to the size of the device,
2955  * and clear the end of the buffer head manually.
2956  *
2957  * Truly out-of-range accesses will turn into actual IO
2958  * errors, this only handles the "we need to be able to
2959  * do IO at the final sector" case.
2960  */
2961 static void guard_bh_eod(int rw, struct bio *bio, struct buffer_head *bh)
2962 {
2963         sector_t maxsector;
2964         unsigned bytes;
2965
2966         maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
2967         if (!maxsector)
2968                 return;
2969
2970         /*
2971          * If the *whole* IO is past the end of the device,
2972          * let it through, and the IO layer will turn it into
2973          * an EIO.
2974          */
2975         if (unlikely(bio->bi_sector >= maxsector))
2976                 return;
2977
2978         maxsector -= bio->bi_sector;
2979         bytes = bio->bi_size;
2980         if (likely((bytes >> 9) <= maxsector))
2981                 return;
2982
2983         /* Uhhuh. We've got a bh that straddles the device size! */
2984         bytes = maxsector << 9;
2985
2986         /* Truncate the bio.. */
2987         bio->bi_size = bytes;
2988         bio->bi_io_vec[0].bv_len = bytes;
2989
2990         /* ..and clear the end of the buffer for reads */
2991         if ((rw & RW_MASK) == READ) {
2992                 void *kaddr = kmap_atomic(bh->b_page);
2993                 memset(kaddr + bh_offset(bh) + bytes, 0, bh->b_size - bytes);
2994                 kunmap_atomic(kaddr);
2995                 flush_dcache_page(bh->b_page);
2996         }
2997 }
2998
2999 int _submit_bh(int rw, struct buffer_head *bh, unsigned long bio_flags)
3000 {
3001         struct bio *bio;
3002         int ret = 0;
3003
3004         BUG_ON(!buffer_locked(bh));
3005         BUG_ON(!buffer_mapped(bh));
3006         BUG_ON(!bh->b_end_io);
3007         BUG_ON(buffer_delay(bh));
3008         BUG_ON(buffer_unwritten(bh));
3009
3010         /*
3011          * Only clear out a write error when rewriting
3012          */
3013         if (test_set_buffer_req(bh) && (rw & WRITE))
3014                 clear_buffer_write_io_error(bh);
3015
3016         /*
3017          * from here on down, it's all bio -- do the initial mapping,
3018          * submit_bio -> generic_make_request may further map this bio around
3019          */
3020         bio = bio_alloc(GFP_NOIO, 1);
3021
3022         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3023         bio->bi_bdev = bh->b_bdev;
3024         bio->bi_io_vec[0].bv_page = bh->b_page;
3025         bio->bi_io_vec[0].bv_len = bh->b_size;
3026         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
3027
3028         bio->bi_vcnt = 1;
3029         bio->bi_size = bh->b_size;
3030
3031         bio->bi_end_io = end_bio_bh_io_sync;
3032         bio->bi_private = bh;
3033         bio->bi_flags |= bio_flags;
3034
3035         /* Take care of bh's that straddle the end of the device */
3036         guard_bh_eod(rw, bio, bh);
3037
3038         if (buffer_meta(bh))
3039                 rw |= REQ_META;
3040         if (buffer_prio(bh))
3041                 rw |= REQ_PRIO;
3042
3043         bio_get(bio);
3044         submit_bio(rw, bio);
3045
3046         if (bio_flagged(bio, BIO_EOPNOTSUPP))
3047                 ret = -EOPNOTSUPP;
3048
3049         bio_put(bio);
3050         return ret;
3051 }
3052 EXPORT_SYMBOL_GPL(_submit_bh);
3053
3054 int submit_bh(int rw, struct buffer_head *bh)
3055 {
3056         return _submit_bh(rw, bh, 0);
3057 }
3058 EXPORT_SYMBOL(submit_bh);
3059
3060 /**
3061  * ll_rw_block: low-level access to block devices (DEPRECATED)
3062  * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
3063  * @nr: number of &struct buffer_heads in the array
3064  * @bhs: array of pointers to &struct buffer_head
3065  *
3066  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3067  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
3068  * %READA option is described in the documentation for generic_make_request()
3069  * which ll_rw_block() calls.
