7f0b9b083f778676ed3024fbecf6c8a77fd399bc
[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 EXPORT_SYMBOL(init_buffer);
56
57 static int sync_buffer(void *word)
58 {
59         struct block_device *bd;
60         struct buffer_head *bh
61                 = container_of(word, struct buffer_head, b_state);
62
63         smp_mb();
64         bd = bh->b_bdev;
65         if (bd)
66                 blk_run_address_space(bd->bd_inode->i_mapping);
67         io_schedule();
68         return 0;
69 }
70
71 void __lock_buffer(struct buffer_head *bh)
72 {
73         wait_on_bit_lock(&bh->b_state, BH_Lock, sync_buffer,
74                                                         TASK_UNINTERRUPTIBLE);
75 }
76 EXPORT_SYMBOL(__lock_buffer);
77
78 void unlock_buffer(struct buffer_head *bh)
79 {
80         clear_bit_unlock(BH_Lock, &bh->b_state);
81         smp_mb__after_clear_bit();
82         wake_up_bit(&bh->b_state, BH_Lock);
83 }
84 EXPORT_SYMBOL(unlock_buffer);
85
86 /*
87  * Block until a buffer comes unlocked.  This doesn't stop it
88  * from becoming locked again - you have to lock it yourself
89  * if you want to preserve its state.
90  */
91 void __wait_on_buffer(struct buffer_head * bh)
92 {
93         wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
94 }
95 EXPORT_SYMBOL(__wait_on_buffer);
96
97 static void
98 __clear_page_buffers(struct page *page)
99 {
100         ClearPagePrivate(page);
101         set_page_private(page, 0);
102         page_cache_release(page);
103 }
104
105
106 static int quiet_error(struct buffer_head *bh)
107 {
108         if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
109                 return 0;
110         return 1;
111 }
112
113
114 static void buffer_io_error(struct buffer_head *bh)
115 {
116         char b[BDEVNAME_SIZE];
117         printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
118                         bdevname(bh->b_bdev, b),
119                         (unsigned long long)bh->b_blocknr);
120 }
121
122 /*
123  * End-of-IO handler helper function which does not touch the bh after
124  * unlocking it.
125  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
126  * a race there is benign: unlock_buffer() only use the bh's address for
127  * hashing after unlocking the buffer, so it doesn't actually touch the bh
128  * itself.
129  */
130 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
131 {
132         if (uptodate) {
133                 set_buffer_uptodate(bh);
134         } else {
135                 /* This happens, due to failed READA attempts. */
136                 clear_buffer_uptodate(bh);
137         }
138         unlock_buffer(bh);
139 }
140
141 /*
142  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
143  * unlock the buffer. This is what ll_rw_block uses too.
144  */
145 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
146 {
147         __end_buffer_read_notouch(bh, uptodate);
148         put_bh(bh);
149 }
150 EXPORT_SYMBOL(end_buffer_read_sync);
151
152 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
153 {
154         char b[BDEVNAME_SIZE];
155
156         if (uptodate) {
157                 set_buffer_uptodate(bh);
158         } else {
159                 if (!quiet_error(bh)) {
160                         buffer_io_error(bh);
161                         printk(KERN_WARNING "lost page write due to "
162                                         "I/O error on %s\n",
163                                        bdevname(bh->b_bdev, b));
164                 }
165                 set_buffer_write_io_error(bh);
166                 clear_buffer_uptodate(bh);
167         }
168         unlock_buffer(bh);
169         put_bh(bh);
170 }
171 EXPORT_SYMBOL(end_buffer_write_sync);
172
173 /*
174  * Various filesystems appear to want __find_get_block to be non-blocking.
175  * But it's the page lock which protects the buffers.  To get around this,
176  * we get exclusion from try_to_free_buffers with the blockdev mapping's
177  * private_lock.
178  *
179  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
180  * may be quite high.  This code could TryLock the page, and if that
181  * succeeds, there is no need to take private_lock. (But if
182  * private_lock is contended then so is mapping->tree_lock).
183  */
184 static struct buffer_head *
185 __find_get_block_slow(struct block_device *bdev, sector_t block)
186 {
187         struct inode *bd_inode = bdev->bd_inode;
188         struct address_space *bd_mapping = bd_inode->i_mapping;
189         struct buffer_head *ret = NULL;
190         pgoff_t index;
191         struct buffer_head *bh;
192         struct buffer_head *head;
193         struct page *page;
194         int all_mapped = 1;
195
196         index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
197         page = find_get_page(bd_mapping, index);
198         if (!page)
199                 goto out;
200
201         spin_lock(&bd_mapping->private_lock);
202         if (!page_has_buffers(page))
203                 goto out_unlock;
204         head = page_buffers(page);
205         bh = head;
206         do {
207                 if (!buffer_mapped(bh))
208                         all_mapped = 0;
209                 else if (bh->b_blocknr == block) {
210                         ret = bh;
211                         get_bh(bh);
212                         goto out_unlock;
213                 }
214                 bh = bh->b_this_page;
215         } while (bh != head);
216
217         /* we might be here because some of the buffers on this page are
218          * not mapped.  This is due to various races between
219          * file io on the block device and getblk.  It gets dealt with
220          * elsewhere, don't buffer_error if we had some unmapped buffers
221          */
222         if (all_mapped) {
223                 printk("__find_get_block_slow() failed. "
224                         "block=%llu, b_blocknr=%llu\n",
225                         (unsigned long long)block,
226                         (unsigned long long)bh->b_blocknr);
227                 printk("b_state=0x%08lx, b_size=%zu\n",
228                         bh->b_state, bh->b_size);
229                 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
230         }
231 out_unlock:
232         spin_unlock(&bd_mapping->private_lock);
233         page_cache_release(page);
234 out:
235         return ret;
236 }
237
238 /* If invalidate_buffers() will trash dirty buffers, it means some kind
239    of fs corruption is going on. Trashing dirty data always imply losing
240    information that was supposed to be just stored on the physical layer
241    by the user.
242
243    Thus invalidate_buffers in general usage is not allwowed to trash
244    dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
245    be preserved.  These buffers are simply skipped.
246   
247    We also skip buffers which are still in use.  For example this can
248    happen if a userspace program is reading the block device.
249
250    NOTE: In the case where the user removed a removable-media-disk even if
251    there's still dirty data not synced on disk (due a bug in the device driver
252    or due an error of the user), by not destroying the dirty buffers we could
253    generate corruption also on the next media inserted, thus a parameter is
254    necessary to handle this case in the most safe way possible (trying
255    to not corrupt also the new disk inserted with the data belonging to
256    the old now corrupted disk). Also for the ramdisk the natural thing
257    to do in order to release the ramdisk memory is to destroy dirty buffers.
258
259    These are two special cases. Normal usage imply the device driver
260    to issue a sync on the device (without waiting I/O completion) and
261    then an invalidate_buffers call that doesn't trash dirty buffers.
262
263    For handling cache coherency with the blkdev pagecache the 'update' case
264    is been introduced. It is needed to re-read from disk any pinned
265    buffer. NOTE: re-reading from disk is destructive so we can do it only
266    when we assume nobody is changing the buffercache under our I/O and when
267    we think the disk contains more recent information than the buffercache.
268    The update == 1 pass marks the buffers we need to update, the update == 2
269    pass does the actual I/O. */
270 void invalidate_bdev(struct block_device *bdev)
271 {
272         struct address_space *mapping = bdev->bd_inode->i_mapping;
273
274         if (mapping->nrpages == 0)
275                 return;
276
277         invalidate_bh_lrus();
278         lru_add_drain_all();    /* make sure all lru add caches are flushed */
279         invalidate_mapping_pages(mapping, 0, -1);
280 }
281 EXPORT_SYMBOL(invalidate_bdev);
282
283 /*
284  * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
285  */
286 static void free_more_memory(void)
287 {
288         struct zone *zone;
289         int nid;
290
291         wakeup_flusher_threads(1024);
292         yield();
293
294         for_each_online_node(nid) {
295                 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
296                                                 gfp_zone(GFP_NOFS), NULL,
297                                                 &zone);
298                 if (zone)
299                         try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
300                                                 GFP_NOFS, NULL);
301         }
302 }
303
304 /*
305  * I/O completion handler for block_read_full_page() - pages
306  * which come unlocked at the end of I/O.
307  */
308 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
309 {
310         unsigned long flags;
311         struct buffer_head *first;
312         struct buffer_head *tmp;
313         struct page *page;
314         int page_uptodate = 1;
315
316         BUG_ON(!buffer_async_read(bh));
317
318         page = bh->b_page;
319         if (uptodate) {
320                 set_buffer_uptodate(bh);
321         } else {
322                 clear_buffer_uptodate(bh);
323                 if (!quiet_error(bh))
324                         buffer_io_error(bh);
325                 SetPageError(page);
326         }
327
328         /*
329          * Be _very_ careful from here on. Bad things can happen if
330          * two buffer heads end IO at almost the same time and both
331          * decide that the page is now completely done.
332          */
333         first = page_buffers(page);
334         local_irq_save(flags);
335         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
336         clear_buffer_async_read(bh);
337         unlock_buffer(bh);
338         tmp = bh;
339         do {
340                 if (!buffer_uptodate(tmp))
341                         page_uptodate = 0;
342                 if (buffer_async_read(tmp)) {
343                         BUG_ON(!buffer_locked(tmp));
344                         goto still_busy;
345                 }
346                 tmp = tmp->b_this_page;
347         } while (tmp != bh);
348         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
349         local_irq_restore(flags);
350
351         /*
352          * If none of the buffers had errors and they are all
353          * uptodate then we can set the page uptodate.
354          */
355         if (page_uptodate && !PageError(page))
356                 SetPageUptodate(page);
357         unlock_page(page);
358         return;
359
360 still_busy:
361         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
362         local_irq_restore(flags);
363         return;
364 }
365
366 /*
367  * Completion handler for block_write_full_page() - pages which are unlocked
368  * during I/O, and which have PageWriteback cleared upon I/O completion.
369  */
370 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
371 {
372         char b[BDEVNAME_SIZE];
373         unsigned long flags;
374         struct buffer_head *first;
375         struct buffer_head *tmp;
376         struct page *page;
377
378         BUG_ON(!buffer_async_write(bh));
379
380         page = bh->b_page;
381         if (uptodate) {
382                 set_buffer_uptodate(bh);
383         } else {
384                 if (!quiet_error(bh)) {
385                         buffer_io_error(bh);
386                         printk(KERN_WARNING "lost page write due to "
387                                         "I/O error on %s\n",
388                                bdevname(bh->b_bdev, b));
389                 }
390                 set_bit(AS_EIO, &page->mapping->flags);
391                 set_buffer_write_io_error(bh);
392                 clear_buffer_uptodate(bh);
393                 SetPageError(page);
394         }
395
396         first = page_buffers(page);
397         local_irq_save(flags);
398         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
399
400         clear_buffer_async_write(bh);
401         unlock_buffer(bh);
402         tmp = bh->b_this_page;
403         while (tmp != bh) {
404                 if (buffer_async_write(tmp)) {
405                         BUG_ON(!buffer_locked(tmp));
406                         goto still_busy;
407                 }
408                 tmp = tmp->b_this_page;
409         }
410         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
411         local_irq_restore(flags);
412         end_page_writeback(page);
413         return;
414
415 still_busy:
416         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
417         local_irq_restore(flags);
418         return;
419 }
420 EXPORT_SYMBOL(end_buffer_async_write);
421
422 /*
423  * If a page's buffers are under async readin (end_buffer_async_read
424  * completion) then there is a possibility that another thread of
425  * control could lock one of the buffers after it has completed
426  * but while some of the other buffers have not completed.  This
427  * locked buffer would confuse end_buffer_async_read() into not unlocking
428  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
429  * that this buffer is not under async I/O.
