]> git.karo-electronics.de Git - karo-tx-linux.git/blob - fs/buffer.c
Merge branch 'for-linus' of git://git.kernel.dk/linux-block
[karo-tx-linux.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_atomic();
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_flags(bd_mapping, index, FGP_ACCESSED);
231         if (!page)
232                 goto out;
233
234         spin_lock(&bd_mapping->private_lock);
235         if (!page_has_buffers(page))
236                 goto out_unlock;
237         head = page_buffers(page);
238         bh = head;
239         do {
240                 if (!buffer_mapped(bh))
241                         all_mapped = 0;
242                 else if (bh->b_blocknr == block) {
243                         ret = bh;
244                         get_bh(bh);
245                         goto out_unlock;
246                 }
247                 bh = bh->b_this_page;
248         } while (bh != head);
249
250         /* we might be here because some of the buffers on this page are
251          * not mapped.  This is due to various races between
252          * file io on the block device and getblk.  It gets dealt with
253          * elsewhere, don't buffer_error if we had some unmapped buffers
254          */
255         if (all_mapped) {
256                 char b[BDEVNAME_SIZE];
257
258                 printk("__find_get_block_slow() failed. "
259                         "block=%llu, b_blocknr=%llu\n",
260                         (unsigned long long)block,
261                         (unsigned long long)bh->b_blocknr);
262                 printk("b_state=0x%08lx, b_size=%zu\n",
263                         bh->b_state, bh->b_size);
264                 printk("device %s blocksize: %d\n", bdevname(bdev, b),
265                         1 << bd_inode->i_blkbits);
266         }
267 out_unlock:
268         spin_unlock(&bd_mapping->private_lock);
269         page_cache_release(page);
270 out:
271         return ret;
272 }
273
274 /*
275  * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
276  */
277 static void free_more_memory(void)
278 {
279         struct zone *zone;
280         int nid;
281
282         wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
283         yield();
284
285         for_each_online_node(nid) {
286                 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
287                                                 gfp_zone(GFP_NOFS), NULL,
288                                                 &zone);
289                 if (zone)
290                         try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
291                                                 GFP_NOFS, NULL);
292         }
293 }
294
295 /*
296  * I/O completion handler for block_read_full_page() - pages
297  * which come unlocked at the end of I/O.
298  */
299 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
300 {
301         unsigned long flags;
302         struct buffer_head *first;
303         struct buffer_head *tmp;
304         struct page *page;
305         int page_uptodate = 1;
306
307         BUG_ON(!buffer_async_read(bh));
308
309         page = bh->b_page;
310         if (uptodate) {
311                 set_buffer_uptodate(bh);
312         } else {
313                 clear_buffer_uptodate(bh);
314                 if (!quiet_error(bh))
315                         buffer_io_error(bh);
316                 SetPageError(page);
317         }
318
319         /*
320          * Be _very_ careful from here on. Bad things can happen if
321          * two buffer heads end IO at almost the same time and both
322          * decide that the page is now completely done.
323          */
324         first = page_buffers(page);
325         local_irq_save(flags);
326         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
327         clear_buffer_async_read(bh);
328         unlock_buffer(bh);
329         tmp = bh;
330         do {
331                 if (!buffer_uptodate(tmp))
332                         page_uptodate = 0;
333                 if (buffer_async_read(tmp)) {
334                         BUG_ON(!buffer_locked(tmp));
335                         goto still_busy;
336                 }
337                 tmp = tmp->b_this_page;
338         } while (tmp != bh);
339         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
340         local_irq_restore(flags);
341
342         /*
343          * If none of the buffers had errors and they are all
344          * uptodate then we can set the page uptodate.
345          */
346         if (page_uptodate && !PageError(page))
347                 SetPageUptodate(page);
348         unlock_page(page);
349         return;
350
351 still_busy:
352         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
353         local_irq_restore(flags);
354         return;
355 }
356
357 /*
358  * Completion handler for block_write_full_page() - pages which are unlocked
359  * during I/O, and which have PageWriteback cleared upon I/O completion.
360  */
361 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
362 {
363         char b[BDEVNAME_SIZE];
364         unsigned long flags;
365         struct buffer_head *first;
366         struct buffer_head *tmp;
367         struct page *page;
368
369         BUG_ON(!buffer_async_write(bh));
370
371         page = bh->b_page;
372         if (uptodate) {
373                 set_buffer_uptodate(bh);
374         } else {
375                 if (!quiet_error(bh)) {
376                         buffer_io_error(bh);
377                         printk(KERN_WARNING "lost page write due to "
378                                         "I/O error on %s\n",
379                                bdevname(bh->b_bdev, b));
380                 }
381                 set_bit(AS_EIO, &page->mapping->flags);
382                 set_buffer_write_io_error(bh);
383                 clear_buffer_uptodate(bh);
384                 SetPageError(page);
385         }
386
387         first = page_buffers(page);
388         local_irq_save(flags);
389         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
390
391         clear_buffer_async_write(bh);
392         unlock_buffer(bh);
393         tmp = bh->b_this_page;
394         while (tmp != bh) {
395                 if (buffer_async_write(tmp)) {
396                         BUG_ON(!buffer_locked(tmp));
397                         goto still_busy;
398                 }
399                 tmp = tmp->b_this_page;
400         }
401         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
402         local_irq_restore(flags);
403         end_page_writeback(page);
404         return;
405
406 still_busy:
407         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
408         local_irq_restore(flags);
409         return;
410 }
411 EXPORT_SYMBOL(end_buffer_async_write);
412
413 /*
414  * If a page's buffers are under async readin (end_buffer_async_read
415  * completion) then there is a possibility that another thread of
416  * control could lock one of the buffers after it has completed
417  * but while some of the other buffers have not completed.  This
418  * locked buffer would confuse end_buffer_async_read() into not unlocking
419  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
420  * that this buffer is not under async I/O.
421  *
422  * The page comes unlocked when it has no locked buffer_async buffers
423  * left.
424  *
425  * PageLocked prevents anyone starting new async I/O reads any of
426  * the buffers.
427  *
428  * PageWriteback is used to prevent simultaneous writeout of the same
429  * page.
430  *
431  * PageLocked prevents anyone from starting writeback of a page which is
432  * under read I/O (PageWriteback is only ever set against a locked page).
433  */
434 static void mark_buffer_async_read(struct buffer_head *bh)
435 {
436         bh->b_end_io = end_buffer_async_read;
437         set_buffer_async_read(bh);
438 }
439
440 static void mark_buffer_async_write_endio(struct buffer_head *bh,
441                                           bh_end_io_t *handler)
442 {
443         bh->b_end_io = handler;
444         set_buffer_async_write(bh);
445 }
446
447 void mark_buffer_async_write(struct buffer_head *bh)
448 {
449         mark_buffer_async_write_endio(bh, end_buffer_async_write);
450 }
451 EXPORT_SYMBOL(mark_buffer_async_write);
452
453
454 /*
455  * fs/buffer.c contains helper functions for buffer-backed address space's
456  * fsync functions.  A common requirement for buffer-based filesystems is
457  * that certain data from the backing blockdev needs to be written out for
458  * a successful fsync().  For example, ext2 indirect blocks need to be
459  * written back and waited upon before fsync() returns.
460  *
461  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
462  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
463  * management of a list of dependent buffers at ->i_mapping->private_list.
464  *
465  * Locking is a little subtle: try_to_free_buffers() will remove buffers
466  * from their controlling inode's queue when they are being freed.  But
467  * try_to_free_buffers() will be operating against the *blockdev* mapping
468  * at the time, not against the S_ISREG file which depends on those buffers.
469  * So the locking for private_list is via the private_lock in the address_space
470  * which backs the buffers.  Which is different from the address_space 
471  * against which the buffers are listed.  So for a particular address_space,
472  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
473  * mapping->private_list will always be protected by the backing blockdev's
474  * ->private_lock.
475  *
476  * Which introduces a requirement: all buffers on an address_space's
477  * ->private_list must be from the same address_space: the blockdev's.
478  *
479  * address_spaces which do not place buffers at ->private_list via these
480  * utility functions are free to use private_lock and private_list for
481  * whatever they want.  The only requirement is that list_empty(private_list)
482  * be true at clear_inode() time.
483  *
484  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
485  * filesystems should do that.  invalidate_inode_buffers() should just go
486  * BUG_ON(!list_empty).
487  *
488  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
489  * take an address_space, not an inode.  And it should be called
490  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
491  * queued up.
492  *
493  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
494  * list if it is already on a list.  Because if the buffer is on a list,
495  * it *must* already be on the right one.  If not, the filesystem is being
496  * silly.  This will save a ton of locking.  But first we have to ensure
497  * that buffers are taken *off* the old inode's list when they are freed
498  * (presumably in truncate).  That requires careful auditing of all
499  * filesystems (do it inside bforget()).  It could also be done by bringing
500  * b_inode back.
501  */
502
503 /*
504  * The buffer's backing address_space's private_lock must be held
505  */
506 static void __remove_assoc_queue(struct buffer_head *bh)
507 {
508         list_del_init(&bh->b_assoc_buffers);
509         WARN_ON(!bh->b_assoc_map);
510         if (buffer_write_io_error(bh))
511                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
512         bh->b_assoc_map = NULL;
513 }
514
515 int inode_has_buffers(struct inode *inode)
516 {
517         return !list_empty(&inode->i_data.private_list);
518 }
519
520 /*
521  * osync is designed to support O_SYNC io.  It waits synchronously for
522  * all already-submitted IO to complete, but does not queue any new
523  * writes to the disk.
524  *
525  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
526  * you dirty the buffers, and then use osync_inode_buffers to wait for
527  * completion.  Any other dirty buffers which are not yet queued for
528  * write will not be flushed to disk by the osync.
529  */
530 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
531 {
532         struct buffer_head *bh;
533         struct list_head *p;
534         int err = 0;
535
536         spin_lock(lock);
537 repeat:
538         list_for_each_prev(p, list) {
539                 bh = BH_ENTRY(p);
540                 if (buffer_locked(bh)) {
541                         get_bh(bh);
542                         spin_unlock(lock);
543                         wait_on_buffer(bh);
544                         if (!buffer_uptodate(bh))
545                                 err = -EIO;
546                         brelse(bh);
547                         spin_lock(lock);
548                         goto repeat;
549                 }
550         }
551         spin_unlock(lock);
552         return err;
553 }
554
555 static void do_thaw_one(struct super_block *sb, void *unused)
556 {
557         char b[BDEVNAME_SIZE];
558         while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
559                 printk(KERN_WARNING "Emergency Thaw on %s\n",
560                        bdevname(sb->s_bdev, b));
561 }
562
563 static void do_thaw_all(struct work_struct *work)
564 {
565         iterate_supers(do_thaw_one, NULL);
566         kfree(work);
567         printk(KERN_WARNING "Emergency Thaw complete\n");
568 }
569
570 /**
571  * emergency_thaw_all -- forcibly thaw every frozen filesystem
572  *
573  * Used for emergency unfreeze of all filesystems via SysRq
574  */
575 void emergency_thaw_all(void)
576 {
577         struct work_struct *work;
578
579         work = kmalloc(sizeof(*work), GFP_ATOMIC);
580         if (work) {
581                 INIT_WORK(work, do_thaw_all);
582                 schedule_work(work);
583         }
584 }
585
586 /**
587  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
588  * @mapping: the mapping which wants those buffers written
589  *
590  * Starts I/O against the buffers at mapping->private_list, and waits upon
591  * that I/O.
592  *
593  * Basically, this is a convenience function for fsync().
594  * @mapping is a file or directory which needs those buffers to be written for
595  * a successful fsync().
596  */
597 int sync_mapping_buffers(struct address_space *mapping)
598 {
599         struct address_space *buffer_mapping = mapping->private_data;
600
601         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
602                 return 0;
603
604         return fsync_buffers_list(&buffer_mapping->private_lock,
605                                         &mapping->private_list);
606 }
607 EXPORT_SYMBOL(sync_mapping_buffers);
608
609 /*
610  * Called when we've recently written block `bblock', and it is known that
611  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
612  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
613  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
614  */
615 void write_boundary_block(struct block_device *bdev,
616                         sector_t bblock, unsigned blocksize)
617 {
618         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
619         if (bh) {
620                 if (buffer_dirty(bh))
621                         ll_rw_block(WRITE, 1, &bh);
622                 put_bh(bh);
623         }
624 }
625
626 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
627 {
628         struct address_space *mapping = inode->i_mapping;
629         struct address_space *buffer_mapping = bh->b_page->mapping;
630
631         mark_buffer_dirty(bh);
632         if (!mapping->private_data) {
633                 mapping->private_data = buffer_mapping;
634         } else {
635                 BUG_ON(mapping->private_data != buffer_mapping);
636         }
637         if (!bh->b_assoc_map) {
638                 spin_lock(&buffer_mapping->private_lock);
639                 list_move_tail(&bh->b_assoc_buffers,
640                                 &mapping->private_list);
641                 bh->b_assoc_map = mapping;
642                 spin_unlock(&buffer_mapping->private_lock);
643         }
644 }
645 EXPORT_SYMBOL(mark_buffer_dirty_inode);
646
647 /*
648  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
649  * dirty.
