5930e382959bc504c58bbb428588a372742d4aa4
[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_size = size;
909
910                 /* Link the buffer to its page */
911                 set_bh_page(bh, page, offset);
912
913                 init_buffer(bh, NULL, NULL);
914         }
915         return head;
916 /*
917  * In case anything failed, we just free everything we got.
918  */
919 no_grow:
920         if (head) {
921                 do {
922                         bh = head;
923                         head = head->b_this_page;
924                         free_buffer_head(bh);
925                 } while (head);
926         }
927
928         /*
929          * Return failure for non-async IO requests.  Async IO requests
930          * are not allowed to fail, so we have to wait until buffer heads
931          * become available.  But we don't want tasks sleeping with 
932          * partially complete buffers, so all were released above.
933          */
934         if (!retry)
935                 return NULL;
936
937         /* We're _really_ low on memory. Now we just
938          * wait for old buffer heads to become free due to
939          * finishing IO.  Since this is an async request and
940          * the reserve list is empty, we're sure there are 
941          * async buffer heads in use.
942          */
943         free_more_memory();
944         goto try_again;
945 }
946 EXPORT_SYMBOL_GPL(alloc_page_buffers);
947
948 static inline void
949 link_dev_buffers(struct page *page, struct buffer_head *head)
950 {
951         struct buffer_head *bh, *tail;
952
953         bh = head;
954         do {
955                 tail = bh;
956                 bh = bh->b_this_page;
957         } while (bh);
958         tail->b_this_page = head;
959         attach_page_buffers(page, head);
960 }
961
962 /*
963  * Initialise the state of a blockdev page's buffers.
964  */ 
965 static void
966 init_page_buffers(struct page *page, struct block_device *bdev,
967                         sector_t block, int size)
968 {
969         struct buffer_head *head = page_buffers(page);
970         struct buffer_head *bh = head;
971         int uptodate = PageUptodate(page);
972
973         do {
974                 if (!buffer_mapped(bh)) {
975                         init_buffer(bh, NULL, NULL);
976                         bh->b_bdev = bdev;
977                         bh->b_blocknr = block;
978                         if (uptodate)
979                                 set_buffer_uptodate(bh);
980                         set_buffer_mapped(bh);
981                 }
982                 block++;
983                 bh = bh->b_this_page;
984         } while (bh != head);
985 }
986
987 /*
988  * Create the page-cache page that contains the requested block.
989  *
990  * This is user purely for blockdev mappings.
991  */
992 static struct page *
993 grow_dev_page(struct block_device *bdev, sector_t block,
994                 pgoff_t index, int size)
995 {
996         struct inode *inode = bdev->bd_inode;
997         struct page *page;
998         struct buffer_head *bh;
999
1000         page = find_or_create_page(inode->i_mapping, index,
1001                 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
1002         if (!page)
1003                 return NULL;
1004
1005         BUG_ON(!PageLocked(page));
1006
1007         if (page_has_buffers(page)) {
1008                 bh = page_buffers(page);
1009                 if (bh->b_size == size) {
1010                         init_page_buffers(page, bdev, block, size);
1011                         return page;
1012                 }
1013                 if (!try_to_free_buffers(page))
1014                         goto failed;
1015         }
1016
1017         /*
1018          * Allocate some buffers for this page
1019          */
1020         bh = alloc_page_buffers(page, size, 0);
1021         if (!bh)
1022                 goto failed;
1023
1024         /*
1025          * Link the page to the buffers and initialise them.  Take the
1026          * lock to be atomic wrt __find_get_block(), which does not
1027          * run under the page lock.
1028          */
1029         spin_lock(&inode->i_mapping->private_lock);
1030         link_dev_buffers(page, bh);
1031         init_page_buffers(page, bdev, block, size);
1032         spin_unlock(&inode->i_mapping->private_lock);
1033         return page;
1034
1035 failed:
1036         BUG();
1037         unlock_page(page);
1038         page_cache_release(page);
1039         return NULL;
1040 }
1041
1042 /*
1043  * Create buffers for the specified block device block's page.  If
1044  * that page was dirty, the buffers are set dirty also.
1045  */
1046 static int
1047 grow_buffers(struct block_device *bdev, sector_t block, int size)
1048 {
1049         struct page *page;
1050         pgoff_t index;
1051         int sizebits;
1052
1053         sizebits = -1;
1054         do {
1055                 sizebits++;
1056         } while ((size << sizebits) < PAGE_SIZE);
1057
1058         index = block >> sizebits;
1059
1060         /*
1061          * Check for a block which wants to lie outside our maximum possible
1062          * pagecache index.  (this comparison is done using sector_t types).
1063          */
1064         if (unlikely(index != block >> sizebits)) {
1065                 char b[BDEVNAME_SIZE];
1066
1067                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1068                         "device %s\n",
1069                         __func__, (unsigned long long)block,
1070                         bdevname(bdev, b));
1071                 return -EIO;
1072         }
1073         block = index << sizebits;
1074         /* Create a page with the proper size buffers.. */
1075         page = grow_dev_page(bdev, block, index, size);
1076         if (!page)
1077                 return 0;
1078         unlock_page(page);
1079         page_cache_release(page);
1080         return 1;
1081 }
1082
1083 static struct buffer_head *
1084 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1085 {
1086         /* Size must be multiple of hard sectorsize */
1087         if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1088                         (size < 512 || size > PAGE_SIZE))) {
1089                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1090                                         size);
1091                 printk(KERN_ERR "logical block size: %d\n",
1092                                         bdev_logical_block_size(bdev));
1093
1094                 dump_stack();
1095                 return NULL;
1096         }
1097
1098         for (;;) {
1099                 struct buffer_head * bh;
1100                 int ret;
1101
1102                 bh = __find_get_block(bdev, block, size);
1103                 if (bh)
1104                         return bh;
1105
1106                 ret = grow_buffers(bdev, block, size);
1107                 if (ret < 0)
1108                         return NULL;
1109                 if (ret == 0)
1110                         free_more_memory();
1111         }
1112 }
1113
1114 /*
1115  * The relationship between dirty buffers and dirty pages:
1116  *
1117  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1118  * the page is tagged dirty in its radix tree.
1119  *
1120  * At all times, the dirtiness of the buffers represents the dirtiness of
1121  * subsections of the page.  If the page has buffers, the page dirty bit is
1122  * merely a hint about the true dirty state.
1123  *
1124  * When a page is set dirty in its entirety, all its buffers are marked dirty
1125  * (if the page has buffers).
1126  *
1127  * When a buffer is marked dirty, its page is dirtied, but the page's other
1128  * buffers are not.
1129  *
1130  * Also.  When blockdev buffers are explicitly read with bread(), they
1131  * individually become uptodate.  But their backing page remains not
1132  * uptodate - even if all of its buffers are uptodate.  A subsequent
1133  * block_read_full_page() against that page will discover all the uptodate
1134  * buffers, will set the page uptodate and will perform no I/O.
1135  */
1136
1137 /**
1138  * mark_buffer_dirty - mark a buffer_head as needing writeout
1139  * @bh: the buffer_head to mark dirty
1140  *
1141  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1142  * backing page dirty, then tag the page as dirty in its address_space's radix
1143  * tree and then attach the address_space's inode to its superblock's dirty
1144  * inode list.
1145  *
1146  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1147  * mapping->tree_lock and the global inode_lock.
1148  */
1149 void mark_buffer_dirty(struct buffer_head *bh)
1150 {
1151         WARN_ON_ONCE(!buffer_uptodate(bh));
1152
1153         /*
1154          * Very *carefully* optimize the it-is-already-dirty case.
1155          *
1156          * Don't let the final "is it dirty" escape to before we
1157          * perhaps modified the buffer.
1158          */
1159         if (buffer_dirty(bh)) {
1160                 smp_mb();
1161                 if (buffer_dirty(bh))
1162                         return;
1163         }
1164
1165         if (!test_set_buffer_dirty(bh)) {
1166                 struct page *page = bh->b_page;
1167                 if (!TestSetPageDirty(page)) {
1168                         struct address_space *mapping = page_mapping(page);
1169                         if (mapping)
1170                                 __set_page_dirty(page, mapping, 0);
1171                 }
1172         }
1173 }
1174 EXPORT_SYMBOL(mark_buffer_dirty);
1175
1176 /*
1177  * Decrement a buffer_head's reference count.  If all buffers against a page
1178  * have zero reference count, are clean and unlocked, and if the page is clean
1179  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1180  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1181  * a page but it ends up not being freed, and buffers may later be reattached).