430  *
431  * The page comes unlocked when it has no locked buffer_async buffers
432  * left.
433  *
434  * PageLocked prevents anyone starting new async I/O reads any of
435  * the buffers.
436  *
437  * PageWriteback is used to prevent simultaneous writeout of the same
438  * page.
439  *
440  * PageLocked prevents anyone from starting writeback of a page which is
441  * under read I/O (PageWriteback is only ever set against a locked page).
442  */
443 static void mark_buffer_async_read(struct buffer_head *bh)
444 {
445         bh->b_end_io = end_buffer_async_read;
446         set_buffer_async_read(bh);
447 }
448
449 static void mark_buffer_async_write_endio(struct buffer_head *bh,
450                                           bh_end_io_t *handler)
451 {
452         bh->b_end_io = handler;
453         set_buffer_async_write(bh);
454 }
455
456 void mark_buffer_async_write(struct buffer_head *bh)
457 {
458         mark_buffer_async_write_endio(bh, end_buffer_async_write);
459 }
460 EXPORT_SYMBOL(mark_buffer_async_write);
461
462
463 /*
464  * fs/buffer.c contains helper functions for buffer-backed address space's
465  * fsync functions.  A common requirement for buffer-based filesystems is
466  * that certain data from the backing blockdev needs to be written out for
467  * a successful fsync().  For example, ext2 indirect blocks need to be
468  * written back and waited upon before fsync() returns.
469  *
470  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
471  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
472  * management of a list of dependent buffers at ->i_mapping->private_list.
473  *
474  * Locking is a little subtle: try_to_free_buffers() will remove buffers
475  * from their controlling inode's queue when they are being freed.  But
476  * try_to_free_buffers() will be operating against the *blockdev* mapping
477  * at the time, not against the S_ISREG file which depends on those buffers.
478  * So the locking for private_list is via the private_lock in the address_space
479  * which backs the buffers.  Which is different from the address_space 
480  * against which the buffers are listed.  So for a particular address_space,
481  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
482  * mapping->private_list will always be protected by the backing blockdev's
483  * ->private_lock.
484  *
485  * Which introduces a requirement: all buffers on an address_space's
486  * ->private_list must be from the same address_space: the blockdev's.
487  *
488  * address_spaces which do not place buffers at ->private_list via these
489  * utility functions are free to use private_lock and private_list for
490  * whatever they want.  The only requirement is that list_empty(private_list)
491  * be true at clear_inode() time.
492  *
493  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
494  * filesystems should do that.  invalidate_inode_buffers() should just go
495  * BUG_ON(!list_empty).
496  *
497  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
498  * take an address_space, not an inode.  And it should be called
499  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
500  * queued up.
501  *
502  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
503  * list if it is already on a list.  Because if the buffer is on a list,
504  * it *must* already be on the right one.  If not, the filesystem is being
505  * silly.  This will save a ton of locking.  But first we have to ensure
506  * that buffers are taken *off* the old inode's list when they are freed
507  * (presumably in truncate).  That requires careful auditing of all
508  * filesystems (do it inside bforget()).  It could also be done by bringing
509  * b_inode back.
510  */
511
512 /*
513  * The buffer's backing address_space's private_lock must be held
514  */
515 static void __remove_assoc_queue(struct buffer_head *bh)
516 {
517         list_del_init(&bh->b_assoc_buffers);
518         WARN_ON(!bh->b_assoc_map);
519         if (buffer_write_io_error(bh))
520                 set_bit(AS_EIO, &bh->b_assoc_map->flags);
521         bh->b_assoc_map = NULL;
522 }
523
524 int inode_has_buffers(struct inode *inode)
525 {
526         return !list_empty(&inode->i_data.private_list);
527 }
528
529 /*
530  * osync is designed to support O_SYNC io.  It waits synchronously for
531  * all already-submitted IO to complete, but does not queue any new
532  * writes to the disk.
533  *
534  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
535  * you dirty the buffers, and then use osync_inode_buffers to wait for
536  * completion.  Any other dirty buffers which are not yet queued for
537  * write will not be flushed to disk by the osync.
538  */
539 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
540 {
541         struct buffer_head *bh;
542         struct list_head *p;
543         int err = 0;
544
545         spin_lock(lock);
546 repeat:
547         list_for_each_prev(p, list) {
548                 bh = BH_ENTRY(p);
549                 if (buffer_locked(bh)) {
550                         get_bh(bh);
551                         spin_unlock(lock);
552                         wait_on_buffer(bh);
553                         if (!buffer_uptodate(bh))
554                                 err = -EIO;
555                         brelse(bh);
556                         spin_lock(lock);
557                         goto repeat;
558                 }
559         }
560         spin_unlock(lock);
561         return err;
562 }
563
564 static void do_thaw_one(struct super_block *sb, void *unused)
565 {
566         char b[BDEVNAME_SIZE];
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 }
571
572 static void do_thaw_all(struct work_struct *work)
573 {
574         iterate_supers(do_thaw_one, NULL);
575         kfree(work);
576         printk(KERN_WARNING "Emergency Thaw complete\n");
577 }
578
579 /**
580  * emergency_thaw_all -- forcibly thaw every frozen filesystem
581  *
582  * Used for emergency unfreeze of all filesystems via SysRq
583  */
584 void emergency_thaw_all(void)
585 {
586         struct work_struct *work;
587
588         work = kmalloc(sizeof(*work), GFP_ATOMIC);
589         if (work) {
590                 INIT_WORK(work, do_thaw_all);
591                 schedule_work(work);
592         }
593 }
594
595 /**
596  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
597  * @mapping: the mapping which wants those buffers written
598  *
599  * Starts I/O against the buffers at mapping->private_list, and waits upon
600  * that I/O.
601  *
602  * Basically, this is a convenience function for fsync().
603  * @mapping is a file or directory which needs those buffers to be written for
604  * a successful fsync().
605  */
606 int sync_mapping_buffers(struct address_space *mapping)
607 {
608         struct address_space *buffer_mapping = mapping->assoc_mapping;
609
610         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
611                 return 0;
612
613         return fsync_buffers_list(&buffer_mapping->private_lock,
614                                         &mapping->private_list);
615 }
616 EXPORT_SYMBOL(sync_mapping_buffers);
617
618 /*
619  * Called when we've recently written block `bblock', and it is known that
620  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
621  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
622  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
623  */
624 void write_boundary_block(struct block_device *bdev,
625                         sector_t bblock, unsigned blocksize)
626 {
627         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
628         if (bh) {
629                 if (buffer_dirty(bh))
630                         ll_rw_block(WRITE, 1, &bh);
631                 put_bh(bh);
632         }
633 }
634
635 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
636 {
637         struct address_space *mapping = inode->i_mapping;
638         struct address_space *buffer_mapping = bh->b_page->mapping;
639
640         mark_buffer_dirty(bh);
641         if (!mapping->assoc_mapping) {
642                 mapping->assoc_mapping = buffer_mapping;
643         } else {
644                 BUG_ON(mapping->assoc_mapping != buffer_mapping);
645         }
646         if (!bh->b_assoc_map) {
647                 spin_lock(&buffer_mapping->private_lock);
648                 list_move_tail(&bh->b_assoc_buffers,
649                                 &mapping->private_list);
650                 bh->b_assoc_map = mapping;
651                 spin_unlock(&buffer_mapping->private_lock);
652         }
653 }
654 EXPORT_SYMBOL(mark_buffer_dirty_inode);
655
656 /*
657  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
658  * dirty.
659  *
660  * If warn is true, then emit a warning if the page is not uptodate and has
661  * not been truncated.
662  */
663 static void __set_page_dirty(struct page *page,
664                 struct address_space *mapping, int warn)
665 {
666         spin_lock_irq(&mapping->tree_lock);
667         if (page->mapping) {    /* Race with truncate? */
668                 WARN_ON_ONCE(warn && !PageUptodate(page));
669                 account_page_dirtied(page, mapping);
670                 radix_tree_tag_set(&mapping->page_tree,
671                                 page_index(page), PAGECACHE_TAG_DIRTY);
672         }
673         spin_unlock_irq(&mapping->tree_lock);
674         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
675 }
676
677 /*
678  * Add a page to the dirty page list.
679  *
680  * It is a sad fact of life that this function is called from several places
681  * deeply under spinlocking.  It may not sleep.
682  *
683  * If the page has buffers, the uptodate buffers are set dirty, to preserve
684  * dirty-state coherency between the page and the buffers.  It the page does
685  * not have buffers then when they are later attached they will all be set
686  * dirty.
687  *
688  * The buffers are dirtied before the page is dirtied.  There's a small race
689  * window in which a writepage caller may see the page cleanness but not the
690  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
691  * before the buffers, a concurrent writepage caller could clear the page dirty
692  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
693  * page on the dirty page list.
694  *
695  * We use private_lock to lock against try_to_free_buffers while using the
696  * page's buffer list.  Also use this to protect against clean buffers being
697  * added to the page after it was set dirty.
698  *
699  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
700  * address_space though.
701  */
702 int __set_page_dirty_buffers(struct page *page)
703 {
704         int newly_dirty;
705         struct address_space *mapping = page_mapping(page);
706
707         if (unlikely(!mapping))
708                 return !TestSetPageDirty(page);
709
710         spin_lock(&mapping->private_lock);
711         if (page_has_buffers(page)) {
712                 struct buffer_head *head = page_buffers(page);
713                 struct buffer_head *bh = head;
714
715                 do {
716                         set_buffer_dirty(bh);
717                         bh = bh->b_this_page;
718                 } while (bh != head);
719         }
720         newly_dirty = !TestSetPageDirty(page);
721         spin_unlock(&mapping->private_lock);
722
723         if (newly_dirty)
724                 __set_page_dirty(page, mapping, 1);
725         return newly_dirty;
726 }
727 EXPORT_SYMBOL(__set_page_dirty_buffers);
728
729 /*
730  * Write out and wait upon a list of buffers.
731  *
732  * We have conflicting pressures: we want to make sure that all
733  * initially dirty buffers get waited on, but that any subsequently
734  * dirtied buffers don't.  After all, we don't want fsync to last
735  * forever if somebody is actively writing to the file.
736  *
737  * Do this in two main stages: first we copy dirty buffers to a
738  * temporary inode list, queueing the writes as we go.  Then we clean
739  * up, waiting for those writes to complete.
740  * 
741  * During this second stage, any subsequent updates to the file may end
742  * up refiling the buffer on the original inode's dirty list again, so
743  * there is a chance we will end up with a buffer queued for write but
744  * not yet completed on that list.  So, as a final cleanup we go through
745  * the osync code to catch these locked, dirty buffers without requeuing
746  * any newly dirty buffers for write.
747  */
748 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
749 {
750         struct buffer_head *bh;
751         struct list_head tmp;
752         struct address_space *mapping, *prev_mapping = NULL;
753         int err = 0, err2;
754
755         INIT_LIST_HEAD(&tmp);
756
757         spin_lock(lock);
758         while (!list_empty(list)) {
759                 bh = BH_ENTRY(list->next);
760                 mapping = bh->b_assoc_map;
761                 __remove_assoc_queue(bh);
762                 /* Avoid race with mark_buffer_dirty_inode() which does
763                  * a lockless check and we rely on seeing the dirty bit */
764                 smp_mb();
765                 if (buffer_dirty(bh) || buffer_locked(bh)) {
766                         list_add(&bh->b_assoc_buffers, &tmp);
767                         bh->b_assoc_map = mapping;
768                         if (buffer_dirty(bh)) {
769                                 get_bh(bh);
770                                 spin_unlock(lock);
771                                 /*
772                                  * Ensure any pending I/O completes so that
773                                  * write_dirty_buffer() actually writes the
774                                  * current contents - it is a noop if I/O is
775                                  * still in flight on potentially older
776                                  * contents.
