vfs: Add BUG_ON for delayed and unwritten flags in submit_bh()
[linux-3.10.git] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/capability.h>
28 #include <linux/blkdev.h>
29 #include <linux/file.h>
30 #include <linux/quotaops.h>
31 #include <linux/highmem.h>
32 #include <linux/module.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
56 static int sync_buffer(void *word)
57 {
58         struct block_device *bd;
59         struct buffer_head *bh
60                 = container_of(word, struct buffer_head, b_state);
61
62         smp_mb();
63         bd = bh->b_bdev;
64         if (bd)
65                 blk_run_address_space(bd->bd_inode->i_mapping);
66         io_schedule();
67         return 0;
68 }
69
70 void __lock_buffer(struct buffer_head *bh)
71 {
72         wait_on_bit_lock(&bh->b_state, BH_Lock, sync_buffer,
73                                                         TASK_UNINTERRUPTIBLE);
74 }
75 EXPORT_SYMBOL(__lock_buffer);
76
77 void unlock_buffer(struct buffer_head *bh)
78 {
79         clear_bit_unlock(BH_Lock, &bh->b_state);
80         smp_mb__after_clear_bit();
81         wake_up_bit(&bh->b_state, BH_Lock);
82 }
83
84 /*
85  * Block until a buffer comes unlocked.  This doesn't stop it
86  * from becoming locked again - you have to lock it yourself
87  * if you want to preserve its state.
88  */
89 void __wait_on_buffer(struct buffer_head * bh)
90 {
91         wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
92 }
93
94 static void
95 __clear_page_buffers(struct page *page)
96 {
97         ClearPagePrivate(page);
98         set_page_private(page, 0);
99         page_cache_release(page);
100 }
101
102
103 static int quiet_error(struct buffer_head *bh)
104 {
105         if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
106                 return 0;
107         return 1;
108 }
109
110
111 static void buffer_io_error(struct buffer_head *bh)
112 {
113         char b[BDEVNAME_SIZE];
114         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
115                         bdevname(bh->b_bdev, b),
116                         (unsigned long long)bh->b_blocknr);
117 }
118
119 /*
120  * End-of-IO handler helper function which does not touch the bh after
121  * unlocking it.
122  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
123  * a race there is benign: unlock_buffer() only use the bh's address for
124  * hashing after unlocking the buffer, so it doesn't actually touch the bh
125  * itself.
126  */
127 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
128 {
129         if (uptodate) {
130                 set_buffer_uptodate(bh);
131         } else {
132                 /* This happens, due to failed READA attempts. */
133                 clear_buffer_uptodate(bh);
134         }
135         unlock_buffer(bh);
136 }
137
138 /*
139  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
140  * unlock the buffer. This is what ll_rw_block uses too.
141  */
142 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
143 {
144         __end_buffer_read_notouch(bh, uptodate);
145         put_bh(bh);
146 }
147
148 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
149 {
150         char b[BDEVNAME_SIZE];
151
152         if (uptodate) {
153                 set_buffer_uptodate(bh);
154         } else {
155                 if (!buffer_eopnotsupp(bh) && !quiet_error(bh)) {
156                         buffer_io_error(bh);
157                         printk(KERN_WARNING "lost page write due to "
158                                         "I/O error on %s\n",
159                                        bdevname(bh->b_bdev, b));
160                 }
161                 set_buffer_write_io_error(bh);
162                 clear_buffer_uptodate(bh);
163         }
164         unlock_buffer(bh);
165         put_bh(bh);
166 }
167
168 /*
169  * Various filesystems appear to want __find_get_block to be non-blocking.
170  * But it's the page lock which protects the buffers.  To get around this,
171  * we get exclusion from try_to_free_buffers with the blockdev mapping's
172  * private_lock.
173  *
174  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
175  * may be quite high.  This code could TryLock the page, and if that
176  * succeeds, there is no need to take private_lock. (But if
177  * private_lock is contended then so is mapping->tree_lock).
178  */
179 static struct buffer_head *
180 __find_get_block_slow(struct block_device *bdev, sector_t block)
181 {
182         struct inode *bd_inode = bdev->bd_inode;
183         struct address_space *bd_mapping = bd_inode->i_mapping;
184         struct buffer_head *ret = NULL;
185         pgoff_t index;
186         struct buffer_head *bh;
187         struct buffer_head *head;
188         struct page *page;
189         int all_mapped = 1;
190
191         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
192         page = find_get_page(bd_mapping, index);
193         if (!page)
194                 goto out;
195
196         spin_lock(&bd_mapping->private_lock);
197         if (!page_has_buffers(page))
198                 goto out_unlock;
199         head = page_buffers(page);
200         bh = head;
201         do {
202                 if (!buffer_mapped(bh))
203                         all_mapped = 0;
204                 else if (bh->b_blocknr == block) {
205                         ret = bh;
206                         get_bh(bh);
207                         goto out_unlock;
208                 }
209                 bh = bh->b_this_page;
210         } while (bh != head);
211
212         /* we might be here because some of the buffers on this page are
213          * not mapped.  This is due to various races between
214          * file io on the block device and getblk.  It gets dealt with
215          * elsewhere, don't buffer_error if we had some unmapped buffers
216          */
217         if (all_mapped) {
218                 printk("__find_get_block_slow() failed. "
219                         "block=%llu, b_blocknr=%llu\n",
220                         (unsigned long long)block,
221                         (unsigned long long)bh->b_blocknr);
222                 printk("b_state=0x%08lx, b_size=%zu\n",
223                         bh->b_state, bh->b_size);
224                 printk("device blocksize: %d\n", 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 /* If invalidate_buffers() will trash dirty buffers, it means some kind
234    of fs corruption is going on. Trashing dirty data always imply losing
235    information that was supposed to be just stored on the physical layer
236    by the user.
237
238    Thus invalidate_buffers in general usage is not allwowed to trash
239    dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
240    be preserved.  These buffers are simply skipped.
241   
242    We also skip buffers which are still in use.  For example this can
243    happen if a userspace program is reading the block device.
244
245    NOTE: In the case where the user removed a removable-media-disk even if
246    there's still dirty data not synced on disk (due a bug in the device driver
247    or due an error of the user), by not destroying the dirty buffers we could
248    generate corruption also on the next media inserted, thus a parameter is
249    necessary to handle this case in the most safe way possible (trying
250    to not corrupt also the new disk inserted with the data belonging to
251    the old now corrupted disk). Also for the ramdisk the natural thing
252    to do in order to release the ramdisk memory is to destroy dirty buffers.
253
254    These are two special cases. Normal usage imply the device driver
255    to issue a sync on the device (without waiting I/O completion) and
256    then an invalidate_buffers call that doesn't trash dirty buffers.
257
258    For handling cache coherency with the blkdev pagecache the 'update' case
259    is been introduced. It is needed to re-read from disk any pinned
260    buffer. NOTE: re-reading from disk is destructive so we can do it only
261    when we assume nobody is changing the buffercache under our I/O and when
262    we think the disk contains more recent information than the buffercache.
263    The update == 1 pass marks the buffers we need to update, the update == 2
264    pass does the actual I/O. */
265 void invalidate_bdev(struct block_device *bdev)
266 {
267         struct address_space *mapping = bdev->bd_inode->i_mapping;
268
269         if (mapping->nrpages == 0)
270                 return;
271
272         invalidate_bh_lrus();
273         invalidate_mapping_pages(mapping, 0, -1);
274 }
275
276 /*
277  * Kick pdflush then try to free up some ZONE_NORMAL memory.
278  */
279 static void free_more_memory(void)
280 {
281         struct zone *zone;
282         int nid;
283
284         wakeup_pdflush(1024);
285         yield();
286
287         for_each_online_node(nid) {
288                 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
289                                                 gfp_zone(GFP_NOFS), NULL,
290                                                 &zone);
291                 if (zone)
292                         try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
293                                                 GFP_NOFS, NULL);
294         }
295 }
296
297 /*
298  * I/O completion handler for block_read_full_page() - pages
299  * which come unlocked at the end of I/O.
300  */
301 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
302 {
303         unsigned long flags;
304         struct buffer_head *first;
305         struct buffer_head *tmp;
306         struct page *page;
307         int page_uptodate = 1;
308
309         BUG_ON(!buffer_async_read(bh));
310
311         page = bh->b_page;
312         if (uptodate) {
313                 set_buffer_uptodate(bh);
314         } else {
315                 clear_buffer_uptodate(bh);
316                 if (!quiet_error(bh))
317                         buffer_io_error(bh);
318                 SetPageError(page);
319         }
320
321         /*
322          * Be _very_ careful from here on. Bad things can happen if
323          * two buffer heads end IO at almost the same time and both
324          * decide that the page is now completely done.
325          */
326         first = page_buffers(page);
327         local_irq_save(flags);
328         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
329         clear_buffer_async_read(bh);
330         unlock_buffer(bh);
331         tmp = bh;
332         do {
333                 if (!buffer_uptodate(tmp))
334                         page_uptodate = 0;
335                 if (buffer_async_read(tmp)) {
336                         BUG_ON(!buffer_locked(tmp));
337                         goto still_busy;
338                 }
339                 tmp = tmp->b_this_page;
340         } while (tmp != bh);
341         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
342         local_irq_restore(flags);
343
344         /*
345          * If none of the buffers had errors and they are all
346          * uptodate then we can set the page uptodate.
347          */
348         if (page_uptodate && !PageError(page))
349                 SetPageUptodate(page);
350         unlock_page(page);
351         return;
352
353 still_busy:
354         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
355         local_irq_restore(flags);
356         return;
357 }
358
359 /*
360  * Completion handler for block_write_full_page() - pages which are unlocked
361  * during I/O, and which have PageWriteback cleared upon I/O completion.
362  */
363 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
364 {
365         char b[BDEVNAME_SIZE];
366         unsigned long flags;
367         struct buffer_head *first;
368         struct buffer_head *tmp;
369         struct page *page;
370
371         BUG_ON(!buffer_async_write(bh));
372
373         page = bh->b_page;
374         if (uptodate) {
375                 set_buffer_uptodate(bh);
376         } else {
377                 if (!quiet_error(bh)) {
378                         buffer_io_error(bh);
379                         printk(KERN_WARNING "lost page write due to "
380                                         "I/O error on %s\n",
381                                bdevname(bh->b_bdev, b));
382                 }
383                 set_bit(AS_EIO, &page->mapping->flags);
384                 set_buffer_write_io_error(bh);
385                 clear_buffer_uptodate(bh);
386                 SetPageError(page);
387         }
388
389         first = page_buffers(page);
390         local_irq_save(flags);
391         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
392
393         clear_buffer_async_write(bh);
394         unlock_buffer(bh);
395         tmp = bh->b_this_page;
396         while (tmp != bh) {
397                 if (buffer_async_write(tmp)) {
398                         BUG_ON(!buffer_locked(tmp));
399                         goto still_busy;
400                 }
401                 tmp = tmp->b_this_page;
402         }
403         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
404         local_irq_restore(flags);
405         end_page_writeback(page);
406         return;
407
408 still_busy:
409         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
410         local_irq_restore(flags);
411         return;
412 }
413
414 /*
415  * If a page's buffers are under async readin (end_buffer_async_read
416  * completion) then there is a possibility that another thread of
417  * control could lock one of the buffers after it has completed
418  * but while some of the other buffers have not completed.  This
419  * locked buffer would confuse end_buffer_async_read() into not unlocking
420  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
421  * that this buffer is not under async I/O.
422  *
423  * The page comes unlocked when it has no locked buffer_async buffers
424  * left.
425  *
426  * PageLocked prevents anyone starting new async I/O reads any of
427  * the buffers.
428  *
429  * PageWriteback is used to prevent simultaneous writeout of the same
430  * page.
431  *
432  * PageLocked prevents anyone from starting writeback of a page which is
433  * under read I/O (PageWriteback is only ever set against a locked page).
434  */
435 static void mark_buffer_async_read(struct buffer_head *bh)
436 {
437         bh->b_end_io = end_buffer_async_read;
438         set_buffer_async_read(bh);
439 }
440
441 void mark_buffer_async_write_endio(struct buffer_head *bh,
442                                    bh_end_io_t *handler)
443 {
444         bh->b_end_io = handler;
445         set_buffer_async_write(bh);
446 }
447
448 void mark_buffer_async_write(struct buffer_head *bh)
449 {
450         mark_buffer_async_write_endio(bh, end_buffer_async_write);
451 }
452 EXPORT_SYMBOL(mark_buffer_async_write);
453
454
455 /*
456  * fs/buffer.c contains helper functions for buffer-backed address space's
457  * fsync functions.  A common requirement for buffer-based filesystems is
458  * that certain data from the backing blockdev needs to be written out for
459  * a successful fsync().  For example, ext2 indirect blocks need to be
460  * written back and waited upon before fsync() returns.
461  *
462  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
463  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
464  * management of a list of dependent buffers at ->i_mapping->private_list.
465  *
466  * Locking is a little subtle: try_to_free_buffers() will remove buffers
467  * from their controlling inode's queue when they are being freed.  But
468  * try_to_free_buffers() will be operating against the *blockdev* mapping
469  * at the time, not against the S_ISREG file which depends on those buffers.
470  * So the locking for private_list is via the private_lock in the address_space
471  * which backs the buffers.  Which is different from the address_space 
472  * against which the buffers are listed.  So for a particular address_space,
473  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
474  * mapping->private_list will always be protected by the backing blockdev's
475  * ->private_lock.
476  *
477  * Which introduces a requirement: all buffers on an address_space's
478  * ->private_list must be from the same address_space: the blockdev's.
