writeback: switch to per-bdi threads for flushing data
[linux-2.6.git] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/capability.h>
28 #include <linux/blkdev.h>
29 #include <linux/file.h>
30 #include <linux/quotaops.h>
31 #include <linux/highmem.h>
32 #include <linux/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_flusher_threads(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_logical_block_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 "logical block size: %d\n",
1093                                         bdev_logical_block_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                         struct address_space *mapping = page_mapping(page);
1170                         if (mapping)
1171                                 __set_page_dirty(page, mapping, 0);
1172                 }
1173         }
1174 }
1175
1176 /*
1177  * Decrement a buffer_head's reference count.  If all buffers against a page
1178  * have zero reference count, are clean and unlocked, and if the page is clean
1179  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1180  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1181  * a page but it ends up not being freed, and buffers may later be reattached).
1182  */
1183 void __brelse(struct buffer_head * buf)
1184 {
1185         if (atomic_read(&buf->b_count)) {
1186                 put_bh(buf);
1187                 return;
1188         }
1189         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1190 }
1191
1192 /*
1193  * bforget() is like brelse(), except it discards any
1194  * potentially dirty data.
1195  */
1196 void __bforget(struct buffer_head *bh)
1197 {
1198         clear_buffer_dirty(bh);
1199         if (bh->b_assoc_map) {
1200                 struct address_space *buffer_mapping = bh->b_page->mapping;
1201
1202                 spin_lock(&buffer_mapping->private_lock);
1203                 list_del_init(&bh->b_assoc_buffers);
1204                 bh->b_assoc_map = NULL;
1205                 spin_unlock(&buffer_mapping->private_lock);
1206         }
1207         __brelse(bh);
1208 }
1209
1210 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1211 {
1212         lock_buffer(bh);
1213         if (buffer_uptodate(bh)) {
1214                 unlock_buffer(bh);
1215                 return bh;
1216         } else {
1217                 get_bh(bh);
1218                 bh->b_end_io = end_buffer_read_sync;
1219                 submit_bh(READ, bh);
1220                 wait_on_buffer(bh);
1221                 if (buffer_uptodate(bh))
1222                         return bh;
1223         }
1224         brelse(bh);
1225         return NULL;
1226 }
1227
1228 /*
1229  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1230  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1231  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1232  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1233  * CPU's LRUs at the same time.
1234  *
1235  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1236  * sb_find_get_block().
1237  *
1238  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1239  * a local interrupt disable for that.
1240  */
1241
1242 #define BH_LRU_SIZE     8
1243
1244 struct bh_lru {
1245         struct buffer_head *bhs[BH_LRU_SIZE];
1246 };
1247
1248 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1249
1250 #ifdef CONFIG_SMP
1251 #define bh_lru_lock()   local_irq_disable()
1252 #define bh_lru_unlock() local_irq_enable()
1253 #else
1254 #define bh_lru_lock()   preempt_disable()
1255 #define bh_lru_unlock() preempt_enable()
1256 #endif
1257
1258 static inline void check_irqs_on(void)
1259 {
1260 #ifdef irqs_disabled
1261         BUG_ON(irqs_disabled());
1262 #endif
1263 }
1264
1265 /*
1266  * The LRU management algorithm is dopey-but-simple.  Sorry.
1267  */
1268 static void bh_lru_install(struct buffer_head *bh)
1269 {
1270         struct buffer_head *evictee = NULL;
1271         struct bh_lru *lru;
1272
1273         check_irqs_on();
1274         bh_lru_lock();
1275         lru = &__get_cpu_var(bh_lrus);
1276         if (lru->bhs[0] != bh) {
1277                 struct buffer_head *bhs[BH_LRU_SIZE];
1278                 int in;
1279                 int out = 0;
1280
1281                 get_bh(bh);
1282                 bhs[out++] = bh;
1283                 for (in = 0; in < BH_LRU_SIZE; in++) {
1284                         struct buffer_head *bh2 = lru->bhs[in];
1285
1286                         if (bh2 == bh) {
1287                                 __brelse(bh2);
1288                         } else {
1289                                 if (out >= BH_LRU_SIZE) {
1290                                         BUG_ON(evictee != NULL);
1291                                         evictee = bh2;
1292                                 } else {
1293                                         bhs[out++] = bh2;
1294                                 }
1295                         }
1296                 }
1297                 while (out < BH_LRU_SIZE)
1298                         bhs[out++] = NULL;
1299                 memcpy(lru->bhs, bhs, sizeof(bhs));
1300         }
1301         bh_lru_unlock();
1302
1303         if (evictee)
1304                 __brelse(evictee);
1305 }
1306
1307 /*
1308  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1309  */
1310 static struct buffer_head *
1311 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1312 {
1313         struct buffer_head *ret = NULL;
1314         struct bh_lru *lru;
1315         unsigned int i;
1316
1317         check_irqs_on();
1318         bh_lru_lock();
1319         lru = &__get_cpu_var(bh_lrus);
1320         for (i = 0; i < BH_LRU_SIZE; i++) {
1321                 struct buffer_head *bh = lru->bhs[i];
1322
1323                 if (bh && bh->b_bdev == bdev &&
1324                                 bh->b_blocknr == block && bh->b_size == size) {
1325                         if (i) {
1326                                 while (i) {
1327                                         lru->bhs[i] = lru->bhs[i - 1];
1328                                         i--;
1329                                 }
1330                                 lru->bhs[0] = bh;
1331                         }
1332                         get_bh(bh);
1333                         ret = bh;
1334                         break;
1335                 }
1336         }
1337         bh_lru_unlock();
1338         return ret;
1339 }
1340
1341 /*
1342  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1343  * it in the LRU and mark it as accessed.  If it is not present then return
1344  * NULL
1345  */
1346 struct buffer_head *
1347 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1348 {
1349         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1350
1351         if (bh == NULL) {
1352                 bh = __find_get_block_slow(bdev, block);
1353                 if (bh)
1354                         bh_lru_install(bh);
1355         }
1356         if (bh)
1357                 touch_buffer(bh);
1358         return bh;
1359 }
1360 EXPORT_SYMBOL(__find_get_block);
1361
1362 /*
1363  * __getblk will locate (and, if necessary, create) the buffer_head
1364  * which corresponds to the passed block_device, block and size. The
1365  * returned buffer has its reference count incremented.
1366  *
1367  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1368  * illegal block number, __getblk() will happily return a buffer_head
1369  * which represents the non-existent block.  Very weird.
1370  *
1371  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1372  * attempt is failing.  FIXME, perhaps?
1373  */
1374 struct buffer_head *
1375 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1376 {
1377         struct buffer_head *bh = __find_get_block(bdev, block, size);
1378
1379         might_sleep();
1380         if (bh == NULL)
1381                 bh = __getblk_slow(bdev, block, size);
1382         return bh;
1383 }
1384 EXPORT_SYMBOL(__getblk);
1385
1386 /*
1387  * Do async read-ahead on a buffer..
1388  */
1389 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1390 {
1391         struct buffer_head *bh = __getblk(bdev, block, size);
1392         if (likely(bh)) {
1393                 ll_rw_block(READA, 1, &bh);
1394                 brelse(bh);
1395         }
1396 }
1397 EXPORT_SYMBOL(__breadahead);
1398
1399 /**
1400  *  __bread() - reads a specified block and returns the bh
1401  *  @bdev: the block_device to read from
1402  *  @block: number of block
1403  *  @size: size (in bytes) to read
1404  * 
1405  *  Reads a specified block, and returns buffer head that contains it.
1406  *  It returns NULL if the block was unreadable.
1407  */
1408 struct buffer_head *
1409 __bread(struct block_device *bdev, sector_t block, unsigned size)
1410 {
1411         struct buffer_head *bh = __getblk(bdev, block, size);
1412
1413         if (likely(bh) && !buffer_uptodate(bh))
1414                 bh = __bread_slow(bh);
1415         return bh;
1416 }
1417 EXPORT_SYMBOL(__bread);
1418
1419 /*
1420  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1421  * This doesn't race because it runs in each cpu either in irq
1422  * or with preempt disabled.
1423  */
1424 static void invalidate_bh_lru(void *arg)
1425 {
1426         struct bh_lru *b = &get_cpu_var(bh_lrus);
1427         int i;
1428
1429         for (i = 0; i < BH_LRU_SIZE; i++) {
1430                 brelse(b->bhs[i]);
1431                 b->bhs[i] = NULL;
1432         }
1433         put_cpu_var(bh_lrus);
1434 }
1435         
1436 void invalidate_bh_lrus(void)
1437 {
1438         on_each_cpu(invalidate_bh_lru, NULL, 1);
1439 }
1440 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1441
1442 void set_bh_page(struct buffer_head *bh,
1443                 struct page *page, unsigned long offset)
1444 {
1445         bh->b_page = page;
1446         BUG_ON(offset >= PAGE_SIZE);
1447         if (PageHighMem(page))
1448                 /*
1449                  * This catches illegal uses and preserves the offset:
1450                  */
1451                 bh->b_data = (char *)(0 + offset);
1452         else
1453                 bh->b_data = page_address(page) + offset;
1454 }
1455 EXPORT_SYMBOL(set_bh_page);
1456
1457 /*
1458  * Called when truncating a buffer on a page completely.