777                                  */
778                                 write_dirty_buffer(bh, WRITE_SYNC_PLUG);
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 EXPORT_SYMBOL(mark_buffer_dirty);
1176
1177 /*
1178  * Decrement a buffer_head's reference count.  If all buffers against a page
1179  * have zero reference count, are clean and unlocked, and if the page is clean
1180  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1181  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1182  * a page but it ends up not being freed, and buffers may later be reattached).
1183  */
1184 void __brelse(struct buffer_head * buf)
1185 {
1186         if (atomic_read(&buf->b_count)) {
1187                 put_bh(buf);
1188                 return;
1189         }
1190         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1191 }
1192 EXPORT_SYMBOL(__brelse);
1193
1194 /*
1195  * bforget() is like brelse(), except it discards any
1196  * potentially dirty data.
1197  */
1198 void __bforget(struct buffer_head *bh)
1199 {
1200         clear_buffer_dirty(bh);
1201         if (bh->b_assoc_map) {
1202                 struct address_space *buffer_mapping = bh->b_page->mapping;
1203
1204                 spin_lock(&buffer_mapping->private_lock);
1205                 list_del_init(&bh->b_assoc_buffers);
1206                 bh->b_assoc_map = NULL;
1207                 spin_unlock(&buffer_mapping->private_lock);
1208         }
1209         __brelse(bh);
1210 }
1211 EXPORT_SYMBOL(__bforget);
1212
1213 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1214 {
1215         lock_buffer(bh);
1216         if (buffer_uptodate(bh)) {
1217                 unlock_buffer(bh);
1218                 return bh;
1219         } else {
1220                 get_bh(bh);
1221                 bh->b_end_io = end_buffer_read_sync;
1222                 submit_bh(READ, bh);
1223                 wait_on_buffer(bh);
1224                 if (buffer_uptodate(bh))
1225                         return bh;
1226         }
1227         brelse(bh);
1228         return NULL;
1229 }
1230
1231 /*
1232  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1233  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1234  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1235  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1236  * CPU's LRUs at the same time.
1237  *
1238  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1239  * sb_find_get_block().
1240  *
1241  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1242  * a local interrupt disable for that.
1243  */
1244
1245 #define BH_LRU_SIZE     8
1246
1247 struct bh_lru {
1248         struct buffer_head *bhs[BH_LRU_SIZE];
1249 };
1250
1251 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1252
1253 #ifdef CONFIG_SMP
1254 #define bh_lru_lock()   local_irq_disable()
1255 #define bh_lru_unlock() local_irq_enable()
1256 #else
1257 #define bh_lru_lock()   preempt_disable()
1258 #define bh_lru_unlock() preempt_enable()
1259 #endif
1260
1261 static inline void check_irqs_on(void)
1262 {
1263 #ifdef irqs_disabled
1264         BUG_ON(irqs_disabled());
1265 #endif
1266 }
1267
1268 /*
1269  * The LRU management algorithm is dopey-but-simple.  Sorry.
1270  */
1271 static void bh_lru_install(struct buffer_head *bh)
1272 {
1273         struct buffer_head *evictee = NULL;
1274         struct bh_lru *lru;
1275
1276         check_irqs_on();
1277         bh_lru_lock();
1278         lru = &__get_cpu_var(bh_lrus);
1279         if (lru->bhs[0] != bh) {
1280                 struct buffer_head *bhs[BH_LRU_SIZE];
1281                 int in;
1282                 int out = 0;
1283
1284                 get_bh(bh);
1285                 bhs[out++] = bh;
1286                 for (in = 0; in < BH_LRU_SIZE; in++) {
1287                         struct buffer_head *bh2 = lru->bhs[in];
1288
1289                         if (bh2 == bh) {
1290                                 __brelse(bh2);
1291                         } else {
1292                                 if (out >= BH_LRU_SIZE) {
1293                                         BUG_ON(evictee != NULL);
1294                                         evictee = bh2;
1295                                 } else {
1296                                         bhs[out++] = bh2;
1297                                 }
1298                         }
1299                 }
1300                 while (out < BH_LRU_SIZE)
1301                         bhs[out++] = NULL;
1302                 memcpy(lru->bhs, bhs, sizeof(bhs));
1303         }
1304         bh_lru_unlock();
1305
1306         if (evictee)
1307                 __brelse(evictee);
1308 }
1309
1310 /*
1311  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1312  */
1313 static struct buffer_head *
1314 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1315 {
1316         struct buffer_head *ret = NULL;
1317         struct bh_lru *lru;
1318         unsigned int i;
1319
1320         check_irqs_on();
1321         bh_lru_lock();
1322         lru = &__get_cpu_var(bh_lrus);
1323         for (i = 0; i < BH_LRU_SIZE; i++) {
1324                 struct buffer_head *bh = lru->bhs[i];
1325
1326                 if (bh && bh->b_bdev == bdev &&
1327                                 bh->b_blocknr == block && bh->b_size == size) {
1328                         if (i) {
1329                                 while (i) {
1330                                         lru->bhs[i] = lru->bhs[i - 1];
1331                                         i--;
1332                                 }
1333                                 lru->bhs[0] = bh;
1334                         }
1335                         get_bh(bh);
1336                         ret = bh;
1337                         break;
1338                 }
1339         }
1340         bh_lru_unlock();
1341         return ret;
1342 }
1343
1344 /*
1345  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1346  * it in the LRU and mark it as accessed.  If it is not present then return
1347  * NULL
1348  */
1349 struct buffer_head *
1350 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1351 {
1352         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1353
1354         if (bh == NULL) {
1355                 bh = __find_get_block_slow(bdev, block);
1356                 if (bh)
1357                         bh_lru_install(bh);
1358         }
1359         if (bh)
1360                 touch_buffer(bh);
1361         return bh;
1362 }
1363 EXPORT_SYMBOL(__find_get_block);
1364
1365 /*
1366  * __getblk will locate (and, if necessary, create) the buffer_head
1367  * which corresponds to the passed block_device, block and size. The
1368  * returned buffer has its reference count incremented.
1369  *
1370  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1371  * illegal block number, __getblk() will happily return a buffer_head
1372  * which represents the non-existent block.  Very weird.
1373  *
1374  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1375  * attempt is failing.  FIXME, perhaps?
1376  */
1377 struct buffer_head *
1378 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1379 {
1380         struct buffer_head *bh = __find_get_block(bdev, block, size);
1381
1382         might_sleep();
1383         if (bh == NULL)
1384                 bh = __getblk_slow(bdev, block, size);
1385         return bh;
1386 }
1387 EXPORT_SYMBOL(__getblk);
1388
1389 /*
1390  * Do async read-ahead on a buffer..
1391  */
1392 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1393 {
1394         struct buffer_head *bh = __getblk(bdev, block, size);
1395         if (likely(bh)) {
1396                 ll_rw_block(READA, 1, &bh);
1397                 brelse(bh);
1398         }
1399 }
1400 EXPORT_SYMBOL(__breadahead);
1401
1402 /**
1403  *  __bread() - reads a specified block and returns the bh
1404  *  @bdev: the block_device to read from
1405  *  @block: number of block
1406  *  @size: size (in bytes) to read
1407  * 
1408  *  Reads a specified block, and returns buffer head that contains it.
1409  *  It returns NULL if the block was unreadable.
1410  */
1411 struct buffer_head *
1412 __bread(struct block_device *bdev, sector_t block, unsigned size)
1413 {
1414         struct buffer_head *bh = __getblk(bdev, block, size);
1415
1416         if (likely(bh) && !buffer_uptodate(bh))
1417                 bh = __bread_slow(bh);
1418         return bh;
1419 }
1420 EXPORT_SYMBOL(__bread);
1421
1422 /*
1423  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1424  * This doesn't race because it runs in each cpu either in irq
1425  * or with preempt disabled.
1426  */
1427 static void invalidate_bh_lru(void *arg)
1428 {
1429         struct bh_lru *b = &get_cpu_var(bh_lrus);
1430         int i;
1431
1432         for (i = 0; i < BH_LRU_SIZE; i++) {
1433                 brelse(b->bhs[i]);
1434                 b->bhs[i] = NULL;
1435         }
1436         put_cpu_var(bh_lrus);
1437 }
1438         
1439 void invalidate_bh_lrus(void)
1440 {
1441         on_each_cpu(invalidate_bh_lru, NULL, 1);
1442 }
1443 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1444
1445 void set_bh_page(struct buffer_head *bh,
1446                 struct page *page, unsigned long offset)
1447 {
1448         bh->b_page = page;
1449         BUG_ON(offset >= PAGE_SIZE);
1450         if (PageHighMem(page))
1451                 /*
1452                  * This catches illegal uses and preserves the offset:
1453                  */
1454                 bh->b_data = (char *)(0 + offset);
1455         else
1456                 bh->b_data = page_address(page) + offset;
1457 }
1458 EXPORT_SYMBOL(set_bh_page);
1459
1460 /*
1461  * Called when truncating a buffer on a page completely.
1462  */
1463 static void discard_buffer(struct buffer_head * bh)
1464 {
1465         lock_buffer(bh);
1466         clear_buffer_dirty(bh);
1467         bh->b_bdev = NULL;
1468         clear_buffer_mapped(bh);
1469         clear_buffer_req(bh);
1470         clear_buffer_new(bh);
1471         clear_buffer_delay(bh);
1472         clear_buffer_unwritten(bh);
1473         unlock_buffer(bh);
1474 }
1475
1476 /**
1477  * block_invalidatepage - invalidate part of all of a buffer-backed page
1478  *
1479  * @page: the page which is affected
1480  * @offset: the index of the truncation point
1481  *
1482  * block_invalidatepage() is called when all or part of the page has become
1483  * invalidatedby a truncate operation.
1484  *
1485  * block_invalidatepage() does not have to release all buffers, but it must
1486  * ensure that no dirty buffer is left outside @offset and that no I/O
1487  * is underway against any of the blocks which are outside the truncation
1488  * point.  Because the caller is about to free (and possibly reuse) those
1489  * blocks on-disk.
1490  */
1491 void block_invalidatepage(struct page *page, unsigned long offset)
1492 {
1493         struct buffer_head *head, *bh, *next;
1494         unsigned int curr_off = 0;
1495
1496         BUG_ON(!PageLocked(page));
1497         if (!page_has_buffers(page))
1498                 goto out;
1499
1500         head = page_buffers(page);
1501         bh = head;
1502         do {
1503                 unsigned int next_off = curr_off + bh->b_size;
1504                 next = bh->b_this_page;
1505
1506                 /*
1507                  * is this block fully invalidated?
1508                  */
1509                 if (offset <= curr_off)
1510                         discard_buffer(bh);
1511                 curr_off = next_off;
1512                 bh = next;
1513         } while (bh != head);
1514
1515         /*
1516          * We release buffers only if the entire page is being invalidated.
1517          * The get_block cached value has been unconditionally invalidated,
1518          * so real IO is not possible anymore.
1519          */
1520         if (offset == 0)
1521                 try_to_release_page(page, 0);
1522 out:
1523         return;
1524 }
1525 EXPORT_SYMBOL(block_invalidatepage);
1526
1527 /*
1528  * We attach and possibly dirty the buffers atomically wrt
1529  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1530  * is already excluded via the page lock.