479  *
480  * address_spaces which do not place buffers at ->private_list via these
481  * utility functions are free to use private_lock and private_list for
482  * whatever they want.  The only requirement is that list_empty(private_list)
483  * be true at clear_inode() time.
484  *
485  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
486  * filesystems should do that.  invalidate_inode_buffers() should just go
487  * BUG_ON(!list_empty).
488  *
489  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
490  * take an address_space, not an inode.  And it should be called
491  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
492  * queued up.
493  *
494  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
495  * list if it is already on a list.  Because if the buffer is on a list,
496  * it *must* already be on the right one.  If not, the filesystem is being
497  * silly.  This will save a ton of locking.  But first we have to ensure
498  * that buffers are taken *off* the old inode's list when they are freed
499  * (presumably in truncate).  That requires careful auditing of all
500  * filesystems (do it inside bforget()).  It could also be done by bringing
501  * b_inode back.
502  */
503
504 /*
505  * The buffer's backing address_space's private_lock must be held
506  */
507 static void __remove_assoc_queue(struct buffer_head *bh)
508 {
509         list_del_init(&bh->b_assoc_buffers);
510         WARN_ON(!bh->b_assoc_map);
511         if (buffer_write_io_error(bh))
512                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
513         bh->b_assoc_map = NULL;
514 }
515
516 int inode_has_buffers(struct inode *inode)
517 {
518         return !list_empty(&inode->i_data.private_list);
519 }
520
521 /*
522  * osync is designed to support O_SYNC io.  It waits synchronously for
523  * all already-submitted IO to complete, but does not queue any new
524  * writes to the disk.
525  *
526  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
527  * you dirty the buffers, and then use osync_inode_buffers to wait for
528  * completion.  Any other dirty buffers which are not yet queued for
529  * write will not be flushed to disk by the osync.
530  */
531 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
532 {
533         struct buffer_head *bh;
534         struct list_head *p;
535         int err = 0;
536
537         spin_lock(lock);
538 repeat:
539         list_for_each_prev(p, list) {
540                 bh = BH_ENTRY(p);
541                 if (buffer_locked(bh)) {
542                         get_bh(bh);
543                         spin_unlock(lock);
544                         wait_on_buffer(bh);
545                         if (!buffer_uptodate(bh))
546                                 err = -EIO;
547                         brelse(bh);
548                         spin_lock(lock);
549                         goto repeat;
550                 }
551         }
552         spin_unlock(lock);
553         return err;
554 }
555
556 void do_thaw_all(struct work_struct *work)
557 {
558         struct super_block *sb;
559         char b[BDEVNAME_SIZE];
560
561         spin_lock(&sb_lock);
562 restart:
563         list_for_each_entry(sb, &super_blocks, s_list) {
564                 sb->s_count++;
565                 spin_unlock(&sb_lock);
566                 down_read(&sb->s_umount);
567                 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
568                         printk(KERN_WARNING "Emergency Thaw on %s\n",
569                                bdevname(sb->s_bdev, b));
570                 up_read(&sb->s_umount);
571                 spin_lock(&sb_lock);
572                 if (__put_super_and_need_restart(sb))
573                         goto restart;
574         }
575         spin_unlock(&sb_lock);
576         kfree(work);
577         printk(KERN_WARNING "Emergency Thaw complete\n");
578 }
579
580 /**
581  * emergency_thaw_all -- forcibly thaw every frozen filesystem
582  *
583  * Used for emergency unfreeze of all filesystems via SysRq
584  */
585 void emergency_thaw_all(void)
586 {
587         struct work_struct *work;
588
589         work = kmalloc(sizeof(*work), GFP_ATOMIC);
590         if (work) {
591                 INIT_WORK(work, do_thaw_all);
592                 schedule_work(work);
593         }
594 }
595
596 /**
597  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
598  * @mapping: the mapping which wants those buffers written
599  *
600  * Starts I/O against the buffers at mapping->private_list, and waits upon
601  * that I/O.
602  *
603  * Basically, this is a convenience function for fsync().
604  * @mapping is a file or directory which needs those buffers to be written for
605  * a successful fsync().
606  */
607 int sync_mapping_buffers(struct address_space *mapping)
608 {
609         struct address_space *buffer_mapping = mapping->assoc_mapping;
610
611         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
612                 return 0;
613
614         return fsync_buffers_list(&buffer_mapping->private_lock,
615                                         &mapping->private_list);
616 }
617 EXPORT_SYMBOL(sync_mapping_buffers);
618
619 /*
620  * Called when we've recently written block `bblock', and it is known that
621  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
622  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
623  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
624  */
625 void write_boundary_block(struct block_device *bdev,
626                         sector_t bblock, unsigned blocksize)
627 {
628         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
629         if (bh) {
630                 if (buffer_dirty(bh))
631                         ll_rw_block(WRITE, 1, &bh);
632                 put_bh(bh);
633         }
634 }
635
636 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
637 {
638         struct address_space *mapping = inode->i_mapping;
639         struct address_space *buffer_mapping = bh->b_page->mapping;
640
641         mark_buffer_dirty(bh);
642         if (!mapping->assoc_mapping) {
643                 mapping->assoc_mapping = buffer_mapping;
644         } else {
645                 BUG_ON(mapping->assoc_mapping != buffer_mapping);
646         }
647         if (!bh->b_assoc_map) {
648                 spin_lock(&buffer_mapping->private_lock);
649                 list_move_tail(&bh->b_assoc_buffers,
650                                 &mapping->private_list);
651                 bh->b_assoc_map = mapping;
652                 spin_unlock(&buffer_mapping->private_lock);
653         }
654 }
655 EXPORT_SYMBOL(mark_buffer_dirty_inode);
656
657 /*
658  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
659  * dirty.
660  *
661  * If warn is true, then emit a warning if the page is not uptodate and has
662  * not been truncated.
663  */
664 static void __set_page_dirty(struct page *page,
665                 struct address_space *mapping, int warn)
666 {
667         spin_lock_irq(&mapping->tree_lock);
668         if (page->mapping) {    /* Race with truncate? */
669                 WARN_ON_ONCE(warn && !PageUptodate(page));
670                 account_page_dirtied(page, mapping);
671                 radix_tree_tag_set(&mapping->page_tree,
672                                 page_index(page), PAGECACHE_TAG_DIRTY);
673         }
674         spin_unlock_irq(&mapping->tree_lock);
675         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
676 }
677
678 /*
679  * Add a page to the dirty page list.
680  *
681  * It is a sad fact of life that this function is called from several places
682  * deeply under spinlocking.  It may not sleep.
683  *
684  * If the page has buffers, the uptodate buffers are set dirty, to preserve
685  * dirty-state coherency between the page and the buffers.  It the page does
686  * not have buffers then when they are later attached they will all be set
687  * dirty.
688  *
689  * The buffers are dirtied before the page is dirtied.  There's a small race
690  * window in which a writepage caller may see the page cleanness but not the
691  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
692  * before the buffers, a concurrent writepage caller could clear the page dirty
693  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
694  * page on the dirty page list.
695  *
696  * We use private_lock to lock against try_to_free_buffers while using the
697  * page's buffer list.  Also use this to protect against clean buffers being
698  * added to the page after it was set dirty.
699  *
700  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
701  * address_space though.
702  */
703 int __set_page_dirty_buffers(struct page *page)
704 {
705         int newly_dirty;
706         struct address_space *mapping = page_mapping(page);
707
708         if (unlikely(!mapping))
709                 return !TestSetPageDirty(page);
710
711         spin_lock(&mapping->private_lock);
712         if (page_has_buffers(page)) {
713                 struct buffer_head *head = page_buffers(page);
714                 struct buffer_head *bh = head;
715
716                 do {
717                         set_buffer_dirty(bh);
718                         bh = bh->b_this_page;
719                 } while (bh != head);
720         }
721         newly_dirty = !TestSetPageDirty(page);
722         spin_unlock(&mapping->private_lock);
723
724         if (newly_dirty)
725                 __set_page_dirty(page, mapping, 1);
726         return newly_dirty;
727 }
728 EXPORT_SYMBOL(__set_page_dirty_buffers);
729
730 /*
731  * Write out and wait upon a list of buffers.
732  *
733  * We have conflicting pressures: we want to make sure that all
734  * initially dirty buffers get waited on, but that any subsequently
735  * dirtied buffers don't.  After all, we don't want fsync to last
736  * forever if somebody is actively writing to the file.
737  *
738  * Do this in two main stages: first we copy dirty buffers to a
739  * temporary inode list, queueing the writes as we go.  Then we clean
740  * up, waiting for those writes to complete.
741  * 
742  * During this second stage, any subsequent updates to the file may end
743  * up refiling the buffer on the original inode's dirty list again, so
744  * there is a chance we will end up with a buffer queued for write but
745  * not yet completed on that list.  So, as a final cleanup we go through
746  * the osync code to catch these locked, dirty buffers without requeuing
747  * any newly dirty buffers for write.
748  */
749 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
750 {
751         struct buffer_head *bh;
752         struct list_head tmp;
753         struct address_space *mapping, *prev_mapping = NULL;
754         int err = 0, err2;
755
756         INIT_LIST_HEAD(&tmp);
757
758         spin_lock(lock);
759         while (!list_empty(list)) {
760                 bh = BH_ENTRY(list->next);
761                 mapping = bh->b_assoc_map;
762                 __remove_assoc_queue(bh);
763                 /* Avoid race with mark_buffer_dirty_inode() which does
764                  * a lockless check and we rely on seeing the dirty bit */
765                 smp_mb();
766                 if (buffer_dirty(bh) || buffer_locked(bh)) {
767                         list_add(&bh->b_assoc_buffers, &tmp);
768                         bh->b_assoc_map = mapping;
769                         if (buffer_dirty(bh)) {
770                                 get_bh(bh);
771                                 spin_unlock(lock);
772                                 /*
773                                  * Ensure any pending I/O completes so that
774                                  * ll_rw_block() actually writes the current
775                                  * contents - it is a noop if I/O is still in
776                                  * flight on potentially older contents.
777                                  */
778                                 ll_rw_block(SWRITE_SYNC_PLUG, 1, &bh);
779
780                                 /*
781                                  * Kick off IO for the previous mapping. Note
782                                  * that we will not run the very last mapping,
783                                  * wait_on_buffer() will do that for us
784                                  * through sync_buffer().
785                                  */
786                                 if (prev_mapping && prev_mapping != mapping)
787                                         blk_run_address_space(prev_mapping);
788                                 prev_mapping = mapping;
789
790                                 brelse(bh);
791                                 spin_lock(lock);
792                         }
793                 }
794         }
795
796         while (!list_empty(&tmp)) {
797                 bh = BH_ENTRY(tmp.prev);
798                 get_bh(bh);
799                 mapping = bh->b_assoc_map;
800                 __remove_assoc_queue(bh);
801                 /* Avoid race with mark_buffer_dirty_inode() which does
802                  * a lockless check and we rely on seeing the dirty bit */
803                 smp_mb();
804                 if (buffer_dirty(bh)) {
805                         list_add(&bh->b_assoc_buffers,
806                                  &mapping->private_list);
807                         bh->b_assoc_map = mapping;
808                 }
809                 spin_unlock(lock);
810                 wait_on_buffer(bh);
811                 if (!buffer_uptodate(bh))
812                         err = -EIO;
813                 brelse(bh);
814                 spin_lock(lock);
815         }
816         
817         spin_unlock(lock);
818         err2 = osync_buffers_list(lock, list);
819         if (err)
820                 return err;
821         else
822                 return err2;
823 }
824
825 /*
826  * Invalidate any and all dirty buffers on a given inode.  We are
827  * probably unmounting the fs, but that doesn't mean we have already
828  * done a sync().  Just drop the buffers from the inode list.
829  *
830  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
831  * assumes that all the buffers are against the blockdev.  Not true
832  * for reiserfs.
833  */
834 void invalidate_inode_buffers(struct inode *inode)
835 {
836         if (inode_has_buffers(inode)) {
837                 struct address_space *mapping = &inode->i_data;
838                 struct list_head *list = &mapping->private_list;
839                 struct address_space *buffer_mapping = mapping->assoc_mapping;
840
841                 spin_lock(&buffer_mapping->private_lock);
842                 while (!list_empty(list))
843                         __remove_assoc_queue(BH_ENTRY(list->next));
844                 spin_unlock(&buffer_mapping->private_lock);
845         }
846 }
847 EXPORT_SYMBOL(invalidate_inode_buffers);
848
849 /*
850  * Remove any clean buffers from the inode's buffer list.  This is called
851  * when we're trying to free the inode itself.  Those buffers can pin it.
852  *
853  * Returns true if all buffers were removed.
854  */
855 int remove_inode_buffers(struct inode *inode)
856 {
857         int ret = 1;
858
859         if (inode_has_buffers(inode)) {
860                 struct address_space *mapping = &inode->i_data;
861                 struct list_head *list = &mapping->private_list;
862                 struct address_space *buffer_mapping = mapping->assoc_mapping;
863
864                 spin_lock(&buffer_mapping->private_lock);
865                 while (!list_empty(list)) {
866                         struct buffer_head *bh = BH_ENTRY(list->next);
867                         if (buffer_dirty(bh)) {
868                                 ret = 0;
869                                 break;
870                         }
871                         __remove_assoc_queue(bh);
872                 }
873                 spin_unlock(&buffer_mapping->private_lock);
874         }
875         return ret;
876 }
877
878 /*
879  * Create the appropriate buffers when given a page for data area and
880  * the size of each buffer.. Use the bh->b_this_page linked list to
881  * follow the buffers created.  Return NULL if unable to create more
882  * buffers.