1182  */
1183 void __brelse(struct buffer_head * buf)
1184 {
1185         if (atomic_read(&buf->b_count)) {
1186                 put_bh(buf);
1187                 return;
1188         }
1189         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1190 }
1191 EXPORT_SYMBOL(__brelse);
1192
1193 /*
1194  * bforget() is like brelse(), except it discards any
1195  * potentially dirty data.
1196  */
1197 void __bforget(struct buffer_head *bh)
1198 {
1199         clear_buffer_dirty(bh);
1200         if (bh->b_assoc_map) {
1201                 struct address_space *buffer_mapping = bh->b_page->mapping;
1202
1203                 spin_lock(&buffer_mapping->private_lock);
1204                 list_del_init(&bh->b_assoc_buffers);
1205                 bh->b_assoc_map = NULL;
1206                 spin_unlock(&buffer_mapping->private_lock);
1207         }
1208         __brelse(bh);
1209 }
1210 EXPORT_SYMBOL(__bforget);
1211
1212 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1213 {
1214         lock_buffer(bh);
1215         if (buffer_uptodate(bh)) {
1216                 unlock_buffer(bh);
1217                 return bh;
1218         } else {
1219                 get_bh(bh);
1220                 bh->b_end_io = end_buffer_read_sync;
1221                 submit_bh(READ, bh);
1222                 wait_on_buffer(bh);
1223                 if (buffer_uptodate(bh))
1224                         return bh;
1225         }
1226         brelse(bh);
1227         return NULL;
1228 }
1229
1230 /*
1231  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1232  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1233  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1234  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1235  * CPU's LRUs at the same time.
1236  *
1237  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1238  * sb_find_get_block().
1239  *
1240  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1241  * a local interrupt disable for that.
1242  */
1243
1244 #define BH_LRU_SIZE     8
1245
1246 struct bh_lru {
1247         struct buffer_head *bhs[BH_LRU_SIZE];
1248 };
1249
1250 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1251
1252 #ifdef CONFIG_SMP
1253 #define bh_lru_lock()   local_irq_disable()
1254 #define bh_lru_unlock() local_irq_enable()
1255 #else
1256 #define bh_lru_lock()   preempt_disable()
1257 #define bh_lru_unlock() preempt_enable()
1258 #endif
1259
1260 static inline void check_irqs_on(void)
1261 {
1262 #ifdef irqs_disabled
1263         BUG_ON(irqs_disabled());
1264 #endif
1265 }
1266
1267 /*
1268  * The LRU management algorithm is dopey-but-simple.  Sorry.
1269  */
1270 static void bh_lru_install(struct buffer_head *bh)
1271 {
1272         struct buffer_head *evictee = NULL;
1273         struct bh_lru *lru;
1274
1275         check_irqs_on();
1276         bh_lru_lock();
1277         lru = &__get_cpu_var(bh_lrus);
1278         if (lru->bhs[0] != bh) {
1279                 struct buffer_head *bhs[BH_LRU_SIZE];
1280                 int in;
1281                 int out = 0;
1282
1283                 get_bh(bh);
1284                 bhs[out++] = bh;
1285                 for (in = 0; in < BH_LRU_SIZE; in++) {
1286                         struct buffer_head *bh2 = lru->bhs[in];
1287
1288                         if (bh2 == bh) {
1289                                 __brelse(bh2);
1290                         } else {
1291                                 if (out >= BH_LRU_SIZE) {
1292                                         BUG_ON(evictee != NULL);
1293                                         evictee = bh2;
1294                                 } else {
1295                                         bhs[out++] = bh2;
1296                                 }
1297                         }
1298                 }
1299                 while (out < BH_LRU_SIZE)
1300                         bhs[out++] = NULL;
1301                 memcpy(lru->bhs, bhs, sizeof(bhs));
1302         }
1303         bh_lru_unlock();
1304
1305         if (evictee)
1306                 __brelse(evictee);
1307 }
1308
1309 /*
1310  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1311  */
1312 static struct buffer_head *
1313 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1314 {
1315         struct buffer_head *ret = NULL;
1316         struct bh_lru *lru;
1317         unsigned int i;
1318
1319         check_irqs_on();
1320         bh_lru_lock();
1321         lru = &__get_cpu_var(bh_lrus);
1322         for (i = 0; i < BH_LRU_SIZE; i++) {
1323                 struct buffer_head *bh = lru->bhs[i];
1324
1325                 if (bh && bh->b_bdev == bdev &&
1326                                 bh->b_blocknr == block && bh->b_size == size) {
1327                         if (i) {
1328                                 while (i) {
1329                                         lru->bhs[i] = lru->bhs[i - 1];
1330                                         i--;
1331                                 }
1332                                 lru->bhs[0] = bh;
1333                         }
1334                         get_bh(bh);
1335                         ret = bh;
1336                         break;
1337                 }
1338         }
1339         bh_lru_unlock();
1340         return ret;
1341 }
1342
1343 /*
1344  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1345  * it in the LRU and mark it as accessed.  If it is not present then return
1346  * NULL
1347  */
1348 struct buffer_head *
1349 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1350 {
1351         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1352
1353         if (bh == NULL) {
1354                 bh = __find_get_block_slow(bdev, block);
1355                 if (bh)
1356                         bh_lru_install(bh);
1357         }
1358         if (bh)
1359                 touch_buffer(bh);
1360         return bh;
1361 }
1362 EXPORT_SYMBOL(__find_get_block);
1363
1364 /*
1365  * __getblk will locate (and, if necessary, create) the buffer_head
1366  * which corresponds to the passed block_device, block and size. The
1367  * returned buffer has its reference count incremented.
1368  *
1369  * __getblk() cannot fail - it just keeps trying.  If you pass it an
1370  * illegal block number, __getblk() will happily return a buffer_head
1371  * which represents the non-existent block.  Very weird.
1372  *
1373  * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1374  * attempt is failing.  FIXME, perhaps?
1375  */
1376 struct buffer_head *
1377 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1378 {
1379         struct buffer_head *bh = __find_get_block(bdev, block, size);
1380
1381         might_sleep();
1382         if (bh == NULL)
1383                 bh = __getblk_slow(bdev, block, size);
1384         return bh;
1385 }
1386 EXPORT_SYMBOL(__getblk);
1387
1388 /*
1389  * Do async read-ahead on a buffer..
1390  */
1391 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1392 {
1393         struct buffer_head *bh = __getblk(bdev, block, size);
1394         if (likely(bh)) {
1395                 ll_rw_block(READA, 1, &bh);
1396                 brelse(bh);
1397         }
1398 }
1399 EXPORT_SYMBOL(__breadahead);
1400
1401 /**
1402  *  __bread() - reads a specified block and returns the bh
1403  *  @bdev: the block_device to read from
1404  *  @block: number of block
1405  *  @size: size (in bytes) to read
1406  * 
1407  *  Reads a specified block, and returns buffer head that contains it.
1408  *  It returns NULL if the block was unreadable.
1409  */
1410 struct buffer_head *
1411 __bread(struct block_device *bdev, sector_t block, unsigned size)
1412 {
1413         struct buffer_head *bh = __getblk(bdev, block, size);
1414
1415         if (likely(bh) && !buffer_uptodate(bh))
1416                 bh = __bread_slow(bh);
1417         return bh;
1418 }
1419 EXPORT_SYMBOL(__bread);
1420
1421 /*
1422  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1423  * This doesn't race because it runs in each cpu either in irq
1424  * or with preempt disabled.
1425  */
1426 static void invalidate_bh_lru(void *arg)
1427 {
1428         struct bh_lru *b = &get_cpu_var(bh_lrus);
1429         int i;
1430
1431         for (i = 0; i < BH_LRU_SIZE; i++) {
1432                 brelse(b->bhs[i]);
1433                 b->bhs[i] = NULL;
1434         }
1435         put_cpu_var(bh_lrus);
1436 }
1437         
1438 void invalidate_bh_lrus(void)
1439 {
1440         on_each_cpu(invalidate_bh_lru, NULL, 1);
1441 }
1442 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1443
1444 void set_bh_page(struct buffer_head *bh,
1445                 struct page *page, unsigned long offset)
1446 {
1447         bh->b_page = page;
1448         BUG_ON(offset >= PAGE_SIZE);
1449         if (PageHighMem(page))
1450                 /*
1451                  * This catches illegal uses and preserves the offset:
1452                  */
1453                 bh->b_data = (char *)(0 + offset);
1454         else
1455                 bh->b_data = page_address(page) + offset;
1456 }
1457 EXPORT_SYMBOL(set_bh_page);
1458
1459 /*
1460  * Called when truncating a buffer on a page completely.
