block: fix infinite loop in __getblk_slow
[linux-3.10.git] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/capability.h>
28 #include <linux/blkdev.h>
29 #include <linux/file.h>
30 #include <linux/quotaops.h>
31 #include <linux/highmem.h>
32 #include <linux/export.h>
33 #include <linux/writeback.h>
34 #include <linux/hash.h>
35 #include <linux/suspend.h>
36 #include <linux/buffer_head.h>
37 #include <linux/task_io_accounting_ops.h>
38 #include <linux/bio.h>
39 #include <linux/notifier.h>
40 #include <linux/cpu.h>
41 #include <linux/bitops.h>
42 #include <linux/mpage.h>
43 #include <linux/bit_spinlock.h>
44
45 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
46
47 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
48
49 inline void
50 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 static int sleep_on_buffer(void *word)
58 {
59         io_schedule();
60         return 0;
61 }
62
63 void __lock_buffer(struct buffer_head *bh)
64 {
65         wait_on_bit_lock(&bh->b_state, BH_Lock, sleep_on_buffer,
66                                                         TASK_UNINTERRUPTIBLE);
67 }
68 EXPORT_SYMBOL(__lock_buffer);
69
70 void unlock_buffer(struct buffer_head *bh)
71 {
72         clear_bit_unlock(BH_Lock, &bh->b_state);
73         smp_mb__after_clear_bit();
74         wake_up_bit(&bh->b_state, BH_Lock);
75 }
76 EXPORT_SYMBOL(unlock_buffer);
77
78 /*
79  * Block until a buffer comes unlocked.  This doesn't stop it
80  * from becoming locked again - you have to lock it yourself
81  * if you want to preserve its state.
82  */
83 void __wait_on_buffer(struct buffer_head * bh)
84 {
85         wait_on_bit(&bh->b_state, BH_Lock, sleep_on_buffer, TASK_UNINTERRUPTIBLE);
86 }
87 EXPORT_SYMBOL(__wait_on_buffer);
88
89 static void
90 __clear_page_buffers(struct page *page)
91 {
92         ClearPagePrivate(page);
93         set_page_private(page, 0);
94         page_cache_release(page);
95 }
96
97
98 static int quiet_error(struct buffer_head *bh)
99 {
100         if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
101                 return 0;
102         return 1;
103 }
104
105
106 static void buffer_io_error(struct buffer_head *bh)
107 {
108         char b[BDEVNAME_SIZE];
109         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
110                         bdevname(bh->b_bdev, b),
111                         (unsigned long long)bh->b_blocknr);
112 }
113
114 /*
115  * End-of-IO handler helper function which does not touch the bh after
116  * unlocking it.
117  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
118  * a race there is benign: unlock_buffer() only use the bh's address for
119  * hashing after unlocking the buffer, so it doesn't actually touch the bh
120  * itself.
121  */
122 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
123 {
124         if (uptodate) {
125                 set_buffer_uptodate(bh);
126         } else {
127                 /* This happens, due to failed READA attempts. */
128                 clear_buffer_uptodate(bh);
129         }
130         unlock_buffer(bh);
131 }
132
133 /*
134  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
135  * unlock the buffer. This is what ll_rw_block uses too.
136  */
137 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
138 {
139         __end_buffer_read_notouch(bh, uptodate);
140         put_bh(bh);
141 }
142 EXPORT_SYMBOL(end_buffer_read_sync);
143
144 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
145 {
146         char b[BDEVNAME_SIZE];
147
148         if (uptodate) {
149                 set_buffer_uptodate(bh);
150         } else {
151                 if (!quiet_error(bh)) {
152                         buffer_io_error(bh);
153                         printk(KERN_WARNING "lost page write due to "
154                                         "I/O error on %s\n",
155                                        bdevname(bh->b_bdev, b));
156                 }
157                 set_buffer_write_io_error(bh);
158                 clear_buffer_uptodate(bh);
159         }
160         unlock_buffer(bh);
161         put_bh(bh);
162 }
163 EXPORT_SYMBOL(end_buffer_write_sync);
164
165 /*
166  * Various filesystems appear to want __find_get_block to be non-blocking.
167  * But it's the page lock which protects the buffers.  To get around this,
168  * we get exclusion from try_to_free_buffers with the blockdev mapping's
169  * private_lock.
170  *
171  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
172  * may be quite high.  This code could TryLock the page, and if that
173  * succeeds, there is no need to take private_lock. (But if
174  * private_lock is contended then so is mapping->tree_lock).
175  */
176 static struct buffer_head *
177 __find_get_block_slow(struct block_device *bdev, sector_t block)
178 {
179         struct inode *bd_inode = bdev->bd_inode;
180         struct address_space *bd_mapping = bd_inode->i_mapping;
181         struct buffer_head *ret = NULL;
182         pgoff_t index;
183         struct buffer_head *bh;
184         struct buffer_head *head;
185         struct page *page;
186         int all_mapped = 1;
187
188         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
189         page = find_get_page(bd_mapping, index);
190         if (!page)
191                 goto out;
192
193         spin_lock(&bd_mapping->private_lock);
194         if (!page_has_buffers(page))
195                 goto out_unlock;
196         head = page_buffers(page);
197         bh = head;
198         do {
199                 if (!buffer_mapped(bh))
200                         all_mapped = 0;
201                 else if (bh->b_blocknr == block) {
202                         ret = bh;
203                         get_bh(bh);
204                         goto out_unlock;
205                 }
206                 bh = bh->b_this_page;
207         } while (bh != head);
208
209         /* we might be here because some of the buffers on this page are
210          * not mapped.  This is due to various races between
211          * file io on the block device and getblk.  It gets dealt with
212          * elsewhere, don't buffer_error if we had some unmapped buffers
213          */
214         if (all_mapped) {
215                 char b[BDEVNAME_SIZE];
216
217                 printk("__find_get_block_slow() failed. "
218                         "block=%llu, b_blocknr=%llu\n",
219                         (unsigned long long)block,
220                         (unsigned long long)bh->b_blocknr);
221                 printk("b_state=0x%08lx, b_size=%zu\n",
222                         bh->b_state, bh->b_size);
223                 printk("device %s blocksize: %d\n", bdevname(bdev, b),
224                         1 << bd_inode->i_blkbits);
225         }
226 out_unlock:
227         spin_unlock(&bd_mapping->private_lock);
228         page_cache_release(page);
229 out:
230         return ret;
231 }
232
233 /*
234  * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
235  */
236 static void free_more_memory(void)
237 {
238         struct zone *zone;
239         int nid;
240
241         wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
242         yield();
243
244         for_each_online_node(nid) {
245                 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
246                                                 gfp_zone(GFP_NOFS), NULL,
247                                                 &zone);
248                 if (zone)
249                         try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
250                                                 GFP_NOFS, NULL);
251         }
252 }
253
254 /*
255  * I/O completion handler for block_read_full_page() - pages
256  * which come unlocked at the end of I/O.
257  */
258 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
259 {
260         unsigned long flags;
261         struct buffer_head *first;
262         struct buffer_head *tmp;
263         struct page *page;
264         int page_uptodate = 1;
265
266         BUG_ON(!buffer_async_read(bh));
267
268         page = bh->b_page;
269         if (uptodate) {
270                 set_buffer_uptodate(bh);
271         } else {
272                 clear_buffer_uptodate(bh);
273                 if (!quiet_error(bh))
274                         buffer_io_error(bh);
275                 SetPageError(page);
276         }
277
278         /*
279          * Be _very_ careful from here on. Bad things can happen if
280          * two buffer heads end IO at almost the same time and both
281          * decide that the page is now completely done.
282          */
283         first = page_buffers(page);
284         local_irq_save(flags);
285         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
286         clear_buffer_async_read(bh);
287         unlock_buffer(bh);
288         tmp = bh;
289         do {
290                 if (!buffer_uptodate(tmp))
291                         page_uptodate = 0;
292                 if (buffer_async_read(tmp)) {
293                         BUG_ON(!buffer_locked(tmp));
294                         goto still_busy;
295                 }
296                 tmp = tmp->b_this_page;
297         } while (tmp != bh);
298         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
299         local_irq_restore(flags);
300
301         /*
302          * If none of the buffers had errors and they are all
303          * uptodate then we can set the page uptodate.
304          */
305         if (page_uptodate && !PageError(page))
306                 SetPageUptodate(page);
307         unlock_page(page);
308         return;
309
310 still_busy:
311         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
312         local_irq_restore(flags);
313         return;
314 }
315
316 /*
317  * Completion handler for block_write_full_page() - pages which are unlocked
318  * during I/O, and which have PageWriteback cleared upon I/O completion.
319  */
320 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
321 {
322         char b[BDEVNAME_SIZE];
323         unsigned long flags;
324         struct buffer_head *first;
325         struct buffer_head *tmp;
326         struct page *page;
327
328         BUG_ON(!buffer_async_write(bh));
329
330         page = bh->b_page;
331         if (uptodate) {
332                 set_buffer_uptodate(bh);
333         } else {
334                 if (!quiet_error(bh)) {
335                         buffer_io_error(bh);
336                         printk(KERN_WARNING "lost page write due to "
337                                         "I/O error on %s\n",
338                                bdevname(bh->b_bdev, b));
339                 }
340                 set_bit(AS_EIO, &page->mapping->flags);
341                 set_buffer_write_io_error(bh);
342                 clear_buffer_uptodate(bh);
343                 SetPageError(page);
344         }
345
346         first = page_buffers(page);
347         local_irq_save(flags);
348         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
349
350         clear_buffer_async_write(bh);
351         unlock_buffer(bh);
352         tmp = bh->b_this_page;
353         while (tmp != bh) {
354                 if (buffer_async_write(tmp)) {
355                         BUG_ON(!buffer_locked(tmp));
356                         goto still_busy;
357                 }
358                 tmp = tmp->b_this_page;
359         }
360         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
361         local_irq_restore(flags);
362         end_page_writeback(page);
363         return;
364
365 still_busy:
366         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
367         local_irq_restore(flags);
368         return;
369 }
370 EXPORT_SYMBOL(end_buffer_async_write);
371
372 /*
373  * If a page's buffers are under async readin (end_buffer_async_read
374  * completion) then there is a possibility that another thread of
375  * control could lock one of the buffers after it has completed
376  * but while some of the other buffers have not completed.  This
377  * locked buffer would confuse end_buffer_async_read() into not unlocking
378  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
379  * that this buffer is not under async I/O.
380  *
381  * The page comes unlocked when it has no locked buffer_async buffers
382  * left.
383  *
384  * PageLocked prevents anyone starting new async I/O reads any of
385  * the buffers.
386  *
387  * PageWriteback is used to prevent simultaneous writeout of the same
388  * page.
389  *
390  * PageLocked prevents anyone from starting writeback of a page which is
391  * under read I/O (PageWriteback is only ever set against a locked page).
392  */
393 static void mark_buffer_async_read(struct buffer_head *bh)
394 {
395         bh->b_end_io = end_buffer_async_read;
396         set_buffer_async_read(bh);
397 }
398
399 static void mark_buffer_async_write_endio(struct buffer_head *bh,
400                                           bh_end_io_t *handler)
401 {
402         bh->b_end_io = handler;
403         set_buffer_async_write(bh);
404 }
405
406 void mark_buffer_async_write(struct buffer_head *bh)
407 {
408         mark_buffer_async_write_endio(bh, end_buffer_async_write);
409 }
410 EXPORT_SYMBOL(mark_buffer_async_write);
411
412
413 /*
414  * fs/buffer.c contains helper functions for buffer-backed address space's
415  * fsync functions.  A common requirement for buffer-based filesystems is
416  * that certain data from the backing blockdev needs to be written out for
417  * a successful fsync().  For example, ext2 indirect blocks need to be
418  * written back and waited upon before fsync() returns.
419  *
420  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
421  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
422  * management of a list of dependent buffers at ->i_mapping->private_list.
423  *
424  * Locking is a little subtle: try_to_free_buffers() will remove buffers
425  * from their controlling inode's queue when they are being freed.  But
426  * try_to_free_buffers() will be operating against the *blockdev* mapping
427  * at the time, not against the S_ISREG file which depends on those buffers.
428  * So the locking for private_list is via the private_lock in the address_space
429  * which backs the buffers.  Which is different from the address_space 
430  * against which the buffers are listed.  So for a particular address_space,
431  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
432  * mapping->private_list will always be protected by the backing blockdev's
433  * ->private_lock.
434  *
435  * Which introduces a requirement: all buffers on an address_space's
436  * ->private_list must be from the same address_space: the blockdev's.
437  *
438  * address_spaces which do not place buffers at ->private_list via these
439  * utility functions are free to use private_lock and private_list for
440  * whatever they want.  The only requirement is that list_empty(private_list)
441  * be true at clear_inode() time.
442  *
443  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
444  * filesystems should do that.  invalidate_inode_buffers() should just go
445  * BUG_ON(!list_empty).
446  *
447  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
448  * take an address_space, not an inode.  And it should be called
449  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
450  * queued up.