883  *
884  * The retry flag is used to differentiate async IO (paging, swapping)
885  * which may not fail from ordinary buffer allocations.
886  */
887 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
888                 int retry)
889 {
890         struct buffer_head *bh, *head;
891         long offset;
892
893 try_again:
894         head = NULL;
895         offset = PAGE_SIZE;
896         while ((offset -= size) >= 0) {
897                 bh = alloc_buffer_head(GFP_NOFS);
898                 if (!bh)
899                         goto no_grow;
900
901                 bh->b_bdev = NULL;
902                 bh->b_this_page = head;
903                 bh->b_blocknr = -1;
904                 head = bh;
905
906                 bh->b_state = 0;
907                 atomic_set(&bh->b_count, 0);
908                 bh->b_private = NULL;
909                 bh->b_size = size;
910
911                 /* Link the buffer to its page */
912                 set_bh_page(bh, page, offset);
913
914                 init_buffer(bh, NULL, NULL);
915         }
916         return head;
917 /*
918  * In case anything failed, we just free everything we got.
919  */
920 no_grow:
921         if (head) {
922                 do {
923                         bh = head;
924                         head = head->b_this_page;
925                         free_buffer_head(bh);
926                 } while (head);
927         }
928
929         /*
930          * Return failure for non-async IO requests.  Async IO requests
931          * are not allowed to fail, so we have to wait until buffer heads
932          * become available.  But we don't want tasks sleeping with 
933          * partially complete buffers, so all were released above.
934          */
935         if (!retry)
936                 return NULL;
937
938         /* We're _really_ low on memory. Now we just
939          * wait for old buffer heads to become free due to
940          * finishing IO.  Since this is an async request and
941          * the reserve list is empty, we're sure there are 
942          * async buffer heads in use.
943          */
944         free_more_memory();
945         goto try_again;
946 }
947 EXPORT_SYMBOL_GPL(alloc_page_buffers);
948
949 static inline void
950 link_dev_buffers(struct page *page, struct buffer_head *head)
951 {
952         struct buffer_head *bh, *tail;
953
954         bh = head;
955         do {
956                 tail = bh;
957                 bh = bh->b_this_page;
958         } while (bh);
959         tail->b_this_page = head;
960         attach_page_buffers(page, head);
961 }
962
963 /*
964  * Initialise the state of a blockdev page's buffers.
965  */ 
966 static void
967 init_page_buffers(struct page *page, struct block_device *bdev,
968                         sector_t block, int size)
969 {
970         struct buffer_head *head = page_buffers(page);
971         struct buffer_head *bh = head;
972         int uptodate = PageUptodate(page);
973
974         do {
975                 if (!buffer_mapped(bh)) {
976                         init_buffer(bh, NULL, NULL);
977                         bh->b_bdev = bdev;
978                         bh->b_blocknr = block;
979                         if (uptodate)
980                                 set_buffer_uptodate(bh);
981                         set_buffer_mapped(bh);
982                 }
983                 block++;
984                 bh = bh->b_this_page;
985         } while (bh != head);
986 }
987
988 /*
989  * Create the page-cache page that contains the requested block.
990  *
991  * This is user purely for blockdev mappings.
992  */
993 static struct page *
994 grow_dev_page(struct block_device *bdev, sector_t block,
995                 pgoff_t index, int size)
996 {
997         struct inode *inode = bdev->bd_inode;
998         struct page *page;
999         struct buffer_head *bh;
1000
1001         page = find_or_create_page(inode->i_mapping, index,
1002                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
1003         if (!page)
1004                 return NULL;
1005
1006         BUG_ON(!PageLocked(page));
1007
1008         if (page_has_buffers(page)) {
1009                 bh = page_buffers(page);
1010                 if (bh->b_size == size) {
1011                         init_page_buffers(page, bdev, block, size);
1012                         return page;
1013                 }
1014                 if (!try_to_free_buffers(page))
1015                         goto failed;
1016         }
1017
1018         /*
1019          * Allocate some buffers for this page
1020          */
1021         bh = alloc_page_buffers(page, size, 0);
1022         if (!bh)
1023                 goto failed;
1024
1025         /*
1026          * Link the page to the buffers and initialise them.  Take the
1027          * lock to be atomic wrt __find_get_block(), which does not
1028          * run under the page lock.
1029          */
1030         spin_lock(&inode->i_mapping->private_lock);
1031         link_dev_buffers(page, bh);
1032         init_page_buffers(page, bdev, block, size);
1033         spin_unlock(&inode->i_mapping->private_lock);
1034         return page;
1035
1036 failed:
1037         BUG();
1038         unlock_page(page);
1039         page_cache_release(page);
1040         return NULL;
1041 }
1042
1043 /*
1044  * Create buffers for the specified block device block's page.  If
1045  * that page was dirty, the buffers are set dirty also.
1046  */
1047 static int
1048 grow_buffers(struct block_device *bdev, sector_t block, int size)
1049 {
1050         struct page *page;
1051         pgoff_t index;
1052         int sizebits;
1053
1054         sizebits = -1;
1055         do {
1056                 sizebits++;
1057         } while ((size << sizebits) < PAGE_SIZE);
1058
1059         index = block >> sizebits;
1060
1061         /*
1062          * Check for a block which wants to lie outside our maximum possible
1063          * pagecache index.  (this comparison is done using sector_t types).
1064          */
1065         if (unlikely(index != block >> sizebits)) {
1066                 char b[BDEVNAME_SIZE];
1067
1068                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1069                         "device %s\n",
1070                         __func__, (unsigned long long)block,
1071                         bdevname(bdev, b));
1072                 return -EIO;
1073         }
1074         block = index << sizebits;
1075         /* Create a page with the proper size buffers.. */
1076         page = grow_dev_page(bdev, block, index, size);
1077         if (!page)
1078                 return 0;
1079         unlock_page(page);
1080         page_cache_release(page);
1081         return 1;
1082 }
1083
1084 static struct buffer_head *
1085 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1086 {
1087         /* Size must be multiple of hard sectorsize */
1088         if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1089                         (size < 512 || size > PAGE_SIZE))) {
1090                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1091                                         size);
1092                 printk(KERN_ERR "hardsect size: %d\n",
1093                                         bdev_hardsect_size(bdev));
1094
1095                 dump_stack();
1096                 return NULL;
1097         }
1098
1099         for (;;) {
1100                 struct buffer_head * bh;
1101                 int ret;
1102
1103                 bh = __find_get_block(bdev, block, size);
1104                 if (bh)
1105                         return bh;
1106
1107                 ret = grow_buffers(bdev, block, size);
1108                 if (ret < 0)
1109                         return NULL;
1110                 if (ret == 0)
1111                         free_more_memory();
1112         }
1113 }
1114
1115 /*
1116  * The relationship between dirty buffers and dirty pages:
1117  *
1118  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1119  * the page is tagged dirty in its radix tree.
1120  *
1121  * At all times, the dirtiness of the buffers represents the dirtiness of
1122  * subsections of the page.  If the page has buffers, the page dirty bit is
1123  * merely a hint about the true dirty state.
1124  *
1125  * When a page is set dirty in its entirety, all its buffers are marked dirty
1126  * (if the page has buffers).
1127  *
1128  * When a buffer is marked dirty, its page is dirtied, but the page's other
1129  * buffers are not.
1130  *
1131  * Also.  When blockdev buffers are explicitly read with bread(), they
1132  * individually become uptodate.  But their backing page remains not
1133  * uptodate - even if all of its buffers are uptodate.  A subsequent
1134  * block_read_full_page() against that page will discover all the uptodate
1135  * buffers, will set the page uptodate and will perform no I/O.
1136  */
1137
1138 /**
1139  * mark_buffer_dirty - mark a buffer_head as needing writeout
1140  * @bh: the buffer_head to mark dirty
1141  *
1142  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1143  * backing page dirty, then tag the page as dirty in its address_space's radix
1144  * tree and then attach the address_space's inode to its superblock's dirty
1145  * inode list.
1146  *
1147  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1148  * mapping->tree_lock and the global inode_lock.
1149  */
1150 void mark_buffer_dirty(struct buffer_head *bh)
1151 {
1152         WARN_ON_ONCE(!buffer_uptodate(bh));
1153
1154         /*
1155          * Very *carefully* optimize the it-is-already-dirty case.
1156          *
1157          * Don't let the final "is it dirty" escape to before we
1158          * perhaps modified the buffer.
1159          */
1160         if (buffer_dirty(bh)) {
1161                 smp_mb();
1162                 if (buffer_dirty(bh))
1163                         return;
1164         }
1165
1166         if (!test_set_buffer_dirty(bh)) {
1167                 struct page *page = bh->b_page;
1168                 if (!TestSetPageDirty(page))
1169                         __set_page_dirty(page, page_mapping(page), 0);
1170         }
1171 }
1172
1173 /*
1174  * Decrement a buffer_head's reference count.  If all buffers against a page
1175  * have zero reference count, are clean and unlocked, and if the page is clean
1176  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1177  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1178  * a page but it ends up not being freed, and buffers may later be reattached).
1179  */
1180 void __brelse(struct buffer_head * buf)
1181 {
1182         if (atomic_read(&buf->b_count)) {
1183                 put_bh(buf);
1184                 return;
1185         }
1186         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1187 }
1188
1189 /*
1190  * bforget() is like brelse(), except it discards any
1191  * potentially dirty data.
1192  */
1193 void __bforget(struct buffer_head *bh)
1194 {
1195         clear_buffer_dirty(bh);
1196         if (bh->b_assoc_map) {
1197                 struct address_space *buffer_mapping = bh->b_page->mapping;
1198
1199                 spin_lock(&buffer_mapping->private_lock);
1200                 list_del_init(&bh->b_assoc_buffers);
1201                 bh->b_assoc_map = NULL;
1202                 spin_unlock(&buffer_mapping->private_lock);
1203         }
1204         __brelse(bh);
1205 }
1206
1207 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1208 {
1209         lock_buffer(bh);
1210         if (buffer_uptodate(bh)) {
1211                 unlock_buffer(bh);
1212                 return bh;
1213         } else {
1214                 get_bh(bh);
1215                 bh->b_end_io = end_buffer_read_sync;
1216                 submit_bh(READ, bh);
1217                 wait_on_buffer(bh);
1218                 if (buffer_uptodate(bh))
1219                         return bh;
1220         }
1221         brelse(bh);
1222         return NULL;
1223 }
1224
1225 /*
1226  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1227  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1228  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1229  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1230  * CPU's LRUs at the same time.
1231  *
1232  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1233  * sb_find_get_block().
1234  *
1235  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1236  * a local interrupt disable for that.
1237  */
1238
1239 #define BH_LRU_SIZE     8
1240
1241 struct bh_lru {
1242         struct buffer_head *bhs[BH_LRU_SIZE];
1243 };
1244
1245 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1246
1247 #ifdef CONFIG_SMP
1248 #define bh_lru_lock()   local_irq_disable()
1249 #define bh_lru_unlock() local_irq_enable()
1250 #else
1251 #define bh_lru_lock()   preempt_disable()
1252 #define bh_lru_unlock() preempt_enable()
1253 #endif
1254
1255 static inline void check_irqs_on(void)
1256 {
1257 #ifdef irqs_disabled
1258         BUG_ON(irqs_disabled());
1259 #endif
1260 }
1261
1262 /*
1263  * The LRU management algorithm is dopey-but-simple.  Sorry.
1264  */
1265 static void bh_lru_install(struct buffer_head *bh)
1266 {
1267         struct buffer_head *evictee = NULL;
1268         struct bh_lru *lru;
1269
1270         check_irqs_on();
1271         bh_lru_lock();
1272         lru = &__get_cpu_var(bh_lrus);
1273         if (lru->bhs[0] != bh) {
1274                 struct buffer_head *bhs[BH_LRU_SIZE];
1275                 int in;
1276                 int out = 0;
1277
1278                 get_bh(bh);
1279                 bhs[out++] = bh;
1280                 for (in = 0; in < BH_LRU_SIZE; in++) {
1281                         struct buffer_head *bh2 = lru->bhs[in];
1282
1283                         if (bh2 == bh) {
1284                                 __brelse(bh2);
1285                         } else {
1286                                 if (out >= BH_LRU_SIZE) {
1287                                         BUG_ON(evictee != NULL);
1288                                         evictee = bh2;
1289                                 } else {
1290                                         bhs[out++] = bh2;
1291                                 }
1292                         }
1293                 }
1294                 while (out < BH_LRU_SIZE)
1295                         bhs[out++] = NULL;
1296                 memcpy(lru->bhs, bhs, sizeof(bhs));
1297         }
1298         bh_lru_unlock();
1299
1300         if (evictee)
1301                 __brelse(evictee);
1302 }
1303
1304 /*
1305  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1306  */
1307 static struct buffer_head *
1308 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1309 {
1310         struct buffer_head *ret = NULL;
1311         struct bh_lru *lru;
1312         unsigned int i;
1313
1314         check_irqs_on();
1315         bh_lru_lock();
1316         lru = &__get_cpu_var(bh_lrus);
1317         for (i = 0; i < BH_LRU_SIZE; i++) {
1318                 struct buffer_head *bh = lru->bhs[i];
1319
1320                 if (bh && bh->b_bdev == bdev &&
1321                                 bh->b_blocknr == block && bh->b_size == size) {
1322                         if (i) {
1323                                 while (i) {
1324                                         lru->bhs[i] = lru->bhs[i - 1];
1325                                         i--;
1326                                 }
1327                                 lru->bhs[0] = bh;
1328                         }
1329                         get_bh(bh);
1330                         ret = bh;
1331                         break;
1332                 }
1333         }
1334         bh_lru_unlock();
1335         return ret;
1336 }
1337
1338 /*
1339  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1340  * it in the LRU and mark it as accessed.  If it is not present then return
1341  * NULL
1342  */
1343 struct buffer_head *
1344 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1345 {
1346         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1347
1348         if (bh == NULL) {
1349                 bh = __find_get_block_slow(bdev, block);
1350                 if (bh)
1351                         bh_lru_install(bh);
1352         }
1353         if (bh)
1354                 touch_buffer(bh);
1355         return bh;
1356 }
1357 EXPORT_SYMBOL(__find_get_block);
1358
1359 /*
1360  * __getblk will locate (and, if necessary, create) the buffer_head
1361  * which corresponds to the passed block_device, block and size. The
1362  * returned buffer has its reference count incremented.