1461  */
1462 static void discard_buffer(struct buffer_head * bh)
1463 {
1464         lock_buffer(bh);
1465         clear_buffer_dirty(bh);
1466         bh->b_bdev = NULL;
1467         clear_buffer_mapped(bh);
1468         clear_buffer_req(bh);
1469         clear_buffer_new(bh);
1470         clear_buffer_delay(bh);
1471         clear_buffer_unwritten(bh);
1472         unlock_buffer(bh);
1473 }
1474
1475 /**
1476  * block_invalidatepage - invalidate part of all of a buffer-backed page
1477  *
1478  * @page: the page which is affected
1479  * @offset: the index of the truncation point
1480  *
1481  * block_invalidatepage() is called when all or part of the page has become
1482  * invalidatedby a truncate operation.
1483  *
1484  * block_invalidatepage() does not have to release all buffers, but it must
1485  * ensure that no dirty buffer is left outside @offset and that no I/O
1486  * is underway against any of the blocks which are outside the truncation
1487  * point.  Because the caller is about to free (and possibly reuse) those
1488  * blocks on-disk.
1489  */
1490 void block_invalidatepage(struct page *page, unsigned long offset)
1491 {
1492         struct buffer_head *head, *bh, *next;
1493         unsigned int curr_off = 0;
1494
1495         BUG_ON(!PageLocked(page));
1496         if (!page_has_buffers(page))
1497                 goto out;
1498
1499         head = page_buffers(page);
1500         bh = head;
1501         do {
1502                 unsigned int next_off = curr_off + bh->b_size;
1503                 next = bh->b_this_page;
1504
1505                 /*
1506                  * is this block fully invalidated?
1507                  */
1508                 if (offset <= curr_off)
1509                         discard_buffer(bh);
1510                 curr_off = next_off;
1511                 bh = next;
1512         } while (bh != head);
1513
1514         /*
1515          * We release buffers only if the entire page is being invalidated.
1516          * The get_block cached value has been unconditionally invalidated,
1517          * so real IO is not possible anymore.
1518          */
1519         if (offset == 0)
1520                 try_to_release_page(page, 0);
1521 out:
1522         return;
1523 }
1524 EXPORT_SYMBOL(block_invalidatepage);
1525
1526 /*
1527  * We attach and possibly dirty the buffers atomically wrt
1528  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1529  * is already excluded via the page lock.
1530  */
1531 void create_empty_buffers(struct page *page,
1532                         unsigned long blocksize, unsigned long b_state)
1533 {
1534         struct buffer_head *bh, *head, *tail;
1535
1536         head = alloc_page_buffers(page, blocksize, 1);
1537         bh = head;
1538         do {
1539                 bh->b_state |= b_state;
1540                 tail = bh;
1541                 bh = bh->b_this_page;
1542         } while (bh);
1543         tail->b_this_page = head;
1544
1545         spin_lock(&page->mapping->private_lock);
1546         if (PageUptodate(page) || PageDirty(page)) {
1547                 bh = head;
1548                 do {
1549                         if (PageDirty(page))
1550                                 set_buffer_dirty(bh);
1551                         if (PageUptodate(page))
1552                                 set_buffer_uptodate(bh);
1553                         bh = bh->b_this_page;
1554                 } while (bh != head);
1555         }
1556         attach_page_buffers(page, head);
1557         spin_unlock(&page->mapping->private_lock);
1558 }
1559 EXPORT_SYMBOL(create_empty_buffers);
1560
1561 /*
1562  * We are taking a block for data and we don't want any output from any
1563  * buffer-cache aliases starting from return from that function and
1564  * until the moment when something will explicitly mark the buffer
1565  * dirty (hopefully that will not happen until we will free that block ;-)
1566  * We don't even need to mark it not-uptodate - nobody can expect
1567  * anything from a newly allocated buffer anyway. We used to used
1568  * unmap_buffer() for such invalidation, but that was wrong. We definitely
1569  * don't want to mark the alias unmapped, for example - it would confuse
1570  * anyone who might pick it with bread() afterwards...
1571  *
1572  * Also..  Note that bforget() doesn't lock the buffer.  So there can
1573  * be writeout I/O going on against recently-freed buffers.  We don't
1574  * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1575  * only if we really need to.  That happens here.
1576  */
1577 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1578 {
1579         struct buffer_head *old_bh;
1580
1581         might_sleep();
1582
1583         old_bh = __find_get_block_slow(bdev, block);
1584         if (old_bh) {
1585                 clear_buffer_dirty(old_bh);
1586                 wait_on_buffer(old_bh);
1587                 clear_buffer_req(old_bh);
1588                 __brelse(old_bh);
1589         }
1590 }
1591 EXPORT_SYMBOL(unmap_underlying_metadata);
1592
1593 /*
1594  * NOTE! All mapped/uptodate combinations are valid:
1595  *
1596  *      Mapped  Uptodate        Meaning
1597  *
1598  *      No      No              "unknown" - must do get_block()
1599  *      No      Yes             "hole" - zero-filled
1600  *      Yes     No              "allocated" - allocated on disk, not read in
1601  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1602  *
1603  * "Dirty" is valid only with the last case (mapped+uptodate).
1604  */
1605
1606 /*
1607  * While block_write_full_page is writing back the dirty buffers under
1608  * the page lock, whoever dirtied the buffers may decide to clean them
1609  * again at any time.  We handle that by only looking at the buffer
1610  * state inside lock_buffer().
1611  *
1612  * If block_write_full_page() is called for regular writeback
1613  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1614  * locked buffer.   This only can happen if someone has written the buffer
1615  * directly, with submit_bh().  At the address_space level PageWriteback
1616  * prevents this contention from occurring.
1617  *
1618  * If block_write_full_page() is called with wbc->sync_mode ==
1619  * WB_SYNC_ALL, the writes are posted using WRITE_SYNC_PLUG; this
1620  * causes the writes to be flagged as synchronous writes, but the
1621  * block device queue will NOT be unplugged, since usually many pages
1622  * will be pushed to the out before the higher-level caller actually
1623  * waits for the writes to be completed.  The various wait functions,
1624  * such as wait_on_writeback_range() will ultimately call sync_page()
1625  * which will ultimately call blk_run_backing_dev(), which will end up
1626  * unplugging the device queue.
1627  */
1628 static int __block_write_full_page(struct inode *inode, struct page *page,
1629                         get_block_t *get_block, struct writeback_control *wbc,
1630                         bh_end_io_t *handler)
1631 {
1632         int err;
1633         sector_t block;
1634         sector_t last_block;
1635         struct buffer_head *bh, *head;
1636         const unsigned blocksize = 1 << inode->i_blkbits;
1637         int nr_underway = 0;
1638         int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1639                         WRITE_SYNC_PLUG : WRITE);
1640
1641         BUG_ON(!PageLocked(page));
1642
1643         last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1644
1645         if (!page_has_buffers(page)) {
1646                 create_empty_buffers(page, blocksize,
1647                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1648         }
1649
1650         /*
1651          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1652          * here, and the (potentially unmapped) buffers may become dirty at
1653          * any time.  If a buffer becomes dirty here after we've inspected it
1654          * then we just miss that fact, and the page stays dirty.
1655          *
1656          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1657          * handle that here by just cleaning them.
1658          */
1659
1660         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1661         head = page_buffers(page);
1662         bh = head;
1663
1664         /*
1665          * Get all the dirty buffers mapped to disk addresses and
1666          * handle any aliases from the underlying blockdev's mapping.
1667          */
1668         do {
1669                 if (block > last_block) {
1670                         /*
1671                          * mapped buffers outside i_size will occur, because
1672                          * this page can be outside i_size when there is a
1673                          * truncate in progress.
1674                          */
1675                         /*
1676                          * The buffer was zeroed by block_write_full_page()
1677                          */
1678                         clear_buffer_dirty(bh);
1679                         set_buffer_uptodate(bh);
1680                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1681                            buffer_dirty(bh)) {
1682                         WARN_ON(bh->b_size != blocksize);
1683                         err = get_block(inode, block, bh, 1);
1684                         if (err)
1685                                 goto recover;
1686                         clear_buffer_delay(bh);
1687                         if (buffer_new(bh)) {
1688                                 /* blockdev mappings never come here */
1689                                 clear_buffer_new(bh);
1690                                 unmap_underlying_metadata(bh->b_bdev,
1691                                                         bh->b_blocknr);
1692                         }
1693                 }
1694                 bh = bh->b_this_page;
1695                 block++;
1696         } while (bh != head);
1697
1698         do {
1699                 if (!buffer_mapped(bh))
1700                         continue;
1701                 /*
1702                  * If it's a fully non-blocking write attempt and we cannot
1703                  * lock the buffer then redirty the page.  Note that this can
1704                  * potentially cause a busy-wait loop from writeback threads
1705                  * and kswapd activity, but those code paths have their own
1706                  * higher-level throttling.