1531  */
1532 void create_empty_buffers(struct page *page,
1533                         unsigned long blocksize, unsigned long b_state)
1534 {
1535         struct buffer_head *bh, *head, *tail;
1536
1537         head = alloc_page_buffers(page, blocksize, 1);
1538         bh = head;
1539         do {
1540                 bh->b_state |= b_state;
1541                 tail = bh;
1542                 bh = bh->b_this_page;
1543         } while (bh);
1544         tail->b_this_page = head;
1545
1546         spin_lock(&page->mapping->private_lock);
1547         if (PageUptodate(page) || PageDirty(page)) {
1548                 bh = head;
1549                 do {
1550                         if (PageDirty(page))
1551                                 set_buffer_dirty(bh);
1552                         if (PageUptodate(page))
1553                                 set_buffer_uptodate(bh);
1554                         bh = bh->b_this_page;
1555                 } while (bh != head);
1556         }
1557         attach_page_buffers(page, head);
1558         spin_unlock(&page->mapping->private_lock);
1559 }
1560 EXPORT_SYMBOL(create_empty_buffers);
1561
1562 /*
1563  * We are taking a block for data and we don't want any output from any
1564  * buffer-cache aliases starting from return from that function and
1565  * until the moment when something will explicitly mark the buffer
1566  * dirty (hopefully that will not happen until we will free that block ;-)
1567  * We don't even need to mark it not-uptodate - nobody can expect
1568  * anything from a newly allocated buffer anyway. We used to used
1569  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1570  * don't want to mark the alias unmapped, for example - it would confuse
1571  * anyone who might pick it with bread() afterwards...
1572  *
1573  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1574  * be writeout I/O going on against recently-freed buffers.  We don't
1575  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1576  * only if we really need to.  That happens here.
1577  */
1578 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1579 {
1580         struct buffer_head *old_bh;
1581
1582         might_sleep();
1583
1584         old_bh = __find_get_block_slow(bdev, block);
1585         if (old_bh) {
1586                 clear_buffer_dirty(old_bh);
1587                 wait_on_buffer(old_bh);
1588                 clear_buffer_req(old_bh);
1589                 __brelse(old_bh);
1590         }
1591 }
1592 EXPORT_SYMBOL(unmap_underlying_metadata);
1593
1594 /*
1595  * NOTE! All mapped/uptodate combinations are valid:
1596  *
1597  *      Mapped  Uptodate        Meaning
1598  *
1599  *      No      No              "unknown" - must do get_block()
1600  *      No      Yes             "hole" - zero-filled
1601  *      Yes     No              "allocated" - allocated on disk, not read in
1602  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1603  *
1604  * "Dirty" is valid only with the last case (mapped+uptodate).
1605  */
1606
1607 /*
1608  * While block_write_full_page is writing back the dirty buffers under
1609  * the page lock, whoever dirtied the buffers may decide to clean them
1610  * again at any time.  We handle that by only looking at the buffer
1611  * state inside lock_buffer().
1612  *
1613  * If block_write_full_page() is called for regular writeback
1614  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1615  * locked buffer.   This only can happen if someone has written the buffer
1616  * directly, with submit_bh().  At the address_space level PageWriteback
1617  * prevents this contention from occurring.
1618  *
1619  * If block_write_full_page() is called with wbc->sync_mode ==
1620  * WB_SYNC_ALL, the writes are posted using WRITE_SYNC_PLUG; this
1621  * causes the writes to be flagged as synchronous writes, but the
1622  * block device queue will NOT be unplugged, since usually many pages
1623  * will be pushed to the out before the higher-level caller actually
1624  * waits for the writes to be completed.  The various wait functions,
1625  * such as wait_on_writeback_range() will ultimately call sync_page()
1626  * which will ultimately call blk_run_backing_dev(), which will end up
1627  * unplugging the device queue.
1628  */
1629 static int __block_write_full_page(struct inode *inode, struct page *page,
1630                         get_block_t *get_block, struct writeback_control *wbc,
1631                         bh_end_io_t *handler)
1632 {
1633         int err;
1634         sector_t block;
1635         sector_t last_block;
1636         struct buffer_head *bh, *head;
1637         const unsigned blocksize = 1 << inode->i_blkbits;
1638         int nr_underway = 0;
1639         int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1640                         WRITE_SYNC_PLUG : WRITE);
1641
1642         BUG_ON(!PageLocked(page));
1643
1644         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1645
1646         if (!page_has_buffers(page)) {
1647                 create_empty_buffers(page, blocksize,
1648                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1649         }
1650
1651         /*
1652          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1653          * here, and the (potentially unmapped) buffers may become dirty at
1654          * any time.  If a buffer becomes dirty here after we've inspected it
1655          * then we just miss that fact, and the page stays dirty.
1656          *
1657          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1658          * handle that here by just cleaning them.
1659          */
1660
1661         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1662         head = page_buffers(page);
1663         bh = head;
1664
1665         /*
1666          * Get all the dirty buffers mapped to disk addresses and
1667          * handle any aliases from the underlying blockdev's mapping.
1668          */
1669         do {
1670                 if (block > last_block) {
1671                         /*
1672                          * mapped buffers outside i_size will occur, because
1673                          * this page can be outside i_size when there is a
1674                          * truncate in progress.
1675                          */
1676                         /*
1677                          * The buffer was zeroed by block_write_full_page()
1678                          */
1679                         clear_buffer_dirty(bh);
1680                         set_buffer_uptodate(bh);
1681                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1682                            buffer_dirty(bh)) {
1683                         WARN_ON(bh->b_size != blocksize);
1684                         err = get_block(inode, block, bh, 1);
1685                         if (err)
1686                                 goto recover;
1687                         clear_buffer_delay(bh);
1688                         if (buffer_new(bh)) {
1689                                 /* blockdev mappings never come here */
1690                                 clear_buffer_new(bh);
1691                                 unmap_underlying_metadata(bh->b_bdev,
1692                                                         bh->b_blocknr);
1693                         }
1694                 }
1695                 bh = bh->b_this_page;
1696                 block++;
1697         } while (bh != head);
1698
1699         do {
1700                 if (!buffer_mapped(bh))
1701                         continue;
1702                 /*
1703                  * If it's a fully non-blocking write attempt and we cannot
1704                  * lock the buffer then redirty the page.  Note that this can
1705                  * potentially cause a busy-wait loop from writeback threads
1706                  * and kswapd activity, but those code paths have their own
1707                  * higher-level throttling.
1708                  */
1709                 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1710                         lock_buffer(bh);
1711                 } else if (!trylock_buffer(bh)) {
1712                         redirty_page_for_writepage(wbc, page);
1713                         continue;
1714                 }
1715                 if (test_clear_buffer_dirty(bh)) {
1716                         mark_buffer_async_write_endio(bh, handler);
1717                 } else {
1718                         unlock_buffer(bh);
1719                 }
1720         } while ((bh = bh->b_this_page) != head);
1721
1722         /*
1723          * The page and its buffers are protected by PageWriteback(), so we can
1724          * drop the bh refcounts early.
1725          */
1726         BUG_ON(PageWriteback(page));
1727         set_page_writeback(page);
1728
1729         do {
1730                 struct buffer_head *next = bh->b_this_page;
1731                 if (buffer_async_write(bh)) {
1732                         submit_bh(write_op, bh);
1733                         nr_underway++;
1734                 }
1735                 bh = next;
1736         } while (bh != head);
1737         unlock_page(page);
1738
1739         err = 0;
1740 done:
1741         if (nr_underway == 0) {
1742                 /*
1743                  * The page was marked dirty, but the buffers were
1744                  * clean.  Someone wrote them back by hand with
1745                  * ll_rw_block/submit_bh.  A rare case.
1746                  */
1747                 end_page_writeback(page);
1748
1749                 /*
1750                  * The page and buffer_heads can be released at any time from
1751                  * here on.
1752                  */
1753         }
1754         return err;
1755
1756 recover:
1757         /*
1758          * ENOSPC, or some other error.  We may already have added some
1759          * blocks to the file, so we need to write these out to avoid
1760          * exposing stale data.
1761          * The page is currently locked and not marked for writeback
1762          */
1763         bh = head;
1764         /* Recovery: lock and submit the mapped buffers */
1765         do {
1766                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1767                     !buffer_delay(bh)) {
1768                         lock_buffer(bh);
1769                         mark_buffer_async_write_endio(bh, handler);
1770                 } else {
1771                         /*
1772                          * The buffer may have been set dirty during
1773                          * attachment to a dirty page.
1774                          */
1775                         clear_buffer_dirty(bh);
1776                 }
1777         } while ((bh = bh->b_this_page) != head);
1778         SetPageError(page);
1779         BUG_ON(PageWriteback(page));
1780         mapping_set_error(page->mapping, err);
1781         set_page_writeback(page);
1782         do {
1783                 struct buffer_head *next = bh->b_this_page;
1784                 if (buffer_async_write(bh)) {
1785                         clear_buffer_dirty(bh);
1786                         submit_bh(write_op, bh);
1787                         nr_underway++;
1788                 }
1789                 bh = next;
1790         } while (bh != head);
1791         unlock_page(page);
1792         goto done;
1793 }
1794
1795 /*
1796  * If a page has any new buffers, zero them out here, and mark them uptodate
1797  * and dirty so they'll be written out (in order to prevent uninitialised
1798  * block data from leaking). And clear the new bit.
1799  */
1800 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1801 {
1802         unsigned int block_start, block_end;
1803         struct buffer_head *head, *bh;
1804
1805         BUG_ON(!PageLocked(page));
1806         if (!page_has_buffers(page))
1807                 return;
1808
1809         bh = head = page_buffers(page);
1810         block_start = 0;
1811         do {
1812                 block_end = block_start + bh->b_size;
1813
1814                 if (buffer_new(bh)) {
1815                         if (block_end > from && block_start < to) {
1816                                 if (!PageUptodate(page)) {
1817                                         unsigned start, size;
1818
1819                                         start = max(from, block_start);
1820                                         size = min(to, block_end) - start;
1821
1822                                         zero_user(page, start, size);
1823                                         set_buffer_uptodate(bh);
1824                                 }
1825
1826                                 clear_buffer_new(bh);
1827                                 mark_buffer_dirty(bh);
1828                         }
1829                 }
1830
1831                 block_start = block_end;
1832                 bh = bh->b_this_page;
1833         } while (bh != head);
1834 }
1835 EXPORT_SYMBOL(page_zero_new_buffers);
1836
1837 int block_prepare_write(struct page *page, unsigned from, unsigned to,
1838                 get_block_t *get_block)
1839 {
1840         struct inode *inode = page->mapping->host;
1841         unsigned block_start, block_end;
1842         sector_t block;
1843         int err = 0;
1844         unsigned blocksize, bbits;
1845         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1846
1847         BUG_ON(!PageLocked(page));
1848         BUG_ON(from > PAGE_CACHE_SIZE);
1849         BUG_ON(to > PAGE_CACHE_SIZE);
1850         BUG_ON(from > to);
1851
1852         blocksize = 1 << inode->i_blkbits;
1853         if (!page_has_buffers(page))
1854                 create_empty_buffers(page, blocksize, 0);
1855         head = page_buffers(page);
1856
1857         bbits = inode->i_blkbits;
1858         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1859
1860         for(bh = head, block_start = 0; bh != head || !block_start;
1861             block++, block_start=block_end, bh = bh->b_this_page) {
1862                 block_end = block_start + blocksize;
1863                 if (block_end <= from || block_start >= to) {
1864                         if (PageUptodate(page)) {
1865                                 if (!buffer_uptodate(bh))
1866                                         set_buffer_uptodate(bh);
1867                         }
1868                         continue;
1869                 }
1870                 if (buffer_new(bh))
1871                         clear_buffer_new(bh);
1872                 if (!buffer_mapped(bh)) {
1873                         WARN_ON(bh->b_size != blocksize);
1874                         err = get_block(inode, block, bh, 1);
1875                         if (err)
1876                                 break;
1877                         if (buffer_new(bh)) {
1878                                 unmap_underlying_metadata(bh->b_bdev,
1879                                                         bh->b_blocknr);
1880                                 if (PageUptodate(page)) {
1881                                         clear_buffer_new(bh);
1882                                         set_buffer_uptodate(bh);
1883                                         mark_buffer_dirty(bh);
1884                                         continue;
1885                                 }
1886                                 if (block_end > to || block_start < from)
1887                                         zero_user_segments(page,
1888                                                 to, block_end,
1889                                                 block_start, from);
1890                                 continue;
1891                         }
1892                 }
1893                 if (PageUptodate(page)) {
1894                         if (!buffer_uptodate(bh))
1895                                 set_buffer_uptodate(bh);
1896                         continue; 
1897                 }
1898                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1899                     !buffer_unwritten(bh) &&
1900                      (block_start < from || block_end > to)) {
1901                         ll_rw_block(READ, 1, &bh);
1902                         *wait_bh++=bh;
1903                 }
1904         }
1905         /*
1906          * If we issued read requests - let them complete.
