Merge commit 'v3.4.4' into android-tegra-nv-3.4
[linux-2.6.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         /* Size must be multiple of hard sectorsize */
1040         if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1041                         (size < 512 || size > PAGE_SIZE))) {
1042                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1043                                         size);
1044                 printk(KERN_ERR "logical block size: %d\n",
1045                                         bdev_logical_block_size(bdev));
1046
1047                 dump_stack();
1048                 return NULL;
1049         }
1050
1051         for (;;) {
1052                 struct buffer_head * bh;
1053                 int ret;
1054
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                         return NULL;
1062                 if (ret == 0)
1063                         free_more_memory();
1064         }
1065 }
1066
1067 /*
1068  * The relationship between dirty buffers and dirty pages:
1069  *
1070  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1071  * the page is tagged dirty in its radix tree.
1072  *
1073  * At all times, the dirtiness of the buffers represents the dirtiness of
1074  * subsections of the page.  If the page has buffers, the page dirty bit is
1075  * merely a hint about the true dirty state.
1076  *
1077  * When a page is set dirty in its entirety, all its buffers are marked dirty
1078  * (if the page has buffers).
1079  *
1080  * When a buffer is marked dirty, its page is dirtied, but the page's other
1081  * buffers are not.
1082  *
1083  * Also.  When blockdev buffers are explicitly read with bread(), they
1084  * individually become uptodate.  But their backing page remains not
1085  * uptodate - even if all of its buffers are uptodate.  A subsequent
1086  * block_read_full_page() against that page will discover all the uptodate
1087  * buffers, will set the page uptodate and will perform no I/O.
1088  */
1089
1090 /**
1091  * mark_buffer_dirty - mark a buffer_head as needing writeout
1092  * @bh: the buffer_head to mark dirty
1093  *
1094  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1095  * backing page dirty, then tag the page as dirty in its address_space's radix
1096  * tree and then attach the address_space's inode to its superblock's dirty
1097  * inode list.
1098  *
1099  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1100  * mapping->tree_lock and mapping->host->i_lock.
1101  */
1102 void mark_buffer_dirty(struct buffer_head *bh)
1103 {
1104         WARN_ON_ONCE(!buffer_uptodate(bh));
1105
1106         /*
1107          * Very *carefully* optimize the it-is-already-dirty case.
1108          *
1109          * Don't let the final "is it dirty" escape to before we
1110          * perhaps modified the buffer.
1111          */
1112         if (buffer_dirty(bh)) {
1113                 smp_mb();
1114                 if (buffer_dirty(bh))
1115                         return;
1116         }
1117
1118         if (!test_set_buffer_dirty(bh)) {
1119                 struct page *page = bh->b_page;
1120                 if (!TestSetPageDirty(page)) {
1121                         struct address_space *mapping = page_mapping(page);
1122                         if (mapping)
1123                                 __set_page_dirty(page, mapping, 0);
1124                 }
1125         }
1126 }
1127 EXPORT_SYMBOL(mark_buffer_dirty);
1128
1129 /*
1130  * Decrement a buffer_head's reference count.  If all buffers against a page
1131  * have zero reference count, are clean and unlocked, and if the page is clean
1132  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1133  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1134  * a page but it ends up not being freed, and buffers may later be reattached).
1135  */
1136 void __brelse(struct buffer_head * buf)
1137 {
1138         if (atomic_read(&buf->b_count)) {
1139                 put_bh(buf);
1140                 return;
1141         }
1142         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1143 }
1144 EXPORT_SYMBOL(__brelse);
1145
1146 /*
1147  * bforget() is like brelse(), except it discards any
1148  * potentially dirty data.
1149  */
1150 void __bforget(struct buffer_head *bh)
1151 {
1152         clear_buffer_dirty(bh);
1153         if (bh->b_assoc_map) {
1154                 struct address_space *buffer_mapping = bh->b_page->mapping;
1155
1156                 spin_lock(&buffer_mapping->private_lock);
1157                 list_del_init(&bh->b_assoc_buffers);
1158                 bh->b_assoc_map = NULL;
1159                 spin_unlock(&buffer_mapping->private_lock);
1160         }
1161         __brelse(bh);
1162 }
1163 EXPORT_SYMBOL(__bforget);
1164
1165 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1166 {
1167         lock_buffer(bh);
1168         if (buffer_uptodate(bh)) {
1169                 unlock_buffer(bh);
1170                 return bh;
1171         } else {
1172                 get_bh(bh);
1173                 bh->b_end_io = end_buffer_read_sync;
1174                 submit_bh(READ, bh);
1175                 wait_on_buffer(bh);
1176                 if (buffer_uptodate(bh))
1177                         return bh;
1178         }
1179         brelse(bh);
1180         return NULL;
1181 }
1182
1183 /*
1184  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1185  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1186  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1187  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1188  * CPU's LRUs at the same time.
1189  *
1190  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1191  * sb_find_get_block().
1192  *
1193  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1194  * a local interrupt disable for that.
1195  */
1196
1197 #define BH_LRU_SIZE     8
1198
1199 struct bh_lru {
1200         struct buffer_head *bhs[BH_LRU_SIZE];
1201 };
1202
1203 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1204
1205 #ifdef CONFIG_SMP
1206 #define bh_lru_lock()   local_irq_disable()
1207 #define bh_lru_unlock() local_irq_enable()
1208 #else
1209 #define bh_lru_lock()   preempt_disable()
1210 #define bh_lru_unlock() preempt_enable()
1211 #endif
1212
1213 static inline void check_irqs_on(void)
1214 {
1215 #ifdef irqs_disabled
1216         BUG_ON(irqs_disabled());
1217 #endif
1218 }
1219
1220 /*
1221  * The LRU management algorithm is dopey-but-simple.  Sorry.
1222  */
1223 static void bh_lru_install(struct buffer_head *bh)
1224 {
1225         struct buffer_head *evictee = NULL;
1226
1227         check_irqs_on();
1228         bh_lru_lock();
1229         if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1230                 struct buffer_head *bhs[BH_LRU_SIZE];
1231                 int in;
1232                 int out = 0;
1233
1234                 get_bh(bh);
1235                 bhs[out++] = bh;
1236                 for (in = 0; in < BH_LRU_SIZE; in++) {
1237                         struct buffer_head *bh2 =
1238                                 __this_cpu_read(bh_lrus.bhs[in]);
1239
1240                         if (bh2 == bh) {
1241                                 __brelse(bh2);
1242                         } else {
1243                                 if (out >= BH_LRU_SIZE) {
1244                                         BUG_ON(evictee != NULL);
1245                                         evictee = bh2;
1246                                 } else {
1247                                         bhs[out++] = bh2;
1248                                 }
1249                         }
1250                 }
1251                 while (out < BH_LRU_SIZE)
1252                         bhs[out++] = NULL;
1253                 memcpy(__this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1254         }
1255         bh_lru_unlock();
1256
1257         if (evictee)
1258                 __brelse(evictee);
1259 }
1260
1261 /*
1262  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1263  */
1264 static struct buffer_head *
1265 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1266 {
1267         struct buffer_head *ret = NULL;
1268         unsigned int i;
1269
1270         check_irqs_on();
1271         bh_lru_lock();
1272         for (i = 0; i < BH_LRU_SIZE; i++) {
1273                 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1274
1275                 if (bh && bh->b_bdev == bdev &&
1276                                 bh->b_blocknr == block && bh->b_size == size) {
1277                         if (i) {
1278                                 while (i) {
1279                                         __this_cpu_write(bh_lrus.bhs[i],
1280                                                 __this_cpu_read(bh_lrus.bhs[i - 1]));
1281                                         i--;
1282                                 }
1283                                 __this_cpu_write(bh_lrus.bhs[0], bh);
1284                         }
1285                         get_bh(bh);
1286                         ret = bh;
1287                         break;
1288                 }
1289         }
1290         bh_lru_unlock();
1291         return ret;
1292 }
1293
1294 /*
1295  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1296  * it in the LRU and mark it as accessed.  If it is not present then return
1297  * NULL
1298  */
1299 struct buffer_head *
1300 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1301 {
1302         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1303
1304         if (bh == NULL) {
1305                 bh = __find_get_block_slow(bdev, block);
1306                 if (bh)
1307                         bh_lru_install(bh);
1308         }
1309         if (bh)
1310                 touch_buffer(bh);
1311         return bh;
1312 }
1313 EXPORT_SYMBOL(__find_get_block);
1314
1315 /*
1316  * __getblk will locate (and, if necessary, create) the buffer_head
1317  * which corresponds to the passed block_device, block and size. The
1318  * returned buffer has its reference count incremented.
1319  *
1320  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1321  * illegal block number, __getblk() will happily return a buffer_head
1322  * which represents the non-existent block.  Very weird.
1323  *
1324  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1325  * attempt is failing.  FIXME, perhaps?
1326  */
1327 struct buffer_head *
1328 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1329 {
1330         struct buffer_head *bh = __find_get_block(bdev, block, size);
1331
1332         might_sleep();
1333         if (bh == NULL)
1334                 bh = __getblk_slow(bdev, block, size);
1335         return bh;
1336 }
1337 EXPORT_SYMBOL(__getblk);
1338
1339 /*
1340  * Do async read-ahead on a buffer..