3070  *
3071  * This function drops any buffer that it cannot get a lock on (with the
3072  * BH_Lock state bit), any buffer that appears to be clean when doing a write
3073  * request, and any buffer that appears to be up-to-date when doing read
3074  * request.  Further it marks as clean buffers that are processed for
3075  * writing (the buffer cache won't assume that they are actually clean
3076  * until the buffer gets unlocked).
3077  *
3078  * ll_rw_block sets b_end_io to simple completion handler that marks
3079  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
3080  * any waiters. 
3081  *
3082  * All of the buffers must be for the same device, and must also be a
3083  * multiple of the current approved size for the device.
3084  */
3085 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
3086 {
3087         int i;
3088
3089         for (i = 0; i < nr; i++) {
3090                 struct buffer_head *bh = bhs[i];
3091
3092                 if (!trylock_buffer(bh))
3093                         continue;
3094                 if (rw == WRITE) {
3095                         if (test_clear_buffer_dirty(bh)) {
3096                                 bh->b_end_io = end_buffer_write_sync;
3097                                 get_bh(bh);
3098                                 submit_bh(WRITE, bh);
3099                                 continue;
3100                         }
3101                 } else {
3102                         if (!buffer_uptodate(bh)) {
3103                                 bh->b_end_io = end_buffer_read_sync;
3104                                 get_bh(bh);
3105                                 submit_bh(rw, bh);
3106                                 continue;
3107                         }
3108                 }
3109                 unlock_buffer(bh);
3110         }
3111 }
3112 EXPORT_SYMBOL(ll_rw_block);
3113
3114 void write_dirty_buffer(struct buffer_head *bh, int rw)
3115 {
3116         lock_buffer(bh);
3117         if (!test_clear_buffer_dirty(bh)) {
3118                 unlock_buffer(bh);
3119                 return;
3120         }
3121         bh->b_end_io = end_buffer_write_sync;
3122         get_bh(bh);
3123         submit_bh(rw, bh);
3124 }
3125 EXPORT_SYMBOL(write_dirty_buffer);
3126
3127 /*
3128  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3129  * and then start new I/O and then wait upon it.  The caller must have a ref on
3130  * the buffer_head.
3131  */
3132 int __sync_dirty_buffer(struct buffer_head *bh, int rw)
3133 {
3134         int ret = 0;
3135
3136         WARN_ON(atomic_read(&bh->b_count) < 1);
3137         lock_buffer(bh);
3138         if (test_clear_buffer_dirty(bh)) {
3139                 get_bh(bh);
3140                 bh->b_end_io = end_buffer_write_sync;
3141                 ret = submit_bh(rw, bh);
3142                 wait_on_buffer(bh);
3143                 if (!ret && !buffer_uptodate(bh))
3144                         ret = -EIO;
3145         } else {
3146                 unlock_buffer(bh);
3147         }
3148         return ret;
3149 }
3150 EXPORT_SYMBOL(__sync_dirty_buffer);
3151
3152 int sync_dirty_buffer(struct buffer_head *bh)
3153 {
3154         return __sync_dirty_buffer(bh, WRITE_SYNC);
3155 }
3156 EXPORT_SYMBOL(sync_dirty_buffer);
3157
3158 /*
3159  * try_to_free_buffers() checks if all the buffers on this particular page
3160  * are unused, and releases them if so.
3161  *
3162  * Exclusion against try_to_free_buffers may be obtained by either
3163  * locking the page or by holding its mapping's private_lock.
3164  *
3165  * If the page is dirty but all the buffers are clean then we need to
3166  * be sure to mark the page clean as well.  This is because the page
3167  * may be against a block device, and a later reattachment of buffers
3168  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3169  * filesystem data on the same device.
3170  *
3171  * The same applies to regular filesystem pages: if all the buffers are
3172  * clean then we set the page clean and proceed.  To do that, we require
3173  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3174  * private_lock.