1363  *
1364  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1365  * illegal block number, __getblk() will happily return a buffer_head
1366  * which represents the non-existent block.  Very weird.
1367  *
1368  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1369  * attempt is failing.  FIXME, perhaps?
1370  */
1371 struct buffer_head *
1372 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1373 {
1374         struct buffer_head *bh = __find_get_block(bdev, block, size);
1375
1376         might_sleep();
1377         if (bh == NULL)
1378                 bh = __getblk_slow(bdev, block, size);
1379         return bh;
1380 }
1381 EXPORT_SYMBOL(__getblk);
1382
1383 /*
1384  * Do async read-ahead on a buffer..
1385  */
1386 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1387 {
1388         struct buffer_head *bh = __getblk(bdev, block, size);
1389         if (likely(bh)) {
1390                 ll_rw_block(READA, 1, &bh);
1391                 brelse(bh);
1392         }
1393 }
1394 EXPORT_SYMBOL(__breadahead);
1395
1396 /**
1397  *  __bread() - reads a specified block and returns the bh
1398  *  @bdev: the block_device to read from
1399  *  @block: number of block
1400  *  @size: size (in bytes) to read
1401  * 
1402  *  Reads a specified block, and returns buffer head that contains it.
1403  *  It returns NULL if the block was unreadable.
1404  */
1405 struct buffer_head *
1406 __bread(struct block_device *bdev, sector_t block, unsigned size)
1407 {
1408         struct buffer_head *bh = __getblk(bdev, block, size);
1409
1410         if (likely(bh) && !buffer_uptodate(bh))
1411                 bh = __bread_slow(bh);
1412         return bh;
1413 }
1414 EXPORT_SYMBOL(__bread);
1415
1416 /*
1417  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1418  * This doesn't race because it runs in each cpu either in irq
1419  * or with preempt disabled.
1420  */
1421 static void invalidate_bh_lru(void *arg)
1422 {
1423         struct bh_lru *b = &get_cpu_var(bh_lrus);
1424         int i;
1425
1426         for (i = 0; i < BH_LRU_SIZE; i++) {
1427                 brelse(b->bhs[i]);
1428                 b->bhs[i] = NULL;
1429         }
1430         put_cpu_var(bh_lrus);
1431 }
1432         
1433 void invalidate_bh_lrus(void)
1434 {
1435         on_each_cpu(invalidate_bh_lru, NULL, 1);
1436 }
1437 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1438
1439 void set_bh_page(struct buffer_head *bh,
1440                 struct page *page, unsigned long offset)
1441 {
1442         bh->b_page = page;
1443         BUG_ON(offset >= PAGE_SIZE);
1444         if (PageHighMem(page))
1445                 /*
1446                  * This catches illegal uses and preserves the offset:
1447                  */
1448                 bh->b_data = (char *)(0 + offset);
1449         else
1450                 bh->b_data = page_address(page) + offset;
1451 }
1452 EXPORT_SYMBOL(set_bh_page);
1453
1454 /*
1455  * Called when truncating a buffer on a page completely.
1456  */
1457 static void discard_buffer(struct buffer_head * bh)
1458 {
1459         lock_buffer(bh);
1460         clear_buffer_dirty(bh);
1461         bh->b_bdev = NULL;
1462         clear_buffer_mapped(bh);
1463         clear_buffer_req(bh);
1464         clear_buffer_new(bh);
1465         clear_buffer_delay(bh);
1466         clear_buffer_unwritten(bh);
1467         unlock_buffer(bh);
1468 }
1469
1470 /**
1471  * block_invalidatepage - invalidate part of all of a buffer-backed page
1472  *
1473  * @page: the page which is affected
1474  * @offset: the index of the truncation point
1475  *
1476  * block_invalidatepage() is called when all or part of the page has become
1477  * invalidatedby a truncate operation.
1478  *
1479  * block_invalidatepage() does not have to release all buffers, but it must
1480  * ensure that no dirty buffer is left outside @offset and that no I/O
1481  * is underway against any of the blocks which are outside the truncation
1482  * point.  Because the caller is about to free (and possibly reuse) those
1483  * blocks on-disk.
1484  */
1485 void block_invalidatepage(struct page *page, unsigned long offset)
1486 {
1487         struct buffer_head *head, *bh, *next;
1488         unsigned int curr_off = 0;
1489
1490         BUG_ON(!PageLocked(page));
1491         if (!page_has_buffers(page))
1492                 goto out;
1493
1494         head = page_buffers(page);
1495         bh = head;
1496         do {
1497                 unsigned int next_off = curr_off + bh->b_size;
1498                 next = bh->b_this_page;
1499
1500                 /*
1501                  * is this block fully invalidated?
1502                  */
1503                 if (offset <= curr_off)
1504                         discard_buffer(bh);
1505                 curr_off = next_off;
1506                 bh = next;
1507         } while (bh != head);
1508
1509         /*
1510          * We release buffers only if the entire page is being invalidated.
1511          * The get_block cached value has been unconditionally invalidated,
1512          * so real IO is not possible anymore.
1513          */
1514         if (offset == 0)
1515                 try_to_release_page(page, 0);
1516 out:
1517         return;
1518 }
1519 EXPORT_SYMBOL(block_invalidatepage);
1520
1521 /*
1522  * We attach and possibly dirty the buffers atomically wrt
1523  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1524  * is already excluded via the page lock.
1525  */
1526 void create_empty_buffers(struct page *page,
1527                         unsigned long blocksize, unsigned long b_state)
1528 {
1529         struct buffer_head *bh, *head, *tail;
1530
1531         head = alloc_page_buffers(page, blocksize, 1);
1532         bh = head;
1533         do {
1534                 bh->b_state |= b_state;
1535                 tail = bh;
1536                 bh = bh->b_this_page;
1537         } while (bh);
1538         tail->b_this_page = head;
1539
1540         spin_lock(&page->mapping->private_lock);
1541         if (PageUptodate(page) || PageDirty(page)) {
1542                 bh = head;
1543                 do {
1544                         if (PageDirty(page))
1545                                 set_buffer_dirty(bh);
1546                         if (PageUptodate(page))
1547                                 set_buffer_uptodate(bh);
1548                         bh = bh->b_this_page;
1549                 } while (bh != head);
1550         }
1551         attach_page_buffers(page, head);
1552         spin_unlock(&page->mapping->private_lock);
1553 }
1554 EXPORT_SYMBOL(create_empty_buffers);
1555
1556 /*
1557  * We are taking a block for data and we don't want any output from any
1558  * buffer-cache aliases starting from return from that function and
1559  * until the moment when something will explicitly mark the buffer
1560  * dirty (hopefully that will not happen until we will free that block ;-)
1561  * We don't even need to mark it not-uptodate - nobody can expect
1562  * anything from a newly allocated buffer anyway. We used to used
1563  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1564  * don't want to mark the alias unmapped, for example - it would confuse
1565  * anyone who might pick it with bread() afterwards...
1566  *
1567  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1568  * be writeout I/O going on against recently-freed buffers.  We don't
1569  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1570  * only if we really need to.  That happens here.
1571  */
1572 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1573 {
1574         struct buffer_head *old_bh;
1575
1576         might_sleep();
1577
1578         old_bh = __find_get_block_slow(bdev, block);
1579         if (old_bh) {
1580                 clear_buffer_dirty(old_bh);
1581                 wait_on_buffer(old_bh);
1582                 clear_buffer_req(old_bh);
1583                 __brelse(old_bh);
1584         }
1585 }
1586 EXPORT_SYMBOL(unmap_underlying_metadata);
1587
1588 /*
1589  * NOTE! All mapped/uptodate combinations are valid:
1590  *
1591  *      Mapped  Uptodate        Meaning
1592  *
1593  *      No      No              "unknown" - must do get_block()
1594  *      No      Yes             "hole" - zero-filled
1595  *      Yes     No              "allocated" - allocated on disk, not read in
1596  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1597  *
1598  * "Dirty" is valid only with the last case (mapped+uptodate).
1599  */
1600
1601 /*
1602  * While block_write_full_page is writing back the dirty buffers under
1603  * the page lock, whoever dirtied the buffers may decide to clean them
1604  * again at any time.  We handle that by only looking at the buffer
1605  * state inside lock_buffer().
1606  *
1607  * If block_write_full_page() is called for regular writeback
1608  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1609  * locked buffer.   This only can happen if someone has written the buffer
1610  * directly, with submit_bh().  At the address_space level PageWriteback
1611  * prevents this contention from occurring.
1612  *
1613  * If block_write_full_page() is called with wbc->sync_mode ==
1614  * WB_SYNC_ALL, the writes are posted using WRITE_SYNC_PLUG; this
1615  * causes the writes to be flagged as synchronous writes, but the
1616  * block device queue will NOT be unplugged, since usually many pages
1617  * will be pushed to the out before the higher-level caller actually
1618  * waits for the writes to be completed.  The various wait functions,
1619  * such as wait_on_writeback_range() will ultimately call sync_page()
1620  * which will ultimately call blk_run_backing_dev(), which will end up
1621  * unplugging the device queue.
1622  */
1623 static int __block_write_full_page(struct inode *inode, struct page *page,
1624                         get_block_t *get_block, struct writeback_control *wbc,
1625                         bh_end_io_t *handler)
1626 {
1627         int err;
1628         sector_t block;
1629         sector_t last_block;
1630         struct buffer_head *bh, *head;
1631         const unsigned blocksize = 1 << inode->i_blkbits;
1632         int nr_underway = 0;
1633         int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1634                         WRITE_SYNC_PLUG : WRITE);
1635
1636         BUG_ON(!PageLocked(page));
1637
1638         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1639
1640         if (!page_has_buffers(page)) {
1641                 create_empty_buffers(page, blocksize,
1642                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1643         }
1644
1645         /*
1646          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1647          * here, and the (potentially unmapped) buffers may become dirty at
1648          * any time.  If a buffer becomes dirty here after we've inspected it
1649          * then we just miss that fact, and the page stays dirty.
1650          *
1651          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1652          * handle that here by just cleaning them.
1653          */
1654
1655         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1656         head = page_buffers(page);
1657         bh = head;
1658
1659         /*
1660          * Get all the dirty buffers mapped to disk addresses and
1661          * handle any aliases from the underlying blockdev's mapping.
1662          */
1663         do {
1664                 if (block > last_block) {
1665                         /*
1666                          * mapped buffers outside i_size will occur, because
1667                          * this page can be outside i_size when there is a
1668                          * truncate in progress.
1669                          */
1670                         /*
1671                          * The buffer was zeroed by block_write_full_page()
1672                          */
1673                         clear_buffer_dirty(bh);
1674                         set_buffer_uptodate(bh);
1675                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1676                            buffer_dirty(bh)) {
1677                         WARN_ON(bh->b_size != blocksize);
1678                         err = get_block(inode, block, bh, 1);
1679                         if (err)
1680                                 goto recover;
1681                         clear_buffer_delay(bh);
1682                         if (buffer_new(bh)) {
1683                                 /* blockdev mappings never come here */
1684                                 clear_buffer_new(bh);
1685                                 unmap_underlying_metadata(bh->b_bdev,
1686                                                         bh->b_blocknr);
1687                         }
1688                 }
1689                 bh = bh->b_this_page;
1690                 block++;
1691         } while (bh != head);
1692
1693         do {
1694                 if (!buffer_mapped(bh))
1695                         continue;
1696                 /*
1697                  * If it's a fully non-blocking write attempt and we cannot
1698                  * lock the buffer then redirty the page.  Note that this can
1699                  * potentially cause a busy-wait loop from pdflush and kswapd
1700                  * activity, but those code paths have their own higher-level
1701                  * throttling.
1702                  */
1703                 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1704                         lock_buffer(bh);
1705                 } else if (!trylock_buffer(bh)) {
1706                         redirty_page_for_writepage(wbc, page);
1707                         continue;
1708                 }
1709                 if (test_clear_buffer_dirty(bh)) {
1710                         mark_buffer_async_write_endio(bh, handler);
1711                 } else {
1712                         unlock_buffer(bh);
1713                 }
1714         } while ((bh = bh->b_this_page) != head);
1715
1716         /*
1717          * The page and its buffers are protected by PageWriteback(), so we can
1718          * drop the bh refcounts early.
1719          */
1720         BUG_ON(PageWriteback(page));
1721         set_page_writeback(page);
1722
1723         do {
1724                 struct buffer_head *next = bh->b_this_page;
1725                 if (buffer_async_write(bh)) {
1726                         submit_bh(write_op, bh);
1727                         nr_underway++;
1728                 }
1729                 bh = next;
1730         } while (bh != head);
1731         unlock_page(page);
1732
1733         err = 0;
1734 done:
1735         if (nr_underway == 0) {
1736                 /*
1737                  * The page was marked dirty, but the buffers were
1738                  * clean.  Someone wrote them back by hand with
1739                  * ll_rw_block/submit_bh.  A rare case.