1459  */
1460 static void discard_buffer(struct buffer_head * bh)
1461 {
1462         lock_buffer(bh);
1463         clear_buffer_dirty(bh);
1464         bh->b_bdev = NULL;
1465         clear_buffer_mapped(bh);
1466         clear_buffer_req(bh);
1467         clear_buffer_new(bh);
1468         clear_buffer_delay(bh);
1469         clear_buffer_unwritten(bh);
1470         unlock_buffer(bh);
1471 }
1472
1473 /**
1474  * block_invalidatepage - invalidate part of all of a buffer-backed page
1475  *
1476  * @page: the page which is affected
1477  * @offset: the index of the truncation point
1478  *
1479  * block_invalidatepage() is called when all or part of the page has become
1480  * invalidatedby a truncate operation.
1481  *
1482  * block_invalidatepage() does not have to release all buffers, but it must
1483  * ensure that no dirty buffer is left outside @offset and that no I/O
1484  * is underway against any of the blocks which are outside the truncation
1485  * point.  Because the caller is about to free (and possibly reuse) those
1486  * blocks on-disk.
1487  */
1488 void block_invalidatepage(struct page *page, unsigned long offset)
1489 {
1490         struct buffer_head *head, *bh, *next;
1491         unsigned int curr_off = 0;
1492
1493         BUG_ON(!PageLocked(page));
1494         if (!page_has_buffers(page))
1495                 goto out;
1496
1497         head = page_buffers(page);
1498         bh = head;
1499         do {
1500                 unsigned int next_off = curr_off + bh->b_size;
1501                 next = bh->b_this_page;
1502
1503                 /*
1504                  * is this block fully invalidated?
1505                  */
1506                 if (offset <= curr_off)
1507                         discard_buffer(bh);
1508                 curr_off = next_off;
1509                 bh = next;
1510         } while (bh != head);
1511
1512         /*
1513          * We release buffers only if the entire page is being invalidated.
1514          * The get_block cached value has been unconditionally invalidated,
1515          * so real IO is not possible anymore.
1516          */
1517         if (offset == 0)
1518                 try_to_release_page(page, 0);
1519 out:
1520         return;
1521 }
1522 EXPORT_SYMBOL(block_invalidatepage);
1523
1524 /*
1525  * We attach and possibly dirty the buffers atomically wrt
1526  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1527  * is already excluded via the page lock.
1528  */
1529 void create_empty_buffers(struct page *page,
1530                         unsigned long blocksize, unsigned long b_state)
1531 {
1532         struct buffer_head *bh, *head, *tail;
1533
1534         head = alloc_page_buffers(page, blocksize, 1);
1535         bh = head;
1536         do {
1537                 bh->b_state |= b_state;
1538                 tail = bh;
1539                 bh = bh->b_this_page;
1540         } while (bh);
1541         tail->b_this_page = head;
1542
1543         spin_lock(&page->mapping->private_lock);
1544         if (PageUptodate(page) || PageDirty(page)) {
1545                 bh = head;
1546                 do {
1547                         if (PageDirty(page))
1548                                 set_buffer_dirty(bh);
1549                         if (PageUptodate(page))
1550                                 set_buffer_uptodate(bh);
1551                         bh = bh->b_this_page;
1552                 } while (bh != head);
1553         }
1554         attach_page_buffers(page, head);
1555         spin_unlock(&page->mapping->private_lock);
1556 }
1557 EXPORT_SYMBOL(create_empty_buffers);
1558
1559 /*
1560  * We are taking a block for data and we don't want any output from any
1561  * buffer-cache aliases starting from return from that function and
1562  * until the moment when something will explicitly mark the buffer
1563  * dirty (hopefully that will not happen until we will free that block ;-)
1564  * We don't even need to mark it not-uptodate - nobody can expect
1565  * anything from a newly allocated buffer anyway. We used to used
1566  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1567  * don't want to mark the alias unmapped, for example - it would confuse
1568  * anyone who might pick it with bread() afterwards...
1569  *
1570  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1571  * be writeout I/O going on against recently-freed buffers.  We don't
1572  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1573  * only if we really need to.  That happens here.
1574  */
1575 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1576 {
1577         struct buffer_head *old_bh;
1578
1579         might_sleep();
1580
1581         old_bh = __find_get_block_slow(bdev, block);
1582         if (old_bh) {
1583                 clear_buffer_dirty(old_bh);
1584                 wait_on_buffer(old_bh);
1585                 clear_buffer_req(old_bh);
1586                 __brelse(old_bh);
1587         }
1588 }
1589 EXPORT_SYMBOL(unmap_underlying_metadata);
1590
1591 /*
1592  * NOTE! All mapped/uptodate combinations are valid:
1593  *
1594  *      Mapped  Uptodate        Meaning
1595  *
1596  *      No      No              "unknown" - must do get_block()
1597  *      No      Yes             "hole" - zero-filled
1598  *      Yes     No              "allocated" - allocated on disk, not read in
1599  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1600  *
1601  * "Dirty" is valid only with the last case (mapped+uptodate).
1602  */
1603
1604 /*
1605  * While block_write_full_page is writing back the dirty buffers under
1606  * the page lock, whoever dirtied the buffers may decide to clean them
1607  * again at any time.  We handle that by only looking at the buffer
1608  * state inside lock_buffer().
1609  *
1610  * If block_write_full_page() is called for regular writeback
1611  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1612  * locked buffer.   This only can happen if someone has written the buffer
1613  * directly, with submit_bh().  At the address_space level PageWriteback
1614  * prevents this contention from occurring.
1615  *
1616  * If block_write_full_page() is called with wbc->sync_mode ==
1617  * WB_SYNC_ALL, the writes are posted using WRITE_SYNC_PLUG; this
1618  * causes the writes to be flagged as synchronous writes, but the
1619  * block device queue will NOT be unplugged, since usually many pages
1620  * will be pushed to the out before the higher-level caller actually
1621  * waits for the writes to be completed.  The various wait functions,
1622  * such as wait_on_writeback_range() will ultimately call sync_page()
1623  * which will ultimately call blk_run_backing_dev(), which will end up
1624  * unplugging the device queue.
1625  */
1626 static int __block_write_full_page(struct inode *inode, struct page *page,
1627                         get_block_t *get_block, struct writeback_control *wbc,
1628                         bh_end_io_t *handler)
1629 {
1630         int err;
1631         sector_t block;
1632         sector_t last_block;
1633         struct buffer_head *bh, *head;
1634         const unsigned blocksize = 1 << inode->i_blkbits;
1635         int nr_underway = 0;
1636         int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1637                         WRITE_SYNC_PLUG : WRITE);
1638
1639         BUG_ON(!PageLocked(page));
1640
1641         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1642
1643         if (!page_has_buffers(page)) {
1644                 create_empty_buffers(page, blocksize,
1645                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1646         }
1647
1648         /*
1649          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1650          * here, and the (potentially unmapped) buffers may become dirty at
1651          * any time.  If a buffer becomes dirty here after we've inspected it
1652          * then we just miss that fact, and the page stays dirty.
1653          *
1654          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1655          * handle that here by just cleaning them.
1656          */
1657
1658         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1659         head = page_buffers(page);
1660         bh = head;
1661
1662         /*
1663          * Get all the dirty buffers mapped to disk addresses and
1664          * handle any aliases from the underlying blockdev's mapping.
1665          */
1666         do {
1667                 if (block > last_block) {
1668                         /*
1669                          * mapped buffers outside i_size will occur, because
1670                          * this page can be outside i_size when there is a
1671                          * truncate in progress.
1672                          */
1673                         /*
1674                          * The buffer was zeroed by block_write_full_page()
1675                          */
1676                         clear_buffer_dirty(bh);
1677                         set_buffer_uptodate(bh);
1678                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1679                            buffer_dirty(bh)) {
1680                         WARN_ON(bh->b_size != blocksize);
1681                         err = get_block(inode, block, bh, 1);
1682                         if (err)
1683                                 goto recover;
1684                         clear_buffer_delay(bh);
1685                         if (buffer_new(bh)) {
1686                                 /* blockdev mappings never come here */
1687                                 clear_buffer_new(bh);
1688                                 unmap_underlying_metadata(bh->b_bdev,
1689                                                         bh->b_blocknr);
1690                         }
1691                 }
1692                 bh = bh->b_this_page;
1693                 block++;
1694         } while (bh != head);
1695
1696         do {
1697                 if (!buffer_mapped(bh))
1698                         continue;
1699                 /*
1700                  * If it's a fully non-blocking write attempt and we cannot
1701                  * lock the buffer then redirty the page.  Note that this can
1702                  * potentially cause a busy-wait loop from pdflush and kswapd
1703                  * activity, but those code paths have their own higher-level
1704                  * throttling.
1705                  */
1706                 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1707                         lock_buffer(bh);
1708                 } else if (!trylock_buffer(bh)) {
1709                         redirty_page_for_writepage(wbc, page);
1710                         continue;
1711                 }
1712                 if (test_clear_buffer_dirty(bh)) {
1713                         mark_buffer_async_write_endio(bh, handler);
1714                 } else {
1715                         unlock_buffer(bh);
1716                 }
1717         } while ((bh = bh->b_this_page) != head);
1718
1719         /*
1720          * The page and its buffers are protected by PageWriteback(), so we can
1721          * drop the bh refcounts early.