1341  */
1342 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1343 {
1344         struct buffer_head *bh = __getblk(bdev, block, size);
1345         if (likely(bh)) {
1346                 ll_rw_block(READA, 1, &bh);
1347                 brelse(bh);
1348         }
1349 }
1350 EXPORT_SYMBOL(__breadahead);
1351
1352 /**
1353  *  __bread() - reads a specified block and returns the bh
1354  *  @bdev: the block_device to read from
1355  *  @block: number of block
1356  *  @size: size (in bytes) to read
1357  * 
1358  *  Reads a specified block, and returns buffer head that contains it.
1359  *  It returns NULL if the block was unreadable.
1360  */
1361 struct buffer_head *
1362 __bread(struct block_device *bdev, sector_t block, unsigned size)
1363 {
1364         struct buffer_head *bh = __getblk(bdev, block, size);
1365
1366         if (likely(bh) && !buffer_uptodate(bh))
1367                 bh = __bread_slow(bh);
1368         return bh;
1369 }
1370 EXPORT_SYMBOL(__bread);
1371
1372 /*
1373  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1374  * This doesn't race because it runs in each cpu either in irq
1375  * or with preempt disabled.
1376  */
1377 static void invalidate_bh_lru(void *arg)
1378 {
1379         struct bh_lru *b = &get_cpu_var(bh_lrus);
1380         int i;
1381
1382         for (i = 0; i < BH_LRU_SIZE; i++) {
1383                 brelse(b->bhs[i]);
1384                 b->bhs[i] = NULL;
1385         }
1386         put_cpu_var(bh_lrus);
1387 }
1388
1389 static bool has_bh_in_lru(int cpu, void *dummy)
1390 {
1391         struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1392         int i;
1393         
1394         for (i = 0; i < BH_LRU_SIZE; i++) {
1395                 if (b->bhs[i])
1396                         return 1;
1397         }
1398
1399         return 0;
1400 }
1401
1402 void invalidate_bh_lrus(void)
1403 {
1404         on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1405 }
1406 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1407
1408 void set_bh_page(struct buffer_head *bh,
1409                 struct page *page, unsigned long offset)
1410 {
1411         bh->b_page = page;
1412         BUG_ON(offset >= PAGE_SIZE);
1413         if (PageHighMem(page))
1414                 /*
1415                  * This catches illegal uses and preserves the offset:
1416                  */
1417                 bh->b_data = (char *)(0 + offset);
1418         else
1419                 bh->b_data = page_address(page) + offset;
1420 }
1421 EXPORT_SYMBOL(set_bh_page);
1422
1423 /*
1424  * Called when truncating a buffer on a page completely.
1425  */
1426 static void discard_buffer(struct buffer_head * bh)
1427 {
1428         lock_buffer(bh);
1429         clear_buffer_dirty(bh);
1430         bh->b_bdev = NULL;
1431         clear_buffer_mapped(bh);
1432         clear_buffer_req(bh);
1433         clear_buffer_new(bh);
1434         clear_buffer_delay(bh);
1435         clear_buffer_unwritten(bh);
1436         unlock_buffer(bh);
1437 }
1438
1439 /**
1440  * block_invalidatepage - invalidate part or all of a buffer-backed page
1441  *
1442  * @page: the page which is affected
1443  * @offset: the index of the truncation point
1444  *
1445  * block_invalidatepage() is called when all or part of the page has become
1446  * invalidated by a truncate operation.
1447  *
1448  * block_invalidatepage() does not have to release all buffers, but it must
1449  * ensure that no dirty buffer is left outside @offset and that no I/O
1450  * is underway against any of the blocks which are outside the truncation
1451  * point.  Because the caller is about to free (and possibly reuse) those
1452  * blocks on-disk.
1453  */
1454 void block_invalidatepage(struct page *page, unsigned long offset)
1455 {
1456         struct buffer_head *head, *bh, *next;
1457         unsigned int curr_off = 0;
1458
1459         BUG_ON(!PageLocked(page));
1460         if (!page_has_buffers(page))
1461                 goto out;
1462
1463         head = page_buffers(page);
1464         bh = head;
1465         do {
1466                 unsigned int next_off = curr_off + bh->b_size;
1467                 next = bh->b_this_page;
1468
1469                 /*
1470                  * is this block fully invalidated?
1471                  */
1472                 if (offset <= curr_off)
1473                         discard_buffer(bh);
1474                 curr_off = next_off;
1475                 bh = next;
1476         } while (bh != head);
1477
1478         /*
1479          * We release buffers only if the entire page is being invalidated.
1480          * The get_block cached value has been unconditionally invalidated,
1481          * so real IO is not possible anymore.
1482          */
1483         if (offset == 0)
1484                 try_to_release_page(page, 0);
1485 out:
1486         return;
1487 }
1488 EXPORT_SYMBOL(block_invalidatepage);
1489
1490 /*
1491  * We attach and possibly dirty the buffers atomically wrt
1492  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1493  * is already excluded via the page lock.
1494  */
1495 void create_empty_buffers(struct page *page,
1496                         unsigned long blocksize, unsigned long b_state)
1497 {
1498         struct buffer_head *bh, *head, *tail;
1499
1500         head = alloc_page_buffers(page, blocksize, 1);
1501         bh = head;
1502         do {
1503                 bh->b_state |= b_state;
1504                 tail = bh;
1505                 bh = bh->b_this_page;
1506         } while (bh);
1507         tail->b_this_page = head;
1508
1509         spin_lock(&page->mapping->private_lock);
1510         if (PageUptodate(page) || PageDirty(page)) {
1511                 bh = head;
1512                 do {
1513                         if (PageDirty(page))
1514                                 set_buffer_dirty(bh);
1515                         if (PageUptodate(page))
1516                                 set_buffer_uptodate(bh);
1517                         bh = bh->b_this_page;
1518                 } while (bh != head);
1519         }
1520         attach_page_buffers(page, head);
1521         spin_unlock(&page->mapping->private_lock);
1522 }
1523 EXPORT_SYMBOL(create_empty_buffers);
1524
1525 /*
1526  * We are taking a block for data and we don't want any output from any
1527  * buffer-cache aliases starting from return from that function and
1528  * until the moment when something will explicitly mark the buffer
1529  * dirty (hopefully that will not happen until we will free that block ;-)
1530  * We don't even need to mark it not-uptodate - nobody can expect
1531  * anything from a newly allocated buffer anyway. We used to used
1532  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1533  * don't want to mark the alias unmapped, for example - it would confuse
1534  * anyone who might pick it with bread() afterwards...
1535  *
1536  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1537  * be writeout I/O going on against recently-freed buffers.  We don't
1538  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1539  * only if we really need to.  That happens here.
1540  */
1541 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1542 {
1543         struct buffer_head *old_bh;
1544
1545         might_sleep();
1546
1547         old_bh = __find_get_block_slow(bdev, block);
1548         if (old_bh) {
1549                 clear_buffer_dirty(old_bh);
1550                 wait_on_buffer(old_bh);
1551                 clear_buffer_req(old_bh);
1552                 __brelse(old_bh);
1553         }
1554 }
1555 EXPORT_SYMBOL(unmap_underlying_metadata);
1556
1557 /*
1558  * NOTE! All mapped/uptodate combinations are valid:
1559  *
1560  *      Mapped  Uptodate        Meaning
1561  *
1562  *      No      No              "unknown" - must do get_block()
1563  *      No      Yes             "hole" - zero-filled
1564  *      Yes     No              "allocated" - allocated on disk, not read in
1565  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1566  *
1567  * "Dirty" is valid only with the last case (mapped+uptodate).
1568  */
1569
1570 /*
1571  * While block_write_full_page is writing back the dirty buffers under
1572  * the page lock, whoever dirtied the buffers may decide to clean them
1573  * again at any time.  We handle that by only looking at the buffer
1574  * state inside lock_buffer().
1575  *
1576  * If block_write_full_page() is called for regular writeback
1577  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1578  * locked buffer.   This only can happen if someone has written the buffer
1579  * directly, with submit_bh().  At the address_space level PageWriteback
1580  * prevents this contention from occurring.
1581  *
1582  * If block_write_full_page() is called with wbc->sync_mode ==
1583  * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1584  * causes the writes to be flagged as synchronous writes.
1585  */
1586 static int __block_write_full_page(struct inode *inode, struct page *page,
1587                         get_block_t *get_block, struct writeback_control *wbc,
1588                         bh_end_io_t *handler)
1589 {
1590         int err;
1591         sector_t block;
1592         sector_t last_block;
1593         struct buffer_head *bh, *head;
1594         const unsigned blocksize = 1 << inode->i_blkbits;
1595         int nr_underway = 0;
1596         int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1597                         WRITE_SYNC : WRITE);
1598
1599         BUG_ON(!PageLocked(page));
1600
1601         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1602
1603         if (!page_has_buffers(page)) {
1604                 create_empty_buffers(page, blocksize,
1605                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1606         }
1607
1608         /*
1609          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1610          * here, and the (potentially unmapped) buffers may become dirty at
1611          * any time.  If a buffer becomes dirty here after we've inspected it
1612          * then we just miss that fact, and the page stays dirty.
1613          *
1614          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1615          * handle that here by just cleaning them.