1907          */
1908         while(wait_bh > wait) {
1909                 wait_on_buffer(*--wait_bh);
1910                 if (!buffer_uptodate(*wait_bh))
1911                         err = -EIO;
1912         }
1913         if (unlikely(err)) {
1914                 page_zero_new_buffers(page, from, to);
1915                 ClearPageUptodate(page);
1916         }
1917         return err;
1918 }
1919 EXPORT_SYMBOL(block_prepare_write);
1920
1921 static int __block_commit_write(struct inode *inode, struct page *page,
1922                 unsigned from, unsigned to)
1923 {
1924         unsigned block_start, block_end;
1925         int partial = 0;
1926         unsigned blocksize;
1927         struct buffer_head *bh, *head;
1928
1929         blocksize = 1 << inode->i_blkbits;
1930
1931         for(bh = head = page_buffers(page), block_start = 0;
1932             bh != head || !block_start;
1933             block_start=block_end, bh = bh->b_this_page) {
1934                 block_end = block_start + blocksize;
1935                 if (block_end <= from || block_start >= to) {
1936                         if (!buffer_uptodate(bh))
1937                                 partial = 1;
1938                 } else {
1939                         set_buffer_uptodate(bh);
1940                         mark_buffer_dirty(bh);
1941                 }
1942                 clear_buffer_new(bh);
1943         }
1944
1945         /*
1946          * If this is a partial write which happened to make all buffers
1947          * uptodate then we can optimize away a bogus readpage() for
1948          * the next read(). Here we 'discover' whether the page went
1949          * uptodate as a result of this (potentially partial) write.
1950          */
1951         if (!partial)
1952                 SetPageUptodate(page);
1953         return 0;
1954 }
1955
1956 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1957                 get_block_t *get_block)
1958 {
1959         unsigned start = pos & (PAGE_CACHE_SIZE - 1);
1960
1961         return block_prepare_write(page, start, start + len, get_block);
1962 }
1963 EXPORT_SYMBOL(__block_write_begin);
1964
1965 /*
1966  * block_write_begin takes care of the basic task of block allocation and
1967  * bringing partial write blocks uptodate first.
1968  *
1969  * The filesystem needs to handle block truncation upon failure.
1970  */
1971 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
1972                 unsigned flags, struct page **pagep, get_block_t *get_block)
1973 {
1974         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1975         struct page *page;
1976         int status;
1977
1978         page = grab_cache_page_write_begin(mapping, index, flags);
1979         if (!page)
1980                 return -ENOMEM;
1981
1982         status = __block_write_begin(page, pos, len, get_block);
1983         if (unlikely(status)) {
1984                 unlock_page(page);
1985                 page_cache_release(page);
1986                 page = NULL;
1987         }
1988
1989         *pagep = page;
1990         return status;
1991 }
1992 EXPORT_SYMBOL(block_write_begin);
1993
1994 int block_write_end(struct file *file, struct address_space *mapping,
1995                         loff_t pos, unsigned len, unsigned copied,
1996                         struct page *page, void *fsdata)
1997 {
1998         struct inode *inode = mapping->host;
1999         unsigned start;
2000
2001         start = pos & (PAGE_CACHE_SIZE - 1);
2002
2003         if (unlikely(copied < len)) {
2004                 /*
2005                  * The buffers that were written will now be uptodate, so we
2006                  * don't have to worry about a readpage reading them and
2007                  * overwriting a partial write. However if we have encountered
2008                  * a short write and only partially written into a buffer, it
2009                  * will not be marked uptodate, so a readpage might come in and
2010                  * destroy our partial write.
2011                  *
2012                  * Do the simplest thing, and just treat any short write to a
2013                  * non uptodate page as a zero-length write, and force the
2014                  * caller to redo the whole thing.
2015                  */
2016                 if (!PageUptodate(page))
2017                         copied = 0;
2018
2019                 page_zero_new_buffers(page, start+copied, start+len);
2020         }
2021         flush_dcache_page(page);
2022
2023         /* This could be a short (even 0-length) commit */
2024         __block_commit_write(inode, page, start, start+copied);
2025
2026         return copied;
2027 }
2028 EXPORT_SYMBOL(block_write_end);
2029
2030 int generic_write_end(struct file *file, struct address_space *mapping,
2031                         loff_t pos, unsigned len, unsigned copied,
2032                         struct page *page, void *fsdata)
2033 {
2034         struct inode *inode = mapping->host;
2035         int i_size_changed = 0;
2036
2037         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2038
2039         /*
2040          * No need to use i_size_read() here, the i_size
2041          * cannot change under us because we hold i_mutex.
2042          *
2043          * But it's important to update i_size while still holding page lock:
2044          * page writeout could otherwise come in and zero beyond i_size.
2045          */
2046         if (pos+copied > inode->i_size) {
2047                 i_size_write(inode, pos+copied);
2048                 i_size_changed = 1;
2049         }
2050
2051         unlock_page(page);
2052         page_cache_release(page);
2053
2054         /*
2055          * Don't mark the inode dirty under page lock. First, it unnecessarily
2056          * makes the holding time of page lock longer. Second, it forces lock
2057          * ordering of page lock and transaction start for journaling
2058          * filesystems.
2059          */
2060         if (i_size_changed)
2061                 mark_inode_dirty(inode);
2062
2063         return copied;
2064 }
2065 EXPORT_SYMBOL(generic_write_end);
2066
2067 /*
2068  * block_is_partially_uptodate checks whether buffers within a page are
2069  * uptodate or not.
2070  *
2071  * Returns true if all buffers which correspond to a file portion
2072  * we want to read are uptodate.
2073  */
2074 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2075                                         unsigned long from)
2076 {
2077         struct inode *inode = page->mapping->host;
2078         unsigned block_start, block_end, blocksize;
2079         unsigned to;
2080         struct buffer_head *bh, *head;
2081         int ret = 1;
2082
2083         if (!page_has_buffers(page))
2084                 return 0;
2085
2086         blocksize = 1 << inode->i_blkbits;
2087         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2088         to = from + to;
2089         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2090                 return 0;
2091
2092         head = page_buffers(page);
2093         bh = head;
2094         block_start = 0;
2095         do {
2096                 block_end = block_start + blocksize;
2097                 if (block_end > from && block_start < to) {
2098                         if (!buffer_uptodate(bh)) {
2099                                 ret = 0;
2100                                 break;
2101                         }
2102                         if (block_end >= to)
2103                                 break;
2104                 }
2105                 block_start = block_end;
2106                 bh = bh->b_this_page;
2107         } while (bh != head);
2108
2109         return ret;
2110 }
2111 EXPORT_SYMBOL(block_is_partially_uptodate);
2112
2113 /*
2114  * Generic "read page" function for block devices that have the normal
2115  * get_block functionality. This is most of the block device filesystems.
2116  * Reads the page asynchronously --- the unlock_buffer() and
2117  * set/clear_buffer_uptodate() functions propagate buffer state into the
2118  * page struct once IO has completed.
2119  */
2120 int block_read_full_page(struct page *page, get_block_t *get_block)
2121 {
2122         struct inode *inode = page->mapping->host;
2123         sector_t iblock, lblock;
2124         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2125         unsigned int blocksize;
2126         int nr, i;
2127         int fully_mapped = 1;
2128
2129         BUG_ON(!PageLocked(page));
2130         blocksize = 1 << inode->i_blkbits;
2131         if (!page_has_buffers(page))
2132                 create_empty_buffers(page, blocksize, 0);
2133         head = page_buffers(page);
2134
2135         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2136         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2137         bh = head;
2138         nr = 0;
2139         i = 0;
2140
2141         do {
2142                 if (buffer_uptodate(bh))
2143                         continue;
2144
2145                 if (!buffer_mapped(bh)) {
2146                         int err = 0;
2147
2148                         fully_mapped = 0;
2149                         if (iblock < lblock) {
2150                                 WARN_ON(bh->b_size != blocksize);
2151                                 err = get_block(inode, iblock, bh, 0);
2152                                 if (err)
2153                                         SetPageError(page);
2154                         }
2155                         if (!buffer_mapped(bh)) {
2156                                 zero_user(page, i * blocksize, blocksize);
2157                                 if (!err)
2158                                         set_buffer_uptodate(bh);
2159                                 continue;
2160                         }
2161                         /*
2162                          * get_block() might have updated the buffer
2163                          * synchronously
2164                          */
2165                         if (buffer_uptodate(bh))
2166                                 continue;
2167                 }
2168                 arr[nr++] = bh;
2169         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2170
2171         if (fully_mapped)
2172                 SetPageMappedToDisk(page);
2173
2174         if (!nr) {
2175                 /*
2176                  * All buffers are uptodate - we can set the page uptodate
2177                  * as well. But not if get_block() returned an error.
2178                  */
2179                 if (!PageError(page))
2180                         SetPageUptodate(page);
2181                 unlock_page(page);
2182                 return 0;
2183         }
2184
2185         /* Stage two: lock the buffers */
2186         for (i = 0; i < nr; i++) {
2187                 bh = arr[i];
2188                 lock_buffer(bh);
2189                 mark_buffer_async_read(bh);
2190         }
2191
2192         /*
2193          * Stage 3: start the IO.  Check for uptodateness
2194          * inside the buffer lock in case another process reading
2195          * the underlying blockdev brought it uptodate (the sct fix).
2196          */
2197         for (i = 0; i < nr; i++) {
2198                 bh = arr[i];
2199                 if (buffer_uptodate(bh))
2200                         end_buffer_async_read(bh, 1);
2201                 else
2202                         submit_bh(READ, bh);
2203         }
2204         return 0;
2205 }
2206 EXPORT_SYMBOL(block_read_full_page);
2207
2208 /* utility function for filesystems that need to do work on expanding
2209  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2210  * deal with the hole.  