650  *
651  * If warn is true, then emit a warning if the page is not uptodate and has
652  * not been truncated.
653  */
654 static void __set_page_dirty(struct page *page,
655                 struct address_space *mapping, int warn)
656 {
657         unsigned long flags;
658
659         spin_lock_irqsave(&mapping->tree_lock, flags);
660         if (page->mapping) {    /* Race with truncate? */
661                 WARN_ON_ONCE(warn && !PageUptodate(page));
662                 account_page_dirtied(page, mapping);
663                 radix_tree_tag_set(&mapping->page_tree,
664                                 page_index(page), PAGECACHE_TAG_DIRTY);
665         }
666         spin_unlock_irqrestore(&mapping->tree_lock, flags);
667         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
668 }
669
670 /*
671  * Add a page to the dirty page list.
672  *
673  * It is a sad fact of life that this function is called from several places
674  * deeply under spinlocking.  It may not sleep.
675  *
676  * If the page has buffers, the uptodate buffers are set dirty, to preserve
677  * dirty-state coherency between the page and the buffers.  It the page does
678  * not have buffers then when they are later attached they will all be set
679  * dirty.
680  *
681  * The buffers are dirtied before the page is dirtied.  There's a small race
682  * window in which a writepage caller may see the page cleanness but not the
683  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
684  * before the buffers, a concurrent writepage caller could clear the page dirty
685  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
686  * page on the dirty page list.
687  *
688  * We use private_lock to lock against try_to_free_buffers while using the
689  * page's buffer list.  Also use this to protect against clean buffers being
690  * added to the page after it was set dirty.
691  *
692  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
693  * address_space though.
694  */
695 int __set_page_dirty_buffers(struct page *page)
696 {
697         int newly_dirty;
698         struct address_space *mapping = page_mapping(page);
699
700         if (unlikely(!mapping))
701                 return !TestSetPageDirty(page);
702
703         spin_lock(&mapping->private_lock);
704         if (page_has_buffers(page)) {
705                 struct buffer_head *head = page_buffers(page);
706                 struct buffer_head *bh = head;
707
708                 do {
709                         set_buffer_dirty(bh);
710                         bh = bh->b_this_page;
711                 } while (bh != head);
712         }
713         newly_dirty = !TestSetPageDirty(page);
714         spin_unlock(&mapping->private_lock);
715
716         if (newly_dirty)
717                 __set_page_dirty(page, mapping, 1);
718         return newly_dirty;
719 }
720 EXPORT_SYMBOL(__set_page_dirty_buffers);
721
722 /*
723  * Write out and wait upon a list of buffers.
724  *
725  * We have conflicting pressures: we want to make sure that all
726  * initially dirty buffers get waited on, but that any subsequently
727  * dirtied buffers don't.  After all, we don't want fsync to last
728  * forever if somebody is actively writing to the file.
729  *
730  * Do this in two main stages: first we copy dirty buffers to a
731  * temporary inode list, queueing the writes as we go.  Then we clean
732  * up, waiting for those writes to complete.
733  * 
734  * During this second stage, any subsequent updates to the file may end
735  * up refiling the buffer on the original inode's dirty list again, so
736  * there is a chance we will end up with a buffer queued for write but
737  * not yet completed on that list.  So, as a final cleanup we go through
738  * the osync code to catch these locked, dirty buffers without requeuing
739  * any newly dirty buffers for write.
740  */
741 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
742 {
743         struct buffer_head *bh;
744         struct list_head tmp;
745         struct address_space *mapping;
746         int err = 0, err2;
747         struct blk_plug plug;
748
749         INIT_LIST_HEAD(&tmp);
750         blk_start_plug(&plug);
751
752         spin_lock(lock);
753         while (!list_empty(list)) {
754                 bh = BH_ENTRY(list->next);
755                 mapping = bh->b_assoc_map;
756                 __remove_assoc_queue(bh);
757                 /* Avoid race with mark_buffer_dirty_inode() which does
758                  * a lockless check and we rely on seeing the dirty bit */
759                 smp_mb();
760                 if (buffer_dirty(bh) || buffer_locked(bh)) {
761                         list_add(&bh->b_assoc_buffers, &tmp);
762                         bh->b_assoc_map = mapping;
763                         if (buffer_dirty(bh)) {
764                                 get_bh(bh);
765                                 spin_unlock(lock);
766                                 /*
767                                  * Ensure any pending I/O completes so that
768                                  * write_dirty_buffer() actually writes the
769                                  * current contents - it is a noop if I/O is
770                                  * still in flight on potentially older
771                                  * contents.
772                                  */
773                                 write_dirty_buffer(bh, WRITE_SYNC);
774
775                                 /*
776                                  * Kick off IO for the previous mapping. Note
777                                  * that we will not run the very last mapping,
778                                  * wait_on_buffer() will do that for us
779                                  * through sync_buffer().
780                                  */
781                                 brelse(bh);
782                                 spin_lock(lock);
783                         }
784                 }
785         }
786
787         spin_unlock(lock);
788         blk_finish_plug(&plug);
789         spin_lock(lock);
790
791         while (!list_empty(&tmp)) {
792                 bh = BH_ENTRY(tmp.prev);
793                 get_bh(bh);
794                 mapping = bh->b_assoc_map;
795                 __remove_assoc_queue(bh);
796                 /* Avoid race with mark_buffer_dirty_inode() which does
797                  * a lockless check and we rely on seeing the dirty bit */
798                 smp_mb();
799                 if (buffer_dirty(bh)) {
800                         list_add(&bh->b_assoc_buffers,
801                                  &mapping->private_list);
802                         bh->b_assoc_map = mapping;
803                 }
804                 spin_unlock(lock);
805                 wait_on_buffer(bh);
806                 if (!buffer_uptodate(bh))
807                         err = -EIO;
808                 brelse(bh);
809                 spin_lock(lock);
810         }
811         
812         spin_unlock(lock);
813         err2 = osync_buffers_list(lock, list);
814         if (err)
815                 return err;
816         else
817                 return err2;
818 }
819
820 /*
821  * Invalidate any and all dirty buffers on a given inode.  We are
822  * probably unmounting the fs, but that doesn't mean we have already
823  * done a sync().  Just drop the buffers from the inode list.
824  *
825  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
826  * assumes that all the buffers are against the blockdev.  Not true
827  * for reiserfs.
828  */
829 void invalidate_inode_buffers(struct inode *inode)
830 {
831         if (inode_has_buffers(inode)) {
832                 struct address_space *mapping = &inode->i_data;
833                 struct list_head *list = &mapping->private_list;
834                 struct address_space *buffer_mapping = mapping->private_data;
835
836                 spin_lock(&buffer_mapping->private_lock);
837                 while (!list_empty(list))
838                         __remove_assoc_queue(BH_ENTRY(list->next));
839                 spin_unlock(&buffer_mapping->private_lock);
840         }
841 }
842 EXPORT_SYMBOL(invalidate_inode_buffers);
843
844 /*
845  * Remove any clean buffers from the inode's buffer list.  This is called
846  * when we're trying to free the inode itself.  Those buffers can pin it.
847  *
848  * Returns true if all buffers were removed.
849  */
850 int remove_inode_buffers(struct inode *inode)
851 {
852         int ret = 1;
853
854         if (inode_has_buffers(inode)) {
855                 struct address_space *mapping = &inode->i_data;
856                 struct list_head *list = &mapping->private_list;
857                 struct address_space *buffer_mapping = mapping->private_data;
858
859                 spin_lock(&buffer_mapping->private_lock);
860                 while (!list_empty(list)) {
861                         struct buffer_head *bh = BH_ENTRY(list->next);
862                         if (buffer_dirty(bh)) {
863                                 ret = 0;
864                                 break;
865                         }
866                         __remove_assoc_queue(bh);
867                 }
868                 spin_unlock(&buffer_mapping->private_lock);
869         }
870         return ret;
871 }
872
873 /*
874  * Create the appropriate buffers when given a page for data area and
875  * the size of each buffer.. Use the bh->b_this_page linked list to
876  * follow the buffers created.  Return NULL if unable to create more
877  * buffers.
878  *
879  * The retry flag is used to differentiate async IO (paging, swapping)
880  * which may not fail from ordinary buffer allocations.
881  */
882 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
883                 int retry)
884 {
885         struct buffer_head *bh, *head;
886         long offset;
887
888 try_again:
889         head = NULL;
890         offset = PAGE_SIZE;
891         while ((offset -= size) >= 0) {
892                 bh = alloc_buffer_head(GFP_NOFS);
893                 if (!bh)
894                         goto no_grow;
895
896                 bh->b_this_page = head;
897                 bh->b_blocknr = -1;
898                 head = bh;
899
900                 bh->b_size = size;
901
902                 /* Link the buffer to its page */
903                 set_bh_page(bh, page, offset);
904         }
905         return head;
906 /*
907  * In case anything failed, we just free everything we got.
908  */
909 no_grow:
910         if (head) {
911                 do {
912                         bh = head;
913                         head = head->b_this_page;
914                         free_buffer_head(bh);
915                 } while (head);
916         }
917
918         /*
919          * Return failure for non-async IO requests.  Async IO requests
920          * are not allowed to fail, so we have to wait until buffer heads
921          * become available.  But we don't want tasks sleeping with 
922          * partially complete buffers, so all were released above.
923          */
924         if (!retry)
925                 return NULL;
926
927         /* We're _really_ low on memory. Now we just
928          * wait for old buffer heads to become free due to
929          * finishing IO.  Since this is an async request and
930          * the reserve list is empty, we're sure there are 
931          * async buffer heads in use.
932          */
933         free_more_memory();
934         goto try_again;
935 }
936 EXPORT_SYMBOL_GPL(alloc_page_buffers);
937
938 static inline void
939 link_dev_buffers(struct page *page, struct buffer_head *head)
940 {
941         struct buffer_head *bh, *tail;
942
943         bh = head;
944         do {
945                 tail = bh;
946                 bh = bh->b_this_page;
947         } while (bh);
948         tail->b_this_page = head;
949         attach_page_buffers(page, head);
950 }
951
952 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
953 {
954         sector_t retval = ~((sector_t)0);
955         loff_t sz = i_size_read(bdev->bd_inode);
956
957         if (sz) {
958                 unsigned int sizebits = blksize_bits(size);
959                 retval = (sz >> sizebits);
960         }
961         return retval;
962 }
963
964 /*
965  * Initialise the state of a blockdev page's buffers.
966  */ 
967 static sector_t
968 init_page_buffers(struct page *page, struct block_device *bdev,
969                         sector_t block, int size)
970 {
971         struct buffer_head *head = page_buffers(page);
972         struct buffer_head *bh = head;
973         int uptodate = PageUptodate(page);
974         sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
975
976         do {
977                 if (!buffer_mapped(bh)) {
978                         init_buffer(bh, NULL, NULL);
979                         bh->b_bdev = bdev;
980                         bh->b_blocknr = block;
981                         if (uptodate)
982                                 set_buffer_uptodate(bh);
983                         if (block < end_block)
984                                 set_buffer_mapped(bh);
985                 }
986                 block++;
987                 bh = bh->b_this_page;
988         } while (bh != head);
989
990         /*
991          * Caller needs to validate requested block against end of device.
992          */
993         return end_block;
994 }
995
996 /*
997  * Create the page-cache page that contains the requested block.
998  *
999  * This is used purely for blockdev mappings.
1000  */
1001 static int
1002 grow_dev_page(struct block_device *bdev, sector_t block,
1003                 pgoff_t index, int size, int sizebits)
1004 {
1005         struct inode *inode = bdev->bd_inode;
1006         struct page *page;
1007         struct buffer_head *bh;
1008         sector_t end_block;
1009         int ret = 0;            /* Will call free_more_memory() */
1010         gfp_t gfp_mask;
1011
1012         gfp_mask = mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS;
1013         gfp_mask |= __GFP_MOVABLE;
1014         /*
1015          * XXX: __getblk_slow() can not really deal with failure and
1016          * will endlessly loop on improvised global reclaim.  Prefer
1017          * looping in the allocator rather than here, at least that
1018          * code knows what it's doing.
1019          */
1020         gfp_mask |= __GFP_NOFAIL;
1021
1022         page = find_or_create_page(inode->i_mapping, index, gfp_mask);
1023         if (!page)
1024                 return ret;
1025
1026         BUG_ON(!PageLocked(page));
1027
1028         if (page_has_buffers(page)) {
1029                 bh = page_buffers(page);
1030                 if (bh->b_size == size) {
1031                         end_block = init_page_buffers(page, bdev,
1032                                                 index << sizebits, size);
1033                         goto done;
1034                 }
1035                 if (!try_to_free_buffers(page))
1036                         goto failed;
1037         }
1038
1039         /*
1040          * Allocate some buffers for this page
1041          */
1042         bh = alloc_page_buffers(page, size, 0);
1043         if (!bh)
1044                 goto failed;
1045
1046         /*
1047          * Link the page to the buffers and initialise them.  Take the
1048          * lock to be atomic wrt __find_get_block(), which does not
1049          * run under the page lock.