1616          */
1617
1618         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1619         head = page_buffers(page);
1620         bh = head;
1621
1622         /*
1623          * Get all the dirty buffers mapped to disk addresses and
1624          * handle any aliases from the underlying blockdev's mapping.
1625          */
1626         do {
1627                 if (block > last_block) {
1628                         /*
1629                          * mapped buffers outside i_size will occur, because
1630                          * this page can be outside i_size when there is a
1631                          * truncate in progress.
1632                          */
1633                         /*
1634                          * The buffer was zeroed by block_write_full_page()
1635                          */
1636                         clear_buffer_dirty(bh);
1637                         set_buffer_uptodate(bh);
1638                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1639                            buffer_dirty(bh)) {
1640                         WARN_ON(bh->b_size != blocksize);
1641                         err = get_block(inode, block, bh, 1);
1642                         if (err)
1643                                 goto recover;
1644                         clear_buffer_delay(bh);
1645                         if (buffer_new(bh)) {
1646                                 /* blockdev mappings never come here */
1647                                 clear_buffer_new(bh);
1648                                 unmap_underlying_metadata(bh->b_bdev,
1649                                                         bh->b_blocknr);
1650                         }
1651                 }
1652                 bh = bh->b_this_page;
1653                 block++;
1654         } while (bh != head);
1655
1656         do {
1657                 if (!buffer_mapped(bh))
1658                         continue;
1659                 /*
1660                  * If it's a fully non-blocking write attempt and we cannot
1661                  * lock the buffer then redirty the page.  Note that this can
1662                  * potentially cause a busy-wait loop from writeback threads
1663                  * and kswapd activity, but those code paths have their own
1664                  * higher-level throttling.
1665                  */
1666                 if (wbc->sync_mode != WB_SYNC_NONE) {
1667                         lock_buffer(bh);
1668                 } else if (!trylock_buffer(bh)) {
1669                         redirty_page_for_writepage(wbc, page);
1670                         continue;
1671                 }
1672                 if (test_clear_buffer_dirty(bh)) {
1673                         mark_buffer_async_write_endio(bh, handler);
1674                 } else {
1675                         unlock_buffer(bh);
1676                 }
1677         } while ((bh = bh->b_this_page) != head);
1678
1679         /*
1680          * The page and its buffers are protected by PageWriteback(), so we can
1681          * drop the bh refcounts early.
1682          */
1683         BUG_ON(PageWriteback(page));
1684         set_page_writeback(page);
1685
1686         do {
1687                 struct buffer_head *next = bh->b_this_page;
1688                 if (buffer_async_write(bh)) {
1689                         submit_bh(write_op, bh);
1690                         nr_underway++;
1691                 }
1692                 bh = next;
1693         } while (bh != head);
1694         unlock_page(page);
1695
1696         err = 0;
1697 done:
1698         if (nr_underway == 0) {
1699                 /*
1700                  * The page was marked dirty, but the buffers were
1701                  * clean.  Someone wrote them back by hand with
1702                  * ll_rw_block/submit_bh.  A rare case.
1703                  */
1704                 end_page_writeback(page);
1705
1706                 /*
1707                  * The page and buffer_heads can be released at any time from
1708                  * here on.
1709                  */
1710         }
1711         return err;
1712
1713 recover:
1714         /*
1715          * ENOSPC, or some other error.  We may already have added some
1716          * blocks to the file, so we need to write these out to avoid
1717          * exposing stale data.
1718          * The page is currently locked and not marked for writeback
1719          */
1720         bh = head;
1721         /* Recovery: lock and submit the mapped buffers */
1722         do {
1723                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1724                     !buffer_delay(bh)) {
1725                         lock_buffer(bh);
1726                         mark_buffer_async_write_endio(bh, handler);
1727                 } else {
1728                         /*
1729                          * The buffer may have been set dirty during
1730                          * attachment to a dirty page.
1731                          */
1732                         clear_buffer_dirty(bh);
1733                 }
1734         } while ((bh = bh->b_this_page) != head);
1735         SetPageError(page);
1736         BUG_ON(PageWriteback(page));
1737         mapping_set_error(page->mapping, err);
1738         set_page_writeback(page);
1739         do {
1740                 struct buffer_head *next = bh->b_this_page;
1741                 if (buffer_async_write(bh)) {
1742                         clear_buffer_dirty(bh);
1743                         submit_bh(write_op, bh);
1744                         nr_underway++;
1745                 }
1746                 bh = next;
1747         } while (bh != head);
1748         unlock_page(page);
1749         goto done;
1750 }
1751
1752 /*
1753  * If a page has any new buffers, zero them out here, and mark them uptodate
1754  * and dirty so they'll be written out (in order to prevent uninitialised
1755  * block data from leaking). And clear the new bit.
1756  */
1757 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1758 {
1759         unsigned int block_start, block_end;
1760         struct buffer_head *head, *bh;
1761
1762         BUG_ON(!PageLocked(page));
1763         if (!page_has_buffers(page))
1764                 return;
1765
1766         bh = head = page_buffers(page);
1767         block_start = 0;
1768         do {
1769                 block_end = block_start + bh->b_size;
1770
1771                 if (buffer_new(bh)) {
1772                         if (block_end > from && block_start < to) {
1773                                 if (!PageUptodate(page)) {
1774                                         unsigned start, size;
1775
1776                                         start = max(from, block_start);
1777                                         size = min(to, block_end) - start;
1778
1779                                         zero_user(page, start, size);
1780                                         set_buffer_uptodate(bh);
1781                                 }
1782
1783                                 clear_buffer_new(bh);
1784                                 mark_buffer_dirty(bh);
1785                         }
1786                 }
1787
1788                 block_start = block_end;
1789                 bh = bh->b_this_page;
1790         } while (bh != head);
1791 }
1792 EXPORT_SYMBOL(page_zero_new_buffers);
1793
1794 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1795                 get_block_t *get_block)
1796 {
1797         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1798         unsigned to = from + len;
1799         struct inode *inode = page->mapping->host;
1800         unsigned block_start, block_end;
1801         sector_t block;
1802         int err = 0;
1803         unsigned blocksize, bbits;
1804         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1805
1806         BUG_ON(!PageLocked(page));
1807         BUG_ON(from > PAGE_CACHE_SIZE);
1808         BUG_ON(to > PAGE_CACHE_SIZE);
1809         BUG_ON(from > to);
1810
1811         blocksize = 1 << inode->i_blkbits;
1812         if (!page_has_buffers(page))
1813                 create_empty_buffers(page, blocksize, 0);
1814         head = page_buffers(page);
1815
1816         bbits = inode->i_blkbits;
1817         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1818
1819         for(bh = head, block_start = 0; bh != head || !block_start;
1820             block++, block_start=block_end, bh = bh->b_this_page) {
1821                 block_end = block_start + blocksize;
1822                 if (block_end <= from || block_start >= to) {
1823                         if (PageUptodate(page)) {
1824                                 if (!buffer_uptodate(bh))
1825                                         set_buffer_uptodate(bh);
1826                         }
1827                         continue;
1828                 }
1829                 if (buffer_new(bh))
1830                         clear_buffer_new(bh);
1831                 if (!buffer_mapped(bh)) {
1832                         WARN_ON(bh->b_size != blocksize);
1833                         err = get_block(inode, block, bh, 1);
1834                         if (err)
1835                                 break;
1836                         if (buffer_new(bh)) {
1837                                 unmap_underlying_metadata(bh->b_bdev,
1838                                                         bh->b_blocknr);
1839                                 if (PageUptodate(page)) {
1840                                         clear_buffer_new(bh);
1841                                         set_buffer_uptodate(bh);
1842                                         mark_buffer_dirty(bh);
1843                                         continue;
1844                                 }
1845                                 if (block_end > to || block_start < from)
1846                                         zero_user_segments(page,
1847                                                 to, block_end,
1848                                                 block_start, from);
1849                                 continue;
1850                         }
1851                 }
1852                 if (PageUptodate(page)) {
1853                         if (!buffer_uptodate(bh))
1854                                 set_buffer_uptodate(bh);
1855                         continue; 
1856                 }
1857                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1858                     !buffer_unwritten(bh) &&
1859                      (block_start < from || block_end > to)) {
1860                         ll_rw_block(READ, 1, &bh);
1861                         *wait_bh++=bh;
1862                 }
1863         }
1864         /*
1865          * If we issued read requests - let them complete.
1866          */
1867         while(wait_bh > wait) {
1868                 wait_on_buffer(*--wait_bh);
1869                 if (!buffer_uptodate(*wait_bh))
1870                         err = -EIO;
1871         }
1872         if (unlikely(err))
1873                 page_zero_new_buffers(page, from, to);
1874         return err;
1875 }
1876 EXPORT_SYMBOL(__block_write_begin);
1877
1878 static int __block_commit_write(struct inode *inode, struct page *page,
1879                 unsigned from, unsigned to)
1880 {
1881         unsigned block_start, block_end;
1882         int partial = 0;
1883         unsigned blocksize;
1884         struct buffer_head *bh, *head;
1885
1886         blocksize = 1 << inode->i_blkbits;
1887
1888         for(bh = head = page_buffers(page), block_start = 0;
1889             bh != head || !block_start;
1890             block_start=block_end, bh = bh->b_this_page) {
1891                 block_end = block_start + blocksize;
1892                 if (block_end <= from || block_start >= to) {
1893                         if (!buffer_uptodate(bh))
1894                                 partial = 1;
1895                 } else {
1896                         set_buffer_uptodate(bh);
1897                         mark_buffer_dirty(bh);
1898                 }
1899                 clear_buffer_new(bh);
1900         }
1901
1902         /*
1903          * If this is a partial write which happened to make all buffers
1904          * uptodate then we can optimize away a bogus readpage() for
1905          * the next read(). Here we 'discover' whether the page went
1906          * uptodate as a result of this (potentially partial) write.