1722          */
1723         BUG_ON(PageWriteback(page));
1724         set_page_writeback(page);
1725
1726         do {
1727                 struct buffer_head *next = bh->b_this_page;
1728                 if (buffer_async_write(bh)) {
1729                         submit_bh(write_op, bh);
1730                         nr_underway++;
1731                 }
1732                 bh = next;
1733         } while (bh != head);
1734         unlock_page(page);
1735
1736         err = 0;
1737 done:
1738         if (nr_underway == 0) {
1739                 /*
1740                  * The page was marked dirty, but the buffers were
1741                  * clean.  Someone wrote them back by hand with
1742                  * ll_rw_block/submit_bh.  A rare case.
1743                  */
1744                 end_page_writeback(page);
1745
1746                 /*
1747                  * The page and buffer_heads can be released at any time from
1748                  * here on.
1749                  */
1750         }
1751         return err;
1752
1753 recover:
1754         /*
1755          * ENOSPC, or some other error.  We may already have added some
1756          * blocks to the file, so we need to write these out to avoid
1757          * exposing stale data.
1758          * The page is currently locked and not marked for writeback
1759          */
1760         bh = head;
1761         /* Recovery: lock and submit the mapped buffers */
1762         do {
1763                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1764                     !buffer_delay(bh)) {
1765                         lock_buffer(bh);
1766                         mark_buffer_async_write_endio(bh, handler);
1767                 } else {
1768                         /*
1769                          * The buffer may have been set dirty during
1770                          * attachment to a dirty page.
1771                          */
1772                         clear_buffer_dirty(bh);
1773                 }
1774         } while ((bh = bh->b_this_page) != head);
1775         SetPageError(page);
1776         BUG_ON(PageWriteback(page));
1777         mapping_set_error(page->mapping, err);
1778         set_page_writeback(page);
1779         do {
1780                 struct buffer_head *next = bh->b_this_page;
1781                 if (buffer_async_write(bh)) {
1782                         clear_buffer_dirty(bh);
1783                         submit_bh(write_op, bh);
1784                         nr_underway++;
1785                 }
1786                 bh = next;
1787         } while (bh != head);
1788         unlock_page(page);
1789         goto done;
1790 }
1791
1792 /*
1793  * If a page has any new buffers, zero them out here, and mark them uptodate
1794  * and dirty so they'll be written out (in order to prevent uninitialised
1795  * block data from leaking). And clear the new bit.
1796  */
1797 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1798 {
1799         unsigned int block_start, block_end;
1800         struct buffer_head *head, *bh;
1801
1802         BUG_ON(!PageLocked(page));
1803         if (!page_has_buffers(page))
1804                 return;
1805
1806         bh = head = page_buffers(page);
1807         block_start = 0;
1808         do {
1809                 block_end = block_start + bh->b_size;
1810
1811                 if (buffer_new(bh)) {
1812                         if (block_end > from && block_start < to) {
1813                                 if (!PageUptodate(page)) {
1814                                         unsigned start, size;
1815
1816                                         start = max(from, block_start);
1817                                         size = min(to, block_end) - start;
1818
1819                                         zero_user(page, start, size);
1820                                         set_buffer_uptodate(bh);
1821                                 }
1822
1823                                 clear_buffer_new(bh);
1824                                 mark_buffer_dirty(bh);
1825                         }
1826                 }
1827
1828                 block_start = block_end;
1829                 bh = bh->b_this_page;
1830         } while (bh != head);
1831 }
1832 EXPORT_SYMBOL(page_zero_new_buffers);
1833
1834 static int __block_prepare_write(struct inode *inode, struct page *page,
1835                 unsigned from, unsigned to, get_block_t *get_block)
1836 {
1837         unsigned block_start, block_end;
1838         sector_t block;
1839         int err = 0;
1840         unsigned blocksize, bbits;
1841         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1842
1843         BUG_ON(!PageLocked(page));
1844         BUG_ON(from > PAGE_CACHE_SIZE);
1845         BUG_ON(to > PAGE_CACHE_SIZE);
1846         BUG_ON(from > to);
1847
1848         blocksize = 1 << inode->i_blkbits;
1849         if (!page_has_buffers(page))
1850                 create_empty_buffers(page, blocksize, 0);
1851         head = page_buffers(page);
1852
1853         bbits = inode->i_blkbits;
1854         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1855
1856         for(bh = head, block_start = 0; bh != head || !block_start;
1857             block++, block_start=block_end, bh = bh->b_this_page) {
1858                 block_end = block_start + blocksize;
1859                 if (block_end <= from || block_start >= to) {
1860                         if (PageUptodate(page)) {
1861                                 if (!buffer_uptodate(bh))
1862                                         set_buffer_uptodate(bh);
1863                         }
1864                         continue;
1865                 }
1866                 if (buffer_new(bh))
1867                         clear_buffer_new(bh);
1868                 if (!buffer_mapped(bh)) {
1869                         WARN_ON(bh->b_size != blocksize);
1870                         err = get_block(inode, block, bh, 1);
1871                         if (err)
1872                                 break;
1873                         if (buffer_new(bh)) {
1874                                 unmap_underlying_metadata(bh->b_bdev,
1875                                                         bh->b_blocknr);
1876                                 if (PageUptodate(page)) {
1877                                         clear_buffer_new(bh);
1878                                         set_buffer_uptodate(bh);
1879                                         mark_buffer_dirty(bh);
1880                                         continue;
1881                                 }
1882                                 if (block_end > to || block_start < from)
1883                                         zero_user_segments(page,
1884                                                 to, block_end,
1885                                                 block_start, from);
1886                                 continue;
1887                         }
1888                 }
1889                 if (PageUptodate(page)) {
1890                         if (!buffer_uptodate(bh))
1891                                 set_buffer_uptodate(bh);
1892                         continue; 
1893                 }
1894                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1895                     !buffer_unwritten(bh) &&
1896                      (block_start < from || block_end > to)) {
1897                         ll_rw_block(READ, 1, &bh);
1898                         *wait_bh++=bh;
1899                 }
1900         }
1901         /*
1902          * If we issued read requests - let them complete.
1903          */
1904         while(wait_bh > wait) {
1905                 wait_on_buffer(*--wait_bh);
1906                 if (!buffer_uptodate(*wait_bh))
1907                         err = -EIO;
1908         }
1909         if (unlikely(err))
1910                 page_zero_new_buffers(page, from, to);
1911         return err;
1912 }
1913
1914 static int __block_commit_write(struct inode *inode, struct page *page,
1915                 unsigned from, unsigned to)
1916 {
1917         unsigned block_start, block_end;
1918         int partial = 0;
1919         unsigned blocksize;
1920         struct buffer_head *bh, *head;
1921
1922         blocksize = 1 << inode->i_blkbits;
1923
1924         for(bh = head = page_buffers(page), block_start = 0;
1925             bh != head || !block_start;
1926             block_start=block_end, bh = bh->b_this_page) {
1927                 block_end = block_start + blocksize;
1928                 if (block_end <= from || block_start >= to) {
1929                         if (!buffer_uptodate(bh))
1930                                 partial = 1;
1931                 } else {
1932                         set_buffer_uptodate(bh);
1933                         mark_buffer_dirty(bh);
1934                 }
1935                 clear_buffer_new(bh);
1936         }
1937
1938         /*
1939          * If this is a partial write which happened to make all buffers
1940          * uptodate then we can optimize away a bogus readpage() for
1941          * the next read(). Here we 'discover' whether the page went
1942          * uptodate as a result of this (potentially partial) write.
1943          */
1944         if (!partial)
1945                 SetPageUptodate(page);
1946         return 0;
1947 }
1948
1949 /*
1950  * block_write_begin takes care of the basic task of block allocation and
1951  * bringing partial write blocks uptodate first.
1952  *
1953  * If *pagep is not NULL, then block_write_begin uses the locked page
1954  * at *pagep rather than allocating its own. In this case, the page will
1955  * not be unlocked or deallocated on failure.
1956  */
1957 int block_write_begin(struct file *file, struct address_space *mapping,
1958                         loff_t pos, unsigned len, unsigned flags,
1959                         struct page **pagep, void **fsdata,
1960                         get_block_t *get_block)
1961 {
1962         struct inode *inode = mapping->host;
1963         int status = 0;
1964         struct page *page;
1965         pgoff_t index;
1966         unsigned start, end;
1967         int ownpage = 0;
1968
1969         index = pos >> PAGE_CACHE_SHIFT;
1970         start = pos & (PAGE_CACHE_SIZE - 1);
1971         end = start + len;
1972
1973         page = *pagep;
1974         if (page == NULL) {
1975                 ownpage = 1;
1976                 page = grab_cache_page_write_begin(mapping, index, flags);
1977                 if (!page) {
1978                         status = -ENOMEM;
1979                         goto out;
1980                 }
1981                 *pagep = page;
1982         } else
1983                 BUG_ON(!PageLocked(page));
1984
1985         status = __block_prepare_write(inode, page, start, end, get_block);
1986         if (unlikely(status)) {
1987                 ClearPageUptodate(page);
1988
1989                 if (ownpage) {
1990                         unlock_page(page);
1991                         page_cache_release(page);
1992                         *pagep = NULL;
1993
1994                         /*
1995                          * prepare_write() may have instantiated a few blocks
1996                          * outside i_size.  Trim these off again. Don't need
1997                          * i_size_read because we hold i_mutex.