1740                  */
1741                 end_page_writeback(page);
1742
1743                 /*
1744                  * The page and buffer_heads can be released at any time from
1745                  * here on.
1746                  */
1747         }
1748         return err;
1749
1750 recover:
1751         /*
1752          * ENOSPC, or some other error.  We may already have added some
1753          * blocks to the file, so we need to write these out to avoid
1754          * exposing stale data.
1755          * The page is currently locked and not marked for writeback
1756          */
1757         bh = head;
1758         /* Recovery: lock and submit the mapped buffers */
1759         do {
1760                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1761                     !buffer_delay(bh)) {
1762                         lock_buffer(bh);
1763                         mark_buffer_async_write_endio(bh, handler);
1764                 } else {
1765                         /*
1766                          * The buffer may have been set dirty during
1767                          * attachment to a dirty page.
1768                          */
1769                         clear_buffer_dirty(bh);
1770                 }
1771         } while ((bh = bh->b_this_page) != head);
1772         SetPageError(page);
1773         BUG_ON(PageWriteback(page));
1774         mapping_set_error(page->mapping, err);
1775         set_page_writeback(page);
1776         do {
1777                 struct buffer_head *next = bh->b_this_page;
1778                 if (buffer_async_write(bh)) {
1779                         clear_buffer_dirty(bh);
1780                         submit_bh(write_op, bh);
1781                         nr_underway++;
1782                 }
1783                 bh = next;
1784         } while (bh != head);
1785         unlock_page(page);
1786         goto done;
1787 }
1788
1789 /*
1790  * If a page has any new buffers, zero them out here, and mark them uptodate
1791  * and dirty so they'll be written out (in order to prevent uninitialised
1792  * block data from leaking). And clear the new bit.
1793  */
1794 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1795 {
1796         unsigned int block_start, block_end;
1797         struct buffer_head *head, *bh;
1798
1799         BUG_ON(!PageLocked(page));
1800         if (!page_has_buffers(page))
1801                 return;
1802
1803         bh = head = page_buffers(page);
1804         block_start = 0;
1805         do {
1806                 block_end = block_start + bh->b_size;
1807
1808                 if (buffer_new(bh)) {
1809                         if (block_end > from && block_start < to) {
1810                                 if (!PageUptodate(page)) {
1811                                         unsigned start, size;
1812
1813                                         start = max(from, block_start);
1814                                         size = min(to, block_end) - start;
1815
1816                                         zero_user(page, start, size);
1817                                         set_buffer_uptodate(bh);
1818                                 }
1819
1820                                 clear_buffer_new(bh);
1821                                 mark_buffer_dirty(bh);
1822                         }
1823                 }
1824
1825                 block_start = block_end;
1826                 bh = bh->b_this_page;
1827         } while (bh != head);
1828 }
1829 EXPORT_SYMBOL(page_zero_new_buffers);
1830
1831 static int __block_prepare_write(struct inode *inode, struct page *page,
1832                 unsigned from, unsigned to, get_block_t *get_block)
1833 {
1834         unsigned block_start, block_end;
1835         sector_t block;
1836         int err = 0;
1837         unsigned blocksize, bbits;
1838         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1839
1840         BUG_ON(!PageLocked(page));
1841         BUG_ON(from > PAGE_CACHE_SIZE);
1842         BUG_ON(to > PAGE_CACHE_SIZE);
1843         BUG_ON(from > to);
1844
1845         blocksize = 1 << inode->i_blkbits;
1846         if (!page_has_buffers(page))
1847                 create_empty_buffers(page, blocksize, 0);
1848         head = page_buffers(page);
1849
1850         bbits = inode->i_blkbits;
1851         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1852
1853         for(bh = head, block_start = 0; bh != head || !block_start;
1854             block++, block_start=block_end, bh = bh->b_this_page) {
1855                 block_end = block_start + blocksize;
1856                 if (block_end <= from || block_start >= to) {
1857                         if (PageUptodate(page)) {
1858                                 if (!buffer_uptodate(bh))
1859                                         set_buffer_uptodate(bh);
1860                         }
1861                         continue;
1862                 }
1863                 if (buffer_new(bh))
1864                         clear_buffer_new(bh);
1865                 if (!buffer_mapped(bh)) {
1866                         WARN_ON(bh->b_size != blocksize);
1867                         err = get_block(inode, block, bh, 1);
1868                         if (err)
1869                                 break;
1870                         if (buffer_new(bh)) {
1871                                 unmap_underlying_metadata(bh->b_bdev,
1872                                                         bh->b_blocknr);
1873                                 if (PageUptodate(page)) {
1874                                         clear_buffer_new(bh);
1875                                         set_buffer_uptodate(bh);
1876                                         mark_buffer_dirty(bh);
1877                                         continue;
1878                                 }
1879                                 if (block_end > to || block_start < from)
1880                                         zero_user_segments(page,
1881                                                 to, block_end,
1882                                                 block_start, from);
1883                                 continue;
1884                         }
1885                 }
1886                 if (PageUptodate(page)) {
1887                         if (!buffer_uptodate(bh))
1888                                 set_buffer_uptodate(bh);
1889                         continue; 
1890                 }
1891                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1892                     !buffer_unwritten(bh) &&
1893                      (block_start < from || block_end > to)) {
1894                         ll_rw_block(READ, 1, &bh);
1895                         *wait_bh++=bh;
1896                 }
1897         }
1898         /*
1899          * If we issued read requests - let them complete.
1900          */
1901         while(wait_bh > wait) {
1902                 wait_on_buffer(*--wait_bh);
1903                 if (!buffer_uptodate(*wait_bh))
1904                         err = -EIO;
1905         }
1906         if (unlikely(err))
1907                 page_zero_new_buffers(page, from, to);
1908         return err;
1909 }
1910
1911 static int __block_commit_write(struct inode *inode, struct page *page,
1912                 unsigned from, unsigned to)
1913 {
1914         unsigned block_start, block_end;
1915         int partial = 0;
1916         unsigned blocksize;
1917         struct buffer_head *bh, *head;
1918
1919         blocksize = 1 << inode->i_blkbits;
1920
1921         for(bh = head = page_buffers(page), block_start = 0;
1922             bh != head || !block_start;
1923             block_start=block_end, bh = bh->b_this_page) {
1924                 block_end = block_start + blocksize;
1925                 if (block_end <= from || block_start >= to) {
1926                         if (!buffer_uptodate(bh))
1927                                 partial = 1;
1928                 } else {
1929                         set_buffer_uptodate(bh);
1930                         mark_buffer_dirty(bh);
1931                 }
1932                 clear_buffer_new(bh);
1933         }
1934
1935         /*
1936          * If this is a partial write which happened to make all buffers
1937          * uptodate then we can optimize away a bogus readpage() for
1938          * the next read(). Here we 'discover' whether the page went
1939          * uptodate as a result of this (potentially partial) write.
1940          */
1941         if (!partial)
1942                 SetPageUptodate(page);
1943         return 0;
1944 }
1945
1946 /*
1947  * block_write_begin takes care of the basic task of block allocation and
1948  * bringing partial write blocks uptodate first.
1949  *
1950  * If *pagep is not NULL, then block_write_begin uses the locked page
1951  * at *pagep rather than allocating its own. In this case, the page will
1952  * not be unlocked or deallocated on failure.
1953  */
1954 int block_write_begin(struct file *file, struct address_space *mapping,
1955                         loff_t pos, unsigned len, unsigned flags,
1956                         struct page **pagep, void **fsdata,
1957                         get_block_t *get_block)
1958 {
1959         struct inode *inode = mapping->host;
1960         int status = 0;
1961         struct page *page;
1962         pgoff_t index;
1963         unsigned start, end;
1964         int ownpage = 0;
1965
1966         index = pos >> PAGE_CACHE_SHIFT;
1967         start = pos & (PAGE_CACHE_SIZE - 1);
1968         end = start + len;
1969
1970         page = *pagep;
1971         if (page == NULL) {
1972                 ownpage = 1;
1973                 page = grab_cache_page_write_begin(mapping, index, flags);
1974                 if (!page) {
1975                         status = -ENOMEM;
1976                         goto out;
1977                 }
1978                 *pagep = page;
1979         } else
1980                 BUG_ON(!PageLocked(page));
1981
1982         status = __block_prepare_write(inode, page, start, end, get_block);
1983         if (unlikely(status)) {
1984                 ClearPageUptodate(page);
1985
1986                 if (ownpage) {
1987                         unlock_page(page);
1988                         page_cache_release(page);
1989                         *pagep = NULL;
1990
1991                         /*
1992                          * prepare_write() may have instantiated a few blocks
1993                          * outside i_size.  Trim these off again. Don't need
1994                          * i_size_read because we hold i_mutex.
1995                          */
1996                         if (pos + len > inode->i_size)
1997                                 vmtruncate(inode, inode->i_size);
1998                 }
1999         }
2000
2001 out:
2002         return status;
2003 }
2004 EXPORT_SYMBOL(block_write_begin);
2005
2006 int block_write_end(struct file *file, struct address_space *mapping,
2007                         loff_t pos, unsigned len, unsigned copied,
2008                         struct page *page, void *fsdata)
2009 {
2010         struct inode *inode = mapping->host;
2011         unsigned start;
2012
2013         start = pos & (PAGE_CACHE_SIZE - 1);
2014
2015         if (unlikely(copied < len)) {
2016                 /*
2017                  * The buffers that were written will now be uptodate, so we
2018                  * don't have to worry about a readpage reading them and
2019                  * overwriting a partial write. However if we have encountered
2020                  * a short write and only partially written into a buffer, it
2021                  * will not be marked uptodate, so a readpage might come in and
2022                  * destroy our partial write.
2023                  *
2024                  * Do the simplest thing, and just treat any short write to a
2025                  * non uptodate page as a zero-length write, and force the
2026                  * caller to redo the whole thing.
2027                  */
2028                 if (!PageUptodate(page))
2029                         copied = 0;
2030
2031                 page_zero_new_buffers(page, start+copied, start+len);
2032         }
2033         flush_dcache_page(page);
2034
2035         /* This could be a short (even 0-length) commit */
2036         __block_commit_write(inode, page, start, start+copied);
2037
2038         return copied;
2039 }
2040 EXPORT_SYMBOL(block_write_end);
2041
2042 int generic_write_end(struct file *file, struct address_space *mapping,
2043                         loff_t pos, unsigned len, unsigned copied,
2044                         struct page *page, void *fsdata)
2045 {
2046         struct inode *inode = mapping->host;
2047         int i_size_changed = 0;
2048
2049         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2050
2051         /*
2052          * No need to use i_size_read() here, the i_size
2053          * cannot change under us because we hold i_mutex.
2054          *
2055          * But it's important to update i_size while still holding page lock:
2056          * page writeout could otherwise come in and zero beyond i_size.
2057          */
2058         if (pos+copied > inode->i_size) {
2059                 i_size_write(inode, pos+copied);
2060                 i_size_changed = 1;
2061         }
2062
2063         unlock_page(page);
2064         page_cache_release(page);
2065
2066         /*
2067          * Don't mark the inode dirty under page lock. First, it unnecessarily
2068          * makes the holding time of page lock longer. Second, it forces lock
2069          * ordering of page lock and transaction start for journaling
2070          * filesystems.
2071          */
2072         if (i_size_changed)
2073                 mark_inode_dirty(inode);
2074
2075         return copied;
2076 }
2077 EXPORT_SYMBOL(generic_write_end);
2078
2079 /*
2080  * block_is_partially_uptodate checks whether buffers within a page are
2081  * uptodate or not.
2082  *
2083  * Returns true if all buffers which correspond to a file portion
2084  * we want to read are uptodate.
2085  */
2086 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2087                                         unsigned long from)
2088 {
2089         struct inode *inode = page->mapping->host;
2090         unsigned block_start, block_end, blocksize;
2091         unsigned to;
2092         struct buffer_head *bh, *head;
2093         int ret = 1;
2094
2095         if (!page_has_buffers(page))
2096                 return 0;
2097
2098         blocksize = 1 << inode->i_blkbits;
2099         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2100         to = from + to;
2101         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2102                 return 0;
2103
2104         head = page_buffers(page);
2105         bh = head;
2106         block_start = 0;
2107         do {
2108                 block_end = block_start + blocksize;
2109                 if (block_end > from && block_start < to) {
2110                         if (!buffer_uptodate(bh)) {
2111                                 ret = 0;
2112                                 break;
2113                         }
2114                         if (block_end >= to)
2115                                 break;
2116                 }
2117                 block_start = block_end;
2118                 bh = bh->b_this_page;
2119         } while (bh != head);
2120
2121         return ret;
2122 }
2123 EXPORT_SYMBOL(block_is_partially_uptodate);
2124
2125 /*
2126  * Generic "read page" function for block devices that have the normal
2127  * get_block functionality. This is most of the block device filesystems.
2128  * Reads the page asynchronously --- the unlock_buffer() and
2129  * set/clear_buffer_uptodate() functions propagate buffer state into the
2130  * page struct once IO has completed.