1707                  */
1708                 if (wbc->sync_mode != WB_SYNC_NONE) {
1709                         lock_buffer(bh);
1710                 } else if (!trylock_buffer(bh)) {
1711                         redirty_page_for_writepage(wbc, page);
1712                         continue;
1713                 }
1714                 if (test_clear_buffer_dirty(bh)) {
1715                         mark_buffer_async_write_endio(bh, handler);
1716                 } else {
1717                         unlock_buffer(bh);
1718                 }
1719         } while ((bh = bh->b_this_page) != head);
1720
1721         /*
1722          * The page and its buffers are protected by PageWriteback(), so we can
1723          * drop the bh refcounts early.
1724          */
1725         BUG_ON(PageWriteback(page));
1726         set_page_writeback(page);
1727
1728         do {
1729                 struct buffer_head *next = bh->b_this_page;
1730                 if (buffer_async_write(bh)) {
1731                         submit_bh(write_op, bh);
1732                         nr_underway++;
1733                 }
1734                 bh = next;
1735         } while (bh != head);
1736         unlock_page(page);
1737
1738         err = 0;
1739 done:
1740         if (nr_underway == 0) {
1741                 /*
1742                  * The page was marked dirty, but the buffers were
1743                  * clean.  Someone wrote them back by hand with
1744                  * ll_rw_block/submit_bh.  A rare case.
1745                  */
1746                 end_page_writeback(page);
1747
1748                 /*
1749                  * The page and buffer_heads can be released at any time from
1750                  * here on.
1751                  */
1752         }
1753         return err;
1754
1755 recover:
1756         /*
1757          * ENOSPC, or some other error.  We may already have added some
1758          * blocks to the file, so we need to write these out to avoid
1759          * exposing stale data.
1760          * The page is currently locked and not marked for writeback
1761          */
1762         bh = head;
1763         /* Recovery: lock and submit the mapped buffers */
1764         do {
1765                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1766                     !buffer_delay(bh)) {
1767                         lock_buffer(bh);
1768                         mark_buffer_async_write_endio(bh, handler);
1769                 } else {
1770                         /*
1771                          * The buffer may have been set dirty during
1772                          * attachment to a dirty page.
1773                          */
1774                         clear_buffer_dirty(bh);
1775                 }
1776         } while ((bh = bh->b_this_page) != head);
1777         SetPageError(page);
1778         BUG_ON(PageWriteback(page));
1779         mapping_set_error(page->mapping, err);
1780         set_page_writeback(page);
1781         do {
1782                 struct buffer_head *next = bh->b_this_page;
1783                 if (buffer_async_write(bh)) {
1784                         clear_buffer_dirty(bh);
1785                         submit_bh(write_op, bh);
1786                         nr_underway++;
1787                 }
1788                 bh = next;
1789         } while (bh != head);
1790         unlock_page(page);
1791         goto done;
1792 }
1793
1794 /*
1795  * If a page has any new buffers, zero them out here, and mark them uptodate
1796  * and dirty so they'll be written out (in order to prevent uninitialised
1797  * block data from leaking). And clear the new bit.
1798  */
1799 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1800 {
1801         unsigned int block_start, block_end;
1802         struct buffer_head *head, *bh;
1803
1804         BUG_ON(!PageLocked(page));
1805         if (!page_has_buffers(page))
1806                 return;
1807
1808         bh = head = page_buffers(page);
1809         block_start = 0;
1810         do {
1811                 block_end = block_start + bh->b_size;
1812
1813                 if (buffer_new(bh)) {
1814                         if (block_end > from && block_start < to) {
1815                                 if (!PageUptodate(page)) {
1816                                         unsigned start, size;
1817
1818                                         start = max(from, block_start);
1819                                         size = min(to, block_end) - start;
1820
1821                                         zero_user(page, start, size);
1822                                         set_buffer_uptodate(bh);
1823                                 }
1824
1825                                 clear_buffer_new(bh);
1826                                 mark_buffer_dirty(bh);
1827                         }
1828                 }
1829
1830                 block_start = block_end;
1831                 bh = bh->b_this_page;
1832         } while (bh != head);
1833 }
1834 EXPORT_SYMBOL(page_zero_new_buffers);
1835
1836 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1837                 get_block_t *get_block)
1838 {
1839         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1840         unsigned to = from + len;
1841         struct inode *inode = page->mapping->host;
1842         unsigned block_start, block_end;
1843         sector_t block;
1844         int err = 0;
1845         unsigned blocksize, bbits;
1846         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1847
1848         BUG_ON(!PageLocked(page));
1849         BUG_ON(from > PAGE_CACHE_SIZE);
1850         BUG_ON(to > PAGE_CACHE_SIZE);
1851         BUG_ON(from > to);
1852
1853         blocksize = 1 << inode->i_blkbits;
1854         if (!page_has_buffers(page))
1855                 create_empty_buffers(page, blocksize, 0);
1856         head = page_buffers(page);
1857
1858         bbits = inode->i_blkbits;
1859         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1860
1861         for(bh = head, block_start = 0; bh != head || !block_start;
1862             block++, block_start=block_end, bh = bh->b_this_page) {
1863                 block_end = block_start + blocksize;
1864                 if (block_end <= from || block_start >= to) {
1865                         if (PageUptodate(page)) {
1866                                 if (!buffer_uptodate(bh))
1867                                         set_buffer_uptodate(bh);
1868                         }
1869                         continue;
1870                 }
1871                 if (buffer_new(bh))
1872                         clear_buffer_new(bh);
1873                 if (!buffer_mapped(bh)) {
1874                         WARN_ON(bh->b_size != blocksize);
1875                         err = get_block(inode, block, bh, 1);
1876                         if (err)
1877                                 break;
1878                         if (buffer_new(bh)) {
1879                                 unmap_underlying_metadata(bh->b_bdev,
1880                                                         bh->b_blocknr);
1881                                 if (PageUptodate(page)) {
1882                                         clear_buffer_new(bh);
1883                                         set_buffer_uptodate(bh);
1884                                         mark_buffer_dirty(bh);
1885                                         continue;
1886                                 }
1887                                 if (block_end > to || block_start < from)
1888                                         zero_user_segments(page,
1889                                                 to, block_end,
1890                                                 block_start, from);
1891                                 continue;
1892                         }
1893                 }
1894                 if (PageUptodate(page)) {
1895                         if (!buffer_uptodate(bh))
1896                                 set_buffer_uptodate(bh);
1897                         continue; 
1898                 }
1899                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1900                     !buffer_unwritten(bh) &&
1901                      (block_start < from || block_end > to)) {
1902                         ll_rw_block(READ, 1, &bh);
1903                         *wait_bh++=bh;
1904                 }
1905         }
1906         /*
1907          * If we issued read requests - let them complete.
1908          */
1909         while(wait_bh > wait) {
1910                 wait_on_buffer(*--wait_bh);
1911                 if (!buffer_uptodate(*wait_bh))
1912                         err = -EIO;
1913         }
1914         if (unlikely(err)) {
1915                 page_zero_new_buffers(page, from, to);
1916                 ClearPageUptodate(page);
1917         }
1918         return err;
1919 }
1920 EXPORT_SYMBOL(__block_write_begin);
1921
1922 static int __block_commit_write(struct inode *inode, struct page *page,
1923                 unsigned from, unsigned to)
1924 {
1925         unsigned block_start, block_end;
1926         int partial = 0;
1927         unsigned blocksize;
1928         struct buffer_head *bh, *head;
1929
1930         blocksize = 1 << inode->i_blkbits;
1931
1932         for(bh = head = page_buffers(page), block_start = 0;
1933             bh != head || !block_start;
1934             block_start=block_end, bh = bh->b_this_page) {
1935                 block_end = block_start + blocksize;
1936                 if (block_end <= from || block_start >= to) {
1937                         if (!buffer_uptodate(bh))
1938                                 partial = 1;
1939                 } else {
1940                         set_buffer_uptodate(bh);
1941                         mark_buffer_dirty(bh);
1942                 }
1943                 clear_buffer_new(bh);
1944         }
1945
1946         /*
1947          * If this is a partial write which happened to make all buffers
1948          * uptodate then we can optimize away a bogus readpage() for
1949          * the next read(). Here we 'discover' whether the page went
1950          * uptodate as a result of this (potentially partial) write.
1951          */
1952         if (!partial)
1953                 SetPageUptodate(page);
1954         return 0;
1955 }
1956
1957 /*
1958  * block_write_begin takes care of the basic task of block allocation and
1959  * bringing partial write blocks uptodate first.
1960  *
1961  * The filesystem needs to handle block truncation upon failure.