1907          */
1908         if (!partial)
1909                 SetPageUptodate(page);
1910         return 0;
1911 }
1912
1913 /*
1914  * block_write_begin takes care of the basic task of block allocation and
1915  * bringing partial write blocks uptodate first.
1916  *
1917  * The filesystem needs to handle block truncation upon failure.
1918  */
1919 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
1920                 unsigned flags, struct page **pagep, get_block_t *get_block)
1921 {
1922         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1923         struct page *page;
1924         int status;
1925
1926         page = grab_cache_page_write_begin(mapping, index, flags);
1927         if (!page)
1928                 return -ENOMEM;
1929
1930         status = __block_write_begin(page, pos, len, get_block);
1931         if (unlikely(status)) {
1932                 unlock_page(page);
1933                 page_cache_release(page);
1934                 page = NULL;
1935         }
1936
1937         *pagep = page;
1938         return status;
1939 }
1940 EXPORT_SYMBOL(block_write_begin);
1941
1942 int block_write_end(struct file *file, struct address_space *mapping,
1943                         loff_t pos, unsigned len, unsigned copied,
1944                         struct page *page, void *fsdata)
1945 {
1946         struct inode *inode = mapping->host;
1947         unsigned start;
1948
1949         start = pos & (PAGE_CACHE_SIZE - 1);
1950
1951         if (unlikely(copied < len)) {
1952                 /*
1953                  * The buffers that were written will now be uptodate, so we
1954                  * don't have to worry about a readpage reading them and
1955                  * overwriting a partial write. However if we have encountered
1956                  * a short write and only partially written into a buffer, it
1957                  * will not be marked uptodate, so a readpage might come in and
1958                  * destroy our partial write.
1959                  *
1960                  * Do the simplest thing, and just treat any short write to a
1961                  * non uptodate page as a zero-length write, and force the
1962                  * caller to redo the whole thing.
1963                  */
1964                 if (!PageUptodate(page))
1965                         copied = 0;
1966
1967                 page_zero_new_buffers(page, start+copied, start+len);
1968         }
1969         flush_dcache_page(page);
1970
1971         /* This could be a short (even 0-length) commit */
1972         __block_commit_write(inode, page, start, start+copied);
1973
1974         return copied;
1975 }
1976 EXPORT_SYMBOL(block_write_end);
1977
1978 int generic_write_end(struct file *file, struct address_space *mapping,
1979                         loff_t pos, unsigned len, unsigned copied,
1980                         struct page *page, void *fsdata)
1981 {
1982         struct inode *inode = mapping->host;
1983         int i_size_changed = 0;
1984
1985         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1986
1987         /*
1988          * No need to use i_size_read() here, the i_size
1989          * cannot change under us because we hold i_mutex.
1990          *
1991          * But it's important to update i_size while still holding page lock:
1992          * page writeout could otherwise come in and zero beyond i_size.
1993          */
1994         if (pos+copied > inode->i_size) {
1995                 i_size_write(inode, pos+copied);
1996                 i_size_changed = 1;
1997         }
1998
1999         unlock_page(page);
2000         page_cache_release(page);
2001
2002         /*
2003          * Don't mark the inode dirty under page lock. First, it unnecessarily
2004          * makes the holding time of page lock longer. Second, it forces lock
2005          * ordering of page lock and transaction start for journaling
2006          * filesystems.
2007          */
2008         if (i_size_changed)
2009                 mark_inode_dirty(inode);
2010
2011         return copied;
2012 }
2013 EXPORT_SYMBOL(generic_write_end);
2014
2015 /*
2016  * block_is_partially_uptodate checks whether buffers within a page are
2017  * uptodate or not.
2018  *
2019  * Returns true if all buffers which correspond to a file portion
2020  * we want to read are uptodate.
2021  */
2022 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2023                                         unsigned long from)
2024 {
2025         struct inode *inode = page->mapping->host;
2026         unsigned block_start, block_end, blocksize;
2027         unsigned to;
2028         struct buffer_head *bh, *head;
2029         int ret = 1;
2030
2031         if (!page_has_buffers(page))
2032                 return 0;
2033
2034         blocksize = 1 << inode->i_blkbits;
2035         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2036         to = from + to;
2037         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2038                 return 0;
2039
2040         head = page_buffers(page);
2041         bh = head;
2042         block_start = 0;
2043         do {
2044                 block_end = block_start + blocksize;
2045                 if (block_end > from && block_start < to) {
2046                         if (!buffer_uptodate(bh)) {
2047                                 ret = 0;
2048                                 break;
2049                         }
2050                         if (block_end >= to)
2051                                 break;
2052                 }
2053                 block_start = block_end;
2054                 bh = bh->b_this_page;
2055         } while (bh != head);
2056
2057         return ret;
2058 }
2059 EXPORT_SYMBOL(block_is_partially_uptodate);
2060
2061 /*
2062  * Generic "read page" function for block devices that have the normal
2063  * get_block functionality. This is most of the block device filesystems.
2064  * Reads the page asynchronously --- the unlock_buffer() and
2065  * set/clear_buffer_uptodate() functions propagate buffer state into the
2066  * page struct once IO has completed.
2067  */
2068 int block_read_full_page(struct page *page, get_block_t *get_block)
2069 {
2070         struct inode *inode = page->mapping->host;
2071         sector_t iblock, lblock;
2072         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2073         unsigned int blocksize;
2074         int nr, i;
2075         int fully_mapped = 1;
2076
2077         BUG_ON(!PageLocked(page));
2078         blocksize = 1 << inode->i_blkbits;
2079         if (!page_has_buffers(page))
2080                 create_empty_buffers(page, blocksize, 0);
2081         head = page_buffers(page);
2082
2083         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2084         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2085         bh = head;
2086         nr = 0;
2087         i = 0;
2088
2089         do {
2090                 if (buffer_uptodate(bh))
2091                         continue;
2092
2093                 if (!buffer_mapped(bh)) {
2094                         int err = 0;
2095
2096                         fully_mapped = 0;
2097                         if (iblock < lblock) {
2098                                 WARN_ON(bh->b_size != blocksize);
2099                                 err = get_block(inode, iblock, bh, 0);
2100                                 if (err)
2101                                         SetPageError(page);
2102                         }
2103                         if (!buffer_mapped(bh)) {
2104                                 zero_user(page, i * blocksize, blocksize);
2105                                 if (!err)
2106                                         set_buffer_uptodate(bh);
2107                                 continue;
2108                         }
2109                         /*
2110                          * get_block() might have updated the buffer
2111                          * synchronously
2112                          */
2113                         if (buffer_uptodate(bh))
2114                                 continue;
2115                 }
2116                 arr[nr++] = bh;
2117         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2118
2119         if (fully_mapped)
2120                 SetPageMappedToDisk(page);
2121
2122         if (!nr) {
2123                 /*
2124                  * All buffers are uptodate - we can set the page uptodate
2125                  * as well. But not if get_block() returned an error.
2126                  */
2127                 if (!PageError(page))
2128                         SetPageUptodate(page);
2129                 unlock_page(page);
2130                 return 0;
2131         }
2132
2133         /* Stage two: lock the buffers */
2134         for (i = 0; i < nr; i++) {
2135                 bh = arr[i];
2136                 lock_buffer(bh);
2137                 mark_buffer_async_read(bh);
2138         }
2139
2140         /*
2141          * Stage 3: start the IO.  Check for uptodateness
2142          * inside the buffer lock in case another process reading
2143          * the underlying blockdev brought it uptodate (the sct fix).
2144          */
2145         for (i = 0; i < nr; i++) {
2146                 bh = arr[i];
2147                 if (buffer_uptodate(bh))
2148                         end_buffer_async_read(bh, 1);
2149                 else
2150                         submit_bh(READ, bh);
2151         }
2152         return 0;
2153 }
2154 EXPORT_SYMBOL(block_read_full_page);
2155
2156 /* utility function for filesystems that need to do work on expanding
2157  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2158  * deal with the hole.  