451  *
452  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
453  * list if it is already on a list.  Because if the buffer is on a list,
454  * it *must* already be on the right one.  If not, the filesystem is being
455  * silly.  This will save a ton of locking.  But first we have to ensure
456  * that buffers are taken *off* the old inode's list when they are freed
457  * (presumably in truncate).  That requires careful auditing of all
458  * filesystems (do it inside bforget()).  It could also be done by bringing
459  * b_inode back.
460  */
461
462 /*
463  * The buffer's backing address_space's private_lock must be held
464  */
465 static void __remove_assoc_queue(struct buffer_head *bh)
466 {
467         list_del_init(&bh->b_assoc_buffers);
468         WARN_ON(!bh->b_assoc_map);
469         if (buffer_write_io_error(bh))
470                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
471         bh->b_assoc_map = NULL;
472 }
473
474 int inode_has_buffers(struct inode *inode)
475 {
476         return !list_empty(&inode->i_data.private_list);
477 }
478
479 /*
480  * osync is designed to support O_SYNC io.  It waits synchronously for
481  * all already-submitted IO to complete, but does not queue any new
482  * writes to the disk.
483  *
484  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
485  * you dirty the buffers, and then use osync_inode_buffers to wait for
486  * completion.  Any other dirty buffers which are not yet queued for
487  * write will not be flushed to disk by the osync.
488  */
489 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
490 {
491         struct buffer_head *bh;
492         struct list_head *p;
493         int err = 0;
494
495         spin_lock(lock);
496 repeat:
497         list_for_each_prev(p, list) {
498                 bh = BH_ENTRY(p);
499                 if (buffer_locked(bh)) {
500                         get_bh(bh);
501                         spin_unlock(lock);
502                         wait_on_buffer(bh);
503                         if (!buffer_uptodate(bh))
504                                 err = -EIO;
505                         brelse(bh);
506                         spin_lock(lock);
507                         goto repeat;
508                 }
509         }
510         spin_unlock(lock);
511         return err;
512 }
513
514 static void do_thaw_one(struct super_block *sb, void *unused)
515 {
516         char b[BDEVNAME_SIZE];
517         while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
518                 printk(KERN_WARNING "Emergency Thaw on %s\n",
519                        bdevname(sb->s_bdev, b));
520 }
521
522 static void do_thaw_all(struct work_struct *work)
523 {
524         iterate_supers(do_thaw_one, NULL);
525         kfree(work);
526         printk(KERN_WARNING "Emergency Thaw complete\n");
527 }
528
529 /**
530  * emergency_thaw_all -- forcibly thaw every frozen filesystem
531  *
532  * Used for emergency unfreeze of all filesystems via SysRq
533  */
534 void emergency_thaw_all(void)
535 {
536         struct work_struct *work;
537
538         work = kmalloc(sizeof(*work), GFP_ATOMIC);
539         if (work) {
540                 INIT_WORK(work, do_thaw_all);
541                 schedule_work(work);
542         }
543 }
544
545 /**
546  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
547  * @mapping: the mapping which wants those buffers written
548  *
549  * Starts I/O against the buffers at mapping->private_list, and waits upon
550  * that I/O.
551  *
552  * Basically, this is a convenience function for fsync().
553  * @mapping is a file or directory which needs those buffers to be written for
554  * a successful fsync().
555  */
556 int sync_mapping_buffers(struct address_space *mapping)
557 {
558         struct address_space *buffer_mapping = mapping->assoc_mapping;
559
560         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
561                 return 0;
562
563         return fsync_buffers_list(&buffer_mapping->private_lock,
564                                         &mapping->private_list);
565 }
566 EXPORT_SYMBOL(sync_mapping_buffers);
567
568 /*
569  * Called when we've recently written block `bblock', and it is known that
570  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
571  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
572  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
573  */
574 void write_boundary_block(struct block_device *bdev,
575                         sector_t bblock, unsigned blocksize)
576 {
577         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
578         if (bh) {
579                 if (buffer_dirty(bh))
580                         ll_rw_block(WRITE, 1, &bh);
581                 put_bh(bh);
582         }
583 }
584
585 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
586 {
587         struct address_space *mapping = inode->i_mapping;
588         struct address_space *buffer_mapping = bh->b_page->mapping;
589
590         mark_buffer_dirty(bh);
591         if (!mapping->assoc_mapping) {
592                 mapping->assoc_mapping = buffer_mapping;
593         } else {
594                 BUG_ON(mapping->assoc_mapping != buffer_mapping);
595         }
596         if (!bh->b_assoc_map) {
597                 spin_lock(&buffer_mapping->private_lock);
598                 list_move_tail(&bh->b_assoc_buffers,
599                                 &mapping->private_list);
600                 bh->b_assoc_map = mapping;
601                 spin_unlock(&buffer_mapping->private_lock);
602         }
603 }
604 EXPORT_SYMBOL(mark_buffer_dirty_inode);
605
606 /*
607  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
608  * dirty.
609  *
610  * If warn is true, then emit a warning if the page is not uptodate and has
611  * not been truncated.
612  */
613 static void __set_page_dirty(struct page *page,
614                 struct address_space *mapping, int warn)
615 {
616         spin_lock_irq(&mapping->tree_lock);
617         if (page->mapping) {    /* Race with truncate? */
618                 WARN_ON_ONCE(warn && !PageUptodate(page));
619                 account_page_dirtied(page, mapping);
620                 radix_tree_tag_set(&mapping->page_tree,
621                                 page_index(page), PAGECACHE_TAG_DIRTY);
622         }
623         spin_unlock_irq(&mapping->tree_lock);
624         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
625 }
626
627 /*
628  * Add a page to the dirty page list.
629  *
630  * It is a sad fact of life that this function is called from several places
631  * deeply under spinlocking.  It may not sleep.
632  *
633  * If the page has buffers, the uptodate buffers are set dirty, to preserve
634  * dirty-state coherency between the page and the buffers.  It the page does
635  * not have buffers then when they are later attached they will all be set
636  * dirty.
637  *
638  * The buffers are dirtied before the page is dirtied.  There's a small race
639  * window in which a writepage caller may see the page cleanness but not the
640  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
641  * before the buffers, a concurrent writepage caller could clear the page dirty
642  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
643  * page on the dirty page list.
644  *
645  * We use private_lock to lock against try_to_free_buffers while using the
646  * page's buffer list.  Also use this to protect against clean buffers being
647  * added to the page after it was set dirty.
648  *
649  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
650  * address_space though.
651  */
652 int __set_page_dirty_buffers(struct page *page)
653 {
654         int newly_dirty;
655         struct address_space *mapping = page_mapping(page);
656
657         if (unlikely(!mapping))
658                 return !TestSetPageDirty(page);
659
660         spin_lock(&mapping->private_lock);
661         if (page_has_buffers(page)) {
662                 struct buffer_head *head = page_buffers(page);
663                 struct buffer_head *bh = head;
664
665                 do {
666                         set_buffer_dirty(bh);
667                         bh = bh->b_this_page;
668                 } while (bh != head);
669         }
670         newly_dirty = !TestSetPageDirty(page);
671         spin_unlock(&mapping->private_lock);
672
673         if (newly_dirty)
674                 __set_page_dirty(page, mapping, 1);
675         return newly_dirty;
676 }
677 EXPORT_SYMBOL(__set_page_dirty_buffers);
678
679 /*
680  * Write out and wait upon a list of buffers.
681  *
682  * We have conflicting pressures: we want to make sure that all
683  * initially dirty buffers get waited on, but that any subsequently
684  * dirtied buffers don't.  After all, we don't want fsync to last
685  * forever if somebody is actively writing to the file.
686  *
687  * Do this in two main stages: first we copy dirty buffers to a
688  * temporary inode list, queueing the writes as we go.  Then we clean
689  * up, waiting for those writes to complete.
690  * 
691  * During this second stage, any subsequent updates to the file may end
692  * up refiling the buffer on the original inode's dirty list again, so
693  * there is a chance we will end up with a buffer queued for write but
694  * not yet completed on that list.  So, as a final cleanup we go through
695  * the osync code to catch these locked, dirty buffers without requeuing
696  * any newly dirty buffers for write.
697  */
698 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
699 {
700         struct buffer_head *bh;
701         struct list_head tmp;
702         struct address_space *mapping;
703         int err = 0, err2;
704         struct blk_plug plug;
705
706         INIT_LIST_HEAD(&tmp);
707         blk_start_plug(&plug);
708
709         spin_lock(lock);
710         while (!list_empty(list)) {
711                 bh = BH_ENTRY(list->next);
712                 mapping = bh->b_assoc_map;
713                 __remove_assoc_queue(bh);
714                 /* Avoid race with mark_buffer_dirty_inode() which does
715                  * a lockless check and we rely on seeing the dirty bit */
716                 smp_mb();
717                 if (buffer_dirty(bh) || buffer_locked(bh)) {
718                         list_add(&bh->b_assoc_buffers, &tmp);
719                         bh->b_assoc_map = mapping;
720                         if (buffer_dirty(bh)) {
721                                 get_bh(bh);
722                                 spin_unlock(lock);
723                                 /*
724                                  * Ensure any pending I/O completes so that
725                                  * write_dirty_buffer() actually writes the
726                                  * current contents - it is a noop if I/O is
727                                  * still in flight on potentially older
728                                  * contents.
729                                  */
730                                 write_dirty_buffer(bh, WRITE_SYNC);
731
732                                 /*
733                                  * Kick off IO for the previous mapping. Note
734                                  * that we will not run the very last mapping,
735                                  * wait_on_buffer() will do that for us
736                                  * through sync_buffer().
737                                  */
738                                 brelse(bh);
739                                 spin_lock(lock);
740                         }
741                 }
742         }
743
744         spin_unlock(lock);
745         blk_finish_plug(&plug);
746         spin_lock(lock);
747
748         while (!list_empty(&tmp)) {
749                 bh = BH_ENTRY(tmp.prev);
750                 get_bh(bh);
751                 mapping = bh->b_assoc_map;
752                 __remove_assoc_queue(bh);
753                 /* Avoid race with mark_buffer_dirty_inode() which does
754                  * a lockless check and we rely on seeing the dirty bit */
755                 smp_mb();
756                 if (buffer_dirty(bh)) {
757                         list_add(&bh->b_assoc_buffers,
758                                  &mapping->private_list);
759                         bh->b_assoc_map = mapping;
760                 }
761                 spin_unlock(lock);
762                 wait_on_buffer(bh);
763                 if (!buffer_uptodate(bh))
764                         err = -EIO;
765                 brelse(bh);
766                 spin_lock(lock);
767         }
768         
769         spin_unlock(lock);
770         err2 = osync_buffers_list(lock, list);
771         if (err)
772                 return err;
773         else
774                 return err2;
775 }
776
777 /*
778  * Invalidate any and all dirty buffers on a given inode.  We are
779  * probably unmounting the fs, but that doesn't mean we have already
780  * done a sync().  Just drop the buffers from the inode list.
781  *
782  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
783  * assumes that all the buffers are against the blockdev.  Not true
784  * for reiserfs.
785  */
786 void invalidate_inode_buffers(struct inode *inode)
787 {
788         if (inode_has_buffers(inode)) {
789                 struct address_space *mapping = &inode->i_data;
790                 struct list_head *list = &mapping->private_list;
791                 struct address_space *buffer_mapping = mapping->assoc_mapping;
792
793                 spin_lock(&buffer_mapping->private_lock);
794                 while (!list_empty(list))
795                         __remove_assoc_queue(BH_ENTRY(list->next));
796                 spin_unlock(&buffer_mapping->private_lock);
797         }
798 }
799 EXPORT_SYMBOL(invalidate_inode_buffers);
800
801 /*
802  * Remove any clean buffers from the inode's buffer list.  This is called
803  * when we're trying to free the inode itself.  Those buffers can pin it.
804  *
805  * Returns true if all buffers were removed.
806  */
807 int remove_inode_buffers(struct inode *inode)
808 {
809         int ret = 1;
810
811         if (inode_has_buffers(inode)) {
812                 struct address_space *mapping = &inode->i_data;
813                 struct list_head *list = &mapping->private_list;
814                 struct address_space *buffer_mapping = mapping->assoc_mapping;
815
816                 spin_lock(&buffer_mapping->private_lock);
817                 while (!list_empty(list)) {
818                         struct buffer_head *bh = BH_ENTRY(list->next);
819                         if (buffer_dirty(bh)) {
820                                 ret = 0;
821                                 break;
822                         }
823                         __remove_assoc_queue(bh);
824                 }
825                 spin_unlock(&buffer_mapping->private_lock);
826         }
827         return ret;
828 }
829
830 /*
831  * Create the appropriate buffers when given a page for data area and
832  * the size of each buffer.. Use the bh->b_this_page linked list to
833  * follow the buffers created.  Return NULL if unable to create more
834  * buffers.
835  *
836  * The retry flag is used to differentiate async IO (paging, swapping)
837  * which may not fail from ordinary buffer allocations.