1050          */
1051         spin_lock(&inode->i_mapping->private_lock);
1052         link_dev_buffers(page, bh);
1053         end_block = init_page_buffers(page, bdev, index << sizebits, size);
1054         spin_unlock(&inode->i_mapping->private_lock);
1055 done:
1056         ret = (block < end_block) ? 1 : -ENXIO;
1057 failed:
1058         unlock_page(page);
1059         page_cache_release(page);
1060         return ret;
1061 }
1062
1063 /*
1064  * Create buffers for the specified block device block's page.  If
1065  * that page was dirty, the buffers are set dirty also.
1066  */
1067 static int
1068 grow_buffers(struct block_device *bdev, sector_t block, int size)
1069 {
1070         pgoff_t index;
1071         int sizebits;
1072
1073         sizebits = -1;
1074         do {
1075                 sizebits++;
1076         } while ((size << sizebits) < PAGE_SIZE);
1077
1078         index = block >> sizebits;
1079
1080         /*
1081          * Check for a block which wants to lie outside our maximum possible
1082          * pagecache index.  (this comparison is done using sector_t types).
1083          */
1084         if (unlikely(index != block >> sizebits)) {
1085                 char b[BDEVNAME_SIZE];
1086
1087                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1088                         "device %s\n",
1089                         __func__, (unsigned long long)block,
1090                         bdevname(bdev, b));
1091                 return -EIO;
1092         }
1093
1094         /* Create a page with the proper size buffers.. */
1095         return grow_dev_page(bdev, block, index, size, sizebits);
1096 }
1097
1098 static struct buffer_head *
1099 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1100 {
1101         /* Size must be multiple of hard sectorsize */
1102         if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1103                         (size < 512 || size > PAGE_SIZE))) {
1104                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1105                                         size);
1106                 printk(KERN_ERR "logical block size: %d\n",
1107                                         bdev_logical_block_size(bdev));
1108
1109                 dump_stack();
1110                 return NULL;
1111         }
1112
1113         for (;;) {
1114                 struct buffer_head *bh;
1115                 int ret;
1116
1117                 bh = __find_get_block(bdev, block, size);
1118                 if (bh)
1119                         return bh;
1120
1121                 ret = grow_buffers(bdev, block, size);
1122                 if (ret < 0)
1123                         return NULL;
1124                 if (ret == 0)
1125                         free_more_memory();
1126         }
1127 }
1128
1129 /*
1130  * The relationship between dirty buffers and dirty pages:
1131  *
1132  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1133  * the page is tagged dirty in its radix tree.
1134  *
1135  * At all times, the dirtiness of the buffers represents the dirtiness of
1136  * subsections of the page.  If the page has buffers, the page dirty bit is
1137  * merely a hint about the true dirty state.
1138  *
1139  * When a page is set dirty in its entirety, all its buffers are marked dirty
1140  * (if the page has buffers).
1141  *
1142  * When a buffer is marked dirty, its page is dirtied, but the page's other
1143  * buffers are not.
1144  *
1145  * Also.  When blockdev buffers are explicitly read with bread(), they
1146  * individually become uptodate.  But their backing page remains not
1147  * uptodate - even if all of its buffers are uptodate.  A subsequent
1148  * block_read_full_page() against that page will discover all the uptodate
1149  * buffers, will set the page uptodate and will perform no I/O.
1150  */
1151
1152 /**
1153  * mark_buffer_dirty - mark a buffer_head as needing writeout
1154  * @bh: the buffer_head to mark dirty
1155  *
1156  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1157  * backing page dirty, then tag the page as dirty in its address_space's radix
1158  * tree and then attach the address_space's inode to its superblock's dirty
1159  * inode list.
1160  *
1161  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1162  * mapping->tree_lock and mapping->host->i_lock.
1163  */
1164 void mark_buffer_dirty(struct buffer_head *bh)
1165 {
1166         WARN_ON_ONCE(!buffer_uptodate(bh));
1167
1168         trace_block_dirty_buffer(bh);
1169
1170         /*
1171          * Very *carefully* optimize the it-is-already-dirty case.
1172          *
1173          * Don't let the final "is it dirty" escape to before we
1174          * perhaps modified the buffer.
1175          */
1176         if (buffer_dirty(bh)) {
1177                 smp_mb();
1178                 if (buffer_dirty(bh))
1179                         return;
1180         }
1181
1182         if (!test_set_buffer_dirty(bh)) {
1183                 struct page *page = bh->b_page;
1184                 if (!TestSetPageDirty(page)) {
1185                         struct address_space *mapping = page_mapping(page);
1186                         if (mapping)
1187                                 __set_page_dirty(page, mapping, 0);
1188                 }
1189         }
1190 }
1191 EXPORT_SYMBOL(mark_buffer_dirty);
1192
1193 /*
1194  * Decrement a buffer_head's reference count.  If all buffers against a page
1195  * have zero reference count, are clean and unlocked, and if the page is clean
1196  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1197  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1198  * a page but it ends up not being freed, and buffers may later be reattached).
1199  */
1200 void __brelse(struct buffer_head * buf)
1201 {
1202         if (atomic_read(&buf->b_count)) {
1203                 put_bh(buf);
1204                 return;
1205         }
1206         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1207 }
1208 EXPORT_SYMBOL(__brelse);
1209
1210 /*
1211  * bforget() is like brelse(), except it discards any
1212  * potentially dirty data.
1213  */
1214 void __bforget(struct buffer_head *bh)
1215 {
1216         clear_buffer_dirty(bh);
1217         if (bh->b_assoc_map) {
1218                 struct address_space *buffer_mapping = bh->b_page->mapping;
1219
1220                 spin_lock(&buffer_mapping->private_lock);
1221                 list_del_init(&bh->b_assoc_buffers);
1222                 bh->b_assoc_map = NULL;
1223                 spin_unlock(&buffer_mapping->private_lock);
1224         }
1225         __brelse(bh);
1226 }
1227 EXPORT_SYMBOL(__bforget);
1228
1229 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1230 {
1231         lock_buffer(bh);
1232         if (buffer_uptodate(bh)) {
1233                 unlock_buffer(bh);
1234                 return bh;
1235         } else {
1236                 get_bh(bh);
1237                 bh->b_end_io = end_buffer_read_sync;
1238                 submit_bh(READ, bh);
1239                 wait_on_buffer(bh);
1240                 if (buffer_uptodate(bh))
1241                         return bh;
1242         }
1243         brelse(bh);
1244         return NULL;
1245 }
1246
1247 /*
1248  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1249  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1250  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1251  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1252  * CPU's LRUs at the same time.
1253  *
1254  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1255  * sb_find_get_block().
1256  *
1257  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1258  * a local interrupt disable for that.
1259  */
1260
1261 #define BH_LRU_SIZE     8
1262
1263 struct bh_lru {
1264         struct buffer_head *bhs[BH_LRU_SIZE];
1265 };
1266
1267 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1268
1269 #ifdef CONFIG_SMP
1270 #define bh_lru_lock()   local_irq_disable()
1271 #define bh_lru_unlock() local_irq_enable()
1272 #else
1273 #define bh_lru_lock()   preempt_disable()
1274 #define bh_lru_unlock() preempt_enable()
1275 #endif
1276
1277 static inline void check_irqs_on(void)
1278 {
1279 #ifdef irqs_disabled
1280         BUG_ON(irqs_disabled());
1281 #endif
1282 }
1283
1284 /*
1285  * The LRU management algorithm is dopey-but-simple.  Sorry.
1286  */
1287 static void bh_lru_install(struct buffer_head *bh)
1288 {
1289         struct buffer_head *evictee = NULL;
1290
1291         check_irqs_on();
1292         bh_lru_lock();
1293         if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1294                 struct buffer_head *bhs[BH_LRU_SIZE];
1295                 int in;
1296                 int out = 0;
1297
1298                 get_bh(bh);
1299                 bhs[out++] = bh;
1300                 for (in = 0; in < BH_LRU_SIZE; in++) {
1301                         struct buffer_head *bh2 =
1302                                 __this_cpu_read(bh_lrus.bhs[in]);
1303
1304                         if (bh2 == bh) {
1305                                 __brelse(bh2);
1306                         } else {
1307                                 if (out >= BH_LRU_SIZE) {
1308                                         BUG_ON(evictee != NULL);
1309                                         evictee = bh2;
1310                                 } else {
1311                                         bhs[out++] = bh2;
1312                                 }
1313                         }
1314                 }
1315                 while (out < BH_LRU_SIZE)
1316                         bhs[out++] = NULL;
1317                 memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1318         }
1319         bh_lru_unlock();
1320
1321         if (evictee)
1322                 __brelse(evictee);
1323 }
1324
1325 /*
1326  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1327  */
1328 static struct buffer_head *
1329 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1330 {
1331         struct buffer_head *ret = NULL;
1332         unsigned int i;
1333
1334         check_irqs_on();
1335         bh_lru_lock();
1336         for (i = 0; i < BH_LRU_SIZE; i++) {
1337                 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1338
1339                 if (bh && bh->b_bdev == bdev &&
1340                                 bh->b_blocknr == block && bh->b_size == size) {
1341                         if (i) {
1342                                 while (i) {
1343                                         __this_cpu_write(bh_lrus.bhs[i],
1344                                                 __this_cpu_read(bh_lrus.bhs[i - 1]));
1345                                         i--;
1346                                 }
1347                                 __this_cpu_write(bh_lrus.bhs[0], bh);
1348                         }
1349                         get_bh(bh);
1350                         ret = bh;
1351                         break;
1352                 }
1353         }
1354         bh_lru_unlock();
1355         return ret;
1356 }
1357
1358 /*
1359  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1360  * it in the LRU and mark it as accessed.  If it is not present then return
1361  * NULL
1362  */
1363 struct buffer_head *
1364 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1365 {
1366         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1367
1368         if (bh == NULL) {
1369                 /* __find_get_block_slow will mark the page accessed */
1370                 bh = __find_get_block_slow(bdev, block);
1371                 if (bh)
1372                         bh_lru_install(bh);
1373         } else
1374                 touch_buffer(bh);
1375
1376         return bh;
1377 }
1378 EXPORT_SYMBOL(__find_get_block);
1379
1380 /*
1381  * __getblk will locate (and, if necessary, create) the buffer_head
1382  * which corresponds to the passed block_device, block and size. The
1383  * returned buffer has its reference count incremented.
1384  *
1385  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1386  * attempt is failing.  FIXME, perhaps?
1387  */
1388 struct buffer_head *
1389 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1390 {
1391         struct buffer_head *bh = __find_get_block(bdev, block, size);
1392
1393         might_sleep();
1394         if (bh == NULL)
1395                 bh = __getblk_slow(bdev, block, size);
1396         return bh;
1397 }
1398 EXPORT_SYMBOL(__getblk);
1399
1400 /*
1401  * Do async read-ahead on a buffer..
1402  */
1403 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1404 {
1405         struct buffer_head *bh = __getblk(bdev, block, size);
1406         if (likely(bh)) {
1407                 ll_rw_block(READA, 1, &bh);
1408                 brelse(bh);
1409         }
1410 }
1411 EXPORT_SYMBOL(__breadahead);
1412
1413 /**
1414  *  __bread() - reads a specified block and returns the bh
1415  *  @bdev: the block_device to read from
1416  *  @block: number of block
1417  *  @size: size (in bytes) to read
1418  * 
1419  *  Reads a specified block, and returns buffer head that contains it.
1420  *  It returns NULL if the block was unreadable.
1421  */
1422 struct buffer_head *
1423 __bread(struct block_device *bdev, sector_t block, unsigned size)
1424 {
1425         struct buffer_head *bh = __getblk(bdev, block, size);
1426
1427         if (likely(bh) && !buffer_uptodate(bh))
1428                 bh = __bread_slow(bh);
1429         return bh;
1430 }
1431 EXPORT_SYMBOL(__bread);
1432
1433 /*
1434  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1435  * This doesn't race because it runs in each cpu either in irq
1436  * or with preempt disabled.
1437  */
1438 static void invalidate_bh_lru(void *arg)
1439 {
1440         struct bh_lru *b = &get_cpu_var(bh_lrus);
1441         int i;
1442
1443         for (i = 0; i < BH_LRU_SIZE; i++) {
1444                 brelse(b->bhs[i]);
1445                 b->bhs[i] = NULL;
1446         }
1447         put_cpu_var(bh_lrus);
1448 }
1449
1450 static bool has_bh_in_lru(int cpu, void *dummy)
1451 {
1452         struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1453         int i;
1454         
1455         for (i = 0; i < BH_LRU_SIZE; i++) {
1456                 if (b->bhs[i])
1457                         return 1;
1458         }
1459
1460         return 0;
1461 }
1462
1463 void invalidate_bh_lrus(void)
1464 {
1465         on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1466 }
1467 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1468
1469 void set_bh_page(struct buffer_head *bh,
1470                 struct page *page, unsigned long offset)
1471 {
1472         bh->b_page = page;
1473         BUG_ON(offset >= PAGE_SIZE);
1474         if (PageHighMem(page))
1475                 /*
1476                  * This catches illegal uses and preserves the offset:
1477                  */
1478                 bh->b_data = (char *)(0 + offset);
1479         else
1480                 bh->b_data = page_address(page) + offset;
1481 }
1482 EXPORT_SYMBOL(set_bh_page);
1483
1484 /*
1485  * Called when truncating a buffer on a page completely.