2131  */
2132 int block_read_full_page(struct page *page, get_block_t *get_block)
2133 {
2134         struct inode *inode = page->mapping->host;
2135         sector_t iblock, lblock;
2136         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2137         unsigned int blocksize;
2138         int nr, i;
2139         int fully_mapped = 1;
2140
2141         BUG_ON(!PageLocked(page));
2142         blocksize = 1 << inode->i_blkbits;
2143         if (!page_has_buffers(page))
2144                 create_empty_buffers(page, blocksize, 0);
2145         head = page_buffers(page);
2146
2147         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2148         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2149         bh = head;
2150         nr = 0;
2151         i = 0;
2152
2153         do {
2154                 if (buffer_uptodate(bh))
2155                         continue;
2156
2157                 if (!buffer_mapped(bh)) {
2158                         int err = 0;
2159
2160                         fully_mapped = 0;
2161                         if (iblock < lblock) {
2162                                 WARN_ON(bh->b_size != blocksize);
2163                                 err = get_block(inode, iblock, bh, 0);
2164                                 if (err)
2165                                         SetPageError(page);
2166                         }
2167                         if (!buffer_mapped(bh)) {
2168                                 zero_user(page, i * blocksize, blocksize);
2169                                 if (!err)
2170                                         set_buffer_uptodate(bh);
2171                                 continue;
2172                         }
2173                         /*
2174                          * get_block() might have updated the buffer
2175                          * synchronously
2176                          */
2177                         if (buffer_uptodate(bh))
2178                                 continue;
2179                 }
2180                 arr[nr++] = bh;
2181         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2182
2183         if (fully_mapped)
2184                 SetPageMappedToDisk(page);
2185
2186         if (!nr) {
2187                 /*
2188                  * All buffers are uptodate - we can set the page uptodate
2189                  * as well. But not if get_block() returned an error.
2190                  */
2191                 if (!PageError(page))
2192                         SetPageUptodate(page);
2193                 unlock_page(page);
2194                 return 0;
2195         }
2196
2197         /* Stage two: lock the buffers */
2198         for (i = 0; i < nr; i++) {
2199                 bh = arr[i];
2200                 lock_buffer(bh);
2201                 mark_buffer_async_read(bh);
2202         }
2203
2204         /*
2205          * Stage 3: start the IO.  Check for uptodateness
2206          * inside the buffer lock in case another process reading
2207          * the underlying blockdev brought it uptodate (the sct fix).
2208          */
2209         for (i = 0; i < nr; i++) {
2210                 bh = arr[i];
2211                 if (buffer_uptodate(bh))
2212                         end_buffer_async_read(bh, 1);
2213                 else
2214                         submit_bh(READ, bh);
2215         }
2216         return 0;
2217 }
2218
2219 /* utility function for filesystems that need to do work on expanding
2220  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2221  * deal with the hole.  
2222  */
2223 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2224 {
2225         struct address_space *mapping = inode->i_mapping;
2226         struct page *page;
2227         void *fsdata;
2228         unsigned long limit;
2229         int err;
2230
2231         err = -EFBIG;
2232         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2233         if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2234                 send_sig(SIGXFSZ, current, 0);
2235                 goto out;
2236         }
2237         if (size > inode->i_sb->s_maxbytes)
2238                 goto out;
2239
2240         err = pagecache_write_begin(NULL, mapping, size, 0,
2241                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2242                                 &page, &fsdata);
2243         if (err)
2244                 goto out;
2245
2246         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2247         BUG_ON(err > 0);
2248
2249 out:
2250         return err;
2251 }
2252
2253 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2254                             loff_t pos, loff_t *bytes)
2255 {
2256         struct inode *inode = mapping->host;
2257         unsigned blocksize = 1 << inode->i_blkbits;
2258         struct page *page;
2259         void *fsdata;
2260         pgoff_t index, curidx;
2261         loff_t curpos;
2262         unsigned zerofrom, offset, len;
2263         int err = 0;
2264
2265         index = pos >> PAGE_CACHE_SHIFT;
2266         offset = pos & ~PAGE_CACHE_MASK;
2267
2268         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2269                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2270                 if (zerofrom & (blocksize-1)) {
2271                         *bytes |= (blocksize-1);
2272                         (*bytes)++;
2273                 }
2274                 len = PAGE_CACHE_SIZE - zerofrom;
2275
2276                 err = pagecache_write_begin(file, mapping, curpos, len,
2277                                                 AOP_FLAG_UNINTERRUPTIBLE,
2278                                                 &page, &fsdata);
2279                 if (err)
2280                         goto out;
2281                 zero_user(page, zerofrom, len);
2282                 err = pagecache_write_end(file, mapping, curpos, len, len,
2283                                                 page, fsdata);
2284                 if (err < 0)
2285                         goto out;
2286                 BUG_ON(err != len);
2287                 err = 0;
2288
2289                 balance_dirty_pages_ratelimited(mapping);
2290         }
2291
2292         /* page covers the boundary, find the boundary offset */
2293         if (index == curidx) {
2294                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2295                 /* if we will expand the thing last block will be filled */
2296                 if (offset <= zerofrom) {
2297                         goto out;
2298                 }
2299                 if (zerofrom & (blocksize-1)) {
2300                         *bytes |= (blocksize-1);
2301                         (*bytes)++;
2302                 }
2303                 len = offset - zerofrom;
2304
2305                 err = pagecache_write_begin(file, mapping, curpos, len,
2306                                                 AOP_FLAG_UNINTERRUPTIBLE,
2307                                                 &page, &fsdata);
2308                 if (err)
2309                         goto out;
2310                 zero_user(page, zerofrom, len);
2311                 err = pagecache_write_end(file, mapping, curpos, len, len,
2312                                                 page, fsdata);
2313                 if (err < 0)
2314                         goto out;
2315                 BUG_ON(err != len);
2316                 err = 0;
2317         }
2318 out:
2319         return err;
2320 }
2321
2322 /*
2323  * For moronic filesystems that do not allow holes in file.
2324  * We may have to extend the file.
2325  */
2326 int cont_write_begin(struct file *file, struct address_space *mapping,
2327                         loff_t pos, unsigned len, unsigned flags,
2328                         struct page **pagep, void **fsdata,
2329                         get_block_t *get_block, loff_t *bytes)
2330 {
2331         struct inode *inode = mapping->host;
2332         unsigned blocksize = 1 << inode->i_blkbits;
2333         unsigned zerofrom;
2334         int err;
2335
2336         err = cont_expand_zero(file, mapping, pos, bytes);
2337         if (err)
2338                 goto out;
2339
2340         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2341         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2342                 *bytes |= (blocksize-1);
2343                 (*bytes)++;
2344         }
2345
2346         *pagep = NULL;
2347         err = block_write_begin(file, mapping, pos, len,
2348                                 flags, pagep, fsdata, get_block);
2349 out:
2350         return err;
2351 }
2352
2353 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2354                         get_block_t *get_block)
2355 {
2356         struct inode *inode = page->mapping->host;
2357         int err = __block_prepare_write(inode, page, from, to, get_block);
2358         if (err)
2359                 ClearPageUptodate(page);
2360         return err;
2361 }
2362
2363 int block_commit_write(struct page *page, unsigned from, unsigned to)
2364 {
2365         struct inode *inode = page->mapping->host;
2366         __block_commit_write(inode,page,from,to);
2367         return 0;
2368 }
2369
2370 /*
2371  * block_page_mkwrite() is not allowed to change the file size as it gets
2372  * called from a page fault handler when a page is first dirtied. Hence we must
2373  * be careful to check for EOF conditions here. We set the page up correctly
2374  * for a written page which means we get ENOSPC checking when writing into
2375  * holes and correct delalloc and unwritten extent mapping on filesystems that
2376  * support these features.
2377  *
2378  * We are not allowed to take the i_mutex here so we have to play games to
2379  * protect against truncate races as the page could now be beyond EOF.  Because
2380  * vmtruncate() writes the inode size before removing pages, once we have the
2381  * page lock we can determine safely if the page is beyond EOF. If it is not
2382  * beyond EOF, then the page is guaranteed safe against truncation until we
2383  * unlock the page.
2384  */
2385 int
2386 block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2387                    get_block_t get_block)
2388 {
2389         struct page *page = vmf->page;
2390         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2391         unsigned long end;
2392         loff_t size;
2393         int ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
2394
2395         lock_page(page);
2396         size = i_size_read(inode);
2397         if ((page->mapping != inode->i_mapping) ||
2398             (page_offset(page) > size)) {
2399                 /* page got truncated out from underneath us */
2400                 unlock_page(page);
2401                 goto out;
2402         }
2403
2404         /* page is wholly or partially inside EOF */
2405         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2406                 end = size & ~PAGE_CACHE_MASK;
2407         else
2408                 end = PAGE_CACHE_SIZE;
2409
2410         ret = block_prepare_write(page, 0, end, get_block);
2411         if (!ret)
2412                 ret = block_commit_write(page, 0, end);
2413
2414         if (unlikely(ret)) {
2415                 unlock_page(page);
2416                 if (ret == -ENOMEM)
2417                         ret = VM_FAULT_OOM;
2418                 else /* -ENOSPC, -EIO, etc */
2419                         ret = VM_FAULT_SIGBUS;
2420         } else
2421                 ret = VM_FAULT_LOCKED;
2422
2423 out:
2424         return ret;
2425 }
2426
2427 /*
2428  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2429  * immediately, while under the page lock.  So it needs a special end_io
2430  * handler which does not touch the bh after unlocking it.
2431  */
2432 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2433 {
2434         __end_buffer_read_notouch(bh, uptodate);
2435 }
2436
2437 /*
2438  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2439  * the page (converting it to circular linked list and taking care of page
2440  * dirty races).
2441  */
2442 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2443 {
2444         struct buffer_head *bh;
2445
2446         BUG_ON(!PageLocked(page));
2447
2448         spin_lock(&page->mapping->private_lock);
2449         bh = head;
2450         do {
2451                 if (PageDirty(page))
2452                         set_buffer_dirty(bh);
2453                 if (!bh->b_this_page)
2454                         bh->b_this_page = head;
2455                 bh = bh->b_this_page;
2456         } while (bh != head);
2457         attach_page_buffers(page, head);
2458         spin_unlock(&page->mapping->private_lock);
2459 }
2460
2461 /*
2462  * On entry, the page is fully not uptodate.
2463  * On exit the page is fully uptodate in the areas outside (from,to)
2464  */
2465 int nobh_write_begin(struct file *file, struct address_space *mapping,
2466                         loff_t pos, unsigned len, unsigned flags,
2467                         struct page **pagep, void **fsdata,
2468                         get_block_t *get_block)
2469 {
2470         struct inode *inode = mapping->host;
2471         const unsigned blkbits = inode->i_blkbits;
2472         const unsigned blocksize = 1 << blkbits;
2473         struct buffer_head *head, *bh;
2474         struct page *page;
2475         pgoff_t index;
2476         unsigned from, to;
2477         unsigned block_in_page;
2478         unsigned block_start, block_end;
2479         sector_t block_in_file;
2480         int nr_reads = 0;
2481         int ret = 0;
2482         int is_mapped_to_disk = 1;
2483
2484         index = pos >> PAGE_CACHE_SHIFT;
2485         from = pos & (PAGE_CACHE_SIZE - 1);
2486         to = from + len;
2487
2488         page = grab_cache_page_write_begin(mapping, index, flags);
2489         if (!page)
2490                 return -ENOMEM;
2491         *pagep = page;
2492         *fsdata = NULL;
2493
2494         if (page_has_buffers(page)) {
2495                 unlock_page(page);
2496                 page_cache_release(page);
2497                 *pagep = NULL;
2498                 return block_write_begin(file, mapping, pos, len, flags, pagep,
2499                                         fsdata, get_block);
2500         }
2501
2502         if (PageMappedToDisk(page))
2503                 return 0;
2504
2505         /*
2506          * Allocate buffers so that we can keep track of state, and potentially
2507          * attach them to the page if an error occurs. In the common case of
2508          * no error, they will just be freed again without ever being attached
2509          * to the page (which is all OK, because we're under the page lock).
2510          *
2511          * Be careful: the buffer linked list is a NULL terminated one, rather
2512          * than the circular one we're used to.
2513          */
2514         head = alloc_page_buffers(page, blocksize, 0);
2515         if (!head) {
2516                 ret = -ENOMEM;
2517                 goto out_release;
2518         }
2519
2520         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2521
2522         /*
2523          * We loop across all blocks in the page, whether or not they are
2524          * part of the affected region.  This is so we can discover if the
2525          * page is fully mapped-to-disk.
2526          */
2527         for (block_start = 0, block_in_page = 0, bh = head;
2528                   block_start < PAGE_CACHE_SIZE;
2529                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2530                 int create;
2531
2532                 block_end = block_start + blocksize;
2533                 bh->b_state = 0;
2534                 create = 1;
2535                 if (block_start >= to)
2536                         create = 0;
2537                 ret = get_block(inode, block_in_file + block_in_page,
2538                                         bh, create);
2539                 if (ret)
2540                         goto failed;
2541                 if (!buffer_mapped(bh))
2542                         is_mapped_to_disk = 0;
2543                 if (buffer_new(bh))
2544                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2545                 if (PageUptodate(page)) {
2546                         set_buffer_uptodate(bh);
2547                         continue;
2548                 }
2549                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2550                         zero_user_segments(page, block_start, from,
2551                                                         to, block_end);
2552                         continue;
2553                 }
2554                 if (buffer_uptodate(bh))
2555                         continue;       /* reiserfs does this */
2556                 if (block_start < from || block_end > to) {
2557                         lock_buffer(bh);
2558                         bh->b_end_io = end_buffer_read_nobh;
2559                         submit_bh(READ, bh);
2560                         nr_reads++;
2561                 }
2562         }
2563
2564         if (nr_reads) {
2565                 /*
2566                  * The page is locked, so these buffers are protected from
2567                  * any VM or truncate activity.  Hence we don't need to care
2568                  * for the buffer_head refcounts.