838  */
839 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
840                 int retry)
841 {
842         struct buffer_head *bh, *head;
843         long offset;
844
845 try_again:
846         head = NULL;
847         offset = PAGE_SIZE;
848         while ((offset -= size) >= 0) {
849                 bh = alloc_buffer_head(GFP_NOFS);
850                 if (!bh)
851                         goto no_grow;
852
853                 bh->b_bdev = NULL;
854                 bh->b_this_page = head;
855                 bh->b_blocknr = -1;
856                 head = bh;
857
858                 bh->b_state = 0;
859                 atomic_set(&bh->b_count, 0);
860                 bh->b_size = size;
861
862                 /* Link the buffer to its page */
863                 set_bh_page(bh, page, offset);
864
865                 init_buffer(bh, NULL, NULL);
866         }
867         return head;
868 /*
869  * In case anything failed, we just free everything we got.
870  */
871 no_grow:
872         if (head) {
873                 do {
874                         bh = head;
875                         head = head->b_this_page;
876                         free_buffer_head(bh);
877                 } while (head);
878         }
879
880         /*
881          * Return failure for non-async IO requests.  Async IO requests
882          * are not allowed to fail, so we have to wait until buffer heads
883          * become available.  But we don't want tasks sleeping with 
884          * partially complete buffers, so all were released above.
885          */
886         if (!retry)
887                 return NULL;
888
889         /* We're _really_ low on memory. Now we just
890          * wait for old buffer heads to become free due to
891          * finishing IO.  Since this is an async request and
892          * the reserve list is empty, we're sure there are 
893          * async buffer heads in use.
894          */
895         free_more_memory();
896         goto try_again;
897 }
898 EXPORT_SYMBOL_GPL(alloc_page_buffers);
899
900 static inline void
901 link_dev_buffers(struct page *page, struct buffer_head *head)
902 {
903         struct buffer_head *bh, *tail;
904
905         bh = head;
906         do {
907                 tail = bh;
908                 bh = bh->b_this_page;
909         } while (bh);
910         tail->b_this_page = head;
911         attach_page_buffers(page, head);
912 }
913
914 /*
915  * Initialise the state of a blockdev page's buffers.
916  */ 
917 static void
918 init_page_buffers(struct page *page, struct block_device *bdev,
919                         sector_t block, int size)
920 {
921         struct buffer_head *head = page_buffers(page);
922         struct buffer_head *bh = head;
923         int uptodate = PageUptodate(page);
924         sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode));
925
926         do {
927                 if (!buffer_mapped(bh)) {
928                         init_buffer(bh, NULL, NULL);
929                         bh->b_bdev = bdev;
930                         bh->b_blocknr = block;
931                         if (uptodate)
932                                 set_buffer_uptodate(bh);
933                         if (block < end_block)
934                                 set_buffer_mapped(bh);
935                 }
936                 block++;
937                 bh = bh->b_this_page;
938         } while (bh != head);
939 }
940
941 /*
942  * Create the page-cache page that contains the requested block.
943  *
944  * This is user purely for blockdev mappings.
945  */
946 static struct page *
947 grow_dev_page(struct block_device *bdev, sector_t block,
948                 pgoff_t index, int size)
949 {
950         struct inode *inode = bdev->bd_inode;
951         struct page *page;
952         struct buffer_head *bh;
953
954         page = find_or_create_page(inode->i_mapping, index,
955                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
956         if (!page)
957                 return NULL;
958
959         BUG_ON(!PageLocked(page));
960
961         if (page_has_buffers(page)) {
962                 bh = page_buffers(page);
963                 if (bh->b_size == size) {
964                         init_page_buffers(page, bdev, block, size);
965                         return page;
966                 }
967                 if (!try_to_free_buffers(page))
968                         goto failed;
969         }
970
971         /*
972          * Allocate some buffers for this page
973          */
974         bh = alloc_page_buffers(page, size, 0);
975         if (!bh)
976                 goto failed;
977
978         /*
979          * Link the page to the buffers and initialise them.  Take the
980          * lock to be atomic wrt __find_get_block(), which does not
981          * run under the page lock.
982          */
983         spin_lock(&inode->i_mapping->private_lock);
984         link_dev_buffers(page, bh);
985         init_page_buffers(page, bdev, block, size);
986         spin_unlock(&inode->i_mapping->private_lock);
987         return page;
988
989 failed:
990         unlock_page(page);
991         page_cache_release(page);
992         return NULL;
993 }
994
995 /*
996  * Create buffers for the specified block device block's page.  If
997  * that page was dirty, the buffers are set dirty also.
998  */
999 static int
1000 grow_buffers(struct block_device *bdev, sector_t block, int size)
1001 {
1002         struct page *page;
1003         pgoff_t index;
1004         int sizebits;
1005
1006         sizebits = -1;
1007         do {
1008                 sizebits++;
1009         } while ((size << sizebits) < PAGE_SIZE);
1010
1011         index = block >> sizebits;
1012
1013         /*
1014          * Check for a block which wants to lie outside our maximum possible
1015          * pagecache index.  (this comparison is done using sector_t types).
1016          */
1017         if (unlikely(index != block >> sizebits)) {
1018                 char b[BDEVNAME_SIZE];
1019
1020                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1021                         "device %s\n",
1022                         __func__, (unsigned long long)block,
1023                         bdevname(bdev, b));
1024                 return -EIO;
1025         }
1026         block = index << sizebits;
1027         /* Create a page with the proper size buffers.. */
1028         page = grow_dev_page(bdev, block, index, size);
1029         if (!page)
1030                 return 0;
1031         unlock_page(page);
1032         page_cache_release(page);
1033         return 1;
1034 }
1035
1036 static struct buffer_head *
1037 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1038 {
1039         int ret;
1040         struct buffer_head *bh;
1041
1042         /* Size must be multiple of hard sectorsize */
1043         if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1044                         (size < 512 || size > PAGE_SIZE))) {
1045                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1046                                         size);
1047                 printk(KERN_ERR "logical block size: %d\n",
1048                                         bdev_logical_block_size(bdev));
1049
1050                 dump_stack();
1051                 return NULL;
1052         }
1053
1054 retry:
1055         bh = __find_get_block(bdev, block, size);
1056         if (bh)
1057                 return bh;
1058
1059         ret = grow_buffers(bdev, block, size);
1060         if (ret == 0) {
1061                 free_more_memory();
1062                 goto retry;
1063         } else if (ret > 0) {
1064                 bh = __find_get_block(bdev, block, size);
1065                 if (bh)
1066                         return bh;
1067         }
1068         return NULL;
1069 }
1070
1071 /*
1072  * The relationship between dirty buffers and dirty pages:
1073  *
1074  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1075  * the page is tagged dirty in its radix tree.
1076  *
1077  * At all times, the dirtiness of the buffers represents the dirtiness of
1078  * subsections of the page.  If the page has buffers, the page dirty bit is
1079  * merely a hint about the true dirty state.
1080  *
1081  * When a page is set dirty in its entirety, all its buffers are marked dirty
1082  * (if the page has buffers).
1083  *
1084  * When a buffer is marked dirty, its page is dirtied, but the page's other
1085  * buffers are not.
1086  *
1087  * Also.  When blockdev buffers are explicitly read with bread(), they
1088  * individually become uptodate.  But their backing page remains not
1089  * uptodate - even if all of its buffers are uptodate.  A subsequent
1090  * block_read_full_page() against that page will discover all the uptodate
1091  * buffers, will set the page uptodate and will perform no I/O.
1092  */
1093
1094 /**
1095  * mark_buffer_dirty - mark a buffer_head as needing writeout
1096  * @bh: the buffer_head to mark dirty
1097  *
1098  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1099  * backing page dirty, then tag the page as dirty in its address_space's radix
1100  * tree and then attach the address_space's inode to its superblock's dirty
1101  * inode list.
1102  *
1103  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1104  * mapping->tree_lock and mapping->host->i_lock.
1105  */
1106 void mark_buffer_dirty(struct buffer_head *bh)
1107 {
1108         WARN_ON_ONCE(!buffer_uptodate(bh));
1109
1110         /*
1111          * Very *carefully* optimize the it-is-already-dirty case.
1112          *
1113          * Don't let the final "is it dirty" escape to before we
1114          * perhaps modified the buffer.
1115          */
1116         if (buffer_dirty(bh)) {
1117                 smp_mb();
1118                 if (buffer_dirty(bh))
1119                         return;
1120         }
1121
1122         if (!test_set_buffer_dirty(bh)) {
1123                 struct page *page = bh->b_page;
1124                 if (!TestSetPageDirty(page)) {
1125                         struct address_space *mapping = page_mapping(page);
1126                         if (mapping)
1127                                 __set_page_dirty(page, mapping, 0);
1128                 }
1129         }
1130 }
1131 EXPORT_SYMBOL(mark_buffer_dirty);
1132
1133 /*
1134  * Decrement a buffer_head's reference count.  If all buffers against a page
1135  * have zero reference count, are clean and unlocked, and if the page is clean
1136  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1137  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1138  * a page but it ends up not being freed, and buffers may later be reattached).
1139  */
1140 void __brelse(struct buffer_head * buf)
1141 {
1142         if (atomic_read(&buf->b_count)) {
1143                 put_bh(buf);
1144                 return;
1145         }
1146         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1147 }
1148 EXPORT_SYMBOL(__brelse);
1149
1150 /*
1151  * bforget() is like brelse(), except it discards any
1152  * potentially dirty data.
1153  */
1154 void __bforget(struct buffer_head *bh)
1155 {
1156         clear_buffer_dirty(bh);
1157         if (bh->b_assoc_map) {
1158                 struct address_space *buffer_mapping = bh->b_page->mapping;
1159
1160                 spin_lock(&buffer_mapping->private_lock);
1161                 list_del_init(&bh->b_assoc_buffers);
1162                 bh->b_assoc_map = NULL;
1163                 spin_unlock(&buffer_mapping->private_lock);
1164         }
1165         __brelse(bh);
1166 }
1167 EXPORT_SYMBOL(__bforget);
1168
1169 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1170 {
1171         lock_buffer(bh);
1172         if (buffer_uptodate(bh)) {
1173                 unlock_buffer(bh);
1174                 return bh;
1175         } else {
1176                 get_bh(bh);
1177                 bh->b_end_io = end_buffer_read_sync;
1178                 submit_bh(READ, bh);
1179                 wait_on_buffer(bh);
1180                 if (buffer_uptodate(bh))
1181                         return bh;
1182         }
1183         brelse(bh);
1184         return NULL;
1185 }
1186
1187 /*
1188  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1189  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1190  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1191  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1192  * CPU's LRUs at the same time.
1193  *
1194  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1195  * sb_find_get_block().
1196  *
1197  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1198  * a local interrupt disable for that.
1199  */
1200
1201 #define BH_LRU_SIZE     8
1202
1203 struct bh_lru {
1204         struct buffer_head *bhs[BH_LRU_SIZE];
1205 };
1206
1207 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1208
1209 #ifdef CONFIG_SMP
1210 #define bh_lru_lock()   local_irq_disable()
1211 #define bh_lru_unlock() local_irq_enable()
1212 #else
1213 #define bh_lru_lock()   preempt_disable()
1214 #define bh_lru_unlock() preempt_enable()
1215 #endif
1216
1217 static inline void check_irqs_on(void)
1218 {
1219 #ifdef irqs_disabled
1220         BUG_ON(irqs_disabled());
1221 #endif
1222 }
1223
1224 /*
1225  * The LRU management algorithm is dopey-but-simple.  Sorry.
1226  */
1227 static void bh_lru_install(struct buffer_head *bh)
1228 {
1229         struct buffer_head *evictee = NULL;
1230
1231         check_irqs_on();
1232         bh_lru_lock();
1233         if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1234                 struct buffer_head *bhs[BH_LRU_SIZE];
1235                 int in;
1236                 int out = 0;
1237
1238                 get_bh(bh);
1239                 bhs[out++] = bh;
1240                 for (in = 0; in < BH_LRU_SIZE; in++) {
1241                         struct buffer_head *bh2 =
1242                                 __this_cpu_read(bh_lrus.bhs[in]);
1243
1244                         if (bh2 == bh) {
1245                                 __brelse(bh2);
1246                         } else {
1247                                 if (out >= BH_LRU_SIZE) {
1248                                         BUG_ON(evictee != NULL);
1249                                         evictee = bh2;
1250                                 } else {
1251                                         bhs[out++] = bh2;
1252                                 }
1253                         }
1254                 }
1255                 while (out < BH_LRU_SIZE)
1256                         bhs[out++] = NULL;
1257                 memcpy(__this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1258         }
1259         bh_lru_unlock();
1260
1261         if (evictee)
1262                 __brelse(evictee);
1263 }
1264
1265 /*
1266  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1267  */
1268 static struct buffer_head *
1269 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1270 {
1271         struct buffer_head *ret = NULL;
1272         unsigned int i;
1273
1274         check_irqs_on();
1275         bh_lru_lock();
1276         for (i = 0; i < BH_LRU_SIZE; i++) {
1277                 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1278
1279                 if (bh && bh->b_bdev == bdev &&
1280                                 bh->b_blocknr == block && bh->b_size == size) {
1281                         if (i) {
1282                                 while (i) {
1283                                         __this_cpu_write(bh_lrus.bhs[i],
1284                                                 __this_cpu_read(bh_lrus.bhs[i - 1]));
1285                                         i--;
1286                                 }
1287                                 __this_cpu_write(bh_lrus.bhs[0], bh);
1288                         }
1289                         get_bh(bh);
1290                         ret = bh;
1291                         break;
1292                 }
1293         }
1294         bh_lru_unlock();
1295         return ret;
1296 }
1297
1298 /*
1299  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1300  * it in the LRU and mark it as accessed.  If it is not present then return
1301  * NULL
1302  */
1303 struct buffer_head *
1304 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1305 {
1306         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1307
1308         if (bh == NULL) {
1309                 bh = __find_get_block_slow(bdev, block);
1310                 if (bh)
1311                         bh_lru_install(bh);
1312         }
1313         if (bh)
1314                 touch_buffer(bh);
1315         return bh;
1316 }
1317 EXPORT_SYMBOL(__find_get_block);
1318
1319 /*
1320  * __getblk will locate (and, if necessary, create) the buffer_head
1321  * which corresponds to the passed block_device, block and size. The
1322  * returned buffer has its reference count incremented.