1486  */
1487
1488 /* Bits that are cleared during an invalidate */
1489 #define BUFFER_FLAGS_DISCARD \
1490         (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1491          1 << BH_Delay | 1 << BH_Unwritten)
1492
1493 static void discard_buffer(struct buffer_head * bh)
1494 {
1495         unsigned long b_state, b_state_old;
1496
1497         lock_buffer(bh);
1498         clear_buffer_dirty(bh);
1499         bh->b_bdev = NULL;
1500         b_state = bh->b_state;
1501         for (;;) {
1502                 b_state_old = cmpxchg(&bh->b_state, b_state,
1503                                       (b_state & ~BUFFER_FLAGS_DISCARD));
1504                 if (b_state_old == b_state)
1505                         break;
1506                 b_state = b_state_old;
1507         }
1508         unlock_buffer(bh);
1509 }
1510
1511 /**
1512  * block_invalidatepage - invalidate part or all of a buffer-backed page
1513  *
1514  * @page: the page which is affected
1515  * @offset: start of the range to invalidate
1516  * @length: length of the range to invalidate
1517  *
1518  * block_invalidatepage() is called when all or part of the page has become
1519  * invalidated by a truncate operation.
1520  *
1521  * block_invalidatepage() does not have to release all buffers, but it must
1522  * ensure that no dirty buffer is left outside @offset and that no I/O
1523  * is underway against any of the blocks which are outside the truncation
1524  * point.  Because the caller is about to free (and possibly reuse) those
1525  * blocks on-disk.
1526  */
1527 void block_invalidatepage(struct page *page, unsigned int offset,
1528                           unsigned int length)
1529 {
1530         struct buffer_head *head, *bh, *next;
1531         unsigned int curr_off = 0;
1532         unsigned int stop = length + offset;
1533
1534         BUG_ON(!PageLocked(page));
1535         if (!page_has_buffers(page))
1536                 goto out;
1537
1538         /*
1539          * Check for overflow
1540          */
1541         BUG_ON(stop > PAGE_CACHE_SIZE || stop < length);
1542
1543         head = page_buffers(page);
1544         bh = head;
1545         do {
1546                 unsigned int next_off = curr_off + bh->b_size;
1547                 next = bh->b_this_page;
1548
1549                 /*
1550                  * Are we still fully in range ?
1551                  */
1552                 if (next_off > stop)
1553                         goto out;
1554
1555                 /*
1556                  * is this block fully invalidated?
1557                  */
1558                 if (offset <= curr_off)
1559                         discard_buffer(bh);
1560                 curr_off = next_off;
1561                 bh = next;
1562         } while (bh != head);
1563
1564         /*
1565          * We release buffers only if the entire page is being invalidated.
1566          * The get_block cached value has been unconditionally invalidated,
1567          * so real IO is not possible anymore.
1568          */
1569         if (offset == 0)
1570                 try_to_release_page(page, 0);
1571 out:
1572         return;
1573 }
1574 EXPORT_SYMBOL(block_invalidatepage);
1575
1576
1577 /*
1578  * We attach and possibly dirty the buffers atomically wrt
1579  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1580  * is already excluded via the page lock.
1581  */
1582 void create_empty_buffers(struct page *page,
1583                         unsigned long blocksize, unsigned long b_state)
1584 {
1585         struct buffer_head *bh, *head, *tail;
1586
1587         head = alloc_page_buffers(page, blocksize, 1);
1588         bh = head;
1589         do {
1590                 bh->b_state |= b_state;
1591                 tail = bh;
1592                 bh = bh->b_this_page;
1593         } while (bh);
1594         tail->b_this_page = head;
1595
1596         spin_lock(&page->mapping->private_lock);
1597         if (PageUptodate(page) || PageDirty(page)) {
1598                 bh = head;
1599                 do {
1600                         if (PageDirty(page))
1601                                 set_buffer_dirty(bh);
1602                         if (PageUptodate(page))
1603                                 set_buffer_uptodate(bh);
1604                         bh = bh->b_this_page;
1605                 } while (bh != head);
1606         }
1607         attach_page_buffers(page, head);
1608         spin_unlock(&page->mapping->private_lock);
1609 }
1610 EXPORT_SYMBOL(create_empty_buffers);
1611
1612 /*
1613  * We are taking a block for data and we don't want any output from any
1614  * buffer-cache aliases starting from return from that function and
1615  * until the moment when something will explicitly mark the buffer
1616  * dirty (hopefully that will not happen until we will free that block ;-)
1617  * We don't even need to mark it not-uptodate - nobody can expect
1618  * anything from a newly allocated buffer anyway. We used to used
1619  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1620  * don't want to mark the alias unmapped, for example - it would confuse
1621  * anyone who might pick it with bread() afterwards...
1622  *
1623  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1624  * be writeout I/O going on against recently-freed buffers.  We don't
1625  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1626  * only if we really need to.  That happens here.
1627  */
1628 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1629 {
1630         struct buffer_head *old_bh;
1631
1632         might_sleep();
1633
1634         old_bh = __find_get_block_slow(bdev, block);
1635         if (old_bh) {
1636                 clear_buffer_dirty(old_bh);
1637                 wait_on_buffer(old_bh);
1638                 clear_buffer_req(old_bh);
1639                 __brelse(old_bh);
1640         }
1641 }
1642 EXPORT_SYMBOL(unmap_underlying_metadata);
1643
1644 /*
1645  * Size is a power-of-two in the range 512..PAGE_SIZE,
1646  * and the case we care about most is PAGE_SIZE.
1647  *
1648  * So this *could* possibly be written with those
1649  * constraints in mind (relevant mostly if some
1650  * architecture has a slow bit-scan instruction)
1651  */
1652 static inline int block_size_bits(unsigned int blocksize)
1653 {
1654         return ilog2(blocksize);
1655 }
1656
1657 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1658 {
1659         BUG_ON(!PageLocked(page));
1660
1661         if (!page_has_buffers(page))
1662                 create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1663         return page_buffers(page);
1664 }
1665
1666 /*
1667  * NOTE! All mapped/uptodate combinations are valid:
1668  *
1669  *      Mapped  Uptodate        Meaning
1670  *
1671  *      No      No              "unknown" - must do get_block()
1672  *      No      Yes             "hole" - zero-filled
1673  *      Yes     No              "allocated" - allocated on disk, not read in
1674  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1675  *
1676  * "Dirty" is valid only with the last case (mapped+uptodate).
1677  */
1678
1679 /*
1680  * While block_write_full_page is writing back the dirty buffers under
1681  * the page lock, whoever dirtied the buffers may decide to clean them
1682  * again at any time.  We handle that by only looking at the buffer
1683  * state inside lock_buffer().
1684  *
1685  * If block_write_full_page() is called for regular writeback
1686  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1687  * locked buffer.   This only can happen if someone has written the buffer
1688  * directly, with submit_bh().  At the address_space level PageWriteback
1689  * prevents this contention from occurring.
1690  *
1691  * If block_write_full_page() is called with wbc->sync_mode ==
1692  * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1693  * causes the writes to be flagged as synchronous writes.
1694  */
1695 static int __block_write_full_page(struct inode *inode, struct page *page,
1696                         get_block_t *get_block, struct writeback_control *wbc,
1697                         bh_end_io_t *handler)
1698 {
1699         int err;
1700         sector_t block;
1701         sector_t last_block;
1702         struct buffer_head *bh, *head;
1703         unsigned int blocksize, bbits;
1704         int nr_underway = 0;
1705         int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1706                         WRITE_SYNC : WRITE);
1707
1708         head = create_page_buffers(page, inode,
1709                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1710
1711         /*
1712          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1713          * here, and the (potentially unmapped) buffers may become dirty at
1714          * any time.  If a buffer becomes dirty here after we've inspected it
1715          * then we just miss that fact, and the page stays dirty.
1716          *
1717          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1718          * handle that here by just cleaning them.
1719          */
1720
1721         bh = head;
1722         blocksize = bh->b_size;
1723         bbits = block_size_bits(blocksize);
1724
1725         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1726         last_block = (i_size_read(inode) - 1) >> bbits;
1727
1728         /*
1729          * Get all the dirty buffers mapped to disk addresses and
1730          * handle any aliases from the underlying blockdev's mapping.
1731          */
1732         do {
1733                 if (block > last_block) {
1734                         /*
1735                          * mapped buffers outside i_size will occur, because
1736                          * this page can be outside i_size when there is a
1737                          * truncate in progress.
1738                          */
1739                         /*
1740                          * The buffer was zeroed by block_write_full_page()
1741                          */
1742                         clear_buffer_dirty(bh);
1743                         set_buffer_uptodate(bh);
1744                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1745                            buffer_dirty(bh)) {
1746                         WARN_ON(bh->b_size != blocksize);
1747                         err = get_block(inode, block, bh, 1);
1748                         if (err)
1749                                 goto recover;
1750                         clear_buffer_delay(bh);
1751                         if (buffer_new(bh)) {
1752                                 /* blockdev mappings never come here */
1753                                 clear_buffer_new(bh);
1754                                 unmap_underlying_metadata(bh->b_bdev,
1755                                                         bh->b_blocknr);
1756                         }
1757                 }
1758                 bh = bh->b_this_page;
1759                 block++;
1760         } while (bh != head);
1761
1762         do {
1763                 if (!buffer_mapped(bh))
1764                         continue;
1765                 /*
1766                  * If it's a fully non-blocking write attempt and we cannot
1767                  * lock the buffer then redirty the page.  Note that this can
1768                  * potentially cause a busy-wait loop from writeback threads
1769                  * and kswapd activity, but those code paths have their own
1770                  * higher-level throttling.
1771                  */
1772                 if (wbc->sync_mode != WB_SYNC_NONE) {
1773                         lock_buffer(bh);
1774                 } else if (!trylock_buffer(bh)) {
1775                         redirty_page_for_writepage(wbc, page);
1776                         continue;
1777                 }
1778                 if (test_clear_buffer_dirty(bh)) {
1779                         mark_buffer_async_write_endio(bh, handler);
1780                 } else {
1781                         unlock_buffer(bh);
1782                 }
1783         } while ((bh = bh->b_this_page) != head);
1784
1785         /*
1786          * The page and its buffers are protected by PageWriteback(), so we can
1787          * drop the bh refcounts early.
1788          */
1789         BUG_ON(PageWriteback(page));
1790         set_page_writeback(page);
1791
1792         do {
1793                 struct buffer_head *next = bh->b_this_page;
1794                 if (buffer_async_write(bh)) {
1795                         submit_bh(write_op, bh);
1796                         nr_underway++;
1797                 }
1798                 bh = next;
1799         } while (bh != head);
1800         unlock_page(page);
1801
1802         err = 0;
1803 done:
1804         if (nr_underway == 0) {
1805                 /*
1806                  * The page was marked dirty, but the buffers were
1807                  * clean.  Someone wrote them back by hand with
1808                  * ll_rw_block/submit_bh.  A rare case.
1809                  */
1810                 end_page_writeback(page);
1811
1812                 /*
1813                  * The page and buffer_heads can be released at any time from
1814                  * here on.
1815                  */
1816         }
1817         return err;
1818
1819 recover:
1820         /*
1821          * ENOSPC, or some other error.  We may already have added some
1822          * blocks to the file, so we need to write these out to avoid
1823          * exposing stale data.
1824          * The page is currently locked and not marked for writeback
1825          */
1826         bh = head;
1827         /* Recovery: lock and submit the mapped buffers */
1828         do {
1829                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1830                     !buffer_delay(bh)) {
1831                         lock_buffer(bh);
1832                         mark_buffer_async_write_endio(bh, handler);
1833                 } else {
1834                         /*
1835                          * The buffer may have been set dirty during
1836                          * attachment to a dirty page.
1837                          */
1838                         clear_buffer_dirty(bh);
1839                 }
1840         } while ((bh = bh->b_this_page) != head);
1841         SetPageError(page);
1842         BUG_ON(PageWriteback(page));
1843         mapping_set_error(page->mapping, err);
1844         set_page_writeback(page);
1845         do {
1846                 struct buffer_head *next = bh->b_this_page;
1847                 if (buffer_async_write(bh)) {
1848                         clear_buffer_dirty(bh);
1849                         submit_bh(write_op, bh);
1850                         nr_underway++;
1851                 }
1852                 bh = next;
1853         } while (bh != head);
1854         unlock_page(page);
1855         goto done;
1856 }
1857
1858 /*
1859  * If a page has any new buffers, zero them out here, and mark them uptodate
1860  * and dirty so they'll be written out (in order to prevent uninitialised
1861  * block data from leaking). And clear the new bit.