1998                          */
1999                         if (pos + len > inode->i_size)
2000                                 vmtruncate(inode, inode->i_size);
2001                 }
2002         }
2003
2004 out:
2005         return status;
2006 }
2007 EXPORT_SYMBOL(block_write_begin);
2008
2009 int block_write_end(struct file *file, struct address_space *mapping,
2010                         loff_t pos, unsigned len, unsigned copied,
2011                         struct page *page, void *fsdata)
2012 {
2013         struct inode *inode = mapping->host;
2014         unsigned start;
2015
2016         start = pos & (PAGE_CACHE_SIZE - 1);
2017
2018         if (unlikely(copied < len)) {
2019                 /*
2020                  * The buffers that were written will now be uptodate, so we
2021                  * don't have to worry about a readpage reading them and
2022                  * overwriting a partial write. However if we have encountered
2023                  * a short write and only partially written into a buffer, it
2024                  * will not be marked uptodate, so a readpage might come in and
2025                  * destroy our partial write.
2026                  *
2027                  * Do the simplest thing, and just treat any short write to a
2028                  * non uptodate page as a zero-length write, and force the
2029                  * caller to redo the whole thing.
2030                  */
2031                 if (!PageUptodate(page))
2032                         copied = 0;
2033
2034                 page_zero_new_buffers(page, start+copied, start+len);
2035         }
2036         flush_dcache_page(page);
2037
2038         /* This could be a short (even 0-length) commit */
2039         __block_commit_write(inode, page, start, start+copied);
2040
2041         return copied;
2042 }
2043 EXPORT_SYMBOL(block_write_end);
2044
2045 int generic_write_end(struct file *file, struct address_space *mapping,
2046                         loff_t pos, unsigned len, unsigned copied,
2047                         struct page *page, void *fsdata)
2048 {
2049         struct inode *inode = mapping->host;
2050         int i_size_changed = 0;
2051
2052         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2053
2054         /*
2055          * No need to use i_size_read() here, the i_size
2056          * cannot change under us because we hold i_mutex.
2057          *
2058          * But it's important to update i_size while still holding page lock:
2059          * page writeout could otherwise come in and zero beyond i_size.
2060          */
2061         if (pos+copied > inode->i_size) {
2062                 i_size_write(inode, pos+copied);
2063                 i_size_changed = 1;
2064         }
2065
2066         unlock_page(page);
2067         page_cache_release(page);
2068
2069         /*
2070          * Don't mark the inode dirty under page lock. First, it unnecessarily
2071          * makes the holding time of page lock longer. Second, it forces lock
2072          * ordering of page lock and transaction start for journaling
2073          * filesystems.
2074          */
2075         if (i_size_changed)
2076                 mark_inode_dirty(inode);
2077
2078         return copied;
2079 }
2080 EXPORT_SYMBOL(generic_write_end);
2081
2082 /*
2083  * block_is_partially_uptodate checks whether buffers within a page are
2084  * uptodate or not.
2085  *
2086  * Returns true if all buffers which correspond to a file portion
2087  * we want to read are uptodate.
2088  */
2089 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2090                                         unsigned long from)
2091 {
2092         struct inode *inode = page->mapping->host;
2093         unsigned block_start, block_end, blocksize;
2094         unsigned to;
2095         struct buffer_head *bh, *head;
2096         int ret = 1;
2097
2098         if (!page_has_buffers(page))
2099                 return 0;
2100
2101         blocksize = 1 << inode->i_blkbits;
2102         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2103         to = from + to;
2104         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2105                 return 0;
2106
2107         head = page_buffers(page);
2108         bh = head;
2109         block_start = 0;
2110         do {
2111                 block_end = block_start + blocksize;
2112                 if (block_end > from && block_start < to) {
2113                         if (!buffer_uptodate(bh)) {
2114                                 ret = 0;
2115                                 break;
2116                         }
2117                         if (block_end >= to)
2118                                 break;
2119                 }
2120                 block_start = block_end;
2121                 bh = bh->b_this_page;
2122         } while (bh != head);
2123
2124         return ret;
2125 }
2126 EXPORT_SYMBOL(block_is_partially_uptodate);
2127
2128 /*
2129  * Generic "read page" function for block devices that have the normal
2130  * get_block functionality. This is most of the block device filesystems.
2131  * Reads the page asynchronously --- the unlock_buffer() and
2132  * set/clear_buffer_uptodate() functions propagate buffer state into the
2133  * page struct once IO has completed.
2134  */
2135 int block_read_full_page(struct page *page, get_block_t *get_block)
2136 {
2137         struct inode *inode = page->mapping->host;
2138         sector_t iblock, lblock;
2139         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2140         unsigned int blocksize;
2141         int nr, i;
2142         int fully_mapped = 1;
2143
2144         BUG_ON(!PageLocked(page));
2145         blocksize = 1 << inode->i_blkbits;
2146         if (!page_has_buffers(page))
2147                 create_empty_buffers(page, blocksize, 0);
2148         head = page_buffers(page);
2149
2150         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2151         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2152         bh = head;
2153         nr = 0;
2154         i = 0;
2155
2156         do {
2157                 if (buffer_uptodate(bh))
2158                         continue;
2159
2160                 if (!buffer_mapped(bh)) {
2161                         int err = 0;
2162
2163                         fully_mapped = 0;
2164                         if (iblock < lblock) {
2165                                 WARN_ON(bh->b_size != blocksize);
2166                                 err = get_block(inode, iblock, bh, 0);
2167                                 if (err)
2168                                         SetPageError(page);
2169                         }
2170                         if (!buffer_mapped(bh)) {
2171                                 zero_user(page, i * blocksize, blocksize);
2172                                 if (!err)
2173                                         set_buffer_uptodate(bh);
2174                                 continue;
2175                         }
2176                         /*
2177                          * get_block() might have updated the buffer
2178                          * synchronously
2179                          */
2180                         if (buffer_uptodate(bh))
2181                                 continue;
2182                 }
2183                 arr[nr++] = bh;
2184         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2185
2186         if (fully_mapped)
2187                 SetPageMappedToDisk(page);
2188
2189         if (!nr) {
2190                 /*
2191                  * All buffers are uptodate - we can set the page uptodate
2192                  * as well. But not if get_block() returned an error.
2193                  */
2194                 if (!PageError(page))
2195                         SetPageUptodate(page);
2196                 unlock_page(page);
2197                 return 0;
2198         }
2199
2200         /* Stage two: lock the buffers */
2201         for (i = 0; i < nr; i++) {
2202                 bh = arr[i];
2203                 lock_buffer(bh);
2204                 mark_buffer_async_read(bh);
2205         }
2206
2207         /*
2208          * Stage 3: start the IO.  Check for uptodateness
2209          * inside the buffer lock in case another process reading
2210          * the underlying blockdev brought it uptodate (the sct fix).
2211          */
2212         for (i = 0; i < nr; i++) {
2213                 bh = arr[i];
2214                 if (buffer_uptodate(bh))
2215                         end_buffer_async_read(bh, 1);
2216                 else
2217                         submit_bh(READ, bh);
2218         }
2219         return 0;
2220 }
2221
2222 /* utility function for filesystems that need to do work on expanding
2223  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2224  * deal with the hole.  
2225  */
2226 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2227 {
2228         struct address_space *mapping = inode->i_mapping;
2229         struct page *page;
2230         void *fsdata;
2231         unsigned long limit;
2232         int err;
2233
2234         err = -EFBIG;
2235         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2236         if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2237                 send_sig(SIGXFSZ, current, 0);
2238                 goto out;
2239         }
2240         if (size > inode->i_sb->s_maxbytes)
2241                 goto out;
2242
2243         err = pagecache_write_begin(NULL, mapping, size, 0,
2244                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2245                                 &page, &fsdata);
2246         if (err)
2247                 goto out;
2248
2249         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2250         BUG_ON(err > 0);
2251
2252 out:
2253         return err;
2254 }
2255
2256 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2257                             loff_t pos, loff_t *bytes)
2258 {
2259         struct inode *inode = mapping->host;
2260         unsigned blocksize = 1 << inode->i_blkbits;
2261         struct page *page;
2262         void *fsdata;
2263         pgoff_t index, curidx;
2264         loff_t curpos;
2265         unsigned zerofrom, offset, len;
2266         int err = 0;
2267
2268         index = pos >> PAGE_CACHE_SHIFT;
2269         offset = pos & ~PAGE_CACHE_MASK;
2270
2271         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2272                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2273                 if (zerofrom & (blocksize-1)) {
2274                         *bytes |= (blocksize-1);
2275                         (*bytes)++;
2276                 }
2277                 len = PAGE_CACHE_SIZE - zerofrom;
2278
2279                 err = pagecache_write_begin(file, mapping, curpos, len,
2280                                                 AOP_FLAG_UNINTERRUPTIBLE,
2281                                                 &page, &fsdata);
2282                 if (err)
2283                         goto out;
2284                 zero_user(page, zerofrom, len);
2285                 err = pagecache_write_end(file, mapping, curpos, len, len,
2286                                                 page, fsdata);
2287                 if (err < 0)
2288                         goto out;
2289                 BUG_ON(err != len);
2290                 err = 0;
2291
2292                 balance_dirty_pages_ratelimited(mapping);
2293         }
2294
2295         /* page covers the boundary, find the boundary offset */
2296         if (index == curidx) {
2297                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2298                 /* if we will expand the thing last block will be filled */
2299                 if (offset <= zerofrom) {
2300                         goto out;
2301                 }
2302                 if (zerofrom & (blocksize-1)) {
2303                         *bytes |= (blocksize-1);
2304                         (*bytes)++;
2305                 }
2306                 len = offset - zerofrom;
2307
2308                 err = pagecache_write_begin(file, mapping, curpos, len,
2309                                                 AOP_FLAG_UNINTERRUPTIBLE,
2310                                                 &page, &fsdata);
2311                 if (err)
2312                         goto out;
2313                 zero_user(page, zerofrom, len);
2314                 err = pagecache_write_end(file, mapping, curpos, len, len,
2315                                                 page, fsdata);
2316                 if (err < 0)
2317                         goto out;
2318                 BUG_ON(err != len);
2319                 err = 0;
2320         }
2321 out:
2322         return err;
2323 }
2324
2325 /*
2326  * For moronic filesystems that do not allow holes in file.