1962  */
1963 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
1964                 unsigned flags, struct page **pagep, get_block_t *get_block)
1965 {
1966         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1967         struct page *page;
1968         int status;
1969
1970         page = grab_cache_page_write_begin(mapping, index, flags);
1971         if (!page)
1972                 return -ENOMEM;
1973
1974         status = __block_write_begin(page, pos, len, get_block);
1975         if (unlikely(status)) {
1976                 unlock_page(page);
1977                 page_cache_release(page);
1978                 page = NULL;
1979         }
1980
1981         *pagep = page;
1982         return status;
1983 }
1984 EXPORT_SYMBOL(block_write_begin);
1985
1986 int block_write_end(struct file *file, struct address_space *mapping,
1987                         loff_t pos, unsigned len, unsigned copied,
1988                         struct page *page, void *fsdata)
1989 {
1990         struct inode *inode = mapping->host;
1991         unsigned start;
1992
1993         start = pos & (PAGE_CACHE_SIZE - 1);
1994
1995         if (unlikely(copied < len)) {
1996                 /*
1997                  * The buffers that were written will now be uptodate, so we
1998                  * don't have to worry about a readpage reading them and
1999                  * overwriting a partial write. However if we have encountered
2000                  * a short write and only partially written into a buffer, it
2001                  * will not be marked uptodate, so a readpage might come in and
2002                  * destroy our partial write.
2003                  *
2004                  * Do the simplest thing, and just treat any short write to a
2005                  * non uptodate page as a zero-length write, and force the
2006                  * caller to redo the whole thing.
2007                  */
2008                 if (!PageUptodate(page))
2009                         copied = 0;
2010
2011                 page_zero_new_buffers(page, start+copied, start+len);
2012         }
2013         flush_dcache_page(page);
2014
2015         /* This could be a short (even 0-length) commit */
2016         __block_commit_write(inode, page, start, start+copied);
2017
2018         return copied;
2019 }
2020 EXPORT_SYMBOL(block_write_end);
2021
2022 int generic_write_end(struct file *file, struct address_space *mapping,
2023                         loff_t pos, unsigned len, unsigned copied,
2024                         struct page *page, void *fsdata)
2025 {
2026         struct inode *inode = mapping->host;
2027         int i_size_changed = 0;
2028
2029         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2030
2031         /*
2032          * No need to use i_size_read() here, the i_size
2033          * cannot change under us because we hold i_mutex.
2034          *
2035          * But it's important to update i_size while still holding page lock:
2036          * page writeout could otherwise come in and zero beyond i_size.
2037          */
2038         if (pos+copied > inode->i_size) {
2039                 i_size_write(inode, pos+copied);
2040                 i_size_changed = 1;
2041         }
2042
2043         unlock_page(page);
2044         page_cache_release(page);
2045
2046         /*
2047          * Don't mark the inode dirty under page lock. First, it unnecessarily
2048          * makes the holding time of page lock longer. Second, it forces lock
2049          * ordering of page lock and transaction start for journaling
2050          * filesystems.
2051          */
2052         if (i_size_changed)
2053                 mark_inode_dirty(inode);
2054
2055         return copied;
2056 }
2057 EXPORT_SYMBOL(generic_write_end);
2058
2059 /*
2060  * block_is_partially_uptodate checks whether buffers within a page are
2061  * uptodate or not.
2062  *
2063  * Returns true if all buffers which correspond to a file portion
2064  * we want to read are uptodate.
2065  */
2066 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2067                                         unsigned long from)
2068 {
2069         struct inode *inode = page->mapping->host;
2070         unsigned block_start, block_end, blocksize;
2071         unsigned to;
2072         struct buffer_head *bh, *head;
2073         int ret = 1;
2074
2075         if (!page_has_buffers(page))
2076                 return 0;
2077
2078         blocksize = 1 << inode->i_blkbits;
2079         to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2080         to = from + to;
2081         if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2082                 return 0;
2083
2084         head = page_buffers(page);
2085         bh = head;
2086         block_start = 0;
2087         do {
2088                 block_end = block_start + blocksize;
2089                 if (block_end > from && block_start < to) {
2090                         if (!buffer_uptodate(bh)) {
2091                                 ret = 0;
2092                                 break;
2093                         }
2094                         if (block_end >= to)
2095                                 break;
2096                 }
2097                 block_start = block_end;
2098                 bh = bh->b_this_page;
2099         } while (bh != head);
2100
2101         return ret;
2102 }
2103 EXPORT_SYMBOL(block_is_partially_uptodate);
2104
2105 /*
2106  * Generic "read page" function for block devices that have the normal
2107  * get_block functionality. This is most of the block device filesystems.
2108  * Reads the page asynchronously --- the unlock_buffer() and
2109  * set/clear_buffer_uptodate() functions propagate buffer state into the
2110  * page struct once IO has completed.
2111  */
2112 int block_read_full_page(struct page *page, get_block_t *get_block)
2113 {
2114         struct inode *inode = page->mapping->host;
2115         sector_t iblock, lblock;
2116         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2117         unsigned int blocksize;
2118         int nr, i;
2119         int fully_mapped = 1;
2120
2121         BUG_ON(!PageLocked(page));
2122         blocksize = 1 << inode->i_blkbits;
2123         if (!page_has_buffers(page))
2124                 create_empty_buffers(page, blocksize, 0);
2125         head = page_buffers(page);
2126
2127         iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2128         lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2129         bh = head;
2130         nr = 0;
2131         i = 0;
2132
2133         do {
2134                 if (buffer_uptodate(bh))
2135                         continue;
2136
2137                 if (!buffer_mapped(bh)) {
2138                         int err = 0;
2139
2140                         fully_mapped = 0;
2141                         if (iblock < lblock) {
2142                                 WARN_ON(bh->b_size != blocksize);
2143                                 err = get_block(inode, iblock, bh, 0);
2144                                 if (err)
2145                                         SetPageError(page);
2146                         }
2147                         if (!buffer_mapped(bh)) {
2148                                 zero_user(page, i * blocksize, blocksize);
2149                                 if (!err)
2150                                         set_buffer_uptodate(bh);
2151                                 continue;
2152                         }
2153                         /*
2154                          * get_block() might have updated the buffer
2155                          * synchronously
2156                          */
2157                         if (buffer_uptodate(bh))
2158                                 continue;
2159                 }
2160                 arr[nr++] = bh;
2161         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2162
2163         if (fully_mapped)
2164                 SetPageMappedToDisk(page);
2165
2166         if (!nr) {
2167                 /*
2168                  * All buffers are uptodate - we can set the page uptodate
2169                  * as well. But not if get_block() returned an error.
2170                  */
2171                 if (!PageError(page))
2172                         SetPageUptodate(page);
2173                 unlock_page(page);
2174                 return 0;
2175         }
2176
2177         /* Stage two: lock the buffers */
2178         for (i = 0; i < nr; i++) {
2179                 bh = arr[i];
2180                 lock_buffer(bh);
2181                 mark_buffer_async_read(bh);
2182         }
2183
2184         /*
2185          * Stage 3: start the IO.  Check for uptodateness
2186          * inside the buffer lock in case another process reading
2187          * the underlying blockdev brought it uptodate (the sct fix).
2188          */
2189         for (i = 0; i < nr; i++) {
2190                 bh = arr[i];
2191                 if (buffer_uptodate(bh))
2192                         end_buffer_async_read(bh, 1);
2193                 else
2194                         submit_bh(READ, bh);
2195         }
2196         return 0;
2197 }
2198 EXPORT_SYMBOL(block_read_full_page);
2199
2200 /* utility function for filesystems that need to do work on expanding
2201  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2202  * deal with the hole.  