1862  */
1863 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1864 {
1865         unsigned int block_start, block_end;
1866         struct buffer_head *head, *bh;
1867
1868         BUG_ON(!PageLocked(page));
1869         if (!page_has_buffers(page))
1870                 return;
1871
1872         bh = head = page_buffers(page);
1873         block_start = 0;
1874         do {
1875                 block_end = block_start + bh->b_size;
1876
1877                 if (buffer_new(bh)) {
1878                         if (block_end > from && block_start < to) {
1879                                 if (!PageUptodate(page)) {
1880                                         unsigned start, size;
1881
1882                                         start = max(from, block_start);
1883                                         size = min(to, block_end) - start;
1884
1885                                         zero_user(page, start, size);
1886                                         set_buffer_uptodate(bh);
1887                                 }
1888
1889                                 clear_buffer_new(bh);
1890                                 mark_buffer_dirty(bh);
1891                         }
1892                 }
1893
1894                 block_start = block_end;
1895                 bh = bh->b_this_page;
1896         } while (bh != head);
1897 }
1898 EXPORT_SYMBOL(page_zero_new_buffers);
1899
1900 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1901                 get_block_t *get_block)
1902 {
1903         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1904         unsigned to = from + len;
1905         struct inode *inode = page->mapping->host;
1906         unsigned block_start, block_end;
1907         sector_t block;
1908         int err = 0;
1909         unsigned blocksize, bbits;
1910         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1911
1912         BUG_ON(!PageLocked(page));
1913         BUG_ON(from > PAGE_CACHE_SIZE);
1914         BUG_ON(to > PAGE_CACHE_SIZE);
1915         BUG_ON(from > to);
1916
1917         head = create_page_buffers(page, inode, 0);
1918         blocksize = head->b_size;
1919         bbits = block_size_bits(blocksize);
1920
1921         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1922
1923         for(bh = head, block_start = 0; bh != head || !block_start;
1924             block++, block_start=block_end, bh = bh->b_this_page) {
1925                 block_end = block_start + blocksize;
1926                 if (block_end <= from || block_start >= to) {
1927                         if (PageUptodate(page)) {
1928                                 if (!buffer_uptodate(bh))
1929                                         set_buffer_uptodate(bh);
1930                         }
1931                         continue;
1932                 }
1933                 if (buffer_new(bh))
1934                         clear_buffer_new(bh);
1935                 if (!buffer_mapped(bh)) {
1936                         WARN_ON(bh->b_size != blocksize);
1937                         err = get_block(inode, block, bh, 1);
1938                         if (err)
1939                                 break;
1940                         if (buffer_new(bh)) {
1941                                 unmap_underlying_metadata(bh->b_bdev,
1942                                                         bh->b_blocknr);
1943                                 if (PageUptodate(page)) {
1944                                         clear_buffer_new(bh);
1945                                         set_buffer_uptodate(bh);
1946                                         mark_buffer_dirty(bh);
1947                                         continue;
1948                                 }
1949                                 if (block_end > to || block_start < from)
1950                                         zero_user_segments(page,
1951                                                 to, block_end,
1952                                                 block_start, from);
1953                                 continue;
1954                         }
1955                 }
1956                 if (PageUptodate(page)) {
1957                         if (!buffer_uptodate(bh))
1958                                 set_buffer_uptodate(bh);
1959                         continue; 
1960                 }
1961                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1962                     !buffer_unwritten(bh) &&
1963                      (block_start < from || block_end > to)) {
1964                         ll_rw_block(READ, 1, &bh);
1965                         *wait_bh++=bh;
1966                 }
1967         }
1968         /*
1969          * If we issued read requests - let them complete.
1970          */
1971         while(wait_bh > wait) {
1972                 wait_on_buffer(*--wait_bh);
1973                 if (!buffer_uptodate(*wait_bh))
1974                         err = -EIO;
1975         }
1976         if (unlikely(err))
1977                 page_zero_new_buffers(page, from, to);
1978         return err;
1979 }
1980 EXPORT_SYMBOL(__block_write_begin);
1981
1982 static int __block_commit_write(struct inode *inode, struct page *page,
1983                 unsigned from, unsigned to)
1984 {
1985         unsigned block_start, block_end;
1986         int partial = 0;
1987         unsigned blocksize;
1988         struct buffer_head *bh, *head;
1989
1990         bh = head = page_buffers(page);
1991         blocksize = bh->b_size;
1992
1993         block_start = 0;
1994         do {
1995                 block_end = block_start + blocksize;
1996                 if (block_end <= from || block_start >= to) {
1997                         if (!buffer_uptodate(bh))
1998                                 partial = 1;
1999                 } else {
2000                         set_buffer_uptodate(bh);
2001                         mark_buffer_dirty(bh);
2002                 }
2003                 clear_buffer_new(bh);
2004
2005                 block_start = block_end;
2006                 bh = bh->b_this_page;
2007         } while (bh != head);
2008
2009         /*
2010          * If this is a partial write which happened to make all buffers
2011          * uptodate then we can optimize away a bogus readpage() for
2012          * the next read(). Here we 'discover' whether the page went
2013          * uptodate as a result of this (potentially partial) write.
2014          */
2015         if (!partial)
2016                 SetPageUptodate(page);
2017         return 0;
2018 }
2019
2020 /*
2021  * block_write_begin takes care of the basic task of block allocation and
2022  * bringing partial write blocks uptodate first.
2023  *
2024  * The filesystem needs to handle block truncation upon failure.
2025  */
2026 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2027                 unsigned flags, struct page **pagep, get_block_t *get_block)
2028 {
2029         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
2030         struct page *page;
2031         int status;
2032
2033         page = grab_cache_page_write_begin(mapping, index, flags);
2034         if (!page)
2035                 return -ENOMEM;
2036
2037         status = __block_write_begin(page, pos, len, get_block);
2038         if (unlikely(status)) {
2039                 unlock_page(page);
2040                 page_cache_release(page);
2041                 page = NULL;
2042         }
2043
2044         *pagep = page;
2045         return status;
2046 }
2047 EXPORT_SYMBOL(block_write_begin);
2048
2049 int block_write_end(struct file *file, struct address_space *mapping,
2050                         loff_t pos, unsigned len, unsigned copied,
2051                         struct page *page, void *fsdata)
2052 {
2053         struct inode *inode = mapping->host;
2054         unsigned start;
2055
2056         start = pos & (PAGE_CACHE_SIZE - 1);
2057
2058         if (unlikely(copied < len)) {
2059                 /*
2060                  * The buffers that were written will now be uptodate, so we
2061                  * don't have to worry about a readpage reading them and
2062                  * overwriting a partial write. However if we have encountered
2063                  * a short write and only partially written into a buffer, it
2064                  * will not be marked uptodate, so a readpage might come in and
2065                  * destroy our partial write.
2066                  *
2067                  * Do the simplest thing, and just treat any short write to a
2068                  * non uptodate page as a zero-length write, and force the
2069                  * caller to redo the whole thing.
2070                  */
2071                 if (!PageUptodate(page))
2072                         copied = 0;
2073
2074                 page_zero_new_buffers(page, start+copied, start+len);
2075         }
2076         flush_dcache_page(page);
2077
2078         /* This could be a short (even 0-length) commit */
2079         __block_commit_write(inode, page, start, start+copied);
2080
2081         return copied;
2082 }
2083 EXPORT_SYMBOL(block_write_end);
2084
2085 int generic_write_end(struct file *file, struct address_space *mapping,
2086                         loff_t pos, unsigned len, unsigned copied,
2087                         struct page *page, void *fsdata)
2088 {
2089         struct inode *inode = mapping->host;
2090         int i_size_changed = 0;
2091
2092         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2093
2094         /*
2095          * No need to use i_size_read() here, the i_size
2096          * cannot change under us because we hold i_mutex.
2097          *
2098          * But it's important to update i_size while still holding page lock:
2099          * page writeout could otherwise come in and zero beyond i_size.
2100          */
2101         if (pos+copied > inode->i_size) {
2102                 i_size_write(inode, pos+copied);
2103                 i_size_changed = 1;
2104         }
2105
2106         unlock_page(page);
2107         page_cache_release(page);
2108
2109         /*
2110          * Don't mark the inode dirty under page lock. First, it unnecessarily
2111          * makes the holding time of page lock longer. Second, it forces lock
2112          * ordering of page lock and transaction start for journaling
2113          * filesystems.
2114          */
2115         if (i_size_changed)
2116                 mark_inode_dirty(inode);
2117
2118         return copied;
2119 }
2120 EXPORT_SYMBOL(generic_write_end);
2121
2122 /*
2123  * block_is_partially_uptodate checks whether buffers within a page are
2124  * uptodate or not.
2125  *
2126  * Returns true if all buffers which correspond to a file portion
2127  * we want to read are uptodate.
2128  */
2129 int block_is_partially_uptodate(struct page *page, unsigned long from,
2130                                         unsigned long count)
2131 {
2132         unsigned block_start, block_end, blocksize;
2133         unsigned to;
2134         struct buffer_head *bh, *head;
2135         int ret = 1;
2136
2137         if (!page_has_buffers(page))
2138                 return 0;
2139
2140         head = page_buffers(page);
2141         blocksize = head->b_size;
2142         to = min_t(unsigned, PAGE_CACHE_SIZE - from, count);
2143         to = from + to;
2144         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2145                 return 0;
2146
2147         bh = head;
2148         block_start = 0;
2149         do {
2150                 block_end = block_start + blocksize;
2151                 if (block_end > from && block_start < to) {
2152                         if (!buffer_uptodate(bh)) {
2153                                 ret = 0;
2154                                 break;
2155                         }
2156                         if (block_end >= to)
2157                                 break;
2158                 }
2159                 block_start = block_end;
2160                 bh = bh->b_this_page;
2161         } while (bh != head);
2162
2163         return ret;
2164 }
2165 EXPORT_SYMBOL(block_is_partially_uptodate);
2166
2167 /*
2168  * Generic "read page" function for block devices that have the normal
2169  * get_block functionality. This is most of the block device filesystems.
2170  * Reads the page asynchronously --- the unlock_buffer() and
2171  * set/clear_buffer_uptodate() functions propagate buffer state into the
2172  * page struct once IO has completed.
2173  */
2174 int block_read_full_page(struct page *page, get_block_t *get_block)
2175 {
2176         struct inode *inode = page->mapping->host;
2177         sector_t iblock, lblock;
2178         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2179         unsigned int blocksize, bbits;
2180         int nr, i;
2181         int fully_mapped = 1;
2182
2183         head = create_page_buffers(page, inode, 0);
2184         blocksize = head->b_size;
2185         bbits = block_size_bits(blocksize);
2186
2187         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
2188         lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2189         bh = head;
2190         nr = 0;
2191         i = 0;
2192
2193         do {
2194                 if (buffer_uptodate(bh))
2195                         continue;
2196
2197                 if (!buffer_mapped(bh)) {
2198                         int err = 0;
2199
2200                         fully_mapped = 0;
2201                         if (iblock < lblock) {
2202                                 WARN_ON(bh->b_size != blocksize);
2203                                 err = get_block(inode, iblock, bh, 0);
2204                                 if (err)
2205                                         SetPageError(page);
2206                         }
2207                         if (!buffer_mapped(bh)) {
2208                                 zero_user(page, i * blocksize, blocksize);
2209                                 if (!err)
2210                                         set_buffer_uptodate(bh);
2211                                 continue;
2212                         }
2213                         /*
2214                          * get_block() might have updated the buffer
2215                          * synchronously
2216                          */
2217                         if (buffer_uptodate(bh))
2218                                 continue;
2219                 }
2220                 arr[nr++] = bh;
2221         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2222
2223         if (fully_mapped)
2224                 SetPageMappedToDisk(page);
2225
2226         if (!nr) {
2227                 /*
2228                  * All buffers are uptodate - we can set the page uptodate
2229                  * as well. But not if get_block() returned an error.
2230                  */
2231                 if (!PageError(page))
2232                         SetPageUptodate(page);
2233                 unlock_page(page);
2234                 return 0;
2235         }
2236
2237         /* Stage two: lock the buffers */
2238         for (i = 0; i < nr; i++) {
2239                 bh = arr[i];
2240                 lock_buffer(bh);
2241                 mark_buffer_async_read(bh);
2242         }
2243
2244         /*
2245          * Stage 3: start the IO.  Check for uptodateness
2246          * inside the buffer lock in case another process reading
2247          * the underlying blockdev brought it uptodate (the sct fix).
2248          */
2249         for (i = 0; i < nr; i++) {
2250                 bh = arr[i];
2251                 if (buffer_uptodate(bh))
2252                         end_buffer_async_read(bh, 1);
2253                 else
2254                         submit_bh(READ, bh);
2255         }
2256         return 0;
2257 }
2258 EXPORT_SYMBOL(block_read_full_page);
2259
2260 /* utility function for filesystems that need to do work on expanding
2261  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2262  * deal with the hole.  