2327  * We may have to extend the file.
2328  */
2329 int cont_write_begin(struct file *file, struct address_space *mapping,
2330                         loff_t pos, unsigned len, unsigned flags,
2331                         struct page **pagep, void **fsdata,
2332                         get_block_t *get_block, loff_t *bytes)
2333 {
2334         struct inode *inode = mapping->host;
2335         unsigned blocksize = 1 << inode->i_blkbits;
2336         unsigned zerofrom;
2337         int err;
2338
2339         err = cont_expand_zero(file, mapping, pos, bytes);
2340         if (err)
2341                 goto out;
2342
2343         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2344         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2345                 *bytes |= (blocksize-1);
2346                 (*bytes)++;
2347         }
2348
2349         *pagep = NULL;
2350         err = block_write_begin(file, mapping, pos, len,
2351                                 flags, pagep, fsdata, get_block);
2352 out:
2353         return err;
2354 }
2355
2356 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2357                         get_block_t *get_block)
2358 {
2359         struct inode *inode = page->mapping->host;
2360         int err = __block_prepare_write(inode, page, from, to, get_block);
2361         if (err)
2362                 ClearPageUptodate(page);
2363         return err;
2364 }
2365
2366 int block_commit_write(struct page *page, unsigned from, unsigned to)
2367 {
2368         struct inode *inode = page->mapping->host;
2369         __block_commit_write(inode,page,from,to);
2370         return 0;
2371 }
2372
2373 /*
2374  * block_page_mkwrite() is not allowed to change the file size as it gets
2375  * called from a page fault handler when a page is first dirtied. Hence we must
2376  * be careful to check for EOF conditions here. We set the page up correctly
2377  * for a written page which means we get ENOSPC checking when writing into
2378  * holes and correct delalloc and unwritten extent mapping on filesystems that
2379  * support these features.
2380  *
2381  * We are not allowed to take the i_mutex here so we have to play games to
2382  * protect against truncate races as the page could now be beyond EOF.  Because
2383  * vmtruncate() writes the inode size before removing pages, once we have the
2384  * page lock we can determine safely if the page is beyond EOF. If it is not
2385  * beyond EOF, then the page is guaranteed safe against truncation until we
2386  * unlock the page.
2387  */
2388 int
2389 block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2390                    get_block_t get_block)
2391 {
2392         struct page *page = vmf->page;
2393         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2394         unsigned long end;
2395         loff_t size;
2396         int ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
2397
2398         lock_page(page);
2399         size = i_size_read(inode);
2400         if ((page->mapping != inode->i_mapping) ||
2401             (page_offset(page) > size)) {
2402                 /* page got truncated out from underneath us */
2403                 unlock_page(page);
2404                 goto out;
2405         }
2406
2407         /* page is wholly or partially inside EOF */
2408         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2409                 end = size & ~PAGE_CACHE_MASK;
2410         else
2411                 end = PAGE_CACHE_SIZE;
2412
2413         ret = block_prepare_write(page, 0, end, get_block);
2414         if (!ret)
2415                 ret = block_commit_write(page, 0, end);
2416
2417         if (unlikely(ret)) {
2418                 unlock_page(page);
2419                 if (ret == -ENOMEM)
2420                         ret = VM_FAULT_OOM;
2421                 else /* -ENOSPC, -EIO, etc */
2422                         ret = VM_FAULT_SIGBUS;
2423         } else
2424                 ret = VM_FAULT_LOCKED;
2425
2426 out:
2427         return ret;
2428 }
2429
2430 /*
2431  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2432  * immediately, while under the page lock.  So it needs a special end_io
2433  * handler which does not touch the bh after unlocking it.
2434  */
2435 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2436 {
2437         __end_buffer_read_notouch(bh, uptodate);
2438 }
2439
2440 /*
2441  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2442  * the page (converting it to circular linked list and taking care of page
2443  * dirty races).
2444  */
2445 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2446 {
2447         struct buffer_head *bh;
2448
2449         BUG_ON(!PageLocked(page));
2450
2451         spin_lock(&page->mapping->private_lock);
2452         bh = head;
2453         do {
2454                 if (PageDirty(page))
2455                         set_buffer_dirty(bh);
2456                 if (!bh->b_this_page)
2457                         bh->b_this_page = head;
2458                 bh = bh->b_this_page;
2459         } while (bh != head);
2460         attach_page_buffers(page, head);
2461         spin_unlock(&page->mapping->private_lock);
2462 }
2463
2464 /*
2465  * On entry, the page is fully not uptodate.
2466  * On exit the page is fully uptodate in the areas outside (from,to)
2467  */
2468 int nobh_write_begin(struct file *file, struct address_space *mapping,
2469                         loff_t pos, unsigned len, unsigned flags,
2470                         struct page **pagep, void **fsdata,
2471                         get_block_t *get_block)
2472 {
2473         struct inode *inode = mapping->host;
2474         const unsigned blkbits = inode->i_blkbits;
2475         const unsigned blocksize = 1 << blkbits;
2476         struct buffer_head *head, *bh;
2477         struct page *page;
2478         pgoff_t index;
2479         unsigned from, to;
2480         unsigned block_in_page;
2481         unsigned block_start, block_end;
2482         sector_t block_in_file;
2483         int nr_reads = 0;
2484         int ret = 0;
2485         int is_mapped_to_disk = 1;
2486
2487         index = pos >> PAGE_CACHE_SHIFT;
2488         from = pos & (PAGE_CACHE_SIZE - 1);
2489         to = from + len;
2490
2491         page = grab_cache_page_write_begin(mapping, index, flags);
2492         if (!page)
2493                 return -ENOMEM;
2494         *pagep = page;
2495         *fsdata = NULL;
2496
2497         if (page_has_buffers(page)) {
2498                 unlock_page(page);
2499                 page_cache_release(page);
2500                 *pagep = NULL;
2501                 return block_write_begin(file, mapping, pos, len, flags, pagep,
2502                                         fsdata, get_block);
2503         }
2504
2505         if (PageMappedToDisk(page))
2506                 return 0;
2507
2508         /*
2509          * Allocate buffers so that we can keep track of state, and potentially
2510          * attach them to the page if an error occurs. In the common case of
2511          * no error, they will just be freed again without ever being attached
2512          * to the page (which is all OK, because we're under the page lock).
2513          *
2514          * Be careful: the buffer linked list is a NULL terminated one, rather
2515          * than the circular one we're used to.
2516          */
2517         head = alloc_page_buffers(page, blocksize, 0);
2518         if (!head) {
2519                 ret = -ENOMEM;
2520                 goto out_release;
2521         }
2522
2523         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2524
2525         /*
2526          * We loop across all blocks in the page, whether or not they are
2527          * part of the affected region.  This is so we can discover if the
2528          * page is fully mapped-to-disk.
2529          */
2530         for (block_start = 0, block_in_page = 0, bh = head;
2531                   block_start < PAGE_CACHE_SIZE;
2532                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2533                 int create;
2534
2535                 block_end = block_start + blocksize;
2536                 bh->b_state = 0;
2537                 create = 1;
2538                 if (block_start >= to)
2539                         create = 0;
2540                 ret = get_block(inode, block_in_file + block_in_page,
2541                                         bh, create);
2542                 if (ret)
2543                         goto failed;
2544                 if (!buffer_mapped(bh))
2545                         is_mapped_to_disk = 0;
2546                 if (buffer_new(bh))
2547                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2548                 if (PageUptodate(page)) {
2549                         set_buffer_uptodate(bh);
2550                         continue;
2551                 }
2552                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2553                         zero_user_segments(page, block_start, from,
2554                                                         to, block_end);
2555                         continue;
2556                 }
2557                 if (buffer_uptodate(bh))
2558                         continue;       /* reiserfs does this */
2559                 if (block_start < from || block_end > to) {
2560                         lock_buffer(bh);
2561                         bh->b_end_io = end_buffer_read_nobh;
2562                         submit_bh(READ, bh);
2563                         nr_reads++;
2564                 }
2565         }
2566
2567         if (nr_reads) {
2568                 /*
2569                  * The page is locked, so these buffers are protected from
2570                  * any VM or truncate activity.  Hence we don't need to care
2571                  * for the buffer_head refcounts.