2203  */
2204 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2205 {
2206         struct address_space *mapping = inode->i_mapping;
2207         struct page *page;
2208         void *fsdata;
2209         int err;
2210
2211         err = inode_newsize_ok(inode, size);
2212         if (err)
2213                 goto out;
2214
2215         err = pagecache_write_begin(NULL, mapping, size, 0,
2216                                 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2217                                 &page, &fsdata);
2218         if (err)
2219                 goto out;
2220
2221         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2222         BUG_ON(err > 0);
2223
2224 out:
2225         return err;
2226 }
2227 EXPORT_SYMBOL(generic_cont_expand_simple);
2228
2229 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2230                             loff_t pos, loff_t *bytes)
2231 {
2232         struct inode *inode = mapping->host;
2233         unsigned blocksize = 1 << inode->i_blkbits;
2234         struct page *page;
2235         void *fsdata;
2236         pgoff_t index, curidx;
2237         loff_t curpos;
2238         unsigned zerofrom, offset, len;
2239         int err = 0;
2240
2241         index = pos >> PAGE_CACHE_SHIFT;
2242         offset = pos & ~PAGE_CACHE_MASK;
2243
2244         while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2245                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2246                 if (zerofrom & (blocksize-1)) {
2247                         *bytes |= (blocksize-1);
2248                         (*bytes)++;
2249                 }
2250                 len = PAGE_CACHE_SIZE - zerofrom;
2251
2252                 err = pagecache_write_begin(file, mapping, curpos, len,
2253                                                 AOP_FLAG_UNINTERRUPTIBLE,
2254                                                 &page, &fsdata);
2255                 if (err)
2256                         goto out;
2257                 zero_user(page, zerofrom, len);
2258                 err = pagecache_write_end(file, mapping, curpos, len, len,
2259                                                 page, fsdata);
2260                 if (err < 0)
2261                         goto out;
2262                 BUG_ON(err != len);
2263                 err = 0;
2264
2265                 balance_dirty_pages_ratelimited(mapping);
2266         }
2267
2268         /* page covers the boundary, find the boundary offset */
2269         if (index == curidx) {
2270                 zerofrom = curpos & ~PAGE_CACHE_MASK;
2271                 /* if we will expand the thing last block will be filled */
2272                 if (offset <= zerofrom) {
2273                         goto out;
2274                 }
2275                 if (zerofrom & (blocksize-1)) {
2276                         *bytes |= (blocksize-1);
2277                         (*bytes)++;
2278                 }
2279                 len = offset - zerofrom;
2280
2281                 err = pagecache_write_begin(file, mapping, curpos, len,
2282                                                 AOP_FLAG_UNINTERRUPTIBLE,
2283                                                 &page, &fsdata);
2284                 if (err)
2285                         goto out;
2286                 zero_user(page, zerofrom, len);
2287                 err = pagecache_write_end(file, mapping, curpos, len, len,
2288                                                 page, fsdata);
2289                 if (err < 0)
2290                         goto out;
2291                 BUG_ON(err != len);
2292                 err = 0;
2293         }
2294 out:
2295         return err;
2296 }
2297
2298 /*
2299  * For moronic filesystems that do not allow holes in file.
2300  * We may have to extend the file.
2301  */
2302 int cont_write_begin(struct file *file, struct address_space *mapping,
2303                         loff_t pos, unsigned len, unsigned flags,
2304                         struct page **pagep, void **fsdata,
2305                         get_block_t *get_block, loff_t *bytes)
2306 {
2307         struct inode *inode = mapping->host;
2308         unsigned blocksize = 1 << inode->i_blkbits;
2309         unsigned zerofrom;
2310         int err;
2311
2312         err = cont_expand_zero(file, mapping, pos, bytes);
2313         if (err)
2314                 return err;
2315
2316         zerofrom = *bytes & ~PAGE_CACHE_MASK;
2317         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2318                 *bytes |= (blocksize-1);
2319                 (*bytes)++;
2320         }
2321
2322         return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2323 }
2324 EXPORT_SYMBOL(cont_write_begin);
2325
2326 int block_commit_write(struct page *page, unsigned from, unsigned to)
2327 {
2328         struct inode *inode = page->mapping->host;
2329         __block_commit_write(inode,page,from,to);
2330         return 0;
2331 }
2332 EXPORT_SYMBOL(block_commit_write);
2333
2334 /*
2335  * block_page_mkwrite() is not allowed to change the file size as it gets
2336  * called from a page fault handler when a page is first dirtied. Hence we must
2337  * be careful to check for EOF conditions here. We set the page up correctly
2338  * for a written page which means we get ENOSPC checking when writing into
2339  * holes and correct delalloc and unwritten extent mapping on filesystems that
2340  * support these features.
2341  *
2342  * We are not allowed to take the i_mutex here so we have to play games to
2343  * protect against truncate races as the page could now be beyond EOF.  Because
2344  * truncate writes the inode size before removing pages, once we have the
2345  * page lock we can determine safely if the page is beyond EOF. If it is not
2346  * beyond EOF, then the page is guaranteed safe against truncation until we
2347  * unlock the page.
2348  */
2349 int
2350 block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2351                    get_block_t get_block)
2352 {
2353         struct page *page = vmf->page;
2354         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2355         unsigned long end;
2356         loff_t size;
2357         int ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
2358
2359         lock_page(page);
2360         size = i_size_read(inode);
2361         if ((page->mapping != inode->i_mapping) ||
2362             (page_offset(page) > size)) {
2363                 /* page got truncated out from underneath us */
2364                 unlock_page(page);
2365                 goto out;
2366         }
2367
2368         /* page is wholly or partially inside EOF */
2369         if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2370                 end = size & ~PAGE_CACHE_MASK;
2371         else
2372                 end = PAGE_CACHE_SIZE;
2373
2374         ret = __block_write_begin(page, 0, end, get_block);
2375         if (!ret)
2376                 ret = block_commit_write(page, 0, end);
2377
2378         if (unlikely(ret)) {
2379                 unlock_page(page);
2380                 if (ret == -ENOMEM)
2381                         ret = VM_FAULT_OOM;
2382                 else /* -ENOSPC, -EIO, etc */
2383                         ret = VM_FAULT_SIGBUS;
2384         } else
2385                 ret = VM_FAULT_LOCKED;
2386
2387 out:
2388         return ret;
2389 }
2390 EXPORT_SYMBOL(block_page_mkwrite);
2391
2392 /*
2393  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2394  * immediately, while under the page lock.  So it needs a special end_io
2395  * handler which does not touch the bh after unlocking it.
2396  */
2397 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2398 {
2399         __end_buffer_read_notouch(bh, uptodate);
2400 }
2401
2402 /*
2403  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2404  * the page (converting it to circular linked list and taking care of page
2405  * dirty races).
2406  */
2407 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2408 {
2409         struct buffer_head *bh;
2410
2411         BUG_ON(!PageLocked(page));
2412
2413         spin_lock(&page->mapping->private_lock);
2414         bh = head;
2415         do {
2416                 if (PageDirty(page))
2417                         set_buffer_dirty(bh);
2418                 if (!bh->b_this_page)
2419                         bh->b_this_page = head;
2420                 bh = bh->b_this_page;
2421         } while (bh != head);
2422         attach_page_buffers(page, head);
2423         spin_unlock(&page->mapping->private_lock);
2424 }
2425
2426 /*
2427  * On entry, the page is fully not uptodate.
2428  * On exit the page is fully uptodate in the areas outside (from,to)
2429  * The filesystem needs to handle block truncation upon failure.
2430  */
2431 int nobh_write_begin(struct address_space *mapping,
2432                         loff_t pos, unsigned len, unsigned flags,
2433                         struct page **pagep, void **fsdata,
2434                         get_block_t *get_block)
2435 {
2436         struct inode *inode = mapping->host;
2437         const unsigned blkbits = inode->i_blkbits;
2438         const unsigned blocksize = 1 << blkbits;
2439         struct buffer_head *head, *bh;
2440         struct page *page;
2441         pgoff_t index;
2442         unsigned from, to;
2443         unsigned block_in_page;
2444         unsigned block_start, block_end;
2445         sector_t block_in_file;
2446         int nr_reads = 0;
2447         int ret = 0;
2448         int is_mapped_to_disk = 1;
2449
2450         index = pos >> PAGE_CACHE_SHIFT;
2451         from = pos & (PAGE_CACHE_SIZE - 1);
2452         to = from + len;
2453
2454         page = grab_cache_page_write_begin(mapping, index, flags);
2455         if (!page)
2456                 return -ENOMEM;
2457         *pagep = page;
2458         *fsdata = NULL;
2459
2460         if (page_has_buffers(page)) {
2461                 ret = __block_write_begin(page, pos, len, get_block);
2462                 if (unlikely(ret))
2463                         goto out_release;
2464                 return ret;
2465         }
2466
2467         if (PageMappedToDisk(page))
2468                 return 0;
2469
2470         /*
2471          * Allocate buffers so that we can keep track of state, and potentially
2472          * attach them to the page if an error occurs. In the common case of
2473          * no error, they will just be freed again without ever being attached
2474          * to the page (which is all OK, because we're under the page lock).
2475          *
2476          * Be careful: the buffer linked list is a NULL terminated one, rather
2477          * than the circular one we're used to.
2478          */
2479         head = alloc_page_buffers(page, blocksize, 0);
2480         if (!head) {
2481                 ret = -ENOMEM;
2482                 goto out_release;
2483         }
2484
2485         block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2486
2487         /*
2488          * We loop across all blocks in the page, whether or not they are
2489          * part of the affected region.  This is so we can discover if the
2490          * page is fully mapped-to-disk.