2211  */
2212 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2213 {
2214         struct address_space *mapping = inode->i_mapping;
2215         struct page *page;
2216         void *fsdata;
2217         int err;
2218
2219         err = inode_newsize_ok(inode, size);
2220         if (err)
2221                 goto out;
2222
2223         err = pagecache_write_begin(NULL, mapping, size, 0,
2224                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2225                                 &page, &fsdata);
2226         if (err)
2227                 goto out;
2228
2229         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2230         BUG_ON(err > 0);
2231
2232 out:
2233         return err;
2234 }
2235 EXPORT_SYMBOL(generic_cont_expand_simple);
2236
2237 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2238                             loff_t pos, loff_t *bytes)
2239 {
2240         struct inode *inode = mapping->host;
2241         unsigned blocksize = 1 << inode->i_blkbits;
2242         struct page *page;
2243         void *fsdata;
2244         pgoff_t index, curidx;
2245         loff_t curpos;
2246         unsigned zerofrom, offset, len;
2247         int err = 0;
2248
2249         index = pos >> PAGE_CACHE_SHIFT;
2250         offset = pos & ~PAGE_CACHE_MASK;
2251
2252         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2253                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2254                 if (zerofrom & (blocksize-1)) {
2255                         *bytes |= (blocksize-1);
2256                         (*bytes)++;
2257                 }
2258                 len = PAGE_CACHE_SIZE - zerofrom;
2259
2260                 err = pagecache_write_begin(file, mapping, curpos, len,
2261                                                 AOP_FLAG_UNINTERRUPTIBLE,
2262                                                 &page, &fsdata);
2263                 if (err)
2264                         goto out;
2265                 zero_user(page, zerofrom, len);
2266                 err = pagecache_write_end(file, mapping, curpos, len, len,
2267                                                 page, fsdata);
2268                 if (err < 0)
2269                         goto out;
2270                 BUG_ON(err != len);
2271                 err = 0;
2272
2273                 balance_dirty_pages_ratelimited(mapping);
2274         }
2275
2276         /* page covers the boundary, find the boundary offset */
2277         if (index == curidx) {
2278                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2279                 /* if we will expand the thing last block will be filled */
2280                 if (offset <= zerofrom) {
2281                         goto out;
2282                 }
2283                 if (zerofrom & (blocksize-1)) {
2284                         *bytes |= (blocksize-1);
2285                         (*bytes)++;
2286                 }
2287                 len = offset - zerofrom;
2288
2289                 err = pagecache_write_begin(file, mapping, curpos, len,
2290                                                 AOP_FLAG_UNINTERRUPTIBLE,
2291                                                 &page, &fsdata);
2292                 if (err)
2293                         goto out;
2294                 zero_user(page, zerofrom, len);
2295                 err = pagecache_write_end(file, mapping, curpos, len, len,
2296                                                 page, fsdata);
2297                 if (err < 0)
2298                         goto out;
2299                 BUG_ON(err != len);
2300                 err = 0;
2301         }
2302 out:
2303         return err;
2304 }
2305
2306 /*
2307  * For moronic filesystems that do not allow holes in file.
2308  * We may have to extend the file.
2309  */
2310 int cont_write_begin(struct file *file, struct address_space *mapping,
2311                         loff_t pos, unsigned len, unsigned flags,
2312                         struct page **pagep, void **fsdata,
2313                         get_block_t *get_block, loff_t *bytes)
2314 {
2315         struct inode *inode = mapping->host;
2316         unsigned blocksize = 1 << inode->i_blkbits;
2317         unsigned zerofrom;
2318         int err;
2319
2320         err = cont_expand_zero(file, mapping, pos, bytes);
2321         if (err)
2322                 return err;
2323
2324         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2325         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2326                 *bytes |= (blocksize-1);
2327                 (*bytes)++;
2328         }
2329
2330         return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2331 }
2332 EXPORT_SYMBOL(cont_write_begin);
2333
2334 int block_commit_write(struct page *page, unsigned from, unsigned to)
2335 {
2336         struct inode *inode = page->mapping->host;
2337         __block_commit_write(inode,page,from,to);
2338         return 0;
2339 }
2340 EXPORT_SYMBOL(block_commit_write);
2341
2342 /*
2343  * block_page_mkwrite() is not allowed to change the file size as it gets
2344  * called from a page fault handler when a page is first dirtied. Hence we must
2345  * be careful to check for EOF conditions here. We set the page up correctly
2346  * for a written page which means we get ENOSPC checking when writing into
2347  * holes and correct delalloc and unwritten extent mapping on filesystems that
2348  * support these features.
2349  *
2350  * We are not allowed to take the i_mutex here so we have to play games to
2351  * protect against truncate races as the page could now be beyond EOF.  Because
2352  * truncate writes the inode size before removing pages, once we have the
2353  * page lock we can determine safely if the page is beyond EOF. If it is not
2354  * beyond EOF, then the page is guaranteed safe against truncation until we
2355  * unlock the page.
2356  */
2357 int
2358 block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2359                    get_block_t get_block)
2360 {
2361         struct page *page = vmf->page;
2362         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2363         unsigned long end;
2364         loff_t size;
2365         int ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
2366
2367         lock_page(page);
2368         size = i_size_read(inode);
2369         if ((page->mapping != inode->i_mapping) ||
2370             (page_offset(page) > size)) {
2371                 /* page got truncated out from underneath us */
2372                 unlock_page(page);
2373                 goto out;
2374         }
2375
2376         /* page is wholly or partially inside EOF */
2377         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2378                 end = size & ~PAGE_CACHE_MASK;
2379         else
2380                 end = PAGE_CACHE_SIZE;
2381
2382         ret = block_prepare_write(page, 0, end, get_block);
2383         if (!ret)
2384                 ret = block_commit_write(page, 0, end);
2385
2386         if (unlikely(ret)) {
2387                 unlock_page(page);
2388                 if (ret == -ENOMEM)
2389                         ret = VM_FAULT_OOM;
2390                 else /* -ENOSPC, -EIO, etc */
2391                         ret = VM_FAULT_SIGBUS;
2392         } else
2393                 ret = VM_FAULT_LOCKED;
2394
2395 out:
2396         return ret;
2397 }
2398 EXPORT_SYMBOL(block_page_mkwrite);
2399
2400 /*
2401  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2402  * immediately, while under the page lock.  So it needs a special end_io
2403  * handler which does not touch the bh after unlocking it.
2404  */
2405 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2406 {
2407         __end_buffer_read_notouch(bh, uptodate);
2408 }
2409
2410 /*
2411  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2412  * the page (converting it to circular linked list and taking care of page
2413  * dirty races).
2414  */
2415 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2416 {
2417         struct buffer_head *bh;
2418
2419         BUG_ON(!PageLocked(page));
2420
2421         spin_lock(&page->mapping->private_lock);
2422         bh = head;
2423         do {
2424                 if (PageDirty(page))
2425                         set_buffer_dirty(bh);
2426                 if (!bh->b_this_page)
2427                         bh->b_this_page = head;
2428                 bh = bh->b_this_page;
2429         } while (bh != head);
2430         attach_page_buffers(page, head);
2431         spin_unlock(&page->mapping->private_lock);
2432 }
2433
2434 /*
2435  * On entry, the page is fully not uptodate.
2436  * On exit the page is fully uptodate in the areas outside (from,to)
2437  * The filesystem needs to handle block truncation upon failure.
2438  */
2439 int nobh_write_begin(struct address_space *mapping,
2440                         loff_t pos, unsigned len, unsigned flags,
2441                         struct page **pagep, void **fsdata,
2442                         get_block_t *get_block)
2443 {
2444         struct inode *inode = mapping->host;
2445         const unsigned blkbits = inode->i_blkbits;
2446         const unsigned blocksize = 1 << blkbits;
2447         struct buffer_head *head, *bh;
2448         struct page *page;
2449         pgoff_t index;
2450         unsigned from, to;
2451         unsigned block_in_page;
2452         unsigned block_start, block_end;
2453         sector_t block_in_file;
2454         int nr_reads = 0;
2455         int ret = 0;
2456         int is_mapped_to_disk = 1;
2457
2458         index = pos >> PAGE_CACHE_SHIFT;
2459         from = pos & (PAGE_CACHE_SIZE - 1);
2460         to = from + len;
2461
2462         page = grab_cache_page_write_begin(mapping, index, flags);
2463         if (!page)
2464                 return -ENOMEM;
2465         *pagep = page;
2466         *fsdata = NULL;
2467
2468         if (page_has_buffers(page)) {
2469                 unlock_page(page);
2470                 page_cache_release(page);
2471                 *pagep = NULL;
2472                 return block_write_begin(mapping, pos, len, flags, pagep,
2473                                          get_block);
2474         }
2475
2476         if (PageMappedToDisk(page))
2477                 return 0;
2478
2479         /*
2480          * Allocate buffers so that we can keep track of state, and potentially
2481          * attach them to the page if an error occurs. In the common case of
2482          * no error, they will just be freed again without ever being attached
2483          * to the page (which is all OK, because we're under the page lock).
2484          *
2485          * Be careful: the buffer linked list is a NULL terminated one, rather
2486          * than the circular one we're used to.
2487          */
2488         head = alloc_page_buffers(page, blocksize, 0);
2489         if (!head) {
2490                 ret = -ENOMEM;
2491                 goto out_release;
2492         }
2493
2494         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2495
2496         /*
2497          * We loop across all blocks in the page, whether or not they are
2498          * part of the affected region.  This is so we can discover if the
2499          * page is fully mapped-to-disk.
2500          */
2501         for (block_start = 0, block_in_page = 0, bh = head;
2502                   block_start < PAGE_CACHE_SIZE;
2503                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2504                 int create;
2505
2506                 block_end = block_start + blocksize;
2507                 bh->b_state = 0;
2508                 create = 1;
2509                 if (block_start >= to)
2510                         create = 0;
2511                 ret = get_block(inode, block_in_file + block_in_page,
2512                                         bh, create);
2513                 if (ret)
2514                         goto failed;
2515                 if (!buffer_mapped(bh))
2516                         is_mapped_to_disk = 0;
2517                 if (buffer_new(bh))
2518                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2519                 if (PageUptodate(page)) {
2520                         set_buffer_uptodate(bh);
2521                         continue;
2522                 }
2523                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2524                         zero_user_segments(page, block_start, from,
2525                                                         to, block_end);
2526                         continue;
2527                 }
2528                 if (buffer_uptodate(bh))
2529                         continue;       /* reiserfs does this */
2530                 if (block_start < from || block_end > to) {
2531                         lock_buffer(bh);
2532                         bh->b_end_io = end_buffer_read_nobh;
2533                         submit_bh(READ, bh);
2534                         nr_reads++;
2535                 }
2536         }
2537
2538         if (nr_reads) {
2539                 /*
2540                  * The page is locked, so these buffers are protected from
2541                  * any VM or truncate activity.  Hence we don't need to care
2542                  * for the buffer_head refcounts.
2543                  */
2544                 for (bh = head; bh; bh = bh->b_this_page) {
2545                         wait_on_buffer(bh);
2546                         if (!buffer_uptodate(bh))
2547                                 ret = -EIO;
2548                 }
2549                 if (ret)
2550                         goto failed;
2551         }
2552
2553         if (is_mapped_to_disk)
2554                 SetPageMappedToDisk(page);
2555
2556         *fsdata = head; /* to be released by nobh_write_end */
2557
2558         return 0;
2559
2560 failed:
2561         BUG_ON(!ret);
2562         /*
2563          * Error recovery is a bit difficult. We need to zero out blocks that
2564          * were newly allocated, and dirty them to ensure they get written out.
2565          * Buffers need to be attached to the page at this point, otherwise
2566          * the handling of potential IO errors during writeout would be hard
2567          * (could try doing synchronous writeout, but what if that fails too?)
2568          */
2569         attach_nobh_buffers(page, head);
2570         page_zero_new_buffers(page, from, to);
2571
2572 out_release:
2573         unlock_page(page);
2574         page_cache_release(page);
2575         *pagep = NULL;
2576
2577         return ret;
2578 }
2579 EXPORT_SYMBOL(nobh_write_begin);
2580
2581 int nobh_write_end(struct file *file, struct address_space *mapping,
2582                         loff_t pos, unsigned len, unsigned copied,
2583                         struct page *page, void *fsdata)
2584 {
2585         struct inode *inode = page->mapping->host;
2586         struct buffer_head *head = fsdata;
2587         struct buffer_head *bh;
2588         BUG_ON(fsdata != NULL && page_has_buffers(page));
2589
2590         if (unlikely(copied < len) && head)
2591                 attach_nobh_buffers(page, head);
2592         if (page_has_buffers(page))
2593                 return generic_write_end(file, mapping, pos, len,
2594                                         copied, page, fsdata);
2595
2596         SetPageUptodate(page);
2597         set_page_dirty(page);
2598         if (pos+copied > inode->i_size) {
2599                 i_size_write(inode, pos+copied);
2600                 mark_inode_dirty(inode);
2601         }
2602
2603         unlock_page(page);
2604         page_cache_release(page);
2605
2606         while (head) {
2607                 bh = head;
2608                 head = head->b_this_page;
2609                 free_buffer_head(bh);
2610         }
2611
2612         return copied;
2613 }
2614 EXPORT_SYMBOL(nobh_write_end);
2615
2616 /*
2617  * nobh_writepage() - based on block_full_write_page() except
2618  * that it tries to operate without attaching bufferheads to
2619  * the page.