1323  *
1324  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1325  * illegal block number, __getblk() will happily return a buffer_head
1326  * which represents the non-existent block.  Very weird.
1327  *
1328  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1329  * attempt is failing.  FIXME, perhaps?
1330  */
1331 struct buffer_head *
1332 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1333 {
1334         struct buffer_head *bh = __find_get_block(bdev, block, size);
1335
1336         might_sleep();
1337         if (bh == NULL)
1338                 bh = __getblk_slow(bdev, block, size);
1339         return bh;
1340 }
1341 EXPORT_SYMBOL(__getblk);
1342
1343 /*
1344  * Do async read-ahead on a buffer..
1345  */
1346 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1347 {
1348         struct buffer_head *bh = __getblk(bdev, block, size);
1349         if (likely(bh)) {
1350                 ll_rw_block(READA, 1, &bh);
1351                 brelse(bh);
1352         }
1353 }
1354 EXPORT_SYMBOL(__breadahead);
1355
1356 /**
1357  *  __bread() - reads a specified block and returns the bh
1358  *  @bdev: the block_device to read from
1359  *  @block: number of block
1360  *  @size: size (in bytes) to read
1361  * 
1362  *  Reads a specified block, and returns buffer head that contains it.
1363  *  It returns NULL if the block was unreadable.
1364  */
1365 struct buffer_head *
1366 __bread(struct block_device *bdev, sector_t block, unsigned size)
1367 {
1368         struct buffer_head *bh = __getblk(bdev, block, size);
1369
1370         if (likely(bh) && !buffer_uptodate(bh))
1371                 bh = __bread_slow(bh);
1372         return bh;
1373 }
1374 EXPORT_SYMBOL(__bread);
1375
1376 /*
1377  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1378  * This doesn't race because it runs in each cpu either in irq
1379  * or with preempt disabled.
1380  */
1381 static void invalidate_bh_lru(void *arg)
1382 {
1383         struct bh_lru *b = &get_cpu_var(bh_lrus);
1384         int i;
1385
1386         for (i = 0; i < BH_LRU_SIZE; i++) {
1387                 brelse(b->bhs[i]);
1388                 b->bhs[i] = NULL;
1389         }
1390         put_cpu_var(bh_lrus);
1391 }
1392
1393 static bool has_bh_in_lru(int cpu, void *dummy)
1394 {
1395         struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1396         int i;
1397         
1398         for (i = 0; i < BH_LRU_SIZE; i++) {
1399                 if (b->bhs[i])
1400                         return 1;
1401         }
1402
1403         return 0;
1404 }
1405
1406 void invalidate_bh_lrus(void)
1407 {
1408         on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1409 }
1410 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1411
1412 void set_bh_page(struct buffer_head *bh,
1413                 struct page *page, unsigned long offset)
1414 {
1415         bh->b_page = page;
1416         BUG_ON(offset >= PAGE_SIZE);
1417         if (PageHighMem(page))
1418                 /*
1419                  * This catches illegal uses and preserves the offset:
1420                  */
1421                 bh->b_data = (char *)(0 + offset);
1422         else
1423                 bh->b_data = page_address(page) + offset;
1424 }
1425 EXPORT_SYMBOL(set_bh_page);
1426
1427 /*
1428  * Called when truncating a buffer on a page completely.
1429  */
1430 static void discard_buffer(struct buffer_head * bh)
1431 {
1432         lock_buffer(bh);
1433         clear_buffer_dirty(bh);
1434         bh->b_bdev = NULL;
1435         clear_buffer_mapped(bh);
1436         clear_buffer_req(bh);
1437         clear_buffer_new(bh);
1438         clear_buffer_delay(bh);
1439         clear_buffer_unwritten(bh);
1440         unlock_buffer(bh);
1441 }
1442
1443 /**
1444  * block_invalidatepage - invalidate part or all of a buffer-backed page
1445  *
1446  * @page: the page which is affected
1447  * @offset: the index of the truncation point
1448  *
1449  * block_invalidatepage() is called when all or part of the page has become
1450  * invalidated by a truncate operation.
1451  *
1452  * block_invalidatepage() does not have to release all buffers, but it must
1453  * ensure that no dirty buffer is left outside @offset and that no I/O
1454  * is underway against any of the blocks which are outside the truncation
1455  * point.  Because the caller is about to free (and possibly reuse) those
1456  * blocks on-disk.
1457  */
1458 void block_invalidatepage(struct page *page, unsigned long offset)
1459 {
1460         struct buffer_head *head, *bh, *next;
1461         unsigned int curr_off = 0;
1462
1463         BUG_ON(!PageLocked(page));
1464         if (!page_has_buffers(page))
1465                 goto out;
1466
1467         head = page_buffers(page);
1468         bh = head;
1469         do {
1470                 unsigned int next_off = curr_off + bh->b_size;
1471                 next = bh->b_this_page;
1472
1473                 /*
1474                  * is this block fully invalidated?
1475                  */
1476                 if (offset <= curr_off)
1477                         discard_buffer(bh);
1478                 curr_off = next_off;
1479                 bh = next;
1480         } while (bh != head);
1481
1482         /*
1483          * We release buffers only if the entire page is being invalidated.
1484          * The get_block cached value has been unconditionally invalidated,
1485          * so real IO is not possible anymore.
1486          */
1487         if (offset == 0)
1488                 try_to_release_page(page, 0);
1489 out:
1490         return;
1491 }
1492 EXPORT_SYMBOL(block_invalidatepage);
1493
1494 /*
1495  * We attach and possibly dirty the buffers atomically wrt
1496  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1497  * is already excluded via the page lock.
1498  */
1499 void create_empty_buffers(struct page *page,
1500                         unsigned long blocksize, unsigned long b_state)
1501 {
1502         struct buffer_head *bh, *head, *tail;
1503
1504         head = alloc_page_buffers(page, blocksize, 1);
1505         bh = head;
1506         do {
1507                 bh->b_state |= b_state;
1508                 tail = bh;
1509                 bh = bh->b_this_page;
1510         } while (bh);
1511         tail->b_this_page = head;
1512
1513         spin_lock(&page->mapping->private_lock);
1514         if (PageUptodate(page) || PageDirty(page)) {
1515                 bh = head;
1516                 do {
1517                         if (PageDirty(page))
1518                                 set_buffer_dirty(bh);
1519                         if (PageUptodate(page))
1520                                 set_buffer_uptodate(bh);
1521                         bh = bh->b_this_page;
1522                 } while (bh != head);
1523         }
1524         attach_page_buffers(page, head);
1525         spin_unlock(&page->mapping->private_lock);
1526 }
1527 EXPORT_SYMBOL(create_empty_buffers);
1528
1529 /*
1530  * We are taking a block for data and we don't want any output from any
1531  * buffer-cache aliases starting from return from that function and
1532  * until the moment when something will explicitly mark the buffer
1533  * dirty (hopefully that will not happen until we will free that block ;-)
1534  * We don't even need to mark it not-uptodate - nobody can expect
1535  * anything from a newly allocated buffer anyway. We used to used
1536  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1537  * don't want to mark the alias unmapped, for example - it would confuse
1538  * anyone who might pick it with bread() afterwards...
1539  *
1540  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1541  * be writeout I/O going on against recently-freed buffers.  We don't
1542  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1543  * only if we really need to.  That happens here.
1544  */
1545 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1546 {
1547         struct buffer_head *old_bh;
1548
1549         might_sleep();
1550
1551         old_bh = __find_get_block_slow(bdev, block);
1552         if (old_bh) {
1553                 clear_buffer_dirty(old_bh);
1554                 wait_on_buffer(old_bh);
1555                 clear_buffer_req(old_bh);
1556                 __brelse(old_bh);
1557         }
1558 }
1559 EXPORT_SYMBOL(unmap_underlying_metadata);
1560
1561 /*
1562  * NOTE! All mapped/uptodate combinations are valid:
1563  *
1564  *      Mapped  Uptodate        Meaning
1565  *
1566  *      No      No              "unknown" - must do get_block()
1567  *      No      Yes             "hole" - zero-filled
1568  *      Yes     No              "allocated" - allocated on disk, not read in
1569  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1570  *
1571  * "Dirty" is valid only with the last case (mapped+uptodate).
1572  */
1573
1574 /*
1575  * While block_write_full_page is writing back the dirty buffers under
1576  * the page lock, whoever dirtied the buffers may decide to clean them
1577  * again at any time.  We handle that by only looking at the buffer
1578  * state inside lock_buffer().
1579  *
1580  * If block_write_full_page() is called for regular writeback
1581  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1582  * locked buffer.   This only can happen if someone has written the buffer
1583  * directly, with submit_bh().  At the address_space level PageWriteback
1584  * prevents this contention from occurring.
1585  *
1586  * If block_write_full_page() is called with wbc->sync_mode ==
1587  * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1588  * causes the writes to be flagged as synchronous writes.
1589  */
1590 static int __block_write_full_page(struct inode *inode, struct page *page,
1591                         get_block_t *get_block, struct writeback_control *wbc,
1592                         bh_end_io_t *handler)
1593 {
1594         int err;
1595         sector_t block;
1596         sector_t last_block;
1597         struct buffer_head *bh, *head;
1598         const unsigned blocksize = 1 << inode->i_blkbits;
1599         int nr_underway = 0;
1600         int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1601                         WRITE_SYNC : WRITE);
1602
1603         BUG_ON(!PageLocked(page));
1604
1605         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1606
1607         if (!page_has_buffers(page)) {
1608                 create_empty_buffers(page, blocksize,
1609                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1610         }
1611
1612         /*
1613          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1614          * here, and the (potentially unmapped) buffers may become dirty at
1615          * any time.  If a buffer becomes dirty here after we've inspected it
1616          * then we just miss that fact, and the page stays dirty.
1617          *
1618          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1619          * handle that here by just cleaning them.
1620          */
1621
1622         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1623         head = page_buffers(page);
1624         bh = head;
1625
1626         /*
1627          * Get all the dirty buffers mapped to disk addresses and
1628          * handle any aliases from the underlying blockdev's mapping.
1629          */
1630         do {
1631                 if (block > last_block) {
1632                         /*
1633                          * mapped buffers outside i_size will occur, because
1634                          * this page can be outside i_size when there is a
1635                          * truncate in progress.
1636                          */
1637                         /*
1638                          * The buffer was zeroed by block_write_full_page()
1639                          */
1640                         clear_buffer_dirty(bh);
1641                         set_buffer_uptodate(bh);
1642                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1643                            buffer_dirty(bh)) {
1644                         WARN_ON(bh->b_size != blocksize);
1645                         err = get_block(inode, block, bh, 1);
1646                         if (err)
1647                                 goto recover;
1648                         clear_buffer_delay(bh);
1649                         if (buffer_new(bh)) {
1650                                 /* blockdev mappings never come here */
1651                                 clear_buffer_new(bh);
1652                                 unmap_underlying_metadata(bh->b_bdev,
1653                                                         bh->b_blocknr);
1654                         }
1655                 }
1656                 bh = bh->b_this_page;
1657                 block++;
1658         } while (bh != head);
1659
1660         do {
1661                 if (!buffer_mapped(bh))
1662                         continue;
1663                 /*
1664                  * If it's a fully non-blocking write attempt and we cannot
1665                  * lock the buffer then redirty the page.  Note that this can
1666                  * potentially cause a busy-wait loop from writeback threads
1667                  * and kswapd activity, but those code paths have their own
1668                  * higher-level throttling.
1669                  */
1670                 if (wbc->sync_mode != WB_SYNC_NONE) {
1671                         lock_buffer(bh);
1672                 } else if (!trylock_buffer(bh)) {
1673                         redirty_page_for_writepage(wbc, page);
1674                         continue;
1675                 }
1676                 if (test_clear_buffer_dirty(bh)) {
1677                         mark_buffer_async_write_endio(bh, handler);
1678                 } else {
1679                         unlock_buffer(bh);
1680                 }
1681         } while ((bh = bh->b_this_page) != head);
1682
1683         /*
1684          * The page and its buffers are protected by PageWriteback(), so we can
1685          * drop the bh refcounts early.