2263  */
2264 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2265 {
2266         struct address_space *mapping = inode->i_mapping;
2267         struct page *page;
2268         void *fsdata;
2269         int err;
2270
2271         err = inode_newsize_ok(inode, size);
2272         if (err)
2273                 goto out;
2274
2275         err = pagecache_write_begin(NULL, mapping, size, 0,
2276                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2277                                 &page, &fsdata);
2278         if (err)
2279                 goto out;
2280
2281         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2282         BUG_ON(err > 0);
2283
2284 out:
2285         return err;
2286 }
2287 EXPORT_SYMBOL(generic_cont_expand_simple);
2288
2289 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2290                             loff_t pos, loff_t *bytes)
2291 {
2292         struct inode *inode = mapping->host;
2293         unsigned blocksize = 1 << inode->i_blkbits;
2294         struct page *page;
2295         void *fsdata;
2296         pgoff_t index, curidx;
2297         loff_t curpos;
2298         unsigned zerofrom, offset, len;
2299         int err = 0;
2300
2301         index = pos >> PAGE_CACHE_SHIFT;
2302         offset = pos & ~PAGE_CACHE_MASK;
2303
2304         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2305                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2306                 if (zerofrom & (blocksize-1)) {
2307                         *bytes |= (blocksize-1);
2308                         (*bytes)++;
2309                 }
2310                 len = PAGE_CACHE_SIZE - zerofrom;
2311
2312                 err = pagecache_write_begin(file, mapping, curpos, len,
2313                                                 AOP_FLAG_UNINTERRUPTIBLE,
2314                                                 &page, &fsdata);
2315                 if (err)
2316                         goto out;
2317                 zero_user(page, zerofrom, len);
2318                 err = pagecache_write_end(file, mapping, curpos, len, len,
2319                                                 page, fsdata);
2320                 if (err < 0)
2321                         goto out;
2322                 BUG_ON(err != len);
2323                 err = 0;
2324
2325                 balance_dirty_pages_ratelimited(mapping);
2326         }
2327
2328         /* page covers the boundary, find the boundary offset */
2329         if (index == curidx) {
2330                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2331                 /* if we will expand the thing last block will be filled */
2332                 if (offset <= zerofrom) {
2333                         goto out;
2334                 }
2335                 if (zerofrom & (blocksize-1)) {
2336                         *bytes |= (blocksize-1);
2337                         (*bytes)++;
2338                 }
2339                 len = offset - zerofrom;
2340
2341                 err = pagecache_write_begin(file, mapping, curpos, len,
2342                                                 AOP_FLAG_UNINTERRUPTIBLE,
2343                                                 &page, &fsdata);
2344                 if (err)
2345                         goto out;
2346                 zero_user(page, zerofrom, len);
2347                 err = pagecache_write_end(file, mapping, curpos, len, len,
2348                                                 page, fsdata);
2349                 if (err < 0)
2350                         goto out;
2351                 BUG_ON(err != len);
2352                 err = 0;
2353         }
2354 out:
2355         return err;
2356 }
2357
2358 /*
2359  * For moronic filesystems that do not allow holes in file.
2360  * We may have to extend the file.
2361  */
2362 int cont_write_begin(struct file *file, struct address_space *mapping,
2363                         loff_t pos, unsigned len, unsigned flags,
2364                         struct page **pagep, void **fsdata,
2365                         get_block_t *get_block, loff_t *bytes)
2366 {
2367         struct inode *inode = mapping->host;
2368         unsigned blocksize = 1 << inode->i_blkbits;
2369         unsigned zerofrom;
2370         int err;
2371
2372         err = cont_expand_zero(file, mapping, pos, bytes);
2373         if (err)
2374                 return err;
2375
2376         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2377         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2378                 *bytes |= (blocksize-1);
2379                 (*bytes)++;
2380         }
2381
2382         return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2383 }
2384 EXPORT_SYMBOL(cont_write_begin);
2385
2386 int block_commit_write(struct page *page, unsigned from, unsigned to)
2387 {
2388         struct inode *inode = page->mapping->host;
2389         __block_commit_write(inode,page,from,to);
2390         return 0;
2391 }
2392 EXPORT_SYMBOL(block_commit_write);
2393
2394 /*
2395  * block_page_mkwrite() is not allowed to change the file size as it gets
2396  * called from a page fault handler when a page is first dirtied. Hence we must
2397  * be careful to check for EOF conditions here. We set the page up correctly
2398  * for a written page which means we get ENOSPC checking when writing into
2399  * holes and correct delalloc and unwritten extent mapping on filesystems that
2400  * support these features.
2401  *
2402  * We are not allowed to take the i_mutex here so we have to play games to
2403  * protect against truncate races as the page could now be beyond EOF.  Because
2404  * truncate writes the inode size before removing pages, once we have the
2405  * page lock we can determine safely if the page is beyond EOF. If it is not
2406  * beyond EOF, then the page is guaranteed safe against truncation until we
2407  * unlock the page.
2408  *
2409  * Direct callers of this function should protect against filesystem freezing
2410  * using sb_start_write() - sb_end_write() functions.
2411  */
2412 int __block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2413                          get_block_t get_block)
2414 {
2415         struct page *page = vmf->page;
2416         struct inode *inode = file_inode(vma->vm_file);
2417         unsigned long end;
2418         loff_t size;
2419         int ret;
2420
2421         lock_page(page);
2422         size = i_size_read(inode);
2423         if ((page->mapping != inode->i_mapping) ||
2424             (page_offset(page) > size)) {
2425                 /* We overload EFAULT to mean page got truncated */
2426                 ret = -EFAULT;
2427                 goto out_unlock;
2428         }
2429
2430         /* page is wholly or partially inside EOF */
2431         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2432                 end = size & ~PAGE_CACHE_MASK;
2433         else
2434                 end = PAGE_CACHE_SIZE;
2435
2436         ret = __block_write_begin(page, 0, end, get_block);
2437         if (!ret)
2438                 ret = block_commit_write(page, 0, end);
2439
2440         if (unlikely(ret < 0))
2441                 goto out_unlock;
2442         set_page_dirty(page);
2443         wait_for_stable_page(page);
2444         return 0;
2445 out_unlock:
2446         unlock_page(page);
2447         return ret;
2448 }
2449 EXPORT_SYMBOL(__block_page_mkwrite);
2450
2451 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2452                    get_block_t get_block)
2453 {
2454         int ret;
2455         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
2456
2457         sb_start_pagefault(sb);
2458
2459         /*
2460          * Update file times before taking page lock. We may end up failing the
2461          * fault so this update may be superfluous but who really cares...
2462          */
2463         file_update_time(vma->vm_file);
2464
2465         ret = __block_page_mkwrite(vma, vmf, get_block);
2466         sb_end_pagefault(sb);
2467         return block_page_mkwrite_return(ret);
2468 }
2469 EXPORT_SYMBOL(block_page_mkwrite);
2470
2471 /*
2472  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2473  * immediately, while under the page lock.  So it needs a special end_io
2474  * handler which does not touch the bh after unlocking it.
2475  */
2476 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2477 {
2478         __end_buffer_read_notouch(bh, uptodate);
2479 }
2480
2481 /*
2482  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2483  * the page (converting it to circular linked list and taking care of page
2484  * dirty races).
2485  */
2486 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2487 {
2488         struct buffer_head *bh;
2489
2490         BUG_ON(!PageLocked(page));
2491
2492         spin_lock(&page->mapping->private_lock);
2493         bh = head;
2494         do {
2495                 if (PageDirty(page))
2496                         set_buffer_dirty(bh);
2497                 if (!bh->b_this_page)
2498                         bh->b_this_page = head;
2499                 bh = bh->b_this_page;
2500         } while (bh != head);
2501         attach_page_buffers(page, head);
2502         spin_unlock(&page->mapping->private_lock);
2503 }
2504
2505 /*
2506  * On entry, the page is fully not uptodate.
2507  * On exit the page is fully uptodate in the areas outside (from,to)
2508  * The filesystem needs to handle block truncation upon failure.
2509  */
2510 int nobh_write_begin(struct address_space *mapping,
2511                         loff_t pos, unsigned len, unsigned flags,
2512                         struct page **pagep, void **fsdata,
2513                         get_block_t *get_block)
2514 {
2515         struct inode *inode = mapping->host;
2516         const unsigned blkbits = inode->i_blkbits;
2517         const unsigned blocksize = 1 << blkbits;
2518         struct buffer_head *head, *bh;
2519         struct page *page;
2520         pgoff_t index;
2521         unsigned from, to;
2522         unsigned block_in_page;
2523         unsigned block_start, block_end;
2524         sector_t block_in_file;
2525         int nr_reads = 0;
2526         int ret = 0;
2527         int is_mapped_to_disk = 1;
2528
2529         index = pos >> PAGE_CACHE_SHIFT;
2530         from = pos & (PAGE_CACHE_SIZE - 1);
2531         to = from + len;
2532
2533         page = grab_cache_page_write_begin(mapping, index, flags);
2534         if (!page)
2535                 return -ENOMEM;
2536         *pagep = page;
2537         *fsdata = NULL;
2538
2539         if (page_has_buffers(page)) {
2540                 ret = __block_write_begin(page, pos, len, get_block);
2541                 if (unlikely(ret))
2542                         goto out_release;
2543                 return ret;
2544         }
2545
2546         if (PageMappedToDisk(page))
2547                 return 0;
2548
2549         /*
2550          * Allocate buffers so that we can keep track of state, and potentially
2551          * attach them to the page if an error occurs. In the common case of
2552          * no error, they will just be freed again without ever being attached
2553          * to the page (which is all OK, because we're under the page lock).
2554          *
2555          * Be careful: the buffer linked list is a NULL terminated one, rather
2556          * than the circular one we're used to.
2557          */
2558         head = alloc_page_buffers(page, blocksize, 0);
2559         if (!head) {
2560                 ret = -ENOMEM;
2561                 goto out_release;
2562         }
2563
2564         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2565
2566         /*
2567          * We loop across all blocks in the page, whether or not they are
2568          * part of the affected region.  This is so we can discover if the
2569          * page is fully mapped-to-disk.
2570          */
2571         for (block_start = 0, block_in_page = 0, bh = head;
2572                   block_start < PAGE_CACHE_SIZE;
2573                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2574                 int create;
2575
2576                 block_end = block_start + blocksize;
2577                 bh->b_state = 0;
2578                 create = 1;
2579                 if (block_start >= to)
2580                         create = 0;
2581                 ret = get_block(inode, block_in_file + block_in_page,
2582                                         bh, create);
2583                 if (ret)
2584                         goto failed;
2585                 if (!buffer_mapped(bh))
2586                         is_mapped_to_disk = 0;
2587                 if (buffer_new(bh))
2588                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2589                 if (PageUptodate(page)) {
2590                         set_buffer_uptodate(bh);
2591                         continue;
2592                 }
2593                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2594                         zero_user_segments(page, block_start, from,
2595                                                         to, block_end);
2596                         continue;
2597                 }
2598                 if (buffer_uptodate(bh))
2599                         continue;       /* reiserfs does this */
2600                 if (block_start < from || block_end > to) {
2601                         lock_buffer(bh);
2602                         bh->b_end_io = end_buffer_read_nobh;
2603                         submit_bh(READ, bh);
2604                         nr_reads++;
2605                 }
2606         }
2607
2608         if (nr_reads) {
2609                 /*
2610                  * The page is locked, so these buffers are protected from
2611                  * any VM or truncate activity.  Hence we don't need to care
2612                  * for the buffer_head refcounts.
2613                  */
2614                 for (bh = head; bh; bh = bh->b_this_page) {
2615                         wait_on_buffer(bh);
2616                         if (!buffer_uptodate(bh))
2617                                 ret = -EIO;
2618                 }
2619                 if (ret)
2620                         goto failed;
2621         }
2622
2623         if (is_mapped_to_disk)
2624                 SetPageMappedToDisk(page);
2625
2626         *fsdata = head; /* to be released by nobh_write_end */
2627
2628         return 0;
2629
2630 failed:
2631         BUG_ON(!ret);
2632         /*
2633          * Error recovery is a bit difficult. We need to zero out blocks that
2634          * were newly allocated, and dirty them to ensure they get written out.
2635          * Buffers need to be attached to the page at this point, otherwise
2636          * the handling of potential IO errors during writeout would be hard
2637          * (could try doing synchronous writeout, but what if that fails too?)