3175  *
3176  * try_to_free_buffers() is non-blocking.
3177  */
3178 static inline int buffer_busy(struct buffer_head *bh)
3179 {
3180         return atomic_read(&bh->b_count) |
3181                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3182 }
3183
3184 static int
3185 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3186 {
3187         struct buffer_head *head = page_buffers(page);
3188         struct buffer_head *bh;
3189
3190         bh = head;
3191         do {
3192                 if (buffer_write_io_error(bh) && page->mapping)
3193                         set_bit(AS_EIO, &page->mapping->flags);
3194                 if (buffer_busy(bh))
3195                         goto failed;
3196                 bh = bh->b_this_page;
3197         } while (bh != head);
3198
3199         do {
3200                 struct buffer_head *next = bh->b_this_page;
3201
3202                 if (bh->b_assoc_map)
3203                         __remove_assoc_queue(bh);
3204                 bh = next;
3205         } while (bh != head);
3206         *buffers_to_free = head;
3207         __clear_page_buffers(page);
3208         return 1;
3209 failed:
3210         return 0;
3211 }
3212
3213 int try_to_free_buffers(struct page *page)
3214 {
3215         struct address_space * const mapping = page->mapping;
3216         struct buffer_head *buffers_to_free = NULL;
3217         int ret = 0;
3218
3219         BUG_ON(!PageLocked(page));
3220         if (PageWriteback(page))
3221                 return 0;
3222
3223         if (mapping == NULL) {          /* can this still happen? */
3224                 ret = drop_buffers(page, &buffers_to_free);
3225                 goto out;
3226         }
3227
3228         spin_lock(&mapping->private_lock);
3229         ret = drop_buffers(page, &buffers_to_free);
3230
3231         /*
3232          * If the filesystem writes its buffers by hand (eg ext3)
3233          * then we can have clean buffers against a dirty page.  We
3234          * clean the page here; otherwise the VM will never notice
3235          * that the filesystem did any IO at all.
3236          *
3237          * Also, during truncate, discard_buffer will have marked all
3238          * the page's buffers clean.  We discover that here and clean
3239          * the page also.
3240          *
3241          * private_lock must be held over this entire operation in order
3242          * to synchronise against __set_page_dirty_buffers and prevent the
3243          * dirty bit from being lost.
3244          */
3245         if (ret)
3246                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3247         spin_unlock(&mapping->private_lock);
3248 out:
3249         if (buffers_to_free) {
3250                 struct buffer_head *bh = buffers_to_free;
3251
3252                 do {
3253                         struct buffer_head *next = bh->b_this_page;
3254                         free_buffer_head(bh);
3255                         bh = next;
3256                 } while (bh != buffers_to_free);
3257         }
3258         return ret;
3259 }
3260 EXPORT_SYMBOL(try_to_free_buffers);
3261
3262 /*
3263  * There are no bdflush tunables left.  But distributions are
3264  * still running obsolete flush daemons, so we terminate them here.
3265  *
3266  * Use of bdflush() is deprecated and will be removed in a future kernel.
3267  * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3268  */
3269 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3270 {
3271         static int msg_count;
3272
3273         if (!capable(CAP_SYS_ADMIN))
3274                 return -EPERM;
3275
3276         if (msg_count < 5) {
3277                 msg_count++;
3278                 printk(KERN_INFO
3279                         "warning: process `%s' used the obsolete bdflush"
3280                         " system call\n", current->comm);
3281                 printk(KERN_INFO "Fix your initscripts?\n");
3282         }
3283
3284         if (func == 1)
3285                 do_exit(0);
3286         return 0;
3287 }
3288
3289 /*
3290  * Buffer-head allocation
3291  */
3292 static struct kmem_cache *bh_cachep __read_mostly;
3293
3294 /*
3295  * Once the number of bh's in the machine exceeds this level, we start
3296  * stripping them in writeback.