1686          */
1687         BUG_ON(PageWriteback(page));
1688         set_page_writeback(page);
1689
1690         do {
1691                 struct buffer_head *next = bh->b_this_page;
1692                 if (buffer_async_write(bh)) {
1693                         submit_bh(write_op, bh);
1694                         nr_underway++;
1695                 }
1696                 bh = next;
1697         } while (bh != head);
1698         unlock_page(page);
1699
1700         err = 0;
1701 done:
1702         if (nr_underway == 0) {
1703                 /*
1704                  * The page was marked dirty, but the buffers were
1705                  * clean.  Someone wrote them back by hand with
1706                  * ll_rw_block/submit_bh.  A rare case.
1707                  */
1708                 end_page_writeback(page);
1709
1710                 /*
1711                  * The page and buffer_heads can be released at any time from
1712                  * here on.
1713                  */
1714         }
1715         return err;
1716
1717 recover:
1718         /*
1719          * ENOSPC, or some other error.  We may already have added some
1720          * blocks to the file, so we need to write these out to avoid
1721          * exposing stale data.
1722          * The page is currently locked and not marked for writeback
1723          */
1724         bh = head;
1725         /* Recovery: lock and submit the mapped buffers */
1726         do {
1727                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1728                     !buffer_delay(bh)) {
1729                         lock_buffer(bh);
1730                         mark_buffer_async_write_endio(bh, handler);
1731                 } else {
1732                         /*
1733                          * The buffer may have been set dirty during
1734                          * attachment to a dirty page.
1735                          */
1736                         clear_buffer_dirty(bh);
1737                 }
1738         } while ((bh = bh->b_this_page) != head);
1739         SetPageError(page);
1740         BUG_ON(PageWriteback(page));
1741         mapping_set_error(page->mapping, err);
1742         set_page_writeback(page);
1743         do {
1744                 struct buffer_head *next = bh->b_this_page;
1745                 if (buffer_async_write(bh)) {
1746                         clear_buffer_dirty(bh);
1747                         submit_bh(write_op, bh);
1748                         nr_underway++;
1749                 }
1750                 bh = next;
1751         } while (bh != head);
1752         unlock_page(page);
1753         goto done;
1754 }
1755
1756 /*
1757  * If a page has any new buffers, zero them out here, and mark them uptodate
1758  * and dirty so they'll be written out (in order to prevent uninitialised
1759  * block data from leaking). And clear the new bit.
1760  */
1761 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1762 {
1763         unsigned int block_start, block_end;
1764         struct buffer_head *head, *bh;
1765
1766         BUG_ON(!PageLocked(page));
1767         if (!page_has_buffers(page))
1768                 return;
1769
1770         bh = head = page_buffers(page);
1771         block_start = 0;
1772         do {
1773                 block_end = block_start + bh->b_size;
1774
1775                 if (buffer_new(bh)) {
1776                         if (block_end > from && block_start < to) {
1777                                 if (!PageUptodate(page)) {
1778                                         unsigned start, size;
1779
1780                                         start = max(from, block_start);
1781                                         size = min(to, block_end) - start;
1782
1783                                         zero_user(page, start, size);
1784                                         set_buffer_uptodate(bh);
1785                                 }
1786
1787                                 clear_buffer_new(bh);
1788                                 mark_buffer_dirty(bh);
1789                         }
1790                 }
1791
1792                 block_start = block_end;
1793                 bh = bh->b_this_page;
1794         } while (bh != head);
1795 }
1796 EXPORT_SYMBOL(page_zero_new_buffers);
1797
1798 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1799                 get_block_t *get_block)
1800 {
1801         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1802         unsigned to = from + len;
1803         struct inode *inode = page->mapping->host;
1804         unsigned block_start, block_end;
1805         sector_t block;
1806         int err = 0;
1807         unsigned blocksize, bbits;
1808         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1809
1810         BUG_ON(!PageLocked(page));
1811         BUG_ON(from > PAGE_CACHE_SIZE);
1812         BUG_ON(to > PAGE_CACHE_SIZE);
1813         BUG_ON(from > to);
1814
1815         blocksize = 1 << inode->i_blkbits;
1816         if (!page_has_buffers(page))
1817                 create_empty_buffers(page, blocksize, 0);
1818         head = page_buffers(page);
1819
1820         bbits = inode->i_blkbits;
1821         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1822
1823         for(bh = head, block_start = 0; bh != head || !block_start;
1824             block++, block_start=block_end, bh = bh->b_this_page) {
1825                 block_end = block_start + blocksize;
1826                 if (block_end <= from || block_start >= to) {
1827                         if (PageUptodate(page)) {
1828                                 if (!buffer_uptodate(bh))
1829                                         set_buffer_uptodate(bh);
1830                         }
1831                         continue;
1832                 }
1833                 if (buffer_new(bh))
1834                         clear_buffer_new(bh);
1835                 if (!buffer_mapped(bh)) {
1836                         WARN_ON(bh->b_size != blocksize);
1837                         err = get_block(inode, block, bh, 1);
1838                         if (err)
1839                                 break;
1840                         if (buffer_new(bh)) {
1841                                 unmap_underlying_metadata(bh->b_bdev,
1842                                                         bh->b_blocknr);
1843                                 if (PageUptodate(page)) {
1844                                         clear_buffer_new(bh);
1845                                         set_buffer_uptodate(bh);
1846                                         mark_buffer_dirty(bh);
1847                                         continue;
1848                                 }
1849                                 if (block_end > to || block_start < from)
1850                                         zero_user_segments(page,
1851                                                 to, block_end,
1852                                                 block_start, from);
1853                                 continue;
1854                         }
1855                 }
1856                 if (PageUptodate(page)) {
1857                         if (!buffer_uptodate(bh))
1858                                 set_buffer_uptodate(bh);
1859                         continue; 
1860                 }
1861                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1862                     !buffer_unwritten(bh) &&
1863                      (block_start < from || block_end > to)) {
1864                         ll_rw_block(READ, 1, &bh);
1865                         *wait_bh++=bh;
1866                 }
1867         }
1868         /*
1869          * If we issued read requests - let them complete.
1870          */
1871         while(wait_bh > wait) {
1872                 wait_on_buffer(*--wait_bh);
1873                 if (!buffer_uptodate(*wait_bh))
1874                         err = -EIO;
1875         }
1876         if (unlikely(err))
1877                 page_zero_new_buffers(page, from, to);
1878         return err;
1879 }
1880 EXPORT_SYMBOL(__block_write_begin);
1881
1882 static int __block_commit_write(struct inode *inode, struct page *page,
1883                 unsigned from, unsigned to)
1884 {
1885         unsigned block_start, block_end;
1886         int partial = 0;
1887         unsigned blocksize;
1888         struct buffer_head *bh, *head;
1889
1890         blocksize = 1 << inode->i_blkbits;
1891
1892         for(bh = head = page_buffers(page), block_start = 0;
1893             bh != head || !block_start;
1894             block_start=block_end, bh = bh->b_this_page) {
1895                 block_end = block_start + blocksize;
1896                 if (block_end <= from || block_start >= to) {
1897                         if (!buffer_uptodate(bh))
1898                                 partial = 1;
1899                 } else {
1900                         set_buffer_uptodate(bh);
1901                         mark_buffer_dirty(bh);
1902                 }
1903                 clear_buffer_new(bh);
1904         }
1905
1906         /*
1907          * If this is a partial write which happened to make all buffers
1908          * uptodate then we can optimize away a bogus readpage() for
1909          * the next read(). Here we 'discover' whether the page went
1910          * uptodate as a result of this (potentially partial) write.
1911          */
1912         if (!partial)
1913                 SetPageUptodate(page);
1914         return 0;
1915 }
1916
1917 /*
1918  * block_write_begin takes care of the basic task of block allocation and
1919  * bringing partial write blocks uptodate first.
1920  *
1921  * The filesystem needs to handle block truncation upon failure.
1922  */
1923 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
1924                 unsigned flags, struct page **pagep, get_block_t *get_block)
1925 {
1926         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1927         struct page *page;
1928         int status;
1929
1930         page = grab_cache_page_write_begin(mapping, index, flags);
1931         if (!page)
1932                 return -ENOMEM;
1933
1934         status = __block_write_begin(page, pos, len, get_block);
1935         if (unlikely(status)) {
1936                 unlock_page(page);
1937                 page_cache_release(page);
1938                 page = NULL;
1939         }
1940
1941         *pagep = page;
1942         return status;
1943 }
1944 EXPORT_SYMBOL(block_write_begin);
1945
1946 int block_write_end(struct file *file, struct address_space *mapping,
1947                         loff_t pos, unsigned len, unsigned copied,
1948                         struct page *page, void *fsdata)
1949 {
1950         struct inode *inode = mapping->host;
1951         unsigned start;
1952
1953         start = pos & (PAGE_CACHE_SIZE - 1);
1954
1955         if (unlikely(copied < len)) {
1956                 /*
1957                  * The buffers that were written will now be uptodate, so we
1958                  * don't have to worry about a readpage reading them and
1959                  * overwriting a partial write. However if we have encountered
1960                  * a short write and only partially written into a buffer, it
1961                  * will not be marked uptodate, so a readpage might come in and
1962                  * destroy our partial write.
1963                  *
1964                  * Do the simplest thing, and just treat any short write to a
1965                  * non uptodate page as a zero-length write, and force the
1966                  * caller to redo the whole thing.
1967                  */
1968                 if (!PageUptodate(page))
1969                         copied = 0;
1970
1971                 page_zero_new_buffers(page, start+copied, start+len);
1972         }
1973         flush_dcache_page(page);
1974
1975         /* This could be a short (even 0-length) commit */
1976         __block_commit_write(inode, page, start, start+copied);
1977
1978         return copied;
1979 }
1980 EXPORT_SYMBOL(block_write_end);
1981
1982 int generic_write_end(struct file *file, struct address_space *mapping,
1983                         loff_t pos, unsigned len, unsigned copied,
1984                         struct page *page, void *fsdata)
1985 {
1986         struct inode *inode = mapping->host;
1987         int i_size_changed = 0;
1988
1989         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1990
1991         /*
1992          * No need to use i_size_read() here, the i_size
1993          * cannot change under us because we hold i_mutex.
1994          *
1995          * But it's important to update i_size while still holding page lock:
1996          * page writeout could otherwise come in and zero beyond i_size.
1997          */
1998         if (pos+copied > inode->i_size) {
1999                 i_size_write(inode, pos+copied);
2000                 i_size_changed = 1;
2001         }
2002
2003         unlock_page(page);
2004         page_cache_release(page);
2005
2006         /*
2007          * Don't mark the inode dirty under page lock. First, it unnecessarily
2008          * makes the holding time of page lock longer. Second, it forces lock
2009          * ordering of page lock and transaction start for journaling
2010          * filesystems.
2011          */
2012         if (i_size_changed)
2013                 mark_inode_dirty(inode);
2014
2015         return copied;
2016 }
2017 EXPORT_SYMBOL(generic_write_end);
2018
2019 /*
2020  * block_is_partially_uptodate checks whether buffers within a page are
2021  * uptodate or not.
2022  *
2023  * Returns true if all buffers which correspond to a file portion
2024  * we want to read are uptodate.
2025  */
2026 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2027                                         unsigned long from)
2028 {
2029         struct inode *inode = page->mapping->host;
2030         unsigned block_start, block_end, blocksize;
2031         unsigned to;
2032         struct buffer_head *bh, *head;
2033         int ret = 1;
2034
2035         if (!page_has_buffers(page))
2036                 return 0;
2037
2038         blocksize = 1 << inode->i_blkbits;
2039         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2040         to = from + to;
2041         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2042                 return 0;
2043
2044         head = page_buffers(page);
2045         bh = head;
2046         block_start = 0;
2047         do {
2048                 block_end = block_start + blocksize;
2049                 if (block_end > from && block_start < to) {
2050                         if (!buffer_uptodate(bh)) {
2051                                 ret = 0;
2052                                 break;
2053                         }
2054                         if (block_end >= to)
2055                                 break;
2056                 }
2057                 block_start = block_end;
2058                 bh = bh->b_this_page;
2059         } while (bh != head);
2060
2061         return ret;
2062 }
2063 EXPORT_SYMBOL(block_is_partially_uptodate);
2064
2065 /*
2066  * Generic "read page" function for block devices that have the normal
2067  * get_block functionality. This is most of the block device filesystems.
2068  * Reads the page asynchronously --- the unlock_buffer() and
2069  * set/clear_buffer_uptodate() functions propagate buffer state into the
2070  * page struct once IO has completed.