2491          */
2492         for (block_start = 0, block_in_page = 0, bh = head;
2493                   block_start < PAGE_CACHE_SIZE;
2494                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2495                 int create;
2496
2497                 block_end = block_start + blocksize;
2498                 bh->b_state = 0;
2499                 create = 1;
2500                 if (block_start >= to)
2501                         create = 0;
2502                 ret = get_block(inode, block_in_file + block_in_page,
2503                                         bh, create);
2504                 if (ret)
2505                         goto failed;
2506                 if (!buffer_mapped(bh))
2507                         is_mapped_to_disk = 0;
2508                 if (buffer_new(bh))
2509                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2510                 if (PageUptodate(page)) {
2511                         set_buffer_uptodate(bh);
2512                         continue;
2513                 }
2514                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2515                         zero_user_segments(page, block_start, from,
2516                                                         to, block_end);
2517                         continue;
2518                 }
2519                 if (buffer_uptodate(bh))
2520                         continue;       /* reiserfs does this */
2521                 if (block_start < from || block_end > to) {
2522                         lock_buffer(bh);
2523                         bh->b_end_io = end_buffer_read_nobh;
2524                         submit_bh(READ, bh);
2525                         nr_reads++;
2526                 }
2527         }
2528
2529         if (nr_reads) {
2530                 /*
2531                  * The page is locked, so these buffers are protected from
2532                  * any VM or truncate activity.  Hence we don't need to care
2533                  * for the buffer_head refcounts.
2534                  */
2535                 for (bh = head; bh; bh = bh->b_this_page) {
2536                         wait_on_buffer(bh);
2537                         if (!buffer_uptodate(bh))
2538                                 ret = -EIO;
2539                 }
2540                 if (ret)
2541                         goto failed;
2542         }
2543
2544         if (is_mapped_to_disk)
2545                 SetPageMappedToDisk(page);
2546
2547         *fsdata = head; /* to be released by nobh_write_end */
2548
2549         return 0;
2550
2551 failed:
2552         BUG_ON(!ret);
2553         /*
2554          * Error recovery is a bit difficult. We need to zero out blocks that
2555          * were newly allocated, and dirty them to ensure they get written out.
2556          * Buffers need to be attached to the page at this point, otherwise
2557          * the handling of potential IO errors during writeout would be hard
2558          * (could try doing synchronous writeout, but what if that fails too?)
2559          */
2560         attach_nobh_buffers(page, head);
2561         page_zero_new_buffers(page, from, to);
2562
2563 out_release:
2564         unlock_page(page);
2565         page_cache_release(page);
2566         *pagep = NULL;
2567
2568         return ret;
2569 }
2570 EXPORT_SYMBOL(nobh_write_begin);
2571
2572 int nobh_write_end(struct file *file, struct address_space *mapping,
2573                         loff_t pos, unsigned len, unsigned copied,
2574                         struct page *page, void *fsdata)
2575 {
2576         struct inode *inode = page->mapping->host;
2577         struct buffer_head *head = fsdata;
2578         struct buffer_head *bh;
2579         BUG_ON(fsdata != NULL && page_has_buffers(page));
2580
2581         if (unlikely(copied < len) && head)
2582                 attach_nobh_buffers(page, head);
2583         if (page_has_buffers(page))
2584                 return generic_write_end(file, mapping, pos, len,
2585                                         copied, page, fsdata);
2586
2587         SetPageUptodate(page);
2588         set_page_dirty(page);
2589         if (pos+copied > inode->i_size) {
2590                 i_size_write(inode, pos+copied);
2591                 mark_inode_dirty(inode);
2592         }
2593
2594         unlock_page(page);
2595         page_cache_release(page);
2596
2597         while (head) {
2598                 bh = head;
2599                 head = head->b_this_page;
2600                 free_buffer_head(bh);
2601         }
2602
2603         return copied;
2604 }
2605 EXPORT_SYMBOL(nobh_write_end);
2606
2607 /*
2608  * nobh_writepage() - based on block_full_write_page() except
2609  * that it tries to operate without attaching bufferheads to
2610  * the page.
2611  */
2612 int nobh_writepage(struct page *page, get_block_t *get_block,
2613                         struct writeback_control *wbc)
2614 {
2615         struct inode * const inode = page->mapping->host;
2616         loff_t i_size = i_size_read(inode);
2617         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2618         unsigned offset;
2619         int ret;
2620
2621         /* Is the page fully inside i_size? */
2622         if (page->index < end_index)
2623                 goto out;
2624
2625         /* Is the page fully outside i_size? (truncate in progress) */
2626         offset = i_size & (PAGE_CACHE_SIZE-1);
2627         if (page->index >= end_index+1 || !offset) {
2628                 /*
2629                  * The page may have dirty, unmapped buffers.  For example,
2630                  * they may have been added in ext3_writepage().  Make them
2631                  * freeable here, so the page does not leak.
2632                  */
2633 #if 0
2634                 /* Not really sure about this  - do we need this ? */
2635                 if (page->mapping->a_ops->invalidatepage)
2636                         page->mapping->a_ops->invalidatepage(page, offset);
2637 #endif
2638                 unlock_page(page);
2639                 return 0; /* don't care */
2640         }
2641
2642         /*
2643          * The page straddles i_size.  It must be zeroed out on each and every
2644          * writepage invocation because it may be mmapped.  "A file is mapped
2645          * in multiples of the page size.  For a file that is not a multiple of
2646          * the  page size, the remaining memory is zeroed when mapped, and
2647          * writes to that region are not written out to the file."
2648          */
2649         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2650 out:
2651         ret = mpage_writepage(page, get_block, wbc);
2652         if (ret == -EAGAIN)
2653                 ret = __block_write_full_page(inode, page, get_block, wbc,
2654                                               end_buffer_async_write);
2655         return ret;
2656 }
2657 EXPORT_SYMBOL(nobh_writepage);
2658
2659 int nobh_truncate_page(struct address_space *mapping,
2660                         loff_t from, get_block_t *get_block)
2661 {
2662         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2663         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2664         unsigned blocksize;
2665         sector_t iblock;
2666         unsigned length, pos;
2667         struct inode *inode = mapping->host;
2668         struct page *page;
2669         struct buffer_head map_bh;
2670         int err;
2671
2672         blocksize = 1 << inode->i_blkbits;
2673         length = offset & (blocksize - 1);
2674
2675         /* Block boundary? Nothing to do */
2676         if (!length)
2677                 return 0;
2678
2679         length = blocksize - length;
2680         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2681
2682         page = grab_cache_page(mapping, index);
2683         err = -ENOMEM;
2684         if (!page)
2685                 goto out;
2686
2687         if (page_has_buffers(page)) {
2688 has_buffers:
2689                 unlock_page(page);
2690                 page_cache_release(page);
2691                 return block_truncate_page(mapping, from, get_block);
2692         }
2693
2694         /* Find the buffer that contains "offset" */
2695         pos = blocksize;
2696         while (offset >= pos) {
2697                 iblock++;
2698                 pos += blocksize;
2699         }
2700
2701         map_bh.b_size = blocksize;
2702         map_bh.b_state = 0;
2703         err = get_block(inode, iblock, &map_bh, 0);
2704         if (err)
2705                 goto unlock;
2706         /* unmapped? It's a hole - nothing to do */
2707         if (!buffer_mapped(&map_bh))
2708                 goto unlock;
2709
2710         /* Ok, it's mapped. Make sure it's up-to-date */
2711         if (!PageUptodate(page)) {
2712                 err = mapping->a_ops->readpage(NULL, page);
2713                 if (err) {
2714                         page_cache_release(page);
2715                         goto out;
2716                 }
2717                 lock_page(page);
2718                 if (!PageUptodate(page)) {
2719                         err = -EIO;
2720                         goto unlock;
2721                 }
2722                 if (page_has_buffers(page))
2723                         goto has_buffers;
2724         }
2725         zero_user(page, offset, length);
2726         set_page_dirty(page);
2727         err = 0;
2728
2729 unlock:
2730         unlock_page(page);
2731         page_cache_release(page);
2732 out:
2733         return err;
2734 }
2735 EXPORT_SYMBOL(nobh_truncate_page);
2736
2737 int block_truncate_page(struct address_space *mapping,
2738                         loff_t from, get_block_t *get_block)
2739 {
2740         pgoff_t index = from >> PAGE_CACHE_SHIFT;
2741         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2742         unsigned blocksize;
2743         sector_t iblock;
2744         unsigned length, pos;
2745         struct inode *inode = mapping->host;
2746         struct page *page;
2747         struct buffer_head *bh;
2748         int err;
2749
2750         blocksize = 1 << inode->i_blkbits;
2751         length = offset & (blocksize - 1);
2752
2753         /* Block boundary? Nothing to do */
2754         if (!length)
2755                 return 0;
2756
2757         length = blocksize - length;
2758         iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2759         
2760         page = grab_cache_page(mapping, index);
2761         err = -ENOMEM;
2762         if (!page)
2763                 goto out;
2764
2765         if (!page_has_buffers(page))
2766                 create_empty_buffers(page, blocksize, 0);
2767
2768         /* Find the buffer that contains "offset" */
2769         bh = page_buffers(page);
2770         pos = blocksize;
2771         while (offset >= pos) {
2772                 bh = bh->b_this_page;
2773                 iblock++;
2774                 pos += blocksize;
2775         }
2776
2777         err = 0;
2778         if (!buffer_mapped(bh)) {
2779                 WARN_ON(bh->b_size != blocksize);
2780                 err = get_block(inode, iblock, bh, 0);
2781                 if (err)
2782                         goto unlock;
2783                 /* unmapped? It's a hole - nothing to do */
2784                 if (!buffer_mapped(bh))
2785                         goto unlock;
2786         }
2787
2788         /* Ok, it's mapped. Make sure it's up-to-date */
2789         if (PageUptodate(page))
2790                 set_buffer_uptodate(bh);
2791
2792         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2793                 err = -EIO;
2794                 ll_rw_block(READ, 1, &bh);
2795                 wait_on_buffer(bh);
2796                 /* Uhhuh. Read error. Complain and punt. */
2797                 if (!buffer_uptodate(bh))
2798                         goto unlock;
2799         }
2800
2801         zero_user(page, offset, length);
2802         mark_buffer_dirty(bh);
2803         err = 0;
2804
2805 unlock:
2806         unlock_page(page);
2807         page_cache_release(page);
2808 out:
2809         return err;
2810 }
2811 EXPORT_SYMBOL(block_truncate_page);
2812
2813 /*
2814  * The generic ->writepage function for buffer-backed address_spaces
2815  * this form passes in the end_io handler used to finish the IO.