2159  */
2160 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2161 {
2162         struct address_space *mapping = inode->i_mapping;
2163         struct page *page;
2164         void *fsdata;
2165         int err;
2166
2167         err = inode_newsize_ok(inode, size);
2168         if (err)
2169                 goto out;
2170
2171         err = pagecache_write_begin(NULL, mapping, size, 0,
2172                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2173                                 &page, &fsdata);
2174         if (err)
2175                 goto out;
2176
2177         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2178         BUG_ON(err > 0);
2179
2180 out:
2181         return err;
2182 }
2183 EXPORT_SYMBOL(generic_cont_expand_simple);
2184
2185 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2186                             loff_t pos, loff_t *bytes)
2187 {
2188         struct inode *inode = mapping->host;
2189         unsigned blocksize = 1 << inode->i_blkbits;
2190         struct page *page;
2191         void *fsdata;
2192         pgoff_t index, curidx;
2193         loff_t curpos;
2194         unsigned zerofrom, offset, len;
2195         int err = 0;
2196
2197         index = pos >> PAGE_CACHE_SHIFT;
2198         offset = pos & ~PAGE_CACHE_MASK;
2199
2200         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2201                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2202                 if (zerofrom & (blocksize-1)) {
2203                         *bytes |= (blocksize-1);
2204                         (*bytes)++;
2205                 }
2206                 len = PAGE_CACHE_SIZE - zerofrom;
2207
2208                 err = pagecache_write_begin(file, mapping, curpos, len,
2209                                                 AOP_FLAG_UNINTERRUPTIBLE,
2210                                                 &page, &fsdata);
2211                 if (err)
2212                         goto out;
2213                 zero_user(page, zerofrom, len);
2214                 err = pagecache_write_end(file, mapping, curpos, len, len,
2215                                                 page, fsdata);
2216                 if (err < 0)
2217                         goto out;
2218                 BUG_ON(err != len);
2219                 err = 0;
2220
2221                 balance_dirty_pages_ratelimited(mapping);
2222         }
2223
2224         /* page covers the boundary, find the boundary offset */
2225         if (index == curidx) {
2226                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2227                 /* if we will expand the thing last block will be filled */
2228                 if (offset <= zerofrom) {
2229                         goto out;
2230                 }
2231                 if (zerofrom & (blocksize-1)) {
2232                         *bytes |= (blocksize-1);
2233                         (*bytes)++;
2234                 }
2235                 len = offset - zerofrom;
2236
2237                 err = pagecache_write_begin(file, mapping, curpos, len,
2238                                                 AOP_FLAG_UNINTERRUPTIBLE,
2239                                                 &page, &fsdata);
2240                 if (err)
2241                         goto out;
2242                 zero_user(page, zerofrom, len);
2243                 err = pagecache_write_end(file, mapping, curpos, len, len,
2244                                                 page, fsdata);
2245                 if (err < 0)
2246                         goto out;
2247                 BUG_ON(err != len);
2248                 err = 0;
2249         }
2250 out:
2251         return err;
2252 }
2253
2254 /*
2255  * For moronic filesystems that do not allow holes in file.
2256  * We may have to extend the file.
2257  */
2258 int cont_write_begin(struct file *file, struct address_space *mapping,
2259                         loff_t pos, unsigned len, unsigned flags,
2260                         struct page **pagep, void **fsdata,
2261                         get_block_t *get_block, loff_t *bytes)
2262 {
2263         struct inode *inode = mapping->host;
2264         unsigned blocksize = 1 << inode->i_blkbits;
2265         unsigned zerofrom;
2266         int err;
2267
2268         err = cont_expand_zero(file, mapping, pos, bytes);
2269         if (err)
2270                 return err;
2271
2272         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2273         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2274                 *bytes |= (blocksize-1);
2275                 (*bytes)++;
2276         }
2277
2278         return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2279 }
2280 EXPORT_SYMBOL(cont_write_begin);
2281
2282 int block_commit_write(struct page *page, unsigned from, unsigned to)
2283 {
2284         struct inode *inode = page->mapping->host;
2285         __block_commit_write(inode,page,from,to);
2286         return 0;
2287 }
2288 EXPORT_SYMBOL(block_commit_write);
2289
2290 /*
2291  * block_page_mkwrite() is not allowed to change the file size as it gets
2292  * called from a page fault handler when a page is first dirtied. Hence we must
2293  * be careful to check for EOF conditions here. We set the page up correctly
2294  * for a written page which means we get ENOSPC checking when writing into
2295  * holes and correct delalloc and unwritten extent mapping on filesystems that
2296  * support these features.
2297  *
2298  * We are not allowed to take the i_mutex here so we have to play games to
2299  * protect against truncate races as the page could now be beyond EOF.  Because
2300  * truncate writes the inode size before removing pages, once we have the
2301  * page lock we can determine safely if the page is beyond EOF. If it is not
2302  * beyond EOF, then the page is guaranteed safe against truncation until we
2303  * unlock the page.
2304  *
2305  * Direct callers of this function should call vfs_check_frozen() so that page
2306  * fault does not busyloop until the fs is thawed.
2307  */
2308 int __block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2309                          get_block_t get_block)
2310 {
2311         struct page *page = vmf->page;
2312         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2313         unsigned long end;
2314         loff_t size;
2315         int ret;
2316
2317         lock_page(page);
2318         size = i_size_read(inode);
2319         if ((page->mapping != inode->i_mapping) ||
2320             (page_offset(page) > size)) {
2321                 /* We overload EFAULT to mean page got truncated */
2322                 ret = -EFAULT;
2323                 goto out_unlock;
2324         }
2325
2326         /* page is wholly or partially inside EOF */
2327         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2328                 end = size & ~PAGE_CACHE_MASK;
2329         else
2330                 end = PAGE_CACHE_SIZE;
2331
2332         ret = __block_write_begin(page, 0, end, get_block);
2333         if (!ret)
2334                 ret = block_commit_write(page, 0, end);
2335
2336         if (unlikely(ret < 0))
2337                 goto out_unlock;
2338         /*
2339          * Freezing in progress? We check after the page is marked dirty and
2340          * with page lock held so if the test here fails, we are sure freezing
2341          * code will wait during syncing until the page fault is done - at that
2342          * point page will be dirty and unlocked so freezing code will write it
2343          * and writeprotect it again.
2344          */
2345         set_page_dirty(page);
2346         if (inode->i_sb->s_frozen != SB_UNFROZEN) {
2347                 ret = -EAGAIN;
2348                 goto out_unlock;
2349         }
2350         wait_on_page_writeback(page);
2351         return 0;
2352 out_unlock:
2353         unlock_page(page);
2354         return ret;
2355 }
2356 EXPORT_SYMBOL(__block_page_mkwrite);
2357
2358 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2359                    get_block_t get_block)
2360 {
2361         int ret;
2362         struct super_block *sb = vma->vm_file->f_path.dentry->d_inode->i_sb;
2363
2364         /*
2365          * This check is racy but catches the common case. The check in
2366          * __block_page_mkwrite() is reliable.
2367          */
2368         vfs_check_frozen(sb, SB_FREEZE_WRITE);
2369         ret = __block_page_mkwrite(vma, vmf, get_block);
2370         return block_page_mkwrite_return(ret);
2371 }
2372 EXPORT_SYMBOL(block_page_mkwrite);
2373
2374 /*
2375  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2376  * immediately, while under the page lock.  So it needs a special end_io
2377  * handler which does not touch the bh after unlocking it.
2378  */
2379 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2380 {
2381         __end_buffer_read_notouch(bh, uptodate);
2382 }
2383
2384 /*
2385  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2386  * the page (converting it to circular linked list and taking care of page
2387  * dirty races).
2388  */
2389 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2390 {
2391         struct buffer_head *bh;
2392
2393         BUG_ON(!PageLocked(page));
2394
2395         spin_lock(&page->mapping->private_lock);
2396         bh = head;
2397         do {
2398                 if (PageDirty(page))
2399                         set_buffer_dirty(bh);
2400                 if (!bh->b_this_page)
2401                         bh->b_this_page = head;
2402                 bh = bh->b_this_page;
2403         } while (bh != head);
2404         attach_page_buffers(page, head);
2405         spin_unlock(&page->mapping->private_lock);
2406 }
2407
2408 /*
2409  * On entry, the page is fully not uptodate.
2410  * On exit the page is fully uptodate in the areas outside (from,to)
2411  * The filesystem needs to handle block truncation upon failure.
2412  */
2413 int nobh_write_begin(struct address_space *mapping,
2414                         loff_t pos, unsigned len, unsigned flags,
2415                         struct page **pagep, void **fsdata,
2416                         get_block_t *get_block)
2417 {
2418         struct inode *inode = mapping->host;
2419         const unsigned blkbits = inode->i_blkbits;
2420         const unsigned blocksize = 1 << blkbits;
2421         struct buffer_head *head, *bh;
2422         struct page *page;
2423         pgoff_t index;
2424         unsigned from, to;
2425         unsigned block_in_page;
2426         unsigned block_start, block_end;
2427         sector_t block_in_file;
2428         int nr_reads = 0;
2429         int ret = 0;
2430         int is_mapped_to_disk = 1;
2431
2432         index = pos >> PAGE_CACHE_SHIFT;
2433         from = pos & (PAGE_CACHE_SIZE - 1);
2434         to = from + len;
2435
2436         page = grab_cache_page_write_begin(mapping, index, flags);
2437         if (!page)
2438                 return -ENOMEM;
2439         *pagep = page;
2440         *fsdata = NULL;
2441
2442         if (page_has_buffers(page)) {
2443                 ret = __block_write_begin(page, pos, len, get_block);
2444                 if (unlikely(ret))
2445                         goto out_release;
2446                 return ret;
2447         }
2448
2449         if (PageMappedToDisk(page))
2450                 return 0;
2451
2452         /*
2453          * Allocate buffers so that we can keep track of state, and potentially
2454          * attach them to the page if an error occurs. In the common case of
2455          * no error, they will just be freed again without ever being attached
2456          * to the page (which is all OK, because we're under the page lock).
