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