2569                  */
2570                 for (bh = head; bh; bh = bh->b_this_page) {
2571                         wait_on_buffer(bh);
2572                         if (!buffer_uptodate(bh))
2573                                 ret = -EIO;
2574                 }
2575                 if (ret)
2576                         goto failed;
2577         }
2578
2579         if (is_mapped_to_disk)
2580                 SetPageMappedToDisk(page);
2581
2582         *fsdata = head; /* to be released by nobh_write_end */
2583
2584         return 0;
2585
2586 failed:
2587         BUG_ON(!ret);
2588         /*
2589          * Error recovery is a bit difficult. We need to zero out blocks that
2590          * were newly allocated, and dirty them to ensure they get written out.
2591          * Buffers need to be attached to the page at this point, otherwise
2592          * the handling of potential IO errors during writeout would be hard
2593          * (could try doing synchronous writeout, but what if that fails too?)
2594          */
2595         attach_nobh_buffers(page, head);
2596         page_zero_new_buffers(page, from, to);
2597
2598 out_release:
2599         unlock_page(page);
2600         page_cache_release(page);
2601         *pagep = NULL;
2602
2603         if (pos + len > inode->i_size)
2604                 vmtruncate(inode, inode->i_size);
2605
2606         return ret;
2607 }
2608 EXPORT_SYMBOL(nobh_write_begin);
2609
2610 int nobh_write_end(struct file *file, struct address_space *mapping,
2611                         loff_t pos, unsigned len, unsigned copied,
2612                         struct page *page, void *fsdata)
2613 {
2614         struct inode *inode = page->mapping->host;
2615         struct buffer_head *head = fsdata;
2616         struct buffer_head *bh;
2617         BUG_ON(fsdata != NULL && page_has_buffers(page));
2618
2619         if (unlikely(copied < len) && head)
2620                 attach_nobh_buffers(page, head);
2621         if (page_has_buffers(page))
2622                 return generic_write_end(file, mapping, pos, len,
2623                                         copied, page, fsdata);
2624
2625         SetPageUptodate(page);
2626         set_page_dirty(page);
2627         if (pos+copied > inode->i_size) {
2628                 i_size_write(inode, pos+copied);
2629                 mark_inode_dirty(inode);
2630         }
2631
2632         unlock_page(page);
2633         page_cache_release(page);
2634
2635         while (head) {
2636                 bh = head;
2637                 head = head->b_this_page;
2638                 free_buffer_head(bh);
2639         }
2640
2641         return copied;
2642 }
2643 EXPORT_SYMBOL(nobh_write_end);
2644
2645 /*
2646  * nobh_writepage() - based on block_full_write_page() except
2647  * that it tries to operate without attaching bufferheads to
2648  * the page.
2649  */
2650 int nobh_writepage(struct page *page, get_block_t *get_block,
2651                         struct writeback_control *wbc)
2652 {
2653         struct inode * const inode = page->mapping->host;
2654         loff_t i_size = i_size_read(inode);
2655         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2656         unsigned offset;
2657         int ret;
2658
2659         /* Is the page fully inside i_size? */
2660         if (page->index < end_index)
2661                 goto out;
2662
2663         /* Is the page fully outside i_size? (truncate in progress) */
2664         offset = i_size & (PAGE_CACHE_SIZE-1);
2665         if (page->index >= end_index+1 || !offset) {
2666                 /*
2667                  * The page may have dirty, unmapped buffers.  For example,
2668                  * they may have been added in ext3_writepage().  Make them
2669                  * freeable here, so the page does not leak.
2670                  */
2671 #if 0
2672                 /* Not really sure about this  - do we need this ? */
2673                 if (page->mapping->a_ops->invalidatepage)
2674                         page->mapping->a_ops->invalidatepage(page, offset);
2675 #endif
2676                 unlock_page(page);
2677                 return 0; /* don't care */
2678         }
2679
2680         /*
2681          * The page straddles i_size.  It must be zeroed out on each and every
2682          * writepage invocation because it may be mmapped.  "A file is mapped
2683          * in multiples of the page size.  For a file that is not a multiple of
2684          * the  page size, the remaining memory is zeroed when mapped, and
2685          * writes to that region are not written out to the file."
2686          */
2687         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2688 out:
2689         ret = mpage_writepage(page, get_block, wbc);
2690         if (ret == -EAGAIN)
2691                 ret = __block_write_full_page(inode, page, get_block, wbc,
2692                                               end_buffer_async_write);
2693         return ret;
2694 }
2695 EXPORT_SYMBOL(nobh_writepage);
2696
2697 int nobh_truncate_page(struct address_space *mapping,
2698                         loff_t from, get_block_t *get_block)
2699 {
2700         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2701         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2702         unsigned blocksize;
2703         sector_t iblock;
2704         unsigned length, pos;
2705         struct inode *inode = mapping->host;
2706         struct page *page;
2707         struct buffer_head map_bh;
2708         int err;
2709
2710         blocksize = 1 << inode->i_blkbits;
2711         length = offset & (blocksize - 1);
2712
2713         /* Block boundary? Nothing to do */
2714         if (!length)
2715                 return 0;
2716
2717         length = blocksize - length;
2718         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2719
2720         page = grab_cache_page(mapping, index);
2721         err = -ENOMEM;
2722         if (!page)
2723                 goto out;
2724
2725         if (page_has_buffers(page)) {
2726 has_buffers:
2727                 unlock_page(page);
2728                 page_cache_release(page);
2729                 return block_truncate_page(mapping, from, get_block);
2730         }
2731
2732         /* Find the buffer that contains "offset" */
2733         pos = blocksize;
2734         while (offset >= pos) {
2735                 iblock++;
2736                 pos += blocksize;
2737         }
2738
2739         err = get_block(inode, iblock, &map_bh, 0);
2740         if (err)
2741                 goto unlock;
2742         /* unmapped? It's a hole - nothing to do */
2743         if (!buffer_mapped(&map_bh))
2744                 goto unlock;
2745
2746         /* Ok, it's mapped. Make sure it's up-to-date */
2747         if (!PageUptodate(page)) {
2748                 err = mapping->a_ops->readpage(NULL, page);
2749                 if (err) {
2750                         page_cache_release(page);
2751                         goto out;
2752                 }
2753                 lock_page(page);
2754                 if (!PageUptodate(page)) {
2755                         err = -EIO;
2756                         goto unlock;
2757                 }
2758                 if (page_has_buffers(page))
2759                         goto has_buffers;
2760         }
2761         zero_user(page, offset, length);
2762         set_page_dirty(page);
2763         err = 0;
2764
2765 unlock:
2766         unlock_page(page);
2767         page_cache_release(page);
2768 out:
2769         return err;
2770 }
2771 EXPORT_SYMBOL(nobh_truncate_page);
2772
2773 int block_truncate_page(struct address_space *mapping,
2774                         loff_t from, get_block_t *get_block)
2775 {
2776         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2777         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2778         unsigned blocksize;
2779         sector_t iblock;
2780         unsigned length, pos;
2781         struct inode *inode = mapping->host;
2782         struct page *page;
2783         struct buffer_head *bh;
2784         int err;
2785
2786         blocksize = 1 << inode->i_blkbits;
2787         length = offset & (blocksize - 1);
2788
2789         /* Block boundary? Nothing to do */
2790         if (!length)
2791                 return 0;
2792
2793         length = blocksize - length;
2794         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2795         
2796         page = grab_cache_page(mapping, index);
2797         err = -ENOMEM;
2798         if (!page)
2799                 goto out;
2800
2801         if (!page_has_buffers(page))
2802                 create_empty_buffers(page, blocksize, 0);
2803
2804         /* Find the buffer that contains "offset" */
2805         bh = page_buffers(page);
2806         pos = blocksize;
2807         while (offset >= pos) {
2808                 bh = bh->b_this_page;
2809                 iblock++;
2810                 pos += blocksize;
2811         }
2812
2813         err = 0;
2814         if (!buffer_mapped(bh)) {
2815                 WARN_ON(bh->b_size != blocksize);
2816                 err = get_block(inode, iblock, bh, 0);
2817                 if (err)
2818                         goto unlock;
2819                 /* unmapped? It's a hole - nothing to do */
2820                 if (!buffer_mapped(bh))
2821                         goto unlock;
2822         }
2823
2824         /* Ok, it's mapped. Make sure it's up-to-date */
2825         if (PageUptodate(page))
2826                 set_buffer_uptodate(bh);
2827
2828         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2829                 err = -EIO;
2830                 ll_rw_block(READ, 1, &bh);
2831                 wait_on_buffer(bh);
2832                 /* Uhhuh. Read error. Complain and punt. */
2833                 if (!buffer_uptodate(bh))
2834                         goto unlock;
2835         }
2836
2837         zero_user(page, offset, length);
2838         mark_buffer_dirty(bh);
2839         err = 0;
2840
2841 unlock:
2842         unlock_page(page);
2843         page_cache_release(page);
2844 out:
2845         return err;
2846 }
2847
2848 /*
2849  * The generic ->writepage function for buffer-backed address_spaces
2850  * this form passes in the end_io handler used to finish the IO.
2851  */
2852 int block_write_full_page_endio(struct page *page, get_block_t *get_block,
2853                         struct writeback_control *wbc, bh_end_io_t *handler)
2854 {
2855         struct inode * const inode = page->mapping->host;
2856         loff_t i_size = i_size_read(inode);
2857         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2858         unsigned offset;
2859
2860         /* Is the page fully inside i_size? */
2861         if (page->index < end_index)
2862                 return __block_write_full_page(inode, page, get_block, wbc,
2863                                                handler);
2864
2865         /* Is the page fully outside i_size? (truncate in progress) */
2866         offset = i_size & (PAGE_CACHE_SIZE-1);
2867         if (page->index >= end_index+1 || !offset) {
2868                 /*
2869                  * The page may have dirty, unmapped buffers.  For example,
2870                  * they may have been added in ext3_writepage().  Make them
2871                  * freeable here, so the page does not leak.
2872                  */
2873                 do_invalidatepage(page, 0);
2874                 unlock_page(page);
2875                 return 0; /* don't care */
2876         }
2877
2878         /*
2879          * The page straddles i_size.  It must be zeroed out on each and every
2880          * writepage invokation because it may be mmapped.  "A file is mapped
2881          * in multiples of the page size.  For a file that is not a multiple of
2882          * the  page size, the remaining memory is zeroed when mapped, and
2883          * writes to that region are not written out to the file."
2884          */
2885         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2886         return __block_write_full_page(inode, page, get_block, wbc, handler);
2887 }
2888
2889 /*
2890  * The generic ->writepage function for buffer-backed address_spaces
2891  */
2892 int block_write_full_page(struct page *page, get_block_t *get_block,
2893                         struct writeback_control *wbc)
2894 {
2895         return block_write_full_page_endio(page, get_block, wbc,
2896                                            end_buffer_async_write);
2897 }
2898
2899
2900 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2901                             get_block_t *get_block)
2902 {
2903         struct buffer_head tmp;
2904         struct inode *inode = mapping->host;
2905         tmp.b_state = 0;
2906         tmp.b_blocknr = 0;
2907         tmp.b_size = 1 << inode->i_blkbits;
2908         get_block(inode, block, &tmp, 0);
2909         return tmp.b_blocknr;
2910 }
2911
2912 static void end_bio_bh_io_sync(struct bio *bio, int err)
2913 {
2914         struct buffer_head *bh = bio->bi_private;
2915
2916         if (err == -EOPNOTSUPP) {
2917                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2918                 set_bit(BH_Eopnotsupp, &bh->b_state);
2919         }
2920
2921         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2922                 set_bit(BH_Quiet, &bh->b_state);
2923
2924         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2925         bio_put(bio);
2926 }
2927
2928 int submit_bh(int rw, struct buffer_head * bh)
2929 {
2930         struct bio *bio;
2931         int ret = 0;
2932
2933         BUG_ON(!buffer_locked(bh));
2934         BUG_ON(!buffer_mapped(bh));
2935         BUG_ON(!bh->b_end_io);
2936         BUG_ON(buffer_delay(bh));
2937         BUG_ON(buffer_unwritten(bh));
2938
2939         /*
2940          * Mask in barrier bit for a write (could be either a WRITE or a
2941          * WRITE_SYNC
2942          */
2943         if (buffer_ordered(bh) && (rw & WRITE))
2944                 rw |= WRITE_BARRIER;
2945
2946         /*
2947          * Only clear out a write error when rewriting
2948          */
2949         if (test_set_buffer_req(bh) && (rw & WRITE))
2950                 clear_buffer_write_io_error(bh);
2951
2952         /*
2953          * from here on down, it's all bio -- do the initial mapping,
2954          * submit_bio -> generic_make_request may further map this bio around
2955          */
2956         bio = bio_alloc(GFP_NOIO, 1);
2957
2958         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2959         bio->bi_bdev = bh->b_bdev;
2960         bio->bi_io_vec[0].bv_page = bh->b_page;
2961         bio->bi_io_vec[0].bv_len = bh->b_size;
2962         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2963
2964         bio->bi_vcnt = 1;
2965         bio->bi_idx = 0;
2966         bio->bi_size = bh->b_size;
2967
2968         bio->bi_end_io = end_bio_bh_io_sync;
2969         bio->bi_private = bh;
2970
2971         bio_get(bio);
2972         submit_bio(rw, bio);
2973
2974         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2975                 ret = -EOPNOTSUPP;
2976
2977         bio_put(bio);
2978         return ret;
2979 }
2980
2981 /**
2982  * ll_rw_block: low-level access to block devices (DEPRECATED)
2983  * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
2984  * @nr: number of &struct buffer_heads in the array
2985  * @bhs: array of pointers to &struct buffer_head
2986  *
2987  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2988  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2989  * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2990  * are sent to disk. The fourth %READA option is described in the documentation
2991  * for generic_make_request() which ll_rw_block() calls.