2638          */
2639         attach_nobh_buffers(page, head);
2640         page_zero_new_buffers(page, from, to);
2641
2642 out_release:
2643         unlock_page(page);
2644         page_cache_release(page);
2645         *pagep = NULL;
2646
2647         return ret;
2648 }
2649 EXPORT_SYMBOL(nobh_write_begin);
2650
2651 int nobh_write_end(struct file *file, struct address_space *mapping,
2652                         loff_t pos, unsigned len, unsigned copied,
2653                         struct page *page, void *fsdata)
2654 {
2655         struct inode *inode = page->mapping->host;
2656         struct buffer_head *head = fsdata;
2657         struct buffer_head *bh;
2658         BUG_ON(fsdata != NULL && page_has_buffers(page));
2659
2660         if (unlikely(copied < len) && head)
2661                 attach_nobh_buffers(page, head);
2662         if (page_has_buffers(page))
2663                 return generic_write_end(file, mapping, pos, len,
2664                                         copied, page, fsdata);
2665
2666         SetPageUptodate(page);
2667         set_page_dirty(page);
2668         if (pos+copied > inode->i_size) {
2669                 i_size_write(inode, pos+copied);
2670                 mark_inode_dirty(inode);
2671         }
2672
2673         unlock_page(page);
2674         page_cache_release(page);
2675
2676         while (head) {
2677                 bh = head;
2678                 head = head->b_this_page;
2679                 free_buffer_head(bh);
2680         }
2681
2682         return copied;
2683 }
2684 EXPORT_SYMBOL(nobh_write_end);
2685
2686 /*
2687  * nobh_writepage() - based on block_full_write_page() except
2688  * that it tries to operate without attaching bufferheads to
2689  * the page.
2690  */
2691 int nobh_writepage(struct page *page, get_block_t *get_block,
2692                         struct writeback_control *wbc)
2693 {
2694         struct inode * const inode = page->mapping->host;
2695         loff_t i_size = i_size_read(inode);
2696         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2697         unsigned offset;
2698         int ret;
2699
2700         /* Is the page fully inside i_size? */
2701         if (page->index < end_index)
2702                 goto out;
2703
2704         /* Is the page fully outside i_size? (truncate in progress) */
2705         offset = i_size & (PAGE_CACHE_SIZE-1);
2706         if (page->index >= end_index+1 || !offset) {
2707                 /*
2708                  * The page may have dirty, unmapped buffers.  For example,
2709                  * they may have been added in ext3_writepage().  Make them
2710                  * freeable here, so the page does not leak.
2711                  */
2712 #if 0
2713                 /* Not really sure about this  - do we need this ? */
2714                 if (page->mapping->a_ops->invalidatepage)
2715                         page->mapping->a_ops->invalidatepage(page, offset);
2716 #endif
2717                 unlock_page(page);
2718                 return 0; /* don't care */
2719         }
2720
2721         /*
2722          * The page straddles i_size.  It must be zeroed out on each and every
2723          * writepage invocation because it may be mmapped.  "A file is mapped
2724          * in multiples of the page size.  For a file that is not a multiple of
2725          * the  page size, the remaining memory is zeroed when mapped, and
2726          * writes to that region are not written out to the file."
2727          */
2728         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2729 out:
2730         ret = mpage_writepage(page, get_block, wbc);
2731         if (ret == -EAGAIN)
2732                 ret = __block_write_full_page(inode, page, get_block, wbc,
2733                                               end_buffer_async_write);
2734         return ret;
2735 }
2736 EXPORT_SYMBOL(nobh_writepage);
2737
2738 int nobh_truncate_page(struct address_space *mapping,
2739                         loff_t from, get_block_t *get_block)
2740 {
2741         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2742         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2743         unsigned blocksize;
2744         sector_t iblock;
2745         unsigned length, pos;
2746         struct inode *inode = mapping->host;
2747         struct page *page;
2748         struct buffer_head map_bh;
2749         int err;
2750
2751         blocksize = 1 << inode->i_blkbits;
2752         length = offset & (blocksize - 1);
2753
2754         /* Block boundary? Nothing to do */
2755         if (!length)
2756                 return 0;
2757
2758         length = blocksize - length;
2759         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2760
2761         page = grab_cache_page(mapping, index);
2762         err = -ENOMEM;
2763         if (!page)
2764                 goto out;
2765
2766         if (page_has_buffers(page)) {
2767 has_buffers:
2768                 unlock_page(page);
2769                 page_cache_release(page);
2770                 return block_truncate_page(mapping, from, get_block);
2771         }
2772
2773         /* Find the buffer that contains "offset" */
2774         pos = blocksize;
2775         while (offset >= pos) {
2776                 iblock++;
2777                 pos += blocksize;
2778         }
2779
2780         map_bh.b_size = blocksize;
2781         map_bh.b_state = 0;
2782         err = get_block(inode, iblock, &map_bh, 0);
2783         if (err)
2784                 goto unlock;
2785         /* unmapped? It's a hole - nothing to do */
2786         if (!buffer_mapped(&map_bh))
2787                 goto unlock;
2788
2789         /* Ok, it's mapped. Make sure it's up-to-date */
2790         if (!PageUptodate(page)) {
2791                 err = mapping->a_ops->readpage(NULL, page);
2792                 if (err) {
2793                         page_cache_release(page);
2794                         goto out;
2795                 }
2796                 lock_page(page);
2797                 if (!PageUptodate(page)) {
2798                         err = -EIO;
2799                         goto unlock;
2800                 }
2801                 if (page_has_buffers(page))
2802                         goto has_buffers;
2803         }
2804         zero_user(page, offset, length);
2805         set_page_dirty(page);
2806         err = 0;
2807
2808 unlock:
2809         unlock_page(page);
2810         page_cache_release(page);
2811 out:
2812         return err;
2813 }
2814 EXPORT_SYMBOL(nobh_truncate_page);
2815
2816 int block_truncate_page(struct address_space *mapping,
2817                         loff_t from, get_block_t *get_block)
2818 {
2819         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2820         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2821         unsigned blocksize;
2822         sector_t iblock;
2823         unsigned length, pos;
2824         struct inode *inode = mapping->host;
2825         struct page *page;
2826         struct buffer_head *bh;
2827         int err;
2828
2829         blocksize = 1 << inode->i_blkbits;
2830         length = offset & (blocksize - 1);
2831
2832         /* Block boundary? Nothing to do */
2833         if (!length)
2834                 return 0;
2835
2836         length = blocksize - length;
2837         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2838         
2839         page = grab_cache_page(mapping, index);
2840         err = -ENOMEM;
2841         if (!page)
2842                 goto out;
2843
2844         if (!page_has_buffers(page))
2845                 create_empty_buffers(page, blocksize, 0);
2846
2847         /* Find the buffer that contains "offset" */
2848         bh = page_buffers(page);
2849         pos = blocksize;
2850         while (offset >= pos) {
2851                 bh = bh->b_this_page;
2852                 iblock++;
2853                 pos += blocksize;
2854         }
2855
2856         err = 0;
2857         if (!buffer_mapped(bh)) {
2858                 WARN_ON(bh->b_size != blocksize);
2859                 err = get_block(inode, iblock, bh, 0);
2860                 if (err)
2861                         goto unlock;
2862                 /* unmapped? It's a hole - nothing to do */
2863                 if (!buffer_mapped(bh))
2864                         goto unlock;
2865         }
2866
2867         /* Ok, it's mapped. Make sure it's up-to-date */
2868         if (PageUptodate(page))
2869                 set_buffer_uptodate(bh);
2870
2871         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2872                 err = -EIO;
2873                 ll_rw_block(READ, 1, &bh);
2874                 wait_on_buffer(bh);
2875                 /* Uhhuh. Read error. Complain and punt. */
2876                 if (!buffer_uptodate(bh))
2877                         goto unlock;
2878         }
2879
2880         zero_user(page, offset, length);
2881         mark_buffer_dirty(bh);
2882         err = 0;
2883
2884 unlock:
2885         unlock_page(page);
2886         page_cache_release(page);
2887 out:
2888         return err;
2889 }
2890 EXPORT_SYMBOL(block_truncate_page);
2891
2892 /*
2893  * The generic ->writepage function for buffer-backed address_spaces
2894  */
2895 int block_write_full_page(struct page *page, get_block_t *get_block,
2896                         struct writeback_control *wbc)
2897 {
2898         struct inode * const inode = page->mapping->host;
2899         loff_t i_size = i_size_read(inode);
2900         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2901         unsigned offset;
2902
2903         /* Is the page fully inside i_size? */
2904         if (page->index < end_index)
2905                 return __block_write_full_page(inode, page, get_block, wbc,
2906                                                end_buffer_async_write);
2907
2908         /* Is the page fully outside i_size? (truncate in progress) */
2909         offset = i_size & (PAGE_CACHE_SIZE-1);
2910         if (page->index >= end_index+1 || !offset) {
2911                 /*
2912                  * The page may have dirty, unmapped buffers.  For example,
2913                  * they may have been added in ext3_writepage().  Make them
2914                  * freeable here, so the page does not leak.
2915                  */
2916                 do_invalidatepage(page, 0, PAGE_CACHE_SIZE);
2917                 unlock_page(page);
2918                 return 0; /* don't care */
2919         }
2920
2921         /*
2922          * The page straddles i_size.  It must be zeroed out on each and every
2923          * writepage invocation because it may be mmapped.  "A file is mapped
2924          * in multiples of the page size.  For a file that is not a multiple of
2925          * the  page size, the remaining memory is zeroed when mapped, and
2926          * writes to that region are not written out to the file."
2927          */
2928         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2929         return __block_write_full_page(inode, page, get_block, wbc,
2930                                                         end_buffer_async_write);
2931 }
2932 EXPORT_SYMBOL(block_write_full_page);
2933
2934 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2935                             get_block_t *get_block)
2936 {
2937         struct buffer_head tmp;
2938         struct inode *inode = mapping->host;
2939         tmp.b_state = 0;
2940         tmp.b_blocknr = 0;
2941         tmp.b_size = 1 << inode->i_blkbits;
2942         get_block(inode, block, &tmp, 0);
2943         return tmp.b_blocknr;
2944 }
2945 EXPORT_SYMBOL(generic_block_bmap);
2946
2947 static void end_bio_bh_io_sync(struct bio *bio, int err)
2948 {
2949         struct buffer_head *bh = bio->bi_private;
2950
2951         if (err == -EOPNOTSUPP) {
2952                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2953         }
2954
2955         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2956                 set_bit(BH_Quiet, &bh->b_state);
2957
2958         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2959         bio_put(bio);
2960 }
2961
2962 /*
2963  * This allows us to do IO even on the odd last sectors
2964  * of a device, even if the bh block size is some multiple
2965  * of the physical sector size.
2966  *
2967  * We'll just truncate the bio to the size of the device,
2968  * and clear the end of the buffer head manually.
2969  *
2970  * Truly out-of-range accesses will turn into actual IO
2971  * errors, this only handles the "we need to be able to
2972  * do IO at the final sector" case.
2973  */
2974 static void guard_bh_eod(int rw, struct bio *bio, struct buffer_head *bh)
2975 {
2976         sector_t maxsector;
2977         unsigned bytes;
2978
2979         maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
2980         if (!maxsector)
2981                 return;
2982
2983         /*
2984          * If the *whole* IO is past the end of the device,
2985          * let it through, and the IO layer will turn it into
2986          * an EIO.
2987          */
2988         if (unlikely(bio->bi_iter.bi_sector >= maxsector))
2989                 return;
2990
2991         maxsector -= bio->bi_iter.bi_sector;
2992         bytes = bio->bi_iter.bi_size;
2993         if (likely((bytes >> 9) <= maxsector))
2994                 return;
2995
2996         /* Uhhuh. We've got a bh that straddles the device size! */
2997         bytes = maxsector << 9;
2998
2999         /* Truncate the bio.. */
3000         bio->bi_iter.bi_size = bytes;
3001         bio->bi_io_vec[0].bv_len = bytes;
3002
3003         /* ..and clear the end of the buffer for reads */
3004         if ((rw & RW_MASK) == READ) {
3005                 void *kaddr = kmap_atomic(bh->b_page);
3006                 memset(kaddr + bh_offset(bh) + bytes, 0, bh->b_size - bytes);
3007                 kunmap_atomic(kaddr);
3008                 flush_dcache_page(bh->b_page);
3009         }
3010 }
3011
3012 int _submit_bh(int rw, struct buffer_head *bh, unsigned long bio_flags)
3013 {
3014         struct bio *bio;
3015         int ret = 0;
3016
3017         BUG_ON(!buffer_locked(bh));
3018         BUG_ON(!buffer_mapped(bh));
3019         BUG_ON(!bh->b_end_io);
3020         BUG_ON(buffer_delay(bh));
3021         BUG_ON(buffer_unwritten(bh));
3022
3023         /*
3024          * Only clear out a write error when rewriting
3025          */
3026         if (test_set_buffer_req(bh) && (rw & WRITE))
3027                 clear_buffer_write_io_error(bh);
3028
3029         /*
3030          * from here on down, it's all bio -- do the initial mapping,
3031          * submit_bio -> generic_make_request may further map this bio around
3032          */
3033         bio = bio_alloc(GFP_NOIO, 1);
3034
3035         bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3036         bio->bi_bdev = bh->b_bdev;
3037         bio->bi_io_vec[0].bv_page = bh->b_page;
3038         bio->bi_io_vec[0].bv_len = bh->b_size;
3039         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
3040
3041         bio->bi_vcnt = 1;
3042         bio->bi_iter.bi_size = bh->b_size;
3043
3044         bio->bi_end_io = end_bio_bh_io_sync;
3045         bio->bi_private = bh;
3046         bio->bi_flags |= bio_flags;
3047
3048         /* Take care of bh's that straddle the end of the device */
3049         guard_bh_eod(rw, bio, bh);
3050
3051         if (buffer_meta(bh))
3052                 rw |= REQ_META;
3053         if (buffer_prio(bh))
3054                 rw |= REQ_PRIO;
3055
3056         bio_get(bio);
3057         submit_bio(rw, bio);
3058
3059         if (bio_flagged(bio, BIO_EOPNOTSUPP))
3060                 ret = -EOPNOTSUPP;
3061
3062         bio_put(bio);
3063         return ret;
3064 }
3065 EXPORT_SYMBOL_GPL(_submit_bh);
3066
3067 int submit_bh(int rw, struct buffer_head *bh)
3068 {
3069         return _submit_bh(rw, bh, 0);
3070 }
3071 EXPORT_SYMBOL(submit_bh);
3072
3073 /**
3074  * ll_rw_block: low-level access to block devices (DEPRECATED)
3075  * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
3076  * @nr: number of &struct buffer_heads in the array
3077  * @bhs: array of pointers to &struct buffer_head
3078  *
3079  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3080  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
3081  * %READA option is described in the documentation for generic_make_request()
3082  * which ll_rw_block() calls.