2457          *
2458          * Be careful: the buffer linked list is a NULL terminated one, rather
2459          * than the circular one we're used to.
2460          */
2461         head = alloc_page_buffers(page, blocksize, 0);
2462         if (!head) {
2463                 ret = -ENOMEM;
2464                 goto out_release;
2465         }
2466
2467         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2468
2469         /*
2470          * We loop across all blocks in the page, whether or not they are
2471          * part of the affected region.  This is so we can discover if the
2472          * page is fully mapped-to-disk.
2473          */
2474         for (block_start = 0, block_in_page = 0, bh = head;
2475                   block_start < PAGE_CACHE_SIZE;
2476                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2477                 int create;
2478
2479                 block_end = block_start + blocksize;
2480                 bh->b_state = 0;
2481                 create = 1;
2482                 if (block_start >= to)
2483                         create = 0;
2484                 ret = get_block(inode, block_in_file + block_in_page,
2485                                         bh, create);
2486                 if (ret)
2487                         goto failed;
2488                 if (!buffer_mapped(bh))
2489                         is_mapped_to_disk = 0;
2490                 if (buffer_new(bh))
2491                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2492                 if (PageUptodate(page)) {
2493                         set_buffer_uptodate(bh);
2494                         continue;
2495                 }
2496                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2497                         zero_user_segments(page, block_start, from,
2498                                                         to, block_end);
2499                         continue;
2500                 }
2501                 if (buffer_uptodate(bh))
2502                         continue;       /* reiserfs does this */
2503                 if (block_start < from || block_end > to) {
2504                         lock_buffer(bh);
2505                         bh->b_end_io = end_buffer_read_nobh;
2506                         submit_bh(READ, bh);
2507                         nr_reads++;
2508                 }
2509         }
2510
2511         if (nr_reads) {
2512                 /*
2513                  * The page is locked, so these buffers are protected from
2514                  * any VM or truncate activity.  Hence we don't need to care
2515                  * for the buffer_head refcounts.
2516                  */
2517                 for (bh = head; bh; bh = bh->b_this_page) {
2518                         wait_on_buffer(bh);
2519                         if (!buffer_uptodate(bh))
2520                                 ret = -EIO;
2521                 }
2522                 if (ret)
2523                         goto failed;
2524         }
2525
2526         if (is_mapped_to_disk)
2527                 SetPageMappedToDisk(page);
2528
2529         *fsdata = head; /* to be released by nobh_write_end */
2530
2531         return 0;
2532
2533 failed:
2534         BUG_ON(!ret);
2535         /*
2536          * Error recovery is a bit difficult. We need to zero out blocks that
2537          * were newly allocated, and dirty them to ensure they get written out.
2538          * Buffers need to be attached to the page at this point, otherwise
2539          * the handling of potential IO errors during writeout would be hard
2540          * (could try doing synchronous writeout, but what if that fails too?)
2541          */
2542         attach_nobh_buffers(page, head);
2543         page_zero_new_buffers(page, from, to);
2544
2545 out_release:
2546         unlock_page(page);
2547         page_cache_release(page);
2548         *pagep = NULL;
2549
2550         return ret;
2551 }
2552 EXPORT_SYMBOL(nobh_write_begin);
2553
2554 int nobh_write_end(struct file *file, struct address_space *mapping,
2555                         loff_t pos, unsigned len, unsigned copied,
2556                         struct page *page, void *fsdata)
2557 {
2558         struct inode *inode = page->mapping->host;
2559         struct buffer_head *head = fsdata;
2560         struct buffer_head *bh;
2561         BUG_ON(fsdata != NULL && page_has_buffers(page));
2562
2563         if (unlikely(copied < len) && head)
2564                 attach_nobh_buffers(page, head);
2565         if (page_has_buffers(page))
2566                 return generic_write_end(file, mapping, pos, len,
2567                                         copied, page, fsdata);
2568
2569         SetPageUptodate(page);
2570         set_page_dirty(page);
2571         if (pos+copied > inode->i_size) {
2572                 i_size_write(inode, pos+copied);
2573                 mark_inode_dirty(inode);
2574         }
2575
2576         unlock_page(page);
2577         page_cache_release(page);
2578
2579         while (head) {
2580                 bh = head;
2581                 head = head->b_this_page;
2582                 free_buffer_head(bh);
2583         }
2584
2585         return copied;
2586 }
2587 EXPORT_SYMBOL(nobh_write_end);
2588
2589 /*
2590  * nobh_writepage() - based on block_full_write_page() except
2591  * that it tries to operate without attaching bufferheads to
2592  * the page.
2593  */
2594 int nobh_writepage(struct page *page, get_block_t *get_block,
2595                         struct writeback_control *wbc)
2596 {
2597         struct inode * const inode = page->mapping->host;
2598         loff_t i_size = i_size_read(inode);
2599         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2600         unsigned offset;
2601         int ret;
2602
2603         /* Is the page fully inside i_size? */
2604         if (page->index < end_index)
2605                 goto out;
2606
2607         /* Is the page fully outside i_size? (truncate in progress) */
2608         offset = i_size & (PAGE_CACHE_SIZE-1);
2609         if (page->index >= end_index+1 || !offset) {
2610                 /*
2611                  * The page may have dirty, unmapped buffers.  For example,
2612                  * they may have been added in ext3_writepage().  Make them
2613                  * freeable here, so the page does not leak.
2614                  */
2615 #if 0
2616                 /* Not really sure about this  - do we need this ? */
2617                 if (page->mapping->a_ops->invalidatepage)
2618                         page->mapping->a_ops->invalidatepage(page, offset);
2619 #endif
2620                 unlock_page(page);
2621                 return 0; /* don't care */
2622         }
2623
2624         /*
2625          * The page straddles i_size.  It must be zeroed out on each and every
2626          * writepage invocation because it may be mmapped.  "A file is mapped
2627          * in multiples of the page size.  For a file that is not a multiple of
2628          * the  page size, the remaining memory is zeroed when mapped, and
2629          * writes to that region are not written out to the file."
2630          */
2631         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2632 out:
2633         ret = mpage_writepage(page, get_block, wbc);
2634         if (ret == -EAGAIN)
2635                 ret = __block_write_full_page(inode, page, get_block, wbc,
2636                                               end_buffer_async_write);
2637         return ret;
2638 }
2639 EXPORT_SYMBOL(nobh_writepage);
2640
2641 int nobh_truncate_page(struct address_space *mapping,
2642                         loff_t from, get_block_t *get_block)
2643 {
2644         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2645         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2646         unsigned blocksize;
2647         sector_t iblock;
2648         unsigned length, pos;
2649         struct inode *inode = mapping->host;
2650         struct page *page;
2651         struct buffer_head map_bh;
2652         int err;
2653
2654         blocksize = 1 << inode->i_blkbits;
2655         length = offset & (blocksize - 1);
2656
2657         /* Block boundary? Nothing to do */
2658         if (!length)
2659                 return 0;
2660
2661         length = blocksize - length;
2662         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2663
2664         page = grab_cache_page(mapping, index);
2665         err = -ENOMEM;
2666         if (!page)
2667                 goto out;
2668
2669         if (page_has_buffers(page)) {
2670 has_buffers:
2671                 unlock_page(page);
2672                 page_cache_release(page);
2673                 return block_truncate_page(mapping, from, get_block);
2674         }
2675
2676         /* Find the buffer that contains "offset" */
2677         pos = blocksize;
2678         while (offset >= pos) {
2679                 iblock++;
2680                 pos += blocksize;
2681         }
2682
2683         map_bh.b_size = blocksize;
2684         map_bh.b_state = 0;
2685         err = get_block(inode, iblock, &map_bh, 0);
2686         if (err)
2687                 goto unlock;
2688         /* unmapped? It's a hole - nothing to do */
2689         if (!buffer_mapped(&map_bh))
2690                 goto unlock;
2691
2692         /* Ok, it's mapped. Make sure it's up-to-date */
2693         if (!PageUptodate(page)) {
2694                 err = mapping->a_ops->readpage(NULL, page);
2695                 if (err) {
2696                         page_cache_release(page);
2697                         goto out;
2698                 }
2699                 lock_page(page);
2700                 if (!PageUptodate(page)) {
2701                         err = -EIO;
2702                         goto unlock;
2703                 }
2704                 if (page_has_buffers(page))
2705                         goto has_buffers;
2706         }
2707         zero_user(page, offset, length);
2708         set_page_dirty(page);
2709         err = 0;
2710
2711 unlock:
2712         unlock_page(page);
2713         page_cache_release(page);
2714 out:
2715         return err;
2716 }
2717 EXPORT_SYMBOL(nobh_truncate_page);
2718
2719 int block_truncate_page(struct address_space *mapping,
2720                         loff_t from, get_block_t *get_block)
2721 {
2722         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2723         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2724         unsigned blocksize;
2725         sector_t iblock;
2726         unsigned length, pos;
2727         struct inode *inode = mapping->host;
2728         struct page *page;
2729         struct buffer_head *bh;
2730         int err;
2731
2732         blocksize = 1 << inode->i_blkbits;
2733         length = offset & (blocksize - 1);
2734
2735         /* Block boundary? Nothing to do */
2736         if (!length)
2737                 return 0;
2738
2739         length = blocksize - length;
2740         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2741         
2742         page = grab_cache_page(mapping, index);
2743         err = -ENOMEM;
2744         if (!page)
2745                 goto out;
2746
2747         if (!page_has_buffers(page))
2748                 create_empty_buffers(page, blocksize, 0);
2749
2750         /* Find the buffer that contains "offset" */
2751         bh = page_buffers(page);
2752         pos = blocksize;
2753         while (offset >= pos) {
2754                 bh = bh->b_this_page;
2755                 iblock++;
2756                 pos += blocksize;
2757         }
2758
2759         err = 0;
2760         if (!buffer_mapped(bh)) {
2761                 WARN_ON(bh->b_size != blocksize);
2762                 err = get_block(inode, iblock, bh, 0);
2763                 if (err)
2764                         goto unlock;
2765                 /* unmapped? It's a hole - nothing to do */
2766                 if (!buffer_mapped(bh))
2767                         goto unlock;
2768         }
2769
2770         /* Ok, it's mapped. Make sure it's up-to-date */
2771         if (PageUptodate(page))
2772                 set_buffer_uptodate(bh);
2773
2774         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2775                 err = -EIO;
2776                 ll_rw_block(READ, 1, &bh);
2777                 wait_on_buffer(bh);
2778                 /* Uhhuh. Read error. Complain and punt. */
2779                 if (!buffer_uptodate(bh))
2780                         goto unlock;
2781         }
2782
2783         zero_user(page, offset, length);
2784         mark_buffer_dirty(bh);
2785         err = 0;
2786
2787 unlock:
2788         unlock_page(page);
2789         page_cache_release(page);
2790 out:
2791         return err;
2792 }
2793 EXPORT_SYMBOL(block_truncate_page);
2794
2795 /*
2796  * The generic ->writepage function for buffer-backed address_spaces
2797  * this form passes in the end_io handler used to finish the IO.