2572                  */
2573                 for (bh = head; bh; bh = bh->b_this_page) {
2574                         wait_on_buffer(bh);
2575                         if (!buffer_uptodate(bh))
2576                                 ret = -EIO;
2577                 }
2578                 if (ret)
2579                         goto failed;
2580         }
2581
2582         if (is_mapped_to_disk)
2583                 SetPageMappedToDisk(page);
2584
2585         *fsdata = head; /* to be released by nobh_write_end */
2586
2587         return 0;
2588
2589 failed:
2590         BUG_ON(!ret);
2591         /*
2592          * Error recovery is a bit difficult. We need to zero out blocks that
2593          * were newly allocated, and dirty them to ensure they get written out.
2594          * Buffers need to be attached to the page at this point, otherwise
2595          * the handling of potential IO errors during writeout would be hard
2596          * (could try doing synchronous writeout, but what if that fails too?)
2597          */
2598         attach_nobh_buffers(page, head);
2599         page_zero_new_buffers(page, from, to);
2600
2601 out_release:
2602         unlock_page(page);
2603         page_cache_release(page);
2604         *pagep = NULL;
2605
2606         if (pos + len > inode->i_size)
2607                 vmtruncate(inode, inode->i_size);
2608
2609         return ret;
2610 }
2611 EXPORT_SYMBOL(nobh_write_begin);
2612
2613 int nobh_write_end(struct file *file, struct address_space *mapping,
2614                         loff_t pos, unsigned len, unsigned copied,
2615                         struct page *page, void *fsdata)
2616 {
2617         struct inode *inode = page->mapping->host;
2618         struct buffer_head *head = fsdata;
2619         struct buffer_head *bh;
2620         BUG_ON(fsdata != NULL && page_has_buffers(page));
2621
2622         if (unlikely(copied < len) && head)
2623                 attach_nobh_buffers(page, head);
2624         if (page_has_buffers(page))
2625                 return generic_write_end(file, mapping, pos, len,
2626                                         copied, page, fsdata);
2627
2628         SetPageUptodate(page);
2629         set_page_dirty(page);
2630         if (pos+copied > inode->i_size) {
2631                 i_size_write(inode, pos+copied);
2632                 mark_inode_dirty(inode);
2633         }
2634
2635         unlock_page(page);
2636         page_cache_release(page);
2637
2638         while (head) {
2639                 bh = head;
2640                 head = head->b_this_page;
2641                 free_buffer_head(bh);
2642         }
2643
2644         return copied;
2645 }
2646 EXPORT_SYMBOL(nobh_write_end);
2647
2648 /*
2649  * nobh_writepage() - based on block_full_write_page() except
2650  * that it tries to operate without attaching bufferheads to
2651  * the page.
2652  */
2653 int nobh_writepage(struct page *page, get_block_t *get_block,
2654                         struct writeback_control *wbc)
2655 {
2656         struct inode * const inode = page->mapping->host;
2657         loff_t i_size = i_size_read(inode);
2658         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2659         unsigned offset;
2660         int ret;
2661
2662         /* Is the page fully inside i_size? */
2663         if (page->index < end_index)
2664                 goto out;
2665
2666         /* Is the page fully outside i_size? (truncate in progress) */
2667         offset = i_size & (PAGE_CACHE_SIZE-1);
2668         if (page->index >= end_index+1 || !offset) {
2669                 /*
2670                  * The page may have dirty, unmapped buffers.  For example,
2671                  * they may have been added in ext3_writepage().  Make them
2672                  * freeable here, so the page does not leak.
2673                  */
2674 #if 0
2675                 /* Not really sure about this  - do we need this ? */
2676                 if (page->mapping->a_ops->invalidatepage)
2677                         page->mapping->a_ops->invalidatepage(page, offset);
2678 #endif
2679                 unlock_page(page);
2680                 return 0; /* don't care */
2681         }
2682
2683         /*
2684          * The page straddles i_size.  It must be zeroed out on each and every
2685          * writepage invocation because it may be mmapped.  "A file is mapped
2686          * in multiples of the page size.  For a file that is not a multiple of
2687          * the  page size, the remaining memory is zeroed when mapped, and
2688          * writes to that region are not written out to the file."
2689          */
2690         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2691 out:
2692         ret = mpage_writepage(page, get_block, wbc);
2693         if (ret == -EAGAIN)
2694                 ret = __block_write_full_page(inode, page, get_block, wbc,
2695                                               end_buffer_async_write);
2696         return ret;
2697 }
2698 EXPORT_SYMBOL(nobh_writepage);
2699
2700 int nobh_truncate_page(struct address_space *mapping,
2701                         loff_t from, get_block_t *get_block)
2702 {
2703         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2704         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2705         unsigned blocksize;
2706         sector_t iblock;
2707         unsigned length, pos;
2708         struct inode *inode = mapping->host;
2709         struct page *page;
2710         struct buffer_head map_bh;
2711         int err;
2712
2713         blocksize = 1 << inode->i_blkbits;
2714         length = offset & (blocksize - 1);
2715
2716         /* Block boundary? Nothing to do */
2717         if (!length)
2718                 return 0;
2719
2720         length = blocksize - length;
2721         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2722
2723         page = grab_cache_page(mapping, index);
2724         err = -ENOMEM;
2725         if (!page)
2726                 goto out;
2727
2728         if (page_has_buffers(page)) {
2729 has_buffers:
2730                 unlock_page(page);
2731                 page_cache_release(page);
2732                 return block_truncate_page(mapping, from, get_block);
2733         }
2734
2735         /* Find the buffer that contains "offset" */
2736         pos = blocksize;
2737         while (offset >= pos) {
2738                 iblock++;
2739                 pos += blocksize;
2740         }
2741
2742         map_bh.b_size = blocksize;
2743         map_bh.b_state = 0;
2744         err = get_block(inode, iblock, &map_bh, 0);
2745         if (err)
2746                 goto unlock;
2747         /* unmapped? It's a hole - nothing to do */
2748         if (!buffer_mapped(&map_bh))
2749                 goto unlock;
2750
2751         /* Ok, it's mapped. Make sure it's up-to-date */
2752         if (!PageUptodate(page)) {
2753                 err = mapping->a_ops->readpage(NULL, page);
2754                 if (err) {
2755                         page_cache_release(page);
2756                         goto out;
2757                 }
2758                 lock_page(page);
2759                 if (!PageUptodate(page)) {
2760                         err = -EIO;
2761                         goto unlock;
2762                 }
2763                 if (page_has_buffers(page))
2764                         goto has_buffers;
2765         }
2766         zero_user(page, offset, length);
2767         set_page_dirty(page);
2768         err = 0;
2769
2770 unlock:
2771         unlock_page(page);
2772         page_cache_release(page);
2773 out:
2774         return err;
2775 }
2776 EXPORT_SYMBOL(nobh_truncate_page);
2777
2778 int block_truncate_page(struct address_space *mapping,
2779                         loff_t from, get_block_t *get_block)
2780 {
2781         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2782         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2783         unsigned blocksize;
2784         sector_t iblock;
2785         unsigned length, pos;
2786         struct inode *inode = mapping->host;
2787         struct page *page;
2788         struct buffer_head *bh;
2789         int err;
2790
2791         blocksize = 1 << inode->i_blkbits;
2792         length = offset & (blocksize - 1);
2793
2794         /* Block boundary? Nothing to do */
2795         if (!length)
2796                 return 0;
2797
2798         length = blocksize - length;
2799         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2800         
2801         page = grab_cache_page(mapping, index);
2802         err = -ENOMEM;
2803         if (!page)
2804                 goto out;
2805
2806         if (!page_has_buffers(page))
2807                 create_empty_buffers(page, blocksize, 0);
2808
2809         /* Find the buffer that contains "offset" */
2810         bh = page_buffers(page);
2811         pos = blocksize;
2812         while (offset >= pos) {
2813                 bh = bh->b_this_page;
2814                 iblock++;
2815                 pos += blocksize;
2816         }
2817
2818         err = 0;
2819         if (!buffer_mapped(bh)) {
2820                 WARN_ON(bh->b_size != blocksize);
2821                 err = get_block(inode, iblock, bh, 0);
2822                 if (err)
2823                         goto unlock;
2824                 /* unmapped? It's a hole - nothing to do */
2825                 if (!buffer_mapped(bh))
2826                         goto unlock;
2827         }
2828
2829         /* Ok, it's mapped. Make sure it's up-to-date */
2830         if (PageUptodate(page))
2831                 set_buffer_uptodate(bh);
2832
2833         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2834                 err = -EIO;
2835                 ll_rw_block(READ, 1, &bh);
2836                 wait_on_buffer(bh);
2837                 /* Uhhuh. Read error. Complain and punt. */
2838                 if (!buffer_uptodate(bh))
2839                         goto unlock;
2840         }
2841
2842         zero_user(page, offset, length);
2843         mark_buffer_dirty(bh);
2844         err = 0;
2845
2846 unlock:
2847         unlock_page(page);
2848         page_cache_release(page);
2849 out:
2850         return err;
2851 }
2852
2853 /*
2854  * The generic ->writepage function for buffer-backed address_spaces
2855  * this form passes in the end_io handler used to finish the IO.