2071  */
2072 int block_read_full_page(struct page *page, get_block_t *get_block)
2073 {
2074         struct inode *inode = page->mapping->host;
2075         sector_t iblock, lblock;
2076         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2077         unsigned int blocksize;
2078         int nr, i;
2079         int fully_mapped = 1;
2080
2081         BUG_ON(!PageLocked(page));
2082         blocksize = 1 << inode->i_blkbits;
2083         if (!page_has_buffers(page))
2084                 create_empty_buffers(page, blocksize, 0);
2085         head = page_buffers(page);
2086
2087         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2088         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2089         bh = head;
2090         nr = 0;
2091         i = 0;
2092
2093         do {
2094                 if (buffer_uptodate(bh))
2095                         continue;
2096
2097                 if (!buffer_mapped(bh)) {
2098                         int err = 0;
2099
2100                         fully_mapped = 0;
2101                         if (iblock < lblock) {
2102                                 WARN_ON(bh->b_size != blocksize);
2103                                 err = get_block(inode, iblock, bh, 0);
2104                                 if (err)
2105                                         SetPageError(page);
2106                         }
2107                         if (!buffer_mapped(bh)) {
2108                                 zero_user(page, i * blocksize, blocksize);
2109                                 if (!err)
2110                                         set_buffer_uptodate(bh);
2111                                 continue;
2112                         }
2113                         /*
2114                          * get_block() might have updated the buffer
2115                          * synchronously
2116                          */
2117                         if (buffer_uptodate(bh))
2118                                 continue;
2119                 }
2120                 arr[nr++] = bh;
2121         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2122
2123         if (fully_mapped)
2124                 SetPageMappedToDisk(page);
2125
2126         if (!nr) {
2127                 /*
2128                  * All buffers are uptodate - we can set the page uptodate
2129                  * as well. But not if get_block() returned an error.
2130                  */
2131                 if (!PageError(page))
2132                         SetPageUptodate(page);
2133                 unlock_page(page);
2134                 return 0;
2135         }
2136
2137         /* Stage two: lock the buffers */
2138         for (i = 0; i < nr; i++) {
2139                 bh = arr[i];
2140                 lock_buffer(bh);
2141                 mark_buffer_async_read(bh);
2142         }
2143
2144         /*
2145          * Stage 3: start the IO.  Check for uptodateness
2146          * inside the buffer lock in case another process reading
2147          * the underlying blockdev brought it uptodate (the sct fix).
2148          */
2149         for (i = 0; i < nr; i++) {
2150                 bh = arr[i];
2151                 if (buffer_uptodate(bh))
2152                         end_buffer_async_read(bh, 1);
2153                 else
2154                         submit_bh(READ, bh);
2155         }
2156         return 0;
2157 }
2158 EXPORT_SYMBOL(block_read_full_page);
2159
2160 /* utility function for filesystems that need to do work on expanding
2161  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2162  * deal with the hole.  
2163  */
2164 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2165 {
2166         struct address_space *mapping = inode->i_mapping;
2167         struct page *page;
2168         void *fsdata;
2169         int err;
2170
2171         err = inode_newsize_ok(inode, size);
2172         if (err)
2173                 goto out;
2174
2175         err = pagecache_write_begin(NULL, mapping, size, 0,
2176                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2177                                 &page, &fsdata);
2178         if (err)
2179                 goto out;
2180
2181         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2182         BUG_ON(err > 0);
2183
2184 out:
2185         return err;
2186 }
2187 EXPORT_SYMBOL(generic_cont_expand_simple);
2188
2189 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2190                             loff_t pos, loff_t *bytes)
2191 {
2192         struct inode *inode = mapping->host;
2193         unsigned blocksize = 1 << inode->i_blkbits;
2194         struct page *page;
2195         void *fsdata;
2196         pgoff_t index, curidx;
2197         loff_t curpos;
2198         unsigned zerofrom, offset, len;
2199         int err = 0;
2200
2201         index = pos >> PAGE_CACHE_SHIFT;
2202         offset = pos & ~PAGE_CACHE_MASK;
2203
2204         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2205                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2206                 if (zerofrom & (blocksize-1)) {
2207                         *bytes |= (blocksize-1);
2208                         (*bytes)++;
2209                 }
2210                 len = PAGE_CACHE_SIZE - zerofrom;
2211
2212                 err = pagecache_write_begin(file, mapping, curpos, len,
2213                                                 AOP_FLAG_UNINTERRUPTIBLE,
2214                                                 &page, &fsdata);
2215                 if (err)
2216                         goto out;
2217                 zero_user(page, zerofrom, len);
2218                 err = pagecache_write_end(file, mapping, curpos, len, len,
2219                                                 page, fsdata);
2220                 if (err < 0)
2221                         goto out;
2222                 BUG_ON(err != len);
2223                 err = 0;
2224
2225                 balance_dirty_pages_ratelimited(mapping);
2226         }
2227
2228         /* page covers the boundary, find the boundary offset */
2229         if (index == curidx) {
2230                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2231                 /* if we will expand the thing last block will be filled */
2232                 if (offset <= zerofrom) {
2233                         goto out;
2234                 }
2235                 if (zerofrom & (blocksize-1)) {
2236                         *bytes |= (blocksize-1);
2237                         (*bytes)++;
2238                 }
2239                 len = offset - zerofrom;
2240
2241                 err = pagecache_write_begin(file, mapping, curpos, len,
2242                                                 AOP_FLAG_UNINTERRUPTIBLE,
2243                                                 &page, &fsdata);
2244                 if (err)
2245                         goto out;
2246                 zero_user(page, zerofrom, len);
2247                 err = pagecache_write_end(file, mapping, curpos, len, len,
2248                                                 page, fsdata);
2249                 if (err < 0)
2250                         goto out;
2251                 BUG_ON(err != len);
2252                 err = 0;
2253         }
2254 out:
2255         return err;
2256 }
2257
2258 /*
2259  * For moronic filesystems that do not allow holes in file.
2260  * We may have to extend the file.
2261  */
2262 int cont_write_begin(struct file *file, struct address_space *mapping,
2263                         loff_t pos, unsigned len, unsigned flags,
2264                         struct page **pagep, void **fsdata,
2265                         get_block_t *get_block, loff_t *bytes)
2266 {
2267         struct inode *inode = mapping->host;
2268         unsigned blocksize = 1 << inode->i_blkbits;
2269         unsigned zerofrom;
2270         int err;
2271
2272         err = cont_expand_zero(file, mapping, pos, bytes);
2273         if (err)
2274                 return err;
2275
2276         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2277         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2278                 *bytes |= (blocksize-1);
2279                 (*bytes)++;
2280         }
2281
2282         return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2283 }
2284 EXPORT_SYMBOL(cont_write_begin);
2285
2286 int block_commit_write(struct page *page, unsigned from, unsigned to)
2287 {
2288         struct inode *inode = page->mapping->host;
2289         __block_commit_write(inode,page,from,to);
2290         return 0;
2291 }
2292 EXPORT_SYMBOL(block_commit_write);
2293
2294 /*
2295  * block_page_mkwrite() is not allowed to change the file size as it gets
2296  * called from a page fault handler when a page is first dirtied. Hence we must
2297  * be careful to check for EOF conditions here. We set the page up correctly
2298  * for a written page which means we get ENOSPC checking when writing into
2299  * holes and correct delalloc and unwritten extent mapping on filesystems that
2300  * support these features.
2301  *
2302  * We are not allowed to take the i_mutex here so we have to play games to
2303  * protect against truncate races as the page could now be beyond EOF.  Because
2304  * truncate writes the inode size before removing pages, once we have the
2305  * page lock we can determine safely if the page is beyond EOF. If it is not
2306  * beyond EOF, then the page is guaranteed safe against truncation until we
2307  * unlock the page.
2308  *
2309  * Direct callers of this function should call vfs_check_frozen() so that page
2310  * fault does not busyloop until the fs is thawed.
2311  */
2312 int __block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2313                          get_block_t get_block)
2314 {
2315         struct page *page = vmf->page;
2316         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2317         unsigned long end;
2318         loff_t size;
2319         int ret;
2320
2321         lock_page(page);
2322         size = i_size_read(inode);
2323         if ((page->mapping != inode->i_mapping) ||
2324             (page_offset(page) > size)) {
2325                 /* We overload EFAULT to mean page got truncated */
2326                 ret = -EFAULT;
2327                 goto out_unlock;
2328         }
2329
2330         /* page is wholly or partially inside EOF */
2331         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2332                 end = size & ~PAGE_CACHE_MASK;
2333         else
2334                 end = PAGE_CACHE_SIZE;
2335
2336         ret = __block_write_begin(page, 0, end, get_block);
2337         if (!ret)
2338                 ret = block_commit_write(page, 0, end);
2339
2340         if (unlikely(ret < 0))
2341                 goto out_unlock;
2342         /*
2343          * Freezing in progress? We check after the page is marked dirty and
2344          * with page lock held so if the test here fails, we are sure freezing
2345          * code will wait during syncing until the page fault is done - at that
2346          * point page will be dirty and unlocked so freezing code will write it
2347          * and writeprotect it again.
2348          */
2349         set_page_dirty(page);
2350         if (inode->i_sb->s_frozen != SB_UNFROZEN) {
2351                 ret = -EAGAIN;
2352                 goto out_unlock;
2353         }
2354         wait_on_page_writeback(page);
2355         return 0;
2356 out_unlock:
2357         unlock_page(page);
2358         return ret;
2359 }
2360 EXPORT_SYMBOL(__block_page_mkwrite);
2361
2362 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2363                    get_block_t get_block)
2364 {
2365         int ret;
2366         struct super_block *sb = vma->vm_file->f_path.dentry->d_inode->i_sb;
2367
2368         /*
2369          * This check is racy but catches the common case. The check in
2370          * __block_page_mkwrite() is reliable.
2371          */
2372         vfs_check_frozen(sb, SB_FREEZE_WRITE);
2373         ret = __block_page_mkwrite(vma, vmf, get_block);
2374         return block_page_mkwrite_return(ret);
2375 }
2376 EXPORT_SYMBOL(block_page_mkwrite);
2377
2378 /*
2379  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2380  * immediately, while under the page lock.  So it needs a special end_io
2381  * handler which does not touch the bh after unlocking it.
2382  */
2383 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2384 {
2385         __end_buffer_read_notouch(bh, uptodate);
2386 }
2387
2388 /*
2389  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2390  * the page (converting it to circular linked list and taking care of page
2391  * dirty races).
2392  */
2393 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2394 {
2395         struct buffer_head *bh;
2396
2397         BUG_ON(!PageLocked(page));
2398
2399         spin_lock(&page->mapping->private_lock);
2400         bh = head;
2401         do {
2402                 if (PageDirty(page))
2403                         set_buffer_dirty(bh);
2404                 if (!bh->b_this_page)
2405                         bh->b_this_page = head;
2406                 bh = bh->b_this_page;
2407         } while (bh != head);
2408         attach_page_buffers(page, head);
2409         spin_unlock(&page->mapping->private_lock);
2410 }
2411
2412 /*
2413  * On entry, the page is fully not uptodate.
2414  * On exit the page is fully uptodate in the areas outside (from,to)
2415  * The filesystem needs to handle block truncation upon failure.
2416  */
2417 int nobh_write_begin(struct address_space *mapping,
2418                         loff_t pos, unsigned len, unsigned flags,
2419                         struct page **pagep, void **fsdata,
2420                         get_block_t *get_block)
2421 {
2422         struct inode *inode = mapping->host;
2423         const unsigned blkbits = inode->i_blkbits;
2424         const unsigned blocksize = 1 << blkbits;
2425         struct buffer_head *head, *bh;
2426         struct page *page;
2427         pgoff_t index;
2428         unsigned from, to;
2429         unsigned block_in_page;
2430         unsigned block_start, block_end;
2431         sector_t block_in_file;
2432         int nr_reads = 0;
2433         int ret = 0;
2434         int is_mapped_to_disk = 1;
2435
2436         index = pos >> PAGE_CACHE_SHIFT;
2437         from = pos & (PAGE_CACHE_SIZE - 1);
2438         to = from + len;
2439
2440         page = grab_cache_page_write_begin(mapping, index, flags);
2441         if (!page)
2442                 return -ENOMEM;
2443         *pagep = page;
2444         *fsdata = NULL;
2445
2446         if (page_has_buffers(page)) {
2447                 ret = __block_write_begin(page, pos, len, get_block);
2448                 if (unlikely(ret))
2449                         goto out_release;
2450                 return ret;
2451         }
2452
2453         if (PageMappedToDisk(page))
2454                 return 0;
2455
2456         /*
2457          * Allocate buffers so that we can keep track of state, and potentially
2458          * attach them to the page if an error occurs. In the common case of
2459          * no error, they will just be freed again without ever being attached
2460          * to the page (which is all OK, because we're under the page lock).
2461          *
2462          * Be careful: the buffer linked list is a NULL terminated one, rather
2463          * than the circular one we're used to.
2464          */
2465         head = alloc_page_buffers(page, blocksize, 0);
2466         if (!head) {
2467                 ret = -ENOMEM;
2468                 goto out_release;
2469         }
2470
2471         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2472
2473         /*
2474          * We loop across all blocks in the page, whether or not they are
2475          * part of the affected region.  This is so we can discover if the
2476          * page is fully mapped-to-disk.