2620  */
2621 int nobh_writepage(struct page *page, get_block_t *get_block,
2622                         struct writeback_control *wbc)
2623 {
2624         struct inode * const inode = page->mapping->host;
2625         loff_t i_size = i_size_read(inode);
2626         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2627         unsigned offset;
2628         int ret;
2629
2630         /* Is the page fully inside i_size? */
2631         if (page->index < end_index)
2632                 goto out;
2633
2634         /* Is the page fully outside i_size? (truncate in progress) */
2635         offset = i_size & (PAGE_CACHE_SIZE-1);
2636         if (page->index >= end_index+1 || !offset) {
2637                 /*
2638                  * The page may have dirty, unmapped buffers.  For example,
2639                  * they may have been added in ext3_writepage().  Make them
2640                  * freeable here, so the page does not leak.
2641                  */
2642 #if 0
2643                 /* Not really sure about this  - do we need this ? */
2644                 if (page->mapping->a_ops->invalidatepage)
2645                         page->mapping->a_ops->invalidatepage(page, offset);
2646 #endif
2647                 unlock_page(page);
2648                 return 0; /* don't care */
2649         }
2650
2651         /*
2652          * The page straddles i_size.  It must be zeroed out on each and every
2653          * writepage invocation because it may be mmapped.  "A file is mapped
2654          * in multiples of the page size.  For a file that is not a multiple of
2655          * the  page size, the remaining memory is zeroed when mapped, and
2656          * writes to that region are not written out to the file."
2657          */
2658         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2659 out:
2660         ret = mpage_writepage(page, get_block, wbc);
2661         if (ret == -EAGAIN)
2662                 ret = __block_write_full_page(inode, page, get_block, wbc,
2663                                               end_buffer_async_write);
2664         return ret;
2665 }
2666 EXPORT_SYMBOL(nobh_writepage);
2667
2668 int nobh_truncate_page(struct address_space *mapping,
2669                         loff_t from, get_block_t *get_block)
2670 {
2671         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2672         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2673         unsigned blocksize;
2674         sector_t iblock;
2675         unsigned length, pos;
2676         struct inode *inode = mapping->host;
2677         struct page *page;
2678         struct buffer_head map_bh;
2679         int err;
2680
2681         blocksize = 1 << inode->i_blkbits;
2682         length = offset & (blocksize - 1);
2683
2684         /* Block boundary? Nothing to do */
2685         if (!length)
2686                 return 0;
2687
2688         length = blocksize - length;
2689         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2690
2691         page = grab_cache_page(mapping, index);
2692         err = -ENOMEM;
2693         if (!page)
2694                 goto out;
2695
2696         if (page_has_buffers(page)) {
2697 has_buffers:
2698                 unlock_page(page);
2699                 page_cache_release(page);
2700                 return block_truncate_page(mapping, from, get_block);
2701         }
2702
2703         /* Find the buffer that contains "offset" */
2704         pos = blocksize;
2705         while (offset >= pos) {
2706                 iblock++;
2707                 pos += blocksize;
2708         }
2709
2710         map_bh.b_size = blocksize;
2711         map_bh.b_state = 0;
2712         err = get_block(inode, iblock, &map_bh, 0);
2713         if (err)
2714                 goto unlock;
2715         /* unmapped? It's a hole - nothing to do */
2716         if (!buffer_mapped(&map_bh))
2717                 goto unlock;
2718
2719         /* Ok, it's mapped. Make sure it's up-to-date */
2720         if (!PageUptodate(page)) {
2721                 err = mapping->a_ops->readpage(NULL, page);
2722                 if (err) {
2723                         page_cache_release(page);
2724                         goto out;
2725                 }
2726                 lock_page(page);
2727                 if (!PageUptodate(page)) {
2728                         err = -EIO;
2729                         goto unlock;
2730                 }
2731                 if (page_has_buffers(page))
2732                         goto has_buffers;
2733         }
2734         zero_user(page, offset, length);
2735         set_page_dirty(page);
2736         err = 0;
2737
2738 unlock:
2739         unlock_page(page);
2740         page_cache_release(page);
2741 out:
2742         return err;
2743 }
2744 EXPORT_SYMBOL(nobh_truncate_page);
2745
2746 int block_truncate_page(struct address_space *mapping,
2747                         loff_t from, get_block_t *get_block)
2748 {
2749         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2750         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2751         unsigned blocksize;
2752         sector_t iblock;
2753         unsigned length, pos;
2754         struct inode *inode = mapping->host;
2755         struct page *page;
2756         struct buffer_head *bh;
2757         int err;
2758
2759         blocksize = 1 << inode->i_blkbits;
2760         length = offset & (blocksize - 1);
2761
2762         /* Block boundary? Nothing to do */
2763         if (!length)
2764                 return 0;
2765
2766         length = blocksize - length;
2767         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2768         
2769         page = grab_cache_page(mapping, index);
2770         err = -ENOMEM;
2771         if (!page)
2772                 goto out;
2773
2774         if (!page_has_buffers(page))
2775                 create_empty_buffers(page, blocksize, 0);
2776
2777         /* Find the buffer that contains "offset" */
2778         bh = page_buffers(page);
2779         pos = blocksize;
2780         while (offset >= pos) {
2781                 bh = bh->b_this_page;
2782                 iblock++;
2783                 pos += blocksize;
2784         }
2785
2786         err = 0;
2787         if (!buffer_mapped(bh)) {
2788                 WARN_ON(bh->b_size != blocksize);
2789                 err = get_block(inode, iblock, bh, 0);
2790                 if (err)
2791                         goto unlock;
2792                 /* unmapped? It's a hole - nothing to do */
2793                 if (!buffer_mapped(bh))
2794                         goto unlock;
2795         }
2796
2797         /* Ok, it's mapped. Make sure it's up-to-date */
2798         if (PageUptodate(page))
2799                 set_buffer_uptodate(bh);
2800
2801         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2802                 err = -EIO;
2803                 ll_rw_block(READ, 1, &bh);
2804                 wait_on_buffer(bh);
2805                 /* Uhhuh. Read error. Complain and punt. */
2806                 if (!buffer_uptodate(bh))
2807                         goto unlock;
2808         }
2809
2810         zero_user(page, offset, length);
2811         mark_buffer_dirty(bh);
2812         err = 0;
2813
2814 unlock:
2815         unlock_page(page);
2816         page_cache_release(page);
2817 out:
2818         return err;
2819 }
2820 EXPORT_SYMBOL(block_truncate_page);
2821
2822 /*
2823  * The generic ->writepage function for buffer-backed address_spaces
2824  * this form passes in the end_io handler used to finish the IO.
2825  */
2826 int block_write_full_page_endio(struct page *page, get_block_t *get_block,
2827                         struct writeback_control *wbc, bh_end_io_t *handler)
2828 {
2829         struct inode * const inode = page->mapping->host;
2830         loff_t i_size = i_size_read(inode);
2831         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2832         unsigned offset;
2833
2834         /* Is the page fully inside i_size? */
2835         if (page->index < end_index)
2836                 return __block_write_full_page(inode, page, get_block, wbc,
2837                                                handler);
2838
2839         /* Is the page fully outside i_size? (truncate in progress) */
2840         offset = i_size & (PAGE_CACHE_SIZE-1);
2841         if (page->index >= end_index+1 || !offset) {
2842                 /*
2843                  * The page may have dirty, unmapped buffers.  For example,
2844                  * they may have been added in ext3_writepage().  Make them
2845                  * freeable here, so the page does not leak.
2846                  */
2847                 do_invalidatepage(page, 0);
2848                 unlock_page(page);
2849                 return 0; /* don't care */
2850         }
2851
2852         /*
2853          * The page straddles i_size.  It must be zeroed out on each and every
2854          * writepage invocation because it may be mmapped.  "A file is mapped
2855          * in multiples of the page size.  For a file that is not a multiple of
2856          * the  page size, the remaining memory is zeroed when mapped, and
2857          * writes to that region are not written out to the file."
2858          */
2859         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2860         return __block_write_full_page(inode, page, get_block, wbc, handler);
2861 }
2862 EXPORT_SYMBOL(block_write_full_page_endio);
2863
2864 /*
2865  * The generic ->writepage function for buffer-backed address_spaces
2866  */
2867 int block_write_full_page(struct page *page, get_block_t *get_block,
2868                         struct writeback_control *wbc)
2869 {
2870         return block_write_full_page_endio(page, get_block, wbc,
2871                                            end_buffer_async_write);
2872 }
2873 EXPORT_SYMBOL(block_write_full_page);
2874
2875 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2876                             get_block_t *get_block)
2877 {
2878         struct buffer_head tmp;
2879         struct inode *inode = mapping->host;
2880         tmp.b_state = 0;
2881         tmp.b_blocknr = 0;
2882         tmp.b_size = 1 << inode->i_blkbits;
2883         get_block(inode, block, &tmp, 0);
2884         return tmp.b_blocknr;
2885 }
2886 EXPORT_SYMBOL(generic_block_bmap);
2887
2888 static void end_bio_bh_io_sync(struct bio *bio, int err)
2889 {
2890         struct buffer_head *bh = bio->bi_private;
2891
2892         if (err == -EOPNOTSUPP) {
2893                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2894         }
2895
2896         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2897                 set_bit(BH_Quiet, &bh->b_state);
2898
2899         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2900         bio_put(bio);
2901 }
2902
2903 int submit_bh(int rw, struct buffer_head * bh)
2904 {
2905         struct bio *bio;
2906         int ret = 0;
2907
2908         BUG_ON(!buffer_locked(bh));
2909         BUG_ON(!buffer_mapped(bh));
2910         BUG_ON(!bh->b_end_io);
2911         BUG_ON(buffer_delay(bh));
2912         BUG_ON(buffer_unwritten(bh));
2913
2914         /*
2915          * Only clear out a write error when rewriting
2916          */
2917         if (test_set_buffer_req(bh) && (rw & WRITE))
2918                 clear_buffer_write_io_error(bh);
2919
2920         /*
2921          * from here on down, it's all bio -- do the initial mapping,
2922          * submit_bio -> generic_make_request may further map this bio around
2923          */
2924         bio = bio_alloc(GFP_NOIO, 1);
2925
2926         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2927         bio->bi_bdev = bh->b_bdev;
2928         bio->bi_io_vec[0].bv_page = bh->b_page;
2929         bio->bi_io_vec[0].bv_len = bh->b_size;
2930         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2931
2932         bio->bi_vcnt = 1;
2933         bio->bi_idx = 0;
2934         bio->bi_size = bh->b_size;
2935
2936         bio->bi_end_io = end_bio_bh_io_sync;
2937         bio->bi_private = bh;
2938
2939         bio_get(bio);
2940         submit_bio(rw, bio);
2941
2942         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2943                 ret = -EOPNOTSUPP;
2944
2945         bio_put(bio);
2946         return ret;
2947 }
2948 EXPORT_SYMBOL(submit_bh);
2949
2950 /**
2951  * ll_rw_block: low-level access to block devices (DEPRECATED)
2952  * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
2953  * @nr: number of &struct buffer_heads in the array
2954  * @bhs: array of pointers to &struct buffer_head
2955  *
2956  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2957  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2958  * %READA option is described in the documentation for generic_make_request()
2959  * which ll_rw_block() calls.