2798  */
2799 int block_write_full_page_endio(struct page *page, get_block_t *get_block,
2800                         struct writeback_control *wbc, bh_end_io_t *handler)
2801 {
2802         struct inode * const inode = page->mapping->host;
2803         loff_t i_size = i_size_read(inode);
2804         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2805         unsigned offset;
2806
2807         /* Is the page fully inside i_size? */
2808         if (page->index < end_index)
2809                 return __block_write_full_page(inode, page, get_block, wbc,
2810                                                handler);
2811
2812         /* Is the page fully outside i_size? (truncate in progress) */
2813         offset = i_size & (PAGE_CACHE_SIZE-1);
2814         if (page->index >= end_index+1 || !offset) {
2815                 /*
2816                  * The page may have dirty, unmapped buffers.  For example,
2817                  * they may have been added in ext3_writepage().  Make them
2818                  * freeable here, so the page does not leak.
2819                  */
2820                 do_invalidatepage(page, 0);
2821                 unlock_page(page);
2822                 return 0; /* don't care */
2823         }
2824
2825         /*
2826          * The page straddles i_size.  It must be zeroed out on each and every
2827          * writepage invocation because it may be mmapped.  "A file is mapped
2828          * in multiples of the page size.  For a file that is not a multiple of
2829          * the  page size, the remaining memory is zeroed when mapped, and
2830          * writes to that region are not written out to the file."
2831          */
2832         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2833         return __block_write_full_page(inode, page, get_block, wbc, handler);
2834 }
2835 EXPORT_SYMBOL(block_write_full_page_endio);
2836
2837 /*
2838  * The generic ->writepage function for buffer-backed address_spaces
2839  */
2840 int block_write_full_page(struct page *page, get_block_t *get_block,
2841                         struct writeback_control *wbc)
2842 {
2843         return block_write_full_page_endio(page, get_block, wbc,
2844                                            end_buffer_async_write);
2845 }
2846 EXPORT_SYMBOL(block_write_full_page);
2847
2848 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2849                             get_block_t *get_block)
2850 {
2851         struct buffer_head tmp;
2852         struct inode *inode = mapping->host;
2853         tmp.b_state = 0;
2854         tmp.b_blocknr = 0;
2855         tmp.b_size = 1 << inode->i_blkbits;
2856         get_block(inode, block, &tmp, 0);
2857         return tmp.b_blocknr;
2858 }
2859 EXPORT_SYMBOL(generic_block_bmap);
2860
2861 static void end_bio_bh_io_sync(struct bio *bio, int err)
2862 {
2863         struct buffer_head *bh = bio->bi_private;
2864
2865         if (err == -EOPNOTSUPP) {
2866                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2867         }
2868
2869         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2870                 set_bit(BH_Quiet, &bh->b_state);
2871
2872         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2873         bio_put(bio);
2874 }
2875
2876 int submit_bh(int rw, struct buffer_head * bh)
2877 {
2878         struct bio *bio;
2879         int ret = 0;
2880
2881         BUG_ON(!buffer_locked(bh));
2882         BUG_ON(!buffer_mapped(bh));
2883         BUG_ON(!bh->b_end_io);
2884         BUG_ON(buffer_delay(bh));
2885         BUG_ON(buffer_unwritten(bh));
2886
2887         /*
2888          * Only clear out a write error when rewriting
2889          */
2890         if (test_set_buffer_req(bh) && (rw & WRITE))
2891                 clear_buffer_write_io_error(bh);
2892
2893         /*
2894          * from here on down, it's all bio -- do the initial mapping,
2895          * submit_bio -> generic_make_request may further map this bio around
2896          */
2897         bio = bio_alloc(GFP_NOIO, 1);
2898
2899         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2900         bio->bi_bdev = bh->b_bdev;
2901         bio->bi_io_vec[0].bv_page = bh->b_page;
2902         bio->bi_io_vec[0].bv_len = bh->b_size;
2903         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2904
2905         bio->bi_vcnt = 1;
2906         bio->bi_idx = 0;
2907         bio->bi_size = bh->b_size;
2908
2909         bio->bi_end_io = end_bio_bh_io_sync;
2910         bio->bi_private = bh;
2911
2912         bio_get(bio);
2913         submit_bio(rw, bio);
2914
2915         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2916                 ret = -EOPNOTSUPP;
2917
2918         bio_put(bio);
2919         return ret;
2920 }
2921 EXPORT_SYMBOL(submit_bh);
2922
2923 /**
2924  * ll_rw_block: low-level access to block devices (DEPRECATED)
2925  * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
2926  * @nr: number of &struct buffer_heads in the array
2927  * @bhs: array of pointers to &struct buffer_head
2928  *
2929  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2930  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2931  * %READA option is described in the documentation for generic_make_request()
2932  * which ll_rw_block() calls.
2933  *
2934  * This function drops any buffer that it cannot get a lock on (with the
2935  * BH_Lock state bit), any buffer that appears to be clean when doing a write
2936  * request, and any buffer that appears to be up-to-date when doing read
2937  * request.  Further it marks as clean buffers that are processed for
2938  * writing (the buffer cache won't assume that they are actually clean
2939  * until the buffer gets unlocked).
2940  *
2941  * ll_rw_block sets b_end_io to simple completion handler that marks
2942  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2943  * any waiters. 
2944  *
2945  * All of the buffers must be for the same device, and must also be a
2946  * multiple of the current approved size for the device.
2947  */
2948 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2949 {
2950         int i;
2951
2952         for (i = 0; i < nr; i++) {
2953                 struct buffer_head *bh = bhs[i];
2954
2955                 if (!trylock_buffer(bh))
2956                         continue;
2957                 if (rw == WRITE) {
2958                         if (test_clear_buffer_dirty(bh)) {
2959                                 bh->b_end_io = end_buffer_write_sync;
2960                                 get_bh(bh);
2961                                 submit_bh(WRITE, bh);
2962                                 continue;
2963                         }
2964                 } else {
2965                         if (!buffer_uptodate(bh)) {
2966                                 bh->b_end_io = end_buffer_read_sync;
2967                                 get_bh(bh);
2968                                 submit_bh(rw, bh);
2969                                 continue;
2970                         }
2971                 }
2972                 unlock_buffer(bh);
2973         }
2974 }
2975 EXPORT_SYMBOL(ll_rw_block);
2976
2977 void write_dirty_buffer(struct buffer_head *bh, int rw)
2978 {
2979         lock_buffer(bh);
2980         if (!test_clear_buffer_dirty(bh)) {
2981                 unlock_buffer(bh);
2982                 return;
2983         }
2984         bh->b_end_io = end_buffer_write_sync;
2985         get_bh(bh);
2986         submit_bh(rw, bh);
2987 }
2988 EXPORT_SYMBOL(write_dirty_buffer);
2989
2990 /*
2991  * For a data-integrity writeout, we need to wait upon any in-progress I/O
2992  * and then start new I/O and then wait upon it.  The caller must have a ref on
2993  * the buffer_head.