2992  *
2993  * This function drops any buffer that it cannot get a lock on (with the
2994  * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
2995  * clean when doing a write request, and any buffer that appears to be
2996  * up-to-date when doing read request.  Further it marks as clean buffers that
2997  * are processed for writing (the buffer cache won't assume that they are
2998  * actually clean until the buffer gets unlocked).
2999  *
3000  * ll_rw_block sets b_end_io to simple completion handler that marks
3001  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
3002  * any waiters. 
3003  *
3004  * All of the buffers must be for the same device, and must also be a
3005  * multiple of the current approved size for the device.
3006  */
3007 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
3008 {
3009         int i;
3010
3011         for (i = 0; i < nr; i++) {
3012                 struct buffer_head *bh = bhs[i];
3013
3014                 if (rw == SWRITE || rw == SWRITE_SYNC || rw == SWRITE_SYNC_PLUG)
3015                         lock_buffer(bh);
3016                 else if (!trylock_buffer(bh))
3017                         continue;
3018
3019                 if (rw == WRITE || rw == SWRITE || rw == SWRITE_SYNC ||
3020                     rw == SWRITE_SYNC_PLUG) {
3021                         if (test_clear_buffer_dirty(bh)) {
3022                                 bh->b_end_io = end_buffer_write_sync;
3023                                 get_bh(bh);
3024                                 if (rw == SWRITE_SYNC)
3025                                         submit_bh(WRITE_SYNC, bh);
3026                                 else
3027                                         submit_bh(WRITE, bh);
3028                                 continue;
3029                         }
3030                 } else {
3031                         if (!buffer_uptodate(bh)) {
3032                                 bh->b_end_io = end_buffer_read_sync;
3033                                 get_bh(bh);
3034                                 submit_bh(rw, bh);
3035                                 continue;
3036                         }
3037                 }
3038                 unlock_buffer(bh);
3039         }
3040 }
3041
3042 /*
3043  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3044  * and then start new I/O and then wait upon it.  The caller must have a ref on
3045  * the buffer_head.
3046  */
3047 int sync_dirty_buffer(struct buffer_head *bh)
3048 {
3049         int ret = 0;
3050
3051         WARN_ON(atomic_read(&bh->b_count) < 1);
3052         lock_buffer(bh);
3053         if (test_clear_buffer_dirty(bh)) {
3054                 get_bh(bh);
3055                 bh->b_end_io = end_buffer_write_sync;
3056                 ret = submit_bh(WRITE_SYNC, bh);
3057                 wait_on_buffer(bh);
3058                 if (buffer_eopnotsupp(bh)) {
3059                         clear_buffer_eopnotsupp(bh);
3060                         ret = -EOPNOTSUPP;
3061                 }
3062                 if (!ret && !buffer_uptodate(bh))
3063                         ret = -EIO;
3064         } else {
3065                 unlock_buffer(bh);
3066         }
3067         return ret;
3068 }
3069
3070 /*
3071  * try_to_free_buffers() checks if all the buffers on this particular page
3072  * are unused, and releases them if so.
3073  *
3074  * Exclusion against try_to_free_buffers may be obtained by either
3075  * locking the page or by holding its mapping's private_lock.
3076  *
3077  * If the page is dirty but all the buffers are clean then we need to
3078  * be sure to mark the page clean as well.  This is because the page
3079  * may be against a block device, and a later reattachment of buffers
3080  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3081  * filesystem data on the same device.
3082  *
3083  * The same applies to regular filesystem pages: if all the buffers are
3084  * clean then we set the page clean and proceed.  To do that, we require
3085  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3086  * private_lock.
3087  *
3088  * try_to_free_buffers() is non-blocking.
3089  */
3090 static inline int buffer_busy(struct buffer_head *bh)
3091 {
3092         return atomic_read(&bh->b_count) |
3093                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3094 }
3095
3096 static int
3097 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3098 {
3099         struct buffer_head *head = page_buffers(page);
3100         struct buffer_head *bh;
3101
3102         bh = head;
3103         do {
3104                 if (buffer_write_io_error(bh) && page->mapping)
3105                         set_bit(AS_EIO, &page->mapping->flags);
3106                 if (buffer_busy(bh))
3107                         goto failed;
3108                 bh = bh->b_this_page;
3109         } while (bh != head);
3110
3111         do {
3112                 struct buffer_head *next = bh->b_this_page;
3113
3114                 if (bh->b_assoc_map)
3115                         __remove_assoc_queue(bh);
3116                 bh = next;
3117         } while (bh != head);
3118         *buffers_to_free = head;
3119         __clear_page_buffers(page);
3120         return 1;
3121 failed:
3122         return 0;
3123 }
3124
3125 int try_to_free_buffers(struct page *page)
3126 {
3127         struct address_space * const mapping = page->mapping;
3128         struct buffer_head *buffers_to_free = NULL;
3129         int ret = 0;
3130
3131         BUG_ON(!PageLocked(page));
3132         if (PageWriteback(page))
3133                 return 0;
3134
3135         if (mapping == NULL) {          /* can this still happen? */
3136                 ret = drop_buffers(page, &buffers_to_free);
3137                 goto out;
3138         }
3139
3140         spin_lock(&mapping->private_lock);
3141         ret = drop_buffers(page, &buffers_to_free);
3142
3143         /*
3144          * If the filesystem writes its buffers by hand (eg ext3)
3145          * then we can have clean buffers against a dirty page.  We
3146          * clean the page here; otherwise the VM will never notice
3147          * that the filesystem did any IO at all.
3148          *
3149          * Also, during truncate, discard_buffer will have marked all
3150          * the page's buffers clean.  We discover that here and clean
3151          * the page also.
3152          *
3153          * private_lock must be held over this entire operation in order
3154          * to synchronise against __set_page_dirty_buffers and prevent the
3155          * dirty bit from being lost.
3156          */
3157         if (ret)
3158                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3159         spin_unlock(&mapping->private_lock);
3160 out:
3161         if (buffers_to_free) {
3162                 struct buffer_head *bh = buffers_to_free;
3163
3164                 do {
3165                         struct buffer_head *next = bh->b_this_page;
3166                         free_buffer_head(bh);
3167                         bh = next;
3168                 } while (bh != buffers_to_free);
3169         }
3170         return ret;
3171 }
3172 EXPORT_SYMBOL(try_to_free_buffers);
3173
3174 void block_sync_page(struct page *page)
3175 {
3176         struct address_space *mapping;
3177
3178         smp_mb();
3179         mapping = page_mapping(page);
3180         if (mapping)
3181                 blk_run_backing_dev(mapping->backing_dev_info, page);
3182 }
3183
3184 /*
3185  * There are no bdflush tunables left.  But distributions are
3186  * still running obsolete flush daemons, so we terminate them here.
3187  *
3188  * Use of bdflush() is deprecated and will be removed in a future kernel.
3189  * The `pdflush' kernel threads fully replace bdflush daemons and this call.
3190  */
3191 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3192 {
3193         static int msg_count;
3194
3195         if (!capable(CAP_SYS_ADMIN))
3196                 return -EPERM;
3197
3198         if (msg_count < 5) {
3199                 msg_count++;
3200                 printk(KERN_INFO
3201                         "warning: process `%s' used the obsolete bdflush"
3202                         " system call\n", current->comm);
3203                 printk(KERN_INFO "Fix your initscripts?\n");
3204         }
3205
3206         if (func == 1)
3207                 do_exit(0);
3208         return 0;
3209 }
3210
3211 /*
3212  * Buffer-head allocation
3213  */
3214 static struct kmem_cache *bh_cachep;
3215
3216 /*
3217  * Once the number of bh's in the machine exceeds this level, we start
3218  * stripping them in writeback.
3219  */
3220 static int max_buffer_heads;
3221
3222 int buffer_heads_over_limit;
3223
3224 struct bh_accounting {
3225         int nr;                 /* Number of live bh's */
3226         int ratelimit;          /* Limit cacheline bouncing */
3227 };
3228
3229 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3230
3231 static void recalc_bh_state(void)
3232 {
3233         int i;
3234         int tot = 0;
3235
3236         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3237                 return;
3238         __get_cpu_var(bh_accounting).ratelimit = 0;
3239         for_each_online_cpu(i)
3240                 tot += per_cpu(bh_accounting, i).nr;
3241         buffer_heads_over_limit = (tot > max_buffer_heads);
3242 }
3243         
3244 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3245 {
3246         struct buffer_head *ret = kmem_cache_alloc(bh_cachep, gfp_flags);
3247         if (ret) {
3248                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3249                 get_cpu_var(bh_accounting).nr++;
3250                 recalc_bh_state();
3251                 put_cpu_var(bh_accounting);
3252         }
3253         return ret;
3254 }
3255 EXPORT_SYMBOL(alloc_buffer_head);
3256
3257 void free_buffer_head(struct buffer_head *bh)
3258 {
3259         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3260         kmem_cache_free(bh_cachep, bh);
3261         get_cpu_var(bh_accounting).nr--;
3262         recalc_bh_state();
3263         put_cpu_var(bh_accounting);
3264 }
3265 EXPORT_SYMBOL(free_buffer_head);
3266
3267 static void buffer_exit_cpu(int cpu)
3268 {
3269         int i;
3270         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3271
3272         for (i = 0; i < BH_LRU_SIZE; i++) {
3273                 brelse(b->bhs[i]);
3274                 b->bhs[i] = NULL;
3275         }
3276         get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3277         per_cpu(bh_accounting, cpu).nr = 0;
3278         put_cpu_var(bh_accounting);
3279 }
3280
3281 static int buffer_cpu_notify(struct notifier_block *self,
3282                               unsigned long action, void *hcpu)
3283 {
3284         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3285                 buffer_exit_cpu((unsigned long)hcpu);
3286         return NOTIFY_OK;
3287 }
3288
3289 /**
3290  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3291  * @bh: struct buffer_head
3292  *
3293  * Return true if the buffer is up-to-date and false,
3294  * with the buffer locked, if not.
3295  */
3296 int bh_uptodate_or_lock(struct buffer_head *bh)
3297 {
3298         if (!buffer_uptodate(bh)) {
3299                 lock_buffer(bh);
3300                 if (!buffer_uptodate(bh))
3301                         return 0;
3302                 unlock_buffer(bh);
3303         }
3304         return 1;
3305 }
3306 EXPORT_SYMBOL(bh_uptodate_or_lock);
3307
3308 /**
3309  * bh_submit_read - Submit a locked buffer for reading
3310  * @bh: struct buffer_head
3311  *
3312  * Returns zero on success and -EIO on error.
3313  */
3314 int bh_submit_read(struct buffer_head *bh)
3315 {
3316         BUG_ON(!buffer_locked(bh));
3317
3318         if (buffer_uptodate(bh)) {
3319                 unlock_buffer(bh);
3320                 return 0;
3321         }
3322
3323         get_bh(bh);
3324         bh->b_end_io = end_buffer_read_sync;
3325         submit_bh(READ, bh);
3326         wait_on_buffer(bh);
3327         if (buffer_uptodate(bh))
3328                 return 0;
3329         return -EIO;
3330 }
3331 EXPORT_SYMBOL(bh_submit_read);
3332
3333 static void
3334 init_buffer_head(void *data)
3335 {
3336         struct buffer_head *bh = data;
3337
3338         memset(bh, 0, sizeof(*bh));
3339         INIT_LIST_HEAD(&bh->b_assoc_buffers);
3340 }
3341
3342 void __init buffer_init(void)
3343 {
3344         int nrpages;
3345
3346         bh_cachep = kmem_cache_create("buffer_head",
3347                         sizeof(struct buffer_head), 0,
3348                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3349                                 SLAB_MEM_SPREAD),
3350                                 init_buffer_head);
3351
3352         /*
3353          * Limit the bh occupancy to 10% of ZONE_NORMAL
3354          */
3355         nrpages = (nr_free_buffer_pages() * 10) / 100;
3356         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3357         hotcpu_notifier(buffer_cpu_notify, 0);
3358 }
3359
3360 EXPORT_SYMBOL(__bforget);
3361 EXPORT_SYMBOL(__brelse);
3362 EXPORT_SYMBOL(__wait_on_buffer);
3363 EXPORT_SYMBOL(block_commit_write);
3364 EXPORT_SYMBOL(block_prepare_write);
3365 EXPORT_SYMBOL(block_page_mkwrite);
3366 EXPORT_SYMBOL(block_read_full_page);
3367 EXPORT_SYMBOL(block_sync_page);
3368 EXPORT_SYMBOL(block_truncate_page);
3369 EXPORT_SYMBOL(block_write_full_page);
3370 EXPORT_SYMBOL(block_write_full_page_endio);
3371 EXPORT_SYMBOL(cont_write_begin);
3372 EXPORT_SYMBOL(end_buffer_read_sync);
3373 EXPORT_SYMBOL(end_buffer_write_sync);
3374 EXPORT_SYMBOL(end_buffer_async_write);
3375 EXPORT_SYMBOL(file_fsync);
3376 EXPORT_SYMBOL(generic_block_bmap);
3377 EXPORT_SYMBOL(generic_cont_expand_simple);
3378 EXPORT_SYMBOL(init_buffer);
3379 EXPORT_SYMBOL(invalidate_bdev);
3380 EXPORT_SYMBOL(ll_rw_block);
3381 EXPORT_SYMBOL(mark_buffer_dirty);
3382 EXPORT_SYMBOL(submit_bh);
3383 EXPORT_SYMBOL(sync_dirty_buffer);
3384 EXPORT_SYMBOL(unlock_buffer);