2960  *
2961  * This function drops any buffer that it cannot get a lock on (with the
2962  * BH_Lock state bit), any buffer that appears to be clean when doing a write
2963  * request, and any buffer that appears to be up-to-date when doing read
2964  * request.  Further it marks as clean buffers that are processed for
2965  * writing (the buffer cache won't assume that they are actually clean
2966  * until the buffer gets unlocked).
2967  *
2968  * ll_rw_block sets b_end_io to simple completion handler that marks
2969  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2970  * any waiters. 
2971  *
2972  * All of the buffers must be for the same device, and must also be a
2973  * multiple of the current approved size for the device.
2974  */
2975 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2976 {
2977         int i;
2978
2979         for (i = 0; i < nr; i++) {
2980                 struct buffer_head *bh = bhs[i];
2981
2982                 if (!trylock_buffer(bh))
2983                         continue;
2984                 if (rw == WRITE) {
2985                         if (test_clear_buffer_dirty(bh)) {
2986                                 bh->b_end_io = end_buffer_write_sync;
2987                                 get_bh(bh);
2988                                 submit_bh(WRITE, bh);
2989                                 continue;
2990                         }
2991                 } else {
2992                         if (!buffer_uptodate(bh)) {
2993                                 bh->b_end_io = end_buffer_read_sync;
2994                                 get_bh(bh);
2995                                 submit_bh(rw, bh);
2996                                 continue;
2997                         }
2998                 }
2999                 unlock_buffer(bh);
3000         }
3001 }
3002 EXPORT_SYMBOL(ll_rw_block);
3003
3004 void write_dirty_buffer(struct buffer_head *bh, int rw)
3005 {
3006         lock_buffer(bh);
3007         if (!test_clear_buffer_dirty(bh)) {
3008                 unlock_buffer(bh);
3009                 return;
3010         }
3011         bh->b_end_io = end_buffer_write_sync;
3012         get_bh(bh);
3013         submit_bh(rw, bh);
3014 }
3015 EXPORT_SYMBOL(write_dirty_buffer);
3016
3017 /*
3018  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3019  * and then start new I/O and then wait upon it.  The caller must have a ref on
3020  * the buffer_head.
3021  */
3022 int __sync_dirty_buffer(struct buffer_head *bh, int rw)
3023 {
3024         int ret = 0;
3025
3026         WARN_ON(atomic_read(&bh->b_count) < 1);
3027         lock_buffer(bh);
3028         if (test_clear_buffer_dirty(bh)) {
3029                 get_bh(bh);
3030                 bh->b_end_io = end_buffer_write_sync;
3031                 ret = submit_bh(rw, bh);
3032                 wait_on_buffer(bh);
3033                 if (!ret && !buffer_uptodate(bh))
3034                         ret = -EIO;
3035         } else {
3036                 unlock_buffer(bh);
3037         }
3038         return ret;
3039 }
3040 EXPORT_SYMBOL(__sync_dirty_buffer);
3041
3042 int sync_dirty_buffer(struct buffer_head *bh)
3043 {
3044         return __sync_dirty_buffer(bh, WRITE_SYNC);
3045 }
3046 EXPORT_SYMBOL(sync_dirty_buffer);
3047
3048 /*
3049  * try_to_free_buffers() checks if all the buffers on this particular page
3050  * are unused, and releases them if so.
3051  *
3052  * Exclusion against try_to_free_buffers may be obtained by either
3053  * locking the page or by holding its mapping's private_lock.
3054  *
3055  * If the page is dirty but all the buffers are clean then we need to
3056  * be sure to mark the page clean as well.  This is because the page
3057  * may be against a block device, and a later reattachment of buffers
3058  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3059  * filesystem data on the same device.
3060  *
3061  * The same applies to regular filesystem pages: if all the buffers are
3062  * clean then we set the page clean and proceed.  To do that, we require
3063  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3064  * private_lock.
3065  *
3066  * try_to_free_buffers() is non-blocking.
3067  */
3068 static inline int buffer_busy(struct buffer_head *bh)
3069 {
3070         return atomic_read(&bh->b_count) |
3071                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3072 }
3073
3074 static int
3075 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3076 {
3077         struct buffer_head *head = page_buffers(page);
3078         struct buffer_head *bh;
3079
3080         bh = head;
3081         do {
3082                 if (buffer_write_io_error(bh) && page->mapping)
3083                         set_bit(AS_EIO, &page->mapping->flags);
3084                 if (buffer_busy(bh))
3085                         goto failed;
3086                 bh = bh->b_this_page;
3087         } while (bh != head);
3088
3089         do {
3090                 struct buffer_head *next = bh->b_this_page;
3091
3092                 if (bh->b_assoc_map)
3093                         __remove_assoc_queue(bh);
3094                 bh = next;
3095         } while (bh != head);
3096         *buffers_to_free = head;
3097         __clear_page_buffers(page);
3098         return 1;
3099 failed:
3100         return 0;
3101 }
3102
3103 int try_to_free_buffers(struct page *page)
3104 {
3105         struct address_space * const mapping = page->mapping;
3106         struct buffer_head *buffers_to_free = NULL;
3107         int ret = 0;
3108
3109         BUG_ON(!PageLocked(page));
3110         if (PageWriteback(page))
3111                 return 0;
3112
3113         if (mapping == NULL) {          /* can this still happen? */
3114                 ret = drop_buffers(page, &buffers_to_free);
3115                 goto out;
3116         }
3117
3118         spin_lock(&mapping->private_lock);
3119         ret = drop_buffers(page, &buffers_to_free);
3120
3121         /*
3122          * If the filesystem writes its buffers by hand (eg ext3)
3123          * then we can have clean buffers against a dirty page.  We
3124          * clean the page here; otherwise the VM will never notice
3125          * that the filesystem did any IO at all.
3126          *
3127          * Also, during truncate, discard_buffer will have marked all
3128          * the page's buffers clean.  We discover that here and clean
3129          * the page also.
3130          *
3131          * private_lock must be held over this entire operation in order
3132          * to synchronise against __set_page_dirty_buffers and prevent the
3133          * dirty bit from being lost.
3134          */
3135         if (ret)
3136                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3137         spin_unlock(&mapping->private_lock);
3138 out:
3139         if (buffers_to_free) {
3140                 struct buffer_head *bh = buffers_to_free;
3141
3142                 do {
3143                         struct buffer_head *next = bh->b_this_page;
3144                         free_buffer_head(bh);
3145                         bh = next;
3146                 } while (bh != buffers_to_free);
3147         }
3148         return ret;
3149 }
3150 EXPORT_SYMBOL(try_to_free_buffers);
3151
3152 void block_sync_page(struct page *page)
3153 {
3154         struct address_space *mapping;
3155
3156         smp_mb();
3157         mapping = page_mapping(page);
3158         if (mapping)
3159                 blk_run_backing_dev(mapping->backing_dev_info, page);
3160 }
3161 EXPORT_SYMBOL(block_sync_page);
3162
3163 /*
3164  * There are no bdflush tunables left.  But distributions are
3165  * still running obsolete flush daemons, so we terminate them here.
3166  *
3167  * Use of bdflush() is deprecated and will be removed in a future kernel.
3168  * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3169  */
3170 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3171 {
3172         static int msg_count;
3173
3174         if (!capable(CAP_SYS_ADMIN))
3175                 return -EPERM;
3176
3177         if (msg_count < 5) {
3178                 msg_count++;
3179                 printk(KERN_INFO
3180                         "warning: process `%s' used the obsolete bdflush"
3181                         " system call\n", current->comm);
3182                 printk(KERN_INFO "Fix your initscripts?\n");
3183         }
3184
3185         if (func == 1)
3186                 do_exit(0);
3187         return 0;
3188 }
3189
3190 /*
3191  * Buffer-head allocation
3192  */
3193 static struct kmem_cache *bh_cachep;
3194
3195 /*
3196  * Once the number of bh's in the machine exceeds this level, we start
3197  * stripping them in writeback.
3198  */
3199 static int max_buffer_heads;
3200
3201 int buffer_heads_over_limit;
3202
3203 struct bh_accounting {
3204         int nr;                 /* Number of live bh's */
3205         int ratelimit;          /* Limit cacheline bouncing */
3206 };
3207
3208 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3209
3210 static void recalc_bh_state(void)
3211 {
3212         int i;
3213         int tot = 0;
3214
3215         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3216                 return;
3217         __get_cpu_var(bh_accounting).ratelimit = 0;
3218         for_each_online_cpu(i)
3219                 tot += per_cpu(bh_accounting, i).nr;
3220         buffer_heads_over_limit = (tot > max_buffer_heads);
3221 }
3222         
3223 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3224 {
3225         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3226         if (ret) {
3227                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3228                 get_cpu_var(bh_accounting).nr++;
3229                 recalc_bh_state();
3230                 put_cpu_var(bh_accounting);
3231         }
3232         return ret;
3233 }
3234 EXPORT_SYMBOL(alloc_buffer_head);
3235
3236 void free_buffer_head(struct buffer_head *bh)
3237 {
3238         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3239         kmem_cache_free(bh_cachep, bh);
3240         get_cpu_var(bh_accounting).nr--;
3241         recalc_bh_state();
3242         put_cpu_var(bh_accounting);
3243 }
3244 EXPORT_SYMBOL(free_buffer_head);
3245
3246 static void buffer_exit_cpu(int cpu)
3247 {
3248         int i;
3249         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3250
3251         for (i = 0; i < BH_LRU_SIZE; i++) {
3252                 brelse(b->bhs[i]);
3253                 b->bhs[i] = NULL;
3254         }
3255         get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3256         per_cpu(bh_accounting, cpu).nr = 0;
3257         put_cpu_var(bh_accounting);
3258 }
3259
3260 static int buffer_cpu_notify(struct notifier_block *self,
3261                               unsigned long action, void *hcpu)
3262 {
3263         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3264                 buffer_exit_cpu((unsigned long)hcpu);
3265         return NOTIFY_OK;
3266 }
3267
3268 /**
3269  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3270  * @bh: struct buffer_head
3271  *
3272  * Return true if the buffer is up-to-date and false,
3273  * with the buffer locked, if not.
3274  */
3275 int bh_uptodate_or_lock(struct buffer_head *bh)
3276 {
3277         if (!buffer_uptodate(bh)) {
3278                 lock_buffer(bh);
3279                 if (!buffer_uptodate(bh))
3280                         return 0;
3281                 unlock_buffer(bh);
3282         }
3283         return 1;
3284 }
3285 EXPORT_SYMBOL(bh_uptodate_or_lock);
3286
3287 /**
3288  * bh_submit_read - Submit a locked buffer for reading
3289  * @bh: struct buffer_head
3290  *
3291  * Returns zero on success and -EIO on error.
3292  */
3293 int bh_submit_read(struct buffer_head *bh)
3294 {
3295         BUG_ON(!buffer_locked(bh));
3296
3297         if (buffer_uptodate(bh)) {
3298                 unlock_buffer(bh);
3299                 return 0;
3300         }
3301
3302         get_bh(bh);
3303         bh->b_end_io = end_buffer_read_sync;
3304         submit_bh(READ, bh);
3305         wait_on_buffer(bh);
3306         if (buffer_uptodate(bh))
3307                 return 0;
3308         return -EIO;
3309 }
3310 EXPORT_SYMBOL(bh_submit_read);
3311
3312 void __init buffer_init(void)
3313 {
3314         int nrpages;
3315
3316         bh_cachep = kmem_cache_create("buffer_head",
3317                         sizeof(struct buffer_head), 0,
3318                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3319                                 SLAB_MEM_SPREAD),
3320                                 NULL);
3321
3322         /*
3323          * Limit the bh occupancy to 10% of ZONE_NORMAL
3324          */
3325         nrpages = (nr_free_buffer_pages() * 10) / 100;
3326         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3327         hotcpu_notifier(buffer_cpu_notify, 0);
3328 }