2856  */
2857 int block_write_full_page_endio(struct page *page, get_block_t *get_block,
2858                         struct writeback_control *wbc, bh_end_io_t *handler)
2859 {
2860         struct inode * const inode = page->mapping->host;
2861         loff_t i_size = i_size_read(inode);
2862         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2863         unsigned offset;
2864
2865         /* Is the page fully inside i_size? */
2866         if (page->index < end_index)
2867                 return __block_write_full_page(inode, page, get_block, wbc,
2868                                                handler);
2869
2870         /* Is the page fully outside i_size? (truncate in progress) */
2871         offset = i_size & (PAGE_CACHE_SIZE-1);
2872         if (page->index >= end_index+1 || !offset) {
2873                 /*
2874                  * The page may have dirty, unmapped buffers.  For example,
2875                  * they may have been added in ext3_writepage().  Make them
2876                  * freeable here, so the page does not leak.
2877                  */
2878                 do_invalidatepage(page, 0);
2879                 unlock_page(page);
2880                 return 0; /* don't care */
2881         }
2882
2883         /*
2884          * The page straddles i_size.  It must be zeroed out on each and every
2885          * writepage invokation because it may be mmapped.  "A file is mapped
2886          * in multiples of the page size.  For a file that is not a multiple of
2887          * the  page size, the remaining memory is zeroed when mapped, and
2888          * writes to that region are not written out to the file."
2889          */
2890         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2891         return __block_write_full_page(inode, page, get_block, wbc, handler);
2892 }
2893
2894 /*
2895  * The generic ->writepage function for buffer-backed address_spaces
2896  */
2897 int block_write_full_page(struct page *page, get_block_t *get_block,
2898                         struct writeback_control *wbc)
2899 {
2900         return block_write_full_page_endio(page, get_block, wbc,
2901                                            end_buffer_async_write);
2902 }
2903
2904
2905 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2906                             get_block_t *get_block)
2907 {
2908         struct buffer_head tmp;
2909         struct inode *inode = mapping->host;
2910         tmp.b_state = 0;
2911         tmp.b_blocknr = 0;
2912         tmp.b_size = 1 << inode->i_blkbits;
2913         get_block(inode, block, &tmp, 0);
2914         return tmp.b_blocknr;
2915 }
2916
2917 static void end_bio_bh_io_sync(struct bio *bio, int err)
2918 {
2919         struct buffer_head *bh = bio->bi_private;
2920
2921         if (err == -EOPNOTSUPP) {
2922                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2923                 set_bit(BH_Eopnotsupp, &bh->b_state);
2924         }
2925
2926         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2927                 set_bit(BH_Quiet, &bh->b_state);
2928
2929         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2930         bio_put(bio);
2931 }
2932
2933 int submit_bh(int rw, struct buffer_head * bh)
2934 {
2935         struct bio *bio;
2936         int ret = 0;
2937
2938         BUG_ON(!buffer_locked(bh));
2939         BUG_ON(!buffer_mapped(bh));
2940         BUG_ON(!bh->b_end_io);
2941         BUG_ON(buffer_delay(bh));
2942         BUG_ON(buffer_unwritten(bh));
2943
2944         /*
2945          * Mask in barrier bit for a write (could be either a WRITE or a
2946          * WRITE_SYNC
2947          */
2948         if (buffer_ordered(bh) && (rw & WRITE))
2949                 rw |= WRITE_BARRIER;
2950
2951         /*
2952          * Only clear out a write error when rewriting
2953          */
2954         if (test_set_buffer_req(bh) && (rw & WRITE))
2955                 clear_buffer_write_io_error(bh);
2956
2957         /*
2958          * from here on down, it's all bio -- do the initial mapping,
2959          * submit_bio -> generic_make_request may further map this bio around
2960          */
2961         bio = bio_alloc(GFP_NOIO, 1);
2962
2963         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2964         bio->bi_bdev = bh->b_bdev;
2965         bio->bi_io_vec[0].bv_page = bh->b_page;
2966         bio->bi_io_vec[0].bv_len = bh->b_size;
2967         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2968
2969         bio->bi_vcnt = 1;
2970         bio->bi_idx = 0;
2971         bio->bi_size = bh->b_size;
2972
2973         bio->bi_end_io = end_bio_bh_io_sync;
2974         bio->bi_private = bh;
2975
2976         bio_get(bio);
2977         submit_bio(rw, bio);
2978
2979         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2980                 ret = -EOPNOTSUPP;
2981
2982         bio_put(bio);
2983         return ret;
2984 }
2985
2986 /**
2987  * ll_rw_block: low-level access to block devices (DEPRECATED)
2988  * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
2989  * @nr: number of &struct buffer_heads in the array
2990  * @bhs: array of pointers to &struct buffer_head
2991  *
2992  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2993  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2994  * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2995  * are sent to disk. The fourth %READA option is described in the documentation
2996  * for generic_make_request() which ll_rw_block() calls.
2997  *
2998  * This function drops any buffer that it cannot get a lock on (with the
2999  * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
3000  * clean when doing a write request, and any buffer that appears to be
3001  * up-to-date when doing read request.  Further it marks as clean buffers that
3002  * are processed for writing (the buffer cache won't assume that they are
3003  * actually clean until the buffer gets unlocked).
3004  *
3005  * ll_rw_block sets b_end_io to simple completion handler that marks
3006  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
3007  * any waiters. 
3008  *
3009  * All of the buffers must be for the same device, and must also be a
3010  * multiple of the current approved size for the device.
3011  */
3012 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
3013 {
3014         int i;
3015
3016         for (i = 0; i < nr; i++) {
3017                 struct buffer_head *bh = bhs[i];
3018
3019                 if (rw == SWRITE || rw == SWRITE_SYNC || rw == SWRITE_SYNC_PLUG)
3020                         lock_buffer(bh);
3021                 else if (!trylock_buffer(bh))
3022                         continue;
3023
3024                 if (rw == WRITE || rw == SWRITE || rw == SWRITE_SYNC ||
3025                     rw == SWRITE_SYNC_PLUG) {
3026                         if (test_clear_buffer_dirty(bh)) {
3027                                 bh->b_end_io = end_buffer_write_sync;
3028                                 get_bh(bh);
3029                                 if (rw == SWRITE_SYNC)
3030                                         submit_bh(WRITE_SYNC, bh);
3031                                 else
3032                                         submit_bh(WRITE, bh);
3033                                 continue;
3034                         }
3035                 } else {
3036                         if (!buffer_uptodate(bh)) {
3037                                 bh->b_end_io = end_buffer_read_sync;
3038                                 get_bh(bh);
3039                                 submit_bh(rw, bh);
3040                                 continue;
3041                         }
3042                 }
3043                 unlock_buffer(bh);
3044         }
3045 }
3046
3047 /*
3048  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3049  * and then start new I/O and then wait upon it.  The caller must have a ref on
3050  * the buffer_head.
3051  */
3052 int sync_dirty_buffer(struct buffer_head *bh)
3053 {
3054         int ret = 0;
3055
3056         WARN_ON(atomic_read(&bh->b_count) < 1);
3057         lock_buffer(bh);
3058         if (test_clear_buffer_dirty(bh)) {
3059                 get_bh(bh);
3060                 bh->b_end_io = end_buffer_write_sync;
3061                 ret = submit_bh(WRITE_SYNC, bh);
3062                 wait_on_buffer(bh);
3063                 if (buffer_eopnotsupp(bh)) {
3064                         clear_buffer_eopnotsupp(bh);
3065                         ret = -EOPNOTSUPP;
3066                 }
3067                 if (!ret && !buffer_uptodate(bh))
3068                         ret = -EIO;
3069         } else {
3070                 unlock_buffer(bh);
3071         }
3072         return ret;
3073 }
3074
3075 /*
3076  * try_to_free_buffers() checks if all the buffers on this particular page
3077  * are unused, and releases them if so.
3078  *
3079  * Exclusion against try_to_free_buffers may be obtained by either
3080  * locking the page or by holding its mapping's private_lock.
3081  *
3082  * If the page is dirty but all the buffers are clean then we need to
3083  * be sure to mark the page clean as well.  This is because the page
3084  * may be against a block device, and a later reattachment of buffers
3085  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3086  * filesystem data on the same device.
3087  *
3088  * The same applies to regular filesystem pages: if all the buffers are
3089  * clean then we set the page clean and proceed.  To do that, we require
3090  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3091  * private_lock.
3092  *
3093  * try_to_free_buffers() is non-blocking.