3297  */
3298 static unsigned long max_buffer_heads;
3299
3300 int buffer_heads_over_limit;
3301
3302 struct bh_accounting {
3303         int nr;                 /* Number of live bh's */
3304         int ratelimit;          /* Limit cacheline bouncing */
3305 };
3306
3307 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3308
3309 static void recalc_bh_state(void)
3310 {
3311         int i;
3312         int tot = 0;
3313
3314         if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3315                 return;
3316         __this_cpu_write(bh_accounting.ratelimit, 0);
3317         for_each_online_cpu(i)
3318                 tot += per_cpu(bh_accounting, i).nr;
3319         buffer_heads_over_limit = (tot > max_buffer_heads);
3320 }
3321
3322 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3323 {
3324         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3325         if (ret) {
3326                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3327                 preempt_disable();
3328                 __this_cpu_inc(bh_accounting.nr);
3329                 recalc_bh_state();
3330                 preempt_enable();
3331         }
3332         return ret;
3333 }
3334 EXPORT_SYMBOL(alloc_buffer_head);
3335
3336 void free_buffer_head(struct buffer_head *bh)
3337 {
3338         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3339         kmem_cache_free(bh_cachep, bh);
3340         preempt_disable();
3341         __this_cpu_dec(bh_accounting.nr);
3342         recalc_bh_state();
3343         preempt_enable();
3344 }
3345 EXPORT_SYMBOL(free_buffer_head);
3346
3347 static void buffer_exit_cpu(int cpu)
3348 {
3349         int i;
3350         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3351
3352         for (i = 0; i < BH_LRU_SIZE; i++) {
3353                 brelse(b->bhs[i]);
3354                 b->bhs[i] = NULL;
3355         }
3356         this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3357         per_cpu(bh_accounting, cpu).nr = 0;
3358 }
3359
3360 static int buffer_cpu_notify(struct notifier_block *self,
3361                               unsigned long action, void *hcpu)
3362 {
3363         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3364                 buffer_exit_cpu((unsigned long)hcpu);
3365         return NOTIFY_OK;
3366 }
3367
3368 /**
3369  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3370  * @bh: struct buffer_head
3371  *
3372  * Return true if the buffer is up-to-date and false,
3373  * with the buffer locked, if not.
3374  */
3375 int bh_uptodate_or_lock(struct buffer_head *bh)
3376 {
3377         if (!buffer_uptodate(bh)) {
3378                 lock_buffer(bh);
3379                 if (!buffer_uptodate(bh))
3380                         return 0;
3381                 unlock_buffer(bh);
3382         }
3383         return 1;
3384 }
3385 EXPORT_SYMBOL(bh_uptodate_or_lock);
3386
3387 /**
3388  * bh_submit_read - Submit a locked buffer for reading
3389  * @bh: struct buffer_head
3390  *
3391  * Returns zero on success and -EIO on error.
3392  */
3393 int bh_submit_read(struct buffer_head *bh)
3394 {
3395         BUG_ON(!buffer_locked(bh));
3396
3397         if (buffer_uptodate(bh)) {
3398                 unlock_buffer(bh);
3399                 return 0;
3400         }
3401
3402         get_bh(bh);
3403         bh->b_end_io = end_buffer_read_sync;
3404         submit_bh(READ, bh);
3405         wait_on_buffer(bh);
3406         if (buffer_uptodate(bh))
3407                 return 0;
3408         return -EIO;
3409 }
3410 EXPORT_SYMBOL(bh_submit_read);
3411
3412 void __init buffer_init(void)
3413 {
3414         unsigned long nrpages;
3415
3416         bh_cachep = kmem_cache_create("buffer_head",
3417                         sizeof(struct buffer_head), 0,
3418                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3419                                 SLAB_MEM_SPREAD),
3420                                 NULL);
3421
3422         /*
3423          * Limit the bh occupancy to 10% of ZONE_NORMAL
3424          */
3425         nrpages = (nr_free_buffer_pages() * 10) / 100;
3426         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3427         hotcpu_notifier(buffer_cpu_notify, 0);
3428 }