3083  *
3084  * This function drops any buffer that it cannot get a lock on (with the
3085  * BH_Lock state bit), any buffer that appears to be clean when doing a write
3086  * request, and any buffer that appears to be up-to-date when doing read
3087  * request.  Further it marks as clean buffers that are processed for
3088  * writing (the buffer cache won't assume that they are actually clean
3089  * until the buffer gets unlocked).
3090  *
3091  * ll_rw_block sets b_end_io to simple completion handler that marks
3092  * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3093  * any waiters. 
3094  *
3095  * All of the buffers must be for the same device, and must also be a
3096  * multiple of the current approved size for the device.
3097  */
3098 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
3099 {
3100         int i;
3101
3102         for (i = 0; i < nr; i++) {
3103                 struct buffer_head *bh = bhs[i];
3104
3105                 if (!trylock_buffer(bh))
3106                         continue;
3107                 if (rw == WRITE) {
3108                         if (test_clear_buffer_dirty(bh)) {
3109                                 bh->b_end_io = end_buffer_write_sync;
3110                                 get_bh(bh);
3111                                 submit_bh(WRITE, bh);
3112                                 continue;
3113                         }
3114                 } else {
3115                         if (!buffer_uptodate(bh)) {
3116                                 bh->b_end_io = end_buffer_read_sync;
3117                                 get_bh(bh);
3118                                 submit_bh(rw, bh);
3119                                 continue;
3120                         }
3121                 }
3122                 unlock_buffer(bh);
3123         }
3124 }
3125 EXPORT_SYMBOL(ll_rw_block);
3126
3127 void write_dirty_buffer(struct buffer_head *bh, int rw)
3128 {
3129         lock_buffer(bh);
3130         if (!test_clear_buffer_dirty(bh)) {
3131                 unlock_buffer(bh);
3132                 return;
3133         }
3134         bh->b_end_io = end_buffer_write_sync;
3135         get_bh(bh);
3136         submit_bh(rw, bh);
3137 }
3138 EXPORT_SYMBOL(write_dirty_buffer);
3139
3140 /*
3141  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3142  * and then start new I/O and then wait upon it.  The caller must have a ref on
3143  * the buffer_head.
3144  */
3145 int __sync_dirty_buffer(struct buffer_head *bh, int rw)
3146 {
3147         int ret = 0;
3148
3149         WARN_ON(atomic_read(&bh->b_count) < 1);
3150         lock_buffer(bh);
3151         if (test_clear_buffer_dirty(bh)) {
3152                 get_bh(bh);
3153                 bh->b_end_io = end_buffer_write_sync;
3154                 ret = submit_bh(rw, bh);
3155                 wait_on_buffer(bh);
3156                 if (!ret && !buffer_uptodate(bh))
3157                         ret = -EIO;
3158         } else {
3159                 unlock_buffer(bh);
3160         }
3161         return ret;
3162 }
3163 EXPORT_SYMBOL(__sync_dirty_buffer);
3164
3165 int sync_dirty_buffer(struct buffer_head *bh)
3166 {
3167         return __sync_dirty_buffer(bh, WRITE_SYNC);
3168 }
3169 EXPORT_SYMBOL(sync_dirty_buffer);
3170
3171 /*
3172  * try_to_free_buffers() checks if all the buffers on this particular page
3173  * are unused, and releases them if so.
3174  *
3175  * Exclusion against try_to_free_buffers may be obtained by either
3176  * locking the page or by holding its mapping's private_lock.
3177  *
3178  * If the page is dirty but all the buffers are clean then we need to
3179  * be sure to mark the page clean as well.  This is because the page
3180  * may be against a block device, and a later reattachment of buffers
3181  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3182  * filesystem data on the same device.
3183  *
3184  * The same applies to regular filesystem pages: if all the buffers are
3185  * clean then we set the page clean and proceed.  To do that, we require
3186  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3187  * private_lock.
3188  *
3189  * try_to_free_buffers() is non-blocking.
3190  */
3191 static inline int buffer_busy(struct buffer_head *bh)
3192 {
3193         return atomic_read(&bh->b_count) |
3194                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3195 }
3196
3197 static int
3198 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3199 {
3200         struct buffer_head *head = page_buffers(page);
3201         struct buffer_head *bh;
3202
3203         bh = head;
3204         do {
3205                 if (buffer_write_io_error(bh) && page->mapping)
3206                         set_bit(AS_EIO, &page->mapping->flags);
3207                 if (buffer_busy(bh))
3208                         goto failed;
3209                 bh = bh->b_this_page;
3210         } while (bh != head);
3211
3212         do {
3213                 struct buffer_head *next = bh->b_this_page;
3214
3215                 if (bh->b_assoc_map)
3216                         __remove_assoc_queue(bh);
3217                 bh = next;
3218         } while (bh != head);
3219         *buffers_to_free = head;
3220         __clear_page_buffers(page);
3221         return 1;
3222 failed:
3223         return 0;
3224 }
3225
3226 int try_to_free_buffers(struct page *page)
3227 {
3228         struct address_space * const mapping = page->mapping;
3229         struct buffer_head *buffers_to_free = NULL;
3230         int ret = 0;
3231
3232         BUG_ON(!PageLocked(page));
3233         if (PageWriteback(page))
3234                 return 0;
3235
3236         if (mapping == NULL) {          /* can this still happen? */
3237                 ret = drop_buffers(page, &buffers_to_free);
3238                 goto out;
3239         }
3240
3241         spin_lock(&mapping->private_lock);
3242         ret = drop_buffers(page, &buffers_to_free);
3243
3244         /*
3245          * If the filesystem writes its buffers by hand (eg ext3)
3246          * then we can have clean buffers against a dirty page.  We
3247          * clean the page here; otherwise the VM will never notice
3248          * that the filesystem did any IO at all.
3249          *
3250          * Also, during truncate, discard_buffer will have marked all
3251          * the page's buffers clean.  We discover that here and clean
3252          * the page also.
3253          *
3254          * private_lock must be held over this entire operation in order
3255          * to synchronise against __set_page_dirty_buffers and prevent the
3256          * dirty bit from being lost.
3257          */
3258         if (ret)
3259                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3260         spin_unlock(&mapping->private_lock);
3261 out:
3262         if (buffers_to_free) {
3263                 struct buffer_head *bh = buffers_to_free;
3264
3265                 do {
3266                         struct buffer_head *next = bh->b_this_page;
3267                         free_buffer_head(bh);
3268                         bh = next;
3269                 } while (bh != buffers_to_free);
3270         }
3271         return ret;
3272 }
3273 EXPORT_SYMBOL(try_to_free_buffers);
3274
3275 /*
3276  * There are no bdflush tunables left.  But distributions are
3277  * still running obsolete flush daemons, so we terminate them here.
3278  *
3279  * Use of bdflush() is deprecated and will be removed in a future kernel.
3280  * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3281  */
3282 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3283 {
3284         static int msg_count;
3285
3286         if (!capable(CAP_SYS_ADMIN))
3287                 return -EPERM;
3288
3289         if (msg_count < 5) {
3290                 msg_count++;
3291                 printk(KERN_INFO
3292                         "warning: process `%s' used the obsolete bdflush"
3293                         " system call\n", current->comm);
3294                 printk(KERN_INFO "Fix your initscripts?\n");
3295         }
3296
3297         if (func == 1)
3298                 do_exit(0);
3299         return 0;
3300 }
3301
3302 /*
3303  * Buffer-head allocation
3304  */
3305 static struct kmem_cache *bh_cachep __read_mostly;
3306
3307 /*
3308  * Once the number of bh's in the machine exceeds this level, we start
3309  * stripping them in writeback.
3310  */
3311 static unsigned long max_buffer_heads;
3312
3313 int buffer_heads_over_limit;
3314
3315 struct bh_accounting {
3316         int nr;                 /* Number of live bh's */
3317         int ratelimit;          /* Limit cacheline bouncing */
3318 };
3319
3320 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3321
3322 static void recalc_bh_state(void)
3323 {
3324         int i;
3325         int tot = 0;
3326
3327         if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3328                 return;
3329         __this_cpu_write(bh_accounting.ratelimit, 0);
3330         for_each_online_cpu(i)
3331                 tot += per_cpu(bh_accounting, i).nr;
3332         buffer_heads_over_limit = (tot > max_buffer_heads);
3333 }
3334
3335 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3336 {
3337         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3338         if (ret) {
3339                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3340                 preempt_disable();
3341                 __this_cpu_inc(bh_accounting.nr);
3342                 recalc_bh_state();
3343                 preempt_enable();
3344         }
3345         return ret;
3346 }
3347 EXPORT_SYMBOL(alloc_buffer_head);
3348
3349 void free_buffer_head(struct buffer_head *bh)
3350 {
3351         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3352         kmem_cache_free(bh_cachep, bh);
3353         preempt_disable();
3354         __this_cpu_dec(bh_accounting.nr);
3355         recalc_bh_state();
3356         preempt_enable();
3357 }
3358 EXPORT_SYMBOL(free_buffer_head);
3359
3360 static void buffer_exit_cpu(int cpu)
3361 {
3362         int i;
3363         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3364
3365         for (i = 0; i < BH_LRU_SIZE; i++) {
3366                 brelse(b->bhs[i]);
3367                 b->bhs[i] = NULL;
3368         }
3369         this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3370         per_cpu(bh_accounting, cpu).nr = 0;
3371 }
3372
3373 static int buffer_cpu_notify(struct notifier_block *self,
3374                               unsigned long action, void *hcpu)
3375 {
3376         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3377                 buffer_exit_cpu((unsigned long)hcpu);
3378         return NOTIFY_OK;
3379 }
3380
3381 /**
3382  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3383  * @bh: struct buffer_head
3384  *
3385  * Return true if the buffer is up-to-date and false,
3386  * with the buffer locked, if not.
3387  */
3388 int bh_uptodate_or_lock(struct buffer_head *bh)
3389 {
3390         if (!buffer_uptodate(bh)) {
3391                 lock_buffer(bh);
3392                 if (!buffer_uptodate(bh))
3393                         return 0;
3394                 unlock_buffer(bh);
3395         }
3396         return 1;
3397 }
3398 EXPORT_SYMBOL(bh_uptodate_or_lock);
3399
3400 /**
3401  * bh_submit_read - Submit a locked buffer for reading
3402  * @bh: struct buffer_head
3403  *
3404  * Returns zero on success and -EIO on error.
3405  */
3406 int bh_submit_read(struct buffer_head *bh)
3407 {
3408         BUG_ON(!buffer_locked(bh));
3409
3410         if (buffer_uptodate(bh)) {
3411                 unlock_buffer(bh);
3412                 return 0;
3413         }
3414
3415         get_bh(bh);
3416         bh->b_end_io = end_buffer_read_sync;
3417         submit_bh(READ, bh);
3418         wait_on_buffer(bh);
3419         if (buffer_uptodate(bh))
3420                 return 0;
3421         return -EIO;
3422 }
3423 EXPORT_SYMBOL(bh_submit_read);
3424
3425 void __init buffer_init(void)
3426 {
3427         unsigned long nrpages;
3428
3429         bh_cachep = kmem_cache_create("buffer_head",
3430                         sizeof(struct buffer_head), 0,
3431                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3432                                 SLAB_MEM_SPREAD),
3433                                 NULL);
3434
3435         /*
3436          * Limit the bh occupancy to 10% of ZONE_NORMAL
3437          */
3438         nrpages = (nr_free_buffer_pages() * 10) / 100;
3439         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3440         hotcpu_notifier(buffer_cpu_notify, 0);
3441 }