3094  */
3095 static inline int buffer_busy(struct buffer_head *bh)
3096 {
3097         return atomic_read(&bh->b_count) |
3098                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3099 }
3100
3101 static int
3102 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3103 {
3104         struct buffer_head *head = page_buffers(page);
3105         struct buffer_head *bh;
3106
3107         bh = head;
3108         do {
3109                 if (buffer_write_io_error(bh) && page->mapping)
3110                         set_bit(AS_EIO, &page->mapping->flags);
3111                 if (buffer_busy(bh))
3112                         goto failed;
3113                 bh = bh->b_this_page;
3114         } while (bh != head);
3115
3116         do {
3117                 struct buffer_head *next = bh->b_this_page;
3118
3119                 if (bh->b_assoc_map)
3120                         __remove_assoc_queue(bh);
3121                 bh = next;
3122         } while (bh != head);
3123         *buffers_to_free = head;
3124         __clear_page_buffers(page);
3125         return 1;
3126 failed:
3127         return 0;
3128 }
3129
3130 int try_to_free_buffers(struct page *page)
3131 {
3132         struct address_space * const mapping = page->mapping;
3133         struct buffer_head *buffers_to_free = NULL;
3134         int ret = 0;
3135
3136         BUG_ON(!PageLocked(page));
3137         if (PageWriteback(page))
3138                 return 0;
3139
3140         if (mapping == NULL) {          /* can this still happen? */
3141                 ret = drop_buffers(page, &buffers_to_free);
3142                 goto out;
3143         }
3144
3145         spin_lock(&mapping->private_lock);
3146         ret = drop_buffers(page, &buffers_to_free);
3147
3148         /*
3149          * If the filesystem writes its buffers by hand (eg ext3)
3150          * then we can have clean buffers against a dirty page.  We
3151          * clean the page here; otherwise the VM will never notice
3152          * that the filesystem did any IO at all.
3153          *
3154          * Also, during truncate, discard_buffer will have marked all
3155          * the page's buffers clean.  We discover that here and clean
3156          * the page also.
3157          *
3158          * private_lock must be held over this entire operation in order
3159          * to synchronise against __set_page_dirty_buffers and prevent the
3160          * dirty bit from being lost.
3161          */
3162         if (ret)
3163                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3164         spin_unlock(&mapping->private_lock);
3165 out:
3166         if (buffers_to_free) {
3167                 struct buffer_head *bh = buffers_to_free;
3168
3169                 do {
3170                         struct buffer_head *next = bh->b_this_page;
3171                         free_buffer_head(bh);
3172                         bh = next;
3173                 } while (bh != buffers_to_free);
3174         }
3175         return ret;
3176 }
3177 EXPORT_SYMBOL(try_to_free_buffers);
3178
3179 void block_sync_page(struct page *page)
3180 {
3181         struct address_space *mapping;
3182
3183         smp_mb();
3184         mapping = page_mapping(page);
3185         if (mapping)
3186                 blk_run_backing_dev(mapping->backing_dev_info, page);
3187 }
3188
3189 /*
3190  * There are no bdflush tunables left.  But distributions are
3191  * still running obsolete flush daemons, so we terminate them here.
3192  *
3193  * Use of bdflush() is deprecated and will be removed in a future kernel.
3194  * The `pdflush' kernel threads fully replace bdflush daemons and this call.
3195  */
3196 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3197 {
3198         static int msg_count;
3199
3200         if (!capable(CAP_SYS_ADMIN))
3201                 return -EPERM;
3202
3203         if (msg_count < 5) {
3204                 msg_count++;
3205                 printk(KERN_INFO
3206                         "warning: process `%s' used the obsolete bdflush"
3207                         " system call\n", current->comm);
3208                 printk(KERN_INFO "Fix your initscripts?\n");
3209         }
3210
3211         if (func == 1)
3212                 do_exit(0);
3213         return 0;
3214 }
3215
3216 /*
3217  * Buffer-head allocation
3218  */
3219 static struct kmem_cache *bh_cachep;
3220
3221 /*
3222  * Once the number of bh's in the machine exceeds this level, we start
3223  * stripping them in writeback.
3224  */
3225 static int max_buffer_heads;
3226
3227 int buffer_heads_over_limit;
3228
3229 struct bh_accounting {
3230         int nr;                 /* Number of live bh's */
3231         int ratelimit;          /* Limit cacheline bouncing */
3232 };
3233
3234 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3235
3236 static void recalc_bh_state(void)
3237 {
3238         int i;
3239         int tot = 0;
3240
3241         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3242                 return;
3243         __get_cpu_var(bh_accounting).ratelimit = 0;
3244         for_each_online_cpu(i)
3245                 tot += per_cpu(bh_accounting, i).nr;
3246         buffer_heads_over_limit = (tot > max_buffer_heads);
3247 }
3248         
3249 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3250 {
3251         struct buffer_head *ret = kmem_cache_alloc(bh_cachep, gfp_flags);
3252         if (ret) {
3253                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3254                 get_cpu_var(bh_accounting).nr++;
3255                 recalc_bh_state();
3256                 put_cpu_var(bh_accounting);
3257         }
3258         return ret;
3259 }
3260 EXPORT_SYMBOL(alloc_buffer_head);
3261
3262 void free_buffer_head(struct buffer_head *bh)
3263 {
3264         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3265         kmem_cache_free(bh_cachep, bh);
3266         get_cpu_var(bh_accounting).nr--;
3267         recalc_bh_state();
3268         put_cpu_var(bh_accounting);
3269 }
3270 EXPORT_SYMBOL(free_buffer_head);
3271
3272 static void buffer_exit_cpu(int cpu)
3273 {
3274         int i;
3275         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3276
3277         for (i = 0; i < BH_LRU_SIZE; i++) {
3278                 brelse(b->bhs[i]);
3279                 b->bhs[i] = NULL;
3280         }
3281         get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3282         per_cpu(bh_accounting, cpu).nr = 0;
3283         put_cpu_var(bh_accounting);
3284 }
3285
3286 static int buffer_cpu_notify(struct notifier_block *self,
3287                               unsigned long action, void *hcpu)
3288 {
3289         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3290                 buffer_exit_cpu((unsigned long)hcpu);
3291         return NOTIFY_OK;
3292 }
3293
3294 /**
3295  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3296  * @bh: struct buffer_head
3297  *
3298  * Return true if the buffer is up-to-date and false,
3299  * with the buffer locked, if not.
3300  */
3301 int bh_uptodate_or_lock(struct buffer_head *bh)
3302 {
3303         if (!buffer_uptodate(bh)) {
3304                 lock_buffer(bh);
3305                 if (!buffer_uptodate(bh))
3306                         return 0;
3307                 unlock_buffer(bh);
3308         }
3309         return 1;
3310 }
3311 EXPORT_SYMBOL(bh_uptodate_or_lock);
3312
3313 /**
3314  * bh_submit_read - Submit a locked buffer for reading
3315  * @bh: struct buffer_head
3316  *
3317  * Returns zero on success and -EIO on error.
3318  */
3319 int bh_submit_read(struct buffer_head *bh)
3320 {
3321         BUG_ON(!buffer_locked(bh));
3322
3323         if (buffer_uptodate(bh)) {
3324                 unlock_buffer(bh);
3325                 return 0;
3326         }
3327
3328         get_bh(bh);
3329         bh->b_end_io = end_buffer_read_sync;
3330         submit_bh(READ, bh);
3331         wait_on_buffer(bh);
3332         if (buffer_uptodate(bh))
3333                 return 0;
3334         return -EIO;
3335 }
3336 EXPORT_SYMBOL(bh_submit_read);
3337
3338 static void
3339 init_buffer_head(void *data)
3340 {
3341         struct buffer_head *bh = data;
3342
3343         memset(bh, 0, sizeof(*bh));
3344         INIT_LIST_HEAD(&bh->b_assoc_buffers);
3345 }
3346
3347 void __init buffer_init(void)
3348 {
3349         int nrpages;
3350
3351         bh_cachep = kmem_cache_create("buffer_head",
3352                         sizeof(struct buffer_head), 0,
3353                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3354                                 SLAB_MEM_SPREAD),
3355                                 init_buffer_head);
3356
3357         /*
3358          * Limit the bh occupancy to 10% of ZONE_NORMAL
3359          */
3360         nrpages = (nr_free_buffer_pages() * 10) / 100;
3361         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3362         hotcpu_notifier(buffer_cpu_notify, 0);
3363 }
3364
3365 EXPORT_SYMBOL(__bforget);
3366 EXPORT_SYMBOL(__brelse);
3367 EXPORT_SYMBOL(__wait_on_buffer);
3368 EXPORT_SYMBOL(block_commit_write);
3369 EXPORT_SYMBOL(block_prepare_write);
3370 EXPORT_SYMBOL(block_page_mkwrite);
3371 EXPORT_SYMBOL(block_read_full_page);
3372 EXPORT_SYMBOL(block_sync_page);
3373 EXPORT_SYMBOL(block_truncate_page);
3374 EXPORT_SYMBOL(block_write_full_page);
3375 EXPORT_SYMBOL(block_write_full_page_endio);
3376 EXPORT_SYMBOL(cont_write_begin);
3377 EXPORT_SYMBOL(end_buffer_read_sync);
3378 EXPORT_SYMBOL(end_buffer_write_sync);
3379 EXPORT_SYMBOL(end_buffer_async_write);
3380 EXPORT_SYMBOL(file_fsync);
3381 EXPORT_SYMBOL(generic_block_bmap);
3382 EXPORT_SYMBOL(generic_cont_expand_simple);
3383 EXPORT_SYMBOL(init_buffer);
3384 EXPORT_SYMBOL(invalidate_bdev);
3385 EXPORT_SYMBOL(ll_rw_block);
3386 EXPORT_SYMBOL(mark_buffer_dirty);
3387 EXPORT_SYMBOL(submit_bh);
3388 EXPORT_SYMBOL(sync_dirty_buffer);
3389 EXPORT_SYMBOL(unlock_buffer);