2477          */
2478         for (block_start = 0, block_in_page = 0, bh = head;
2479                   block_start < PAGE_CACHE_SIZE;
2480                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2481                 int create;
2482
2483                 block_end = block_start + blocksize;
2484                 bh->b_state = 0;
2485                 create = 1;
2486                 if (block_start >= to)
2487                         create = 0;
2488                 ret = get_block(inode, block_in_file + block_in_page,
2489                                         bh, create);
2490                 if (ret)
2491                         goto failed;
2492                 if (!buffer_mapped(bh))
2493                         is_mapped_to_disk = 0;
2494                 if (buffer_new(bh))
2495                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2496                 if (PageUptodate(page)) {
2497                         set_buffer_uptodate(bh);
2498                         continue;
2499                 }
2500                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2501                         zero_user_segments(page, block_start, from,
2502                                                         to, block_end);
2503                         continue;
2504                 }
2505                 if (buffer_uptodate(bh))
2506                         continue;       /* reiserfs does this */
2507                 if (block_start < from || block_end > to) {
2508                         lock_buffer(bh);
2509                         bh->b_end_io = end_buffer_read_nobh;
2510                         submit_bh(READ, bh);
2511                         nr_reads++;
2512                 }
2513         }
2514
2515         if (nr_reads) {
2516                 /*
2517                  * The page is locked, so these buffers are protected from
2518                  * any VM or truncate activity.  Hence we don't need to care
2519                  * for the buffer_head refcounts.
2520                  */
2521                 for (bh = head; bh; bh = bh->b_this_page) {
2522                         wait_on_buffer(bh);
2523                         if (!buffer_uptodate(bh))
2524                                 ret = -EIO;
2525                 }
2526                 if (ret)
2527                         goto failed;
2528         }
2529
2530         if (is_mapped_to_disk)
2531                 SetPageMappedToDisk(page);
2532
2533         *fsdata = head; /* to be released by nobh_write_end */
2534
2535         return 0;
2536
2537 failed:
2538         BUG_ON(!ret);
2539         /*
2540          * Error recovery is a bit difficult. We need to zero out blocks that
2541          * were newly allocated, and dirty them to ensure they get written out.
2542          * Buffers need to be attached to the page at this point, otherwise
2543          * the handling of potential IO errors during writeout would be hard
2544          * (could try doing synchronous writeout, but what if that fails too?)
2545          */
2546         attach_nobh_buffers(page, head);
2547         page_zero_new_buffers(page, from, to);
2548
2549 out_release:
2550         unlock_page(page);
2551         page_cache_release(page);
2552         *pagep = NULL;
2553
2554         return ret;
2555 }
2556 EXPORT_SYMBOL(nobh_write_begin);
2557
2558 int nobh_write_end(struct file *file, struct address_space *mapping,
2559                         loff_t pos, unsigned len, unsigned copied,
2560                         struct page *page, void *fsdata)
2561 {
2562         struct inode *inode = page->mapping->host;
2563         struct buffer_head *head = fsdata;
2564         struct buffer_head *bh;
2565         BUG_ON(fsdata != NULL && page_has_buffers(page));
2566
2567         if (unlikely(copied < len) && head)
2568                 attach_nobh_buffers(page, head);
2569         if (page_has_buffers(page))
2570                 return generic_write_end(file, mapping, pos, len,
2571                                         copied, page, fsdata);
2572
2573         SetPageUptodate(page);
2574         set_page_dirty(page);
2575         if (pos+copied > inode->i_size) {
2576                 i_size_write(inode, pos+copied);
2577                 mark_inode_dirty(inode);
2578         }
2579
2580         unlock_page(page);
2581         page_cache_release(page);
2582
2583         while (head) {
2584                 bh = head;
2585                 head = head->b_this_page;
2586                 free_buffer_head(bh);
2587         }
2588
2589         return copied;
2590 }
2591 EXPORT_SYMBOL(nobh_write_end);
2592
2593 /*
2594  * nobh_writepage() - based on block_full_write_page() except
2595  * that it tries to operate without attaching bufferheads to
2596  * the page.
2597  */
2598 int nobh_writepage(struct page *page, get_block_t *get_block,
2599                         struct writeback_control *wbc)
2600 {
2601         struct inode * const inode = page->mapping->host;
2602         loff_t i_size = i_size_read(inode);
2603         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2604         unsigned offset;
2605         int ret;
2606
2607         /* Is the page fully inside i_size? */
2608         if (page->index < end_index)
2609                 goto out;
2610
2611         /* Is the page fully outside i_size? (truncate in progress) */
2612         offset = i_size & (PAGE_CACHE_SIZE-1);
2613         if (page->index >= end_index+1 || !offset) {
2614                 /*
2615                  * The page may have dirty, unmapped buffers.  For example,
2616                  * they may have been added in ext3_writepage().  Make them
2617                  * freeable here, so the page does not leak.
2618                  */
2619 #if 0
2620                 /* Not really sure about this  - do we need this ? */
2621                 if (page->mapping->a_ops->invalidatepage)
2622                         page->mapping->a_ops->invalidatepage(page, offset);
2623 #endif
2624                 unlock_page(page);
2625                 return 0; /* don't care */
2626         }
2627
2628         /*
2629          * The page straddles i_size.  It must be zeroed out on each and every
2630          * writepage invocation because it may be mmapped.  "A file is mapped
2631          * in multiples of the page size.  For a file that is not a multiple of
2632          * the  page size, the remaining memory is zeroed when mapped, and
2633          * writes to that region are not written out to the file."
2634          */
2635         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2636 out:
2637         ret = mpage_writepage(page, get_block, wbc);
2638         if (ret == -EAGAIN)
2639                 ret = __block_write_full_page(inode, page, get_block, wbc,
2640                                               end_buffer_async_write);
2641         return ret;
2642 }
2643 EXPORT_SYMBOL(nobh_writepage);
2644
2645 int nobh_truncate_page(struct address_space *mapping,
2646                         loff_t from, get_block_t *get_block)
2647 {
2648         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2649         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2650         unsigned blocksize;
2651         sector_t iblock;
2652         unsigned length, pos;
2653         struct inode *inode = mapping->host;
2654         struct page *page;
2655         struct buffer_head map_bh;
2656         int err;
2657
2658         blocksize = 1 << inode->i_blkbits;
2659         length = offset & (blocksize - 1);
2660
2661         /* Block boundary? Nothing to do */
2662         if (!length)
2663                 return 0;
2664
2665         length = blocksize - length;
2666         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2667
2668         page = grab_cache_page(mapping, index);
2669         err = -ENOMEM;
2670         if (!page)
2671                 goto out;
2672
2673         if (page_has_buffers(page)) {
2674 has_buffers:
2675                 unlock_page(page);
2676                 page_cache_release(page);
2677                 return block_truncate_page(mapping, from, get_block);
2678         }
2679
2680         /* Find the buffer that contains "offset" */
2681         pos = blocksize;
2682         while (offset >= pos) {
2683                 iblock++;
2684                 pos += blocksize;
2685         }
2686
2687         map_bh.b_size = blocksize;
2688         map_bh.b_state = 0;
2689         err = get_block(inode, iblock, &map_bh, 0);
2690         if (err)
2691                 goto unlock;
2692         /* unmapped? It's a hole - nothing to do */
2693         if (!buffer_mapped(&map_bh))
2694                 goto unlock;
2695
2696         /* Ok, it's mapped. Make sure it's up-to-date */
2697         if (!PageUptodate(page)) {
2698                 err = mapping->a_ops->readpage(NULL, page);
2699                 if (err) {
2700                         page_cache_release(page);
2701                         goto out;
2702                 }
2703                 lock_page(page);
2704                 if (!PageUptodate(page)) {
2705                         err = -EIO;
2706                         goto unlock;
2707                 }
2708                 if (page_has_buffers(page))
2709                         goto has_buffers;
2710         }
2711         zero_user(page, offset, length);
2712         set_page_dirty(page);
2713         err = 0;
2714
2715 unlock:
2716         unlock_page(page);
2717         page_cache_release(page);
2718 out:
2719         return err;
2720 }
2721 EXPORT_SYMBOL(nobh_truncate_page);
2722
2723 int block_truncate_page(struct address_space *mapping,
2724                         loff_t from, get_block_t *get_block)
2725 {
2726         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2727         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2728         unsigned blocksize;
2729         sector_t iblock;
2730         unsigned length, pos;
2731         struct inode *inode = mapping->host;
2732         struct page *page;
2733         struct buffer_head *bh;
2734         int err;
2735
2736         blocksize = 1 << inode->i_blkbits;
2737         length = offset & (blocksize - 1);
2738
2739         /* Block boundary? Nothing to do */
2740         if (!length)
2741                 return 0;
2742
2743         length = blocksize - length;
2744         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2745         
2746         page = grab_cache_page(mapping, index);
2747         err = -ENOMEM;
2748         if (!page)
2749                 goto out;
2750
2751         if (!page_has_buffers(page))
2752                 create_empty_buffers(page, blocksize, 0);
2753
2754         /* Find the buffer that contains "offset" */
2755         bh = page_buffers(page);
2756         pos = blocksize;
2757         while (offset >= pos) {
2758                 bh = bh->b_this_page;
2759                 iblock++;
2760                 pos += blocksize;
2761         }
2762
2763         err = 0;
2764         if (!buffer_mapped(bh)) {
2765                 WARN_ON(bh->b_size != blocksize);
2766                 err = get_block(inode, iblock, bh, 0);
2767                 if (err)
2768                         goto unlock;
2769                 /* unmapped? It's a hole - nothing to do */
2770                 if (!buffer_mapped(bh))
2771                         goto unlock;
2772         }
2773
2774         /* Ok, it's mapped. Make sure it's up-to-date */
2775         if (PageUptodate(page))
2776                 set_buffer_uptodate(bh);
2777
2778         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2779                 err = -EIO;
2780                 ll_rw_block(READ, 1, &bh);
2781                 wait_on_buffer(bh);
2782                 /* Uhhuh. Read error. Complain and punt. */
2783                 if (!buffer_uptodate(bh))
2784                         goto unlock;
2785         }
2786
2787         zero_user(page, offset, length);
2788         mark_buffer_dirty(bh);
2789         err = 0;
2790
2791 unlock:
2792         unlock_page(page);
2793         page_cache_release(page);
2794 out:
2795         return err;
2796 }
2797 EXPORT_SYMBOL(block_truncate_page);
2798
2799 /*
2800  * The generic ->writepage function for buffer-backed address_spaces
2801  * this form passes in the end_io handler used to finish the IO.
2802  */
2803 int block_write_full_page_endio(struct page *page, get_block_t *get_block,
2804                         struct writeback_control *wbc, bh_end_io_t *handler)
2805 {
2806         struct inode * const inode = page->mapping->host;
2807         loff_t i_size = i_size_read(inode);
2808         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2809         unsigned offset;
2810
2811         /* Is the page fully inside i_size? */
2812         if (page->index < end_index)
2813                 return __block_write_full_page(inode, page, get_block, wbc,
2814                                                handler);
2815
2816         /* Is the page fully outside i_size? (truncate in progress) */
2817         offset = i_size & (PAGE_CACHE_SIZE-1);
2818         if (page->index >= end_index+1 || !offset) {
2819                 /*
2820                  * The page may have dirty, unmapped buffers.  For example,
2821                  * they may have been added in ext3_writepage().  Make them
2822                  * freeable here, so the page does not leak.
2823                  */
2824                 do_invalidatepage(page, 0);
2825                 unlock_page(page);
2826                 return 0; /* don't care */
2827         }
2828
2829         /*
2830          * The page straddles i_size.  It must be zeroed out on each and every
2831          * writepage invocation because it may be mmapped.  "A file is mapped
2832          * in multiples of the page size.  For a file that is not a multiple of
2833          * the  page size, the remaining memory is zeroed when mapped, and
2834          * writes to that region are not written out to the file."
2835          */
2836         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2837         return __block_write_full_page(inode, page, get_block, wbc, handler);
2838 }
2839 EXPORT_SYMBOL(block_write_full_page_endio);
2840
2841 /*
2842  * The generic ->writepage function for buffer-backed address_spaces
2843  */
2844 int block_write_full_page(struct page *page, get_block_t *get_block,
2845                         struct writeback_control *wbc)
2846 {
2847         return block_write_full_page_endio(page, get_block, wbc,
2848                                            end_buffer_async_write);
2849 }
2850 EXPORT_SYMBOL(block_write_full_page);
2851
2852 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2853                             get_block_t *get_block)
2854 {
2855         struct buffer_head tmp;
2856         struct inode *inode = mapping->host;
2857         tmp.b_state = 0;
2858         tmp.b_blocknr = 0;
2859         tmp.b_size = 1 << inode->i_blkbits;
2860         get_block(inode, block, &tmp, 0);
2861         return tmp.b_blocknr;
2862 }
2863 EXPORT_SYMBOL(generic_block_bmap);
2864
2865 static void end_bio_bh_io_sync(struct bio *bio, int err)
2866 {
2867         struct buffer_head *bh = bio->bi_private;
2868
2869         if (err == -EOPNOTSUPP) {
2870                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2871         }
2872
2873         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2874                 set_bit(BH_Quiet, &bh->b_state);
2875
2876         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2877         bio_put(bio);
2878 }
2879
2880 int submit_bh(int rw, struct buffer_head * bh)
2881 {
2882         struct bio *bio;
2883         int ret = 0;
2884
2885         BUG_ON(!buffer_locked(bh));
2886         BUG_ON(!buffer_mapped(bh));
2887         BUG_ON(!bh->b_end_io);
2888         BUG_ON(buffer_delay(bh));
2889         BUG_ON(buffer_unwritten(bh));
2890
2891         /*
2892          * Only clear out a write error when rewriting
2893          */
2894         if (test_set_buffer_req(bh) && (rw & WRITE))
2895                 clear_buffer_write_io_error(bh);
2896
2897         /*
2898          * from here on down, it's all bio -- do the initial mapping,
2899          * submit_bio -> generic_make_request may further map this bio around
2900          */
2901         bio = bio_alloc(GFP_NOIO, 1);
2902
2903         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2904         bio->bi_bdev = bh->b_bdev;
2905         bio->bi_io_vec[0].bv_page = bh->b_page;
2906         bio->bi_io_vec[0].bv_len = bh->b_size;
2907         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2908
2909         bio->bi_vcnt = 1;
2910         bio->bi_idx = 0;
2911         bio->bi_size = bh->b_size;
2912
2913         bio->bi_end_io = end_bio_bh_io_sync;
2914         bio->bi_private = bh;
2915
2916         bio_get(bio);
2917         submit_bio(rw, bio);
2918
2919         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2920                 ret = -EOPNOTSUPP;
2921
2922         bio_put(bio);
2923         return ret;
2924 }
2925 EXPORT_SYMBOL(submit_bh);
2926
2927 /**
2928  * ll_rw_block: low-level access to block devices (DEPRECATED)
2929  * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
2930  * @nr: number of &struct buffer_heads in the array
2931  * @bhs: array of pointers to &struct buffer_head
2932  *
2933  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2934  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2935  * %READA option is described in the documentation for generic_make_request()
2936  * which ll_rw_block() calls.