2994  */
2995 int __sync_dirty_buffer(struct buffer_head *bh, int rw)
2996 {
2997         int ret = 0;
2998
2999         WARN_ON(atomic_read(&bh->b_count) < 1);
3000         lock_buffer(bh);
3001         if (test_clear_buffer_dirty(bh)) {
3002                 get_bh(bh);
3003                 bh->b_end_io = end_buffer_write_sync;
3004                 ret = submit_bh(rw, bh);
3005                 wait_on_buffer(bh);
3006                 if (!ret && !buffer_uptodate(bh))
3007                         ret = -EIO;
3008         } else {
3009                 unlock_buffer(bh);
3010         }
3011         return ret;
3012 }
3013 EXPORT_SYMBOL(__sync_dirty_buffer);
3014
3015 int sync_dirty_buffer(struct buffer_head *bh)
3016 {
3017         return __sync_dirty_buffer(bh, WRITE_SYNC);
3018 }
3019 EXPORT_SYMBOL(sync_dirty_buffer);
3020
3021 /*
3022  * try_to_free_buffers() checks if all the buffers on this particular page
3023  * are unused, and releases them if so.
3024  *
3025  * Exclusion against try_to_free_buffers may be obtained by either
3026  * locking the page or by holding its mapping's private_lock.
3027  *
3028  * If the page is dirty but all the buffers are clean then we need to
3029  * be sure to mark the page clean as well.  This is because the page
3030  * may be against a block device, and a later reattachment of buffers
3031  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3032  * filesystem data on the same device.
3033  *
3034  * The same applies to regular filesystem pages: if all the buffers are
3035  * clean then we set the page clean and proceed.  To do that, we require
3036  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3037  * private_lock.
3038  *
3039  * try_to_free_buffers() is non-blocking.
3040  */
3041 static inline int buffer_busy(struct buffer_head *bh)
3042 {
3043         return atomic_read(&bh->b_count) |
3044                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3045 }
3046
3047 static int
3048 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3049 {
3050         struct buffer_head *head = page_buffers(page);
3051         struct buffer_head *bh;
3052
3053         bh = head;
3054         do {
3055                 if (buffer_write_io_error(bh) && page->mapping)
3056                         set_bit(AS_EIO, &page->mapping->flags);
3057                 if (buffer_busy(bh))
3058                         goto failed;
3059                 bh = bh->b_this_page;
3060         } while (bh != head);
3061
3062         do {
3063                 struct buffer_head *next = bh->b_this_page;
3064
3065                 if (bh->b_assoc_map)
3066                         __remove_assoc_queue(bh);
3067                 bh = next;
3068         } while (bh != head);
3069         *buffers_to_free = head;
3070         __clear_page_buffers(page);
3071         return 1;
3072 failed:
3073         return 0;
3074 }
3075
3076 int try_to_free_buffers(struct page *page)
3077 {
3078         struct address_space * const mapping = page->mapping;
3079         struct buffer_head *buffers_to_free = NULL;
3080         int ret = 0;
3081
3082         BUG_ON(!PageLocked(page));
3083         if (PageWriteback(page))
3084                 return 0;
3085
3086         if (mapping == NULL) {          /* can this still happen? */
3087                 ret = drop_buffers(page, &buffers_to_free);
3088                 goto out;
3089         }
3090
3091         spin_lock(&mapping->private_lock);
3092         ret = drop_buffers(page, &buffers_to_free);
3093
3094         /*
3095          * If the filesystem writes its buffers by hand (eg ext3)
3096          * then we can have clean buffers against a dirty page.  We
3097          * clean the page here; otherwise the VM will never notice
3098          * that the filesystem did any IO at all.
3099          *
3100          * Also, during truncate, discard_buffer will have marked all
3101          * the page's buffers clean.  We discover that here and clean
3102          * the page also.
3103          *
3104          * private_lock must be held over this entire operation in order
3105          * to synchronise against __set_page_dirty_buffers and prevent the
3106          * dirty bit from being lost.
3107          */
3108         if (ret)
3109                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3110         spin_unlock(&mapping->private_lock);
3111 out:
3112         if (buffers_to_free) {
3113                 struct buffer_head *bh = buffers_to_free;
3114
3115                 do {
3116                         struct buffer_head *next = bh->b_this_page;
3117                         free_buffer_head(bh);
3118                         bh = next;
3119                 } while (bh != buffers_to_free);
3120         }
3121         return ret;
3122 }
3123 EXPORT_SYMBOL(try_to_free_buffers);
3124
3125 /*
3126  * There are no bdflush tunables left.  But distributions are
3127  * still running obsolete flush daemons, so we terminate them here.
3128  *
3129  * Use of bdflush() is deprecated and will be removed in a future kernel.
3130  * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3131  */
3132 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3133 {
3134         static int msg_count;
3135
3136         if (!capable(CAP_SYS_ADMIN))
3137                 return -EPERM;
3138
3139         if (msg_count < 5) {
3140                 msg_count++;
3141                 printk(KERN_INFO
3142                         "warning: process `%s' used the obsolete bdflush"
3143                         " system call\n", current->comm);
3144                 printk(KERN_INFO "Fix your initscripts?\n");
3145         }
3146
3147         if (func == 1)
3148                 do_exit(0);
3149         return 0;
3150 }
3151
3152 /*
3153  * Buffer-head allocation
3154  */
3155 static struct kmem_cache *bh_cachep;
3156
3157 /*
3158  * Once the number of bh's in the machine exceeds this level, we start
3159  * stripping them in writeback.
3160  */
3161 static int max_buffer_heads;
3162
3163 int buffer_heads_over_limit;
3164
3165 struct bh_accounting {
3166         int nr;                 /* Number of live bh's */
3167         int ratelimit;          /* Limit cacheline bouncing */
3168 };
3169
3170 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3171
3172 static void recalc_bh_state(void)
3173 {
3174         int i;
3175         int tot = 0;
3176
3177         if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3178                 return;
3179         __this_cpu_write(bh_accounting.ratelimit, 0);
3180         for_each_online_cpu(i)
3181                 tot += per_cpu(bh_accounting, i).nr;
3182         buffer_heads_over_limit = (tot > max_buffer_heads);
3183 }
3184
3185 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3186 {
3187         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3188         if (ret) {
3189                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3190                 preempt_disable();
3191                 __this_cpu_inc(bh_accounting.nr);
3192                 recalc_bh_state();
3193                 preempt_enable();
3194         }
3195         return ret;
3196 }
3197 EXPORT_SYMBOL(alloc_buffer_head);
3198
3199 void free_buffer_head(struct buffer_head *bh)
3200 {
3201         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3202         kmem_cache_free(bh_cachep, bh);
3203         preempt_disable();
3204         __this_cpu_dec(bh_accounting.nr);
3205         recalc_bh_state();
3206         preempt_enable();
3207 }
3208 EXPORT_SYMBOL(free_buffer_head);
3209
3210 static void buffer_exit_cpu(int cpu)
3211 {
3212         int i;
3213         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3214
3215         for (i = 0; i < BH_LRU_SIZE; i++) {
3216                 brelse(b->bhs[i]);
3217                 b->bhs[i] = NULL;
3218         }
3219         this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3220         per_cpu(bh_accounting, cpu).nr = 0;
3221 }
3222
3223 static int buffer_cpu_notify(struct notifier_block *self,
3224                               unsigned long action, void *hcpu)
3225 {
3226         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3227                 buffer_exit_cpu((unsigned long)hcpu);
3228         return NOTIFY_OK;
3229 }
3230
3231 /**
3232  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3233  * @bh: struct buffer_head
3234  *
3235  * Return true if the buffer is up-to-date and false,
3236  * with the buffer locked, if not.
3237  */
3238 int bh_uptodate_or_lock(struct buffer_head *bh)
3239 {
3240         if (!buffer_uptodate(bh)) {
3241                 lock_buffer(bh);
3242                 if (!buffer_uptodate(bh))
3243                         return 0;
3244                 unlock_buffer(bh);
3245         }
3246         return 1;
3247 }
3248 EXPORT_SYMBOL(bh_uptodate_or_lock);
3249
3250 /**
3251  * bh_submit_read - Submit a locked buffer for reading
3252  * @bh: struct buffer_head
3253  *
3254  * Returns zero on success and -EIO on error.
3255  */
3256 int bh_submit_read(struct buffer_head *bh)
3257 {
3258         BUG_ON(!buffer_locked(bh));
3259
3260         if (buffer_uptodate(bh)) {
3261                 unlock_buffer(bh);
3262                 return 0;
3263         }
3264
3265         get_bh(bh);
3266         bh->b_end_io = end_buffer_read_sync;
3267         submit_bh(READ, bh);
3268         wait_on_buffer(bh);
3269         if (buffer_uptodate(bh))
3270                 return 0;
3271         return -EIO;
3272 }
3273 EXPORT_SYMBOL(bh_submit_read);
3274
3275 void __init buffer_init(void)
3276 {
3277         int nrpages;
3278
3279         bh_cachep = kmem_cache_create("buffer_head",
3280                         sizeof(struct buffer_head), 0,
3281                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3282                                 SLAB_MEM_SPREAD),
3283                                 NULL);
3284
3285         /*
3286          * Limit the bh occupancy to 10% of ZONE_NORMAL
3287          */
3288         nrpages = (nr_free_buffer_pages() * 10) / 100;
3289         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3290         hotcpu_notifier(buffer_cpu_notify, 0);
3291 }