2937  *
2938  * This function drops any buffer that it cannot get a lock on (with the
2939  * BH_Lock state bit), any buffer that appears to be clean when doing a write
2940  * request, and any buffer that appears to be up-to-date when doing read
2941  * request.  Further it marks as clean buffers that are processed for
2942  * writing (the buffer cache won't assume that they are actually clean
2943  * until the buffer gets unlocked).
2944  *
2945  * ll_rw_block sets b_end_io to simple completion handler that marks
2946  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2947  * any waiters. 
2948  *
2949  * All of the buffers must be for the same device, and must also be a
2950  * multiple of the current approved size for the device.
2951  */
2952 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2953 {
2954         int i;
2955
2956         for (i = 0; i < nr; i++) {
2957                 struct buffer_head *bh = bhs[i];
2958
2959                 if (!trylock_buffer(bh))
2960                         continue;
2961                 if (rw == WRITE) {
2962                         if (test_clear_buffer_dirty(bh)) {
2963                                 bh->b_end_io = end_buffer_write_sync;
2964                                 get_bh(bh);
2965                                 submit_bh(WRITE, bh);
2966                                 continue;
2967                         }
2968                 } else {
2969                         if (!buffer_uptodate(bh)) {
2970                                 bh->b_end_io = end_buffer_read_sync;
2971                                 get_bh(bh);
2972                                 submit_bh(rw, bh);
2973                                 continue;
2974                         }
2975                 }
2976                 unlock_buffer(bh);
2977         }
2978 }
2979 EXPORT_SYMBOL(ll_rw_block);
2980
2981 void write_dirty_buffer(struct buffer_head *bh, int rw)
2982 {
2983         lock_buffer(bh);
2984         if (!test_clear_buffer_dirty(bh)) {
2985                 unlock_buffer(bh);
2986                 return;
2987         }
2988         bh->b_end_io = end_buffer_write_sync;
2989         get_bh(bh);
2990         submit_bh(rw, bh);
2991 }
2992 EXPORT_SYMBOL(write_dirty_buffer);
2993
2994 /*
2995  * For a data-integrity writeout, we need to wait upon any in-progress I/O
2996  * and then start new I/O and then wait upon it.  The caller must have a ref on
2997  * the buffer_head.
2998  */
2999 int __sync_dirty_buffer(struct buffer_head *bh, int rw)
3000 {
3001         int ret = 0;
3002
3003         WARN_ON(atomic_read(&bh->b_count) < 1);
3004         lock_buffer(bh);
3005         if (test_clear_buffer_dirty(bh)) {
3006                 get_bh(bh);
3007                 bh->b_end_io = end_buffer_write_sync;
3008                 ret = submit_bh(rw, bh);
3009                 wait_on_buffer(bh);
3010                 if (!ret && !buffer_uptodate(bh))
3011                         ret = -EIO;
3012         } else {
3013                 unlock_buffer(bh);
3014         }
3015         return ret;
3016 }
3017 EXPORT_SYMBOL(__sync_dirty_buffer);
3018
3019 int sync_dirty_buffer(struct buffer_head *bh)
3020 {
3021         return __sync_dirty_buffer(bh, WRITE_SYNC);
3022 }
3023 EXPORT_SYMBOL(sync_dirty_buffer);
3024
3025 /*
3026  * try_to_free_buffers() checks if all the buffers on this particular page
3027  * are unused, and releases them if so.
3028  *
3029  * Exclusion against try_to_free_buffers may be obtained by either
3030  * locking the page or by holding its mapping's private_lock.
3031  *
3032  * If the page is dirty but all the buffers are clean then we need to
3033  * be sure to mark the page clean as well.  This is because the page
3034  * may be against a block device, and a later reattachment of buffers
3035  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3036  * filesystem data on the same device.
3037  *
3038  * The same applies to regular filesystem pages: if all the buffers are
3039  * clean then we set the page clean and proceed.  To do that, we require
3040  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3041  * private_lock.
3042  *
3043  * try_to_free_buffers() is non-blocking.
3044  */
3045 static inline int buffer_busy(struct buffer_head *bh)
3046 {
3047         return atomic_read(&bh->b_count) |
3048                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3049 }
3050
3051 static int
3052 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3053 {
3054         struct buffer_head *head = page_buffers(page);
3055         struct buffer_head *bh;
3056
3057         bh = head;
3058         do {
3059                 if (buffer_write_io_error(bh) && page->mapping)
3060                         set_bit(AS_EIO, &page->mapping->flags);
3061                 if (buffer_busy(bh))
3062                         goto failed;
3063                 bh = bh->b_this_page;
3064         } while (bh != head);
3065
3066         do {
3067                 struct buffer_head *next = bh->b_this_page;
3068
3069                 if (bh->b_assoc_map)
3070                         __remove_assoc_queue(bh);
3071                 bh = next;
3072         } while (bh != head);
3073         *buffers_to_free = head;
3074         __clear_page_buffers(page);
3075         return 1;
3076 failed:
3077         return 0;
3078 }
3079
3080 int try_to_free_buffers(struct page *page)
3081 {
3082         struct address_space * const mapping = page->mapping;
3083         struct buffer_head *buffers_to_free = NULL;
3084         int ret = 0;
3085
3086         BUG_ON(!PageLocked(page));
3087         if (PageWriteback(page))
3088                 return 0;
3089
3090         if (mapping == NULL) {          /* can this still happen? */
3091                 ret = drop_buffers(page, &buffers_to_free);
3092                 goto out;
3093         }
3094
3095         spin_lock(&mapping->private_lock);
3096         ret = drop_buffers(page, &buffers_to_free);
3097
3098         /*
3099          * If the filesystem writes its buffers by hand (eg ext3)
3100          * then we can have clean buffers against a dirty page.  We
3101          * clean the page here; otherwise the VM will never notice
3102          * that the filesystem did any IO at all.
3103          *
3104          * Also, during truncate, discard_buffer will have marked all
3105          * the page's buffers clean.  We discover that here and clean
3106          * the page also.
3107          *
3108          * private_lock must be held over this entire operation in order
3109          * to synchronise against __set_page_dirty_buffers and prevent the
3110          * dirty bit from being lost.
3111          */
3112         if (ret)
3113                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3114         spin_unlock(&mapping->private_lock);
3115 out:
3116         if (buffers_to_free) {
3117                 struct buffer_head *bh = buffers_to_free;
3118
3119                 do {
3120                         struct buffer_head *next = bh->b_this_page;
3121                         free_buffer_head(bh);
3122                         bh = next;
3123                 } while (bh != buffers_to_free);
3124         }
3125         return ret;
3126 }
3127 EXPORT_SYMBOL(try_to_free_buffers);
3128
3129 /*
3130  * There are no bdflush tunables left.  But distributions are
3131  * still running obsolete flush daemons, so we terminate them here.
3132  *
3133  * Use of bdflush() is deprecated and will be removed in a future kernel.
3134  * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3135  */
3136 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3137 {
3138         static int msg_count;
3139
3140         if (!capable(CAP_SYS_ADMIN))
3141                 return -EPERM;
3142
3143         if (msg_count < 5) {
3144                 msg_count++;
3145                 printk(KERN_INFO
3146                         "warning: process `%s' used the obsolete bdflush"
3147                         " system call\n", current->comm);
3148                 printk(KERN_INFO "Fix your initscripts?\n");
3149         }
3150
3151         if (func == 1)
3152                 do_exit(0);
3153         return 0;
3154 }
3155
3156 /*
3157  * Buffer-head allocation
3158  */
3159 static struct kmem_cache *bh_cachep __read_mostly;
3160
3161 /*
3162  * Once the number of bh's in the machine exceeds this level, we start
3163  * stripping them in writeback.
3164  */
3165 static int max_buffer_heads;
3166
3167 int buffer_heads_over_limit;
3168
3169 struct bh_accounting {
3170         int nr;                 /* Number of live bh's */
3171         int ratelimit;          /* Limit cacheline bouncing */
3172 };
3173
3174 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3175
3176 static void recalc_bh_state(void)
3177 {
3178         int i;
3179         int tot = 0;
3180
3181         if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3182                 return;
3183         __this_cpu_write(bh_accounting.ratelimit, 0);
3184         for_each_online_cpu(i)
3185                 tot += per_cpu(bh_accounting, i).nr;
3186         buffer_heads_over_limit = (tot > max_buffer_heads);
3187 }
3188
3189 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3190 {
3191         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3192         if (ret) {
3193                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3194                 preempt_disable();
3195                 __this_cpu_inc(bh_accounting.nr);
3196                 recalc_bh_state();
3197                 preempt_enable();
3198         }
3199         return ret;
3200 }
3201 EXPORT_SYMBOL(alloc_buffer_head);
3202
3203 void free_buffer_head(struct buffer_head *bh)
3204 {
3205         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3206         kmem_cache_free(bh_cachep, bh);
3207         preempt_disable();
3208         __this_cpu_dec(bh_accounting.nr);
3209         recalc_bh_state();
3210         preempt_enable();
3211 }
3212 EXPORT_SYMBOL(free_buffer_head);
3213
3214 static void buffer_exit_cpu(int cpu)
3215 {
3216         int i;
3217         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3218
3219         for (i = 0; i < BH_LRU_SIZE; i++) {
3220                 brelse(b->bhs[i]);
3221                 b->bhs[i] = NULL;
3222         }
3223         this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3224         per_cpu(bh_accounting, cpu).nr = 0;
3225 }
3226
3227 static int buffer_cpu_notify(struct notifier_block *self,
3228                               unsigned long action, void *hcpu)
3229 {
3230         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3231                 buffer_exit_cpu((unsigned long)hcpu);
3232         return NOTIFY_OK;
3233 }
3234
3235 /**
3236  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3237  * @bh: struct buffer_head
3238  *
3239  * Return true if the buffer is up-to-date and false,
3240  * with the buffer locked, if not.
3241  */
3242 int bh_uptodate_or_lock(struct buffer_head *bh)
3243 {
3244         if (!buffer_uptodate(bh)) {
3245                 lock_buffer(bh);
3246                 if (!buffer_uptodate(bh))
3247                         return 0;
3248                 unlock_buffer(bh);
3249         }
3250         return 1;
3251 }
3252 EXPORT_SYMBOL(bh_uptodate_or_lock);
3253
3254 /**
3255  * bh_submit_read - Submit a locked buffer for reading
3256  * @bh: struct buffer_head
3257  *
3258  * Returns zero on success and -EIO on error.
3259  */
3260 int bh_submit_read(struct buffer_head *bh)
3261 {
3262         BUG_ON(!buffer_locked(bh));
3263
3264         if (buffer_uptodate(bh)) {
3265                 unlock_buffer(bh);
3266                 return 0;
3267         }
3268
3269         get_bh(bh);
3270         bh->b_end_io = end_buffer_read_sync;
3271         submit_bh(READ, bh);
3272         wait_on_buffer(bh);
3273         if (buffer_uptodate(bh))
3274                 return 0;
3275         return -EIO;
3276 }
3277 EXPORT_SYMBOL(bh_submit_read);
3278
3279 void __init buffer_init(void)
3280 {
3281         int nrpages;
3282
3283         bh_cachep = kmem_cache_create("buffer_head",
3284                         sizeof(struct buffer_head), 0,
3285                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3286                                 SLAB_MEM_SPREAD),
3287                                 NULL);
3288
3289         /*
3290          * Limit the bh occupancy to 10% of ZONE_NORMAL
3291          */
3292         nrpages = (nr_free_buffer_pages() * 10) / 100;
3293         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3294         hotcpu_notifier(buffer_cpu_notify, 0);
3295 }