2816  */
2817 int block_write_full_page_endio(struct page *page, get_block_t *get_block,
2818                         struct writeback_control *wbc, bh_end_io_t *handler)
2819 {
2820         struct inode * const inode = page->mapping->host;
2821         loff_t i_size = i_size_read(inode);
2822         const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2823         unsigned offset;
2824
2825         /* Is the page fully inside i_size? */
2826         if (page->index < end_index)
2827                 return __block_write_full_page(inode, page, get_block, wbc,
2828                                                handler);
2829
2830         /* Is the page fully outside i_size? (truncate in progress) */
2831         offset = i_size & (PAGE_CACHE_SIZE-1);
2832         if (page->index >= end_index+1 || !offset) {
2833                 /*
2834                  * The page may have dirty, unmapped buffers.  For example,
2835                  * they may have been added in ext3_writepage().  Make them
2836                  * freeable here, so the page does not leak.
2837                  */
2838                 do_invalidatepage(page, 0);
2839                 unlock_page(page);
2840                 return 0; /* don't care */
2841         }
2842
2843         /*
2844          * The page straddles i_size.  It must be zeroed out on each and every
2845          * writepage invocation because it may be mmapped.  "A file is mapped
2846          * in multiples of the page size.  For a file that is not a multiple of
2847          * the  page size, the remaining memory is zeroed when mapped, and
2848          * writes to that region are not written out to the file."
2849          */
2850         zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2851         return __block_write_full_page(inode, page, get_block, wbc, handler);
2852 }
2853 EXPORT_SYMBOL(block_write_full_page_endio);
2854
2855 /*
2856  * The generic ->writepage function for buffer-backed address_spaces
2857  */
2858 int block_write_full_page(struct page *page, get_block_t *get_block,
2859                         struct writeback_control *wbc)
2860 {
2861         return block_write_full_page_endio(page, get_block, wbc,
2862                                            end_buffer_async_write);
2863 }
2864 EXPORT_SYMBOL(block_write_full_page);
2865
2866 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2867                             get_block_t *get_block)
2868 {
2869         struct buffer_head tmp;
2870         struct inode *inode = mapping->host;
2871         tmp.b_state = 0;
2872         tmp.b_blocknr = 0;
2873         tmp.b_size = 1 << inode->i_blkbits;
2874         get_block(inode, block, &tmp, 0);
2875         return tmp.b_blocknr;
2876 }
2877 EXPORT_SYMBOL(generic_block_bmap);
2878
2879 static void end_bio_bh_io_sync(struct bio *bio, int err)
2880 {
2881         struct buffer_head *bh = bio->bi_private;
2882
2883         if (err == -EOPNOTSUPP) {
2884                 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2885         }
2886
2887         if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2888                 set_bit(BH_Quiet, &bh->b_state);
2889
2890         bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2891         bio_put(bio);
2892 }
2893
2894 int submit_bh(int rw, struct buffer_head * bh)
2895 {
2896         struct bio *bio;
2897         int ret = 0;
2898
2899         BUG_ON(!buffer_locked(bh));
2900         BUG_ON(!buffer_mapped(bh));
2901         BUG_ON(!bh->b_end_io);
2902         BUG_ON(buffer_delay(bh));
2903         BUG_ON(buffer_unwritten(bh));
2904
2905         /*
2906          * Only clear out a write error when rewriting
2907          */
2908         if (test_set_buffer_req(bh) && (rw & WRITE))
2909                 clear_buffer_write_io_error(bh);
2910
2911         /*
2912          * from here on down, it's all bio -- do the initial mapping,
2913          * submit_bio -> generic_make_request may further map this bio around
2914          */
2915         bio = bio_alloc(GFP_NOIO, 1);
2916
2917         bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2918         bio->bi_bdev = bh->b_bdev;
2919         bio->bi_io_vec[0].bv_page = bh->b_page;
2920         bio->bi_io_vec[0].bv_len = bh->b_size;
2921         bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2922
2923         bio->bi_vcnt = 1;
2924         bio->bi_idx = 0;
2925         bio->bi_size = bh->b_size;
2926
2927         bio->bi_end_io = end_bio_bh_io_sync;
2928         bio->bi_private = bh;
2929
2930         bio_get(bio);
2931         submit_bio(rw, bio);
2932
2933         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2934                 ret = -EOPNOTSUPP;
2935
2936         bio_put(bio);
2937         return ret;
2938 }
2939 EXPORT_SYMBOL(submit_bh);
2940
2941 /**
2942  * ll_rw_block: low-level access to block devices (DEPRECATED)
2943  * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
2944  * @nr: number of &struct buffer_heads in the array
2945  * @bhs: array of pointers to &struct buffer_head
2946  *
2947  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2948  * requests an I/O operation on them, either a %READ or a %WRITE.  The third
2949  * %READA option is described in the documentation for generic_make_request()
2950  * which ll_rw_block() calls.
2951  *
2952  * This function drops any buffer that it cannot get a lock on (with the
2953  * BH_Lock state bit), any buffer that appears to be clean when doing a write
2954  * request, and any buffer that appears to be up-to-date when doing read
2955  * request.  Further it marks as clean buffers that are processed for
2956  * writing (the buffer cache won't assume that they are actually clean
2957  * until the buffer gets unlocked).
2958  *
2959  * ll_rw_block sets b_end_io to simple completion handler that marks
2960  * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2961  * any waiters. 
2962  *
2963  * All of the buffers must be for the same device, and must also be a
2964  * multiple of the current approved size for the device.
2965  */
2966 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2967 {
2968         int i;
2969
2970         for (i = 0; i < nr; i++) {
2971                 struct buffer_head *bh = bhs[i];
2972
2973                 if (!trylock_buffer(bh))
2974                         continue;
2975                 if (rw == WRITE) {
2976                         if (test_clear_buffer_dirty(bh)) {
2977                                 bh->b_end_io = end_buffer_write_sync;
2978                                 get_bh(bh);
2979                                 submit_bh(WRITE, bh);
2980                                 continue;
2981                         }
2982                 } else {
2983                         if (!buffer_uptodate(bh)) {
2984                                 bh->b_end_io = end_buffer_read_sync;
2985                                 get_bh(bh);
2986                                 submit_bh(rw, bh);
2987                                 continue;
2988                         }
2989                 }
2990                 unlock_buffer(bh);
2991         }
2992 }
2993 EXPORT_SYMBOL(ll_rw_block);
2994
2995 void write_dirty_buffer(struct buffer_head *bh, int rw)
2996 {
2997         lock_buffer(bh);
2998         if (!test_clear_buffer_dirty(bh)) {
2999                 unlock_buffer(bh);
3000                 return;
3001         }
3002         bh->b_end_io = end_buffer_write_sync;
3003         get_bh(bh);
3004         submit_bh(rw, bh);
3005 }
3006 EXPORT_SYMBOL(write_dirty_buffer);
3007
3008 /*
3009  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3010  * and then start new I/O and then wait upon it.  The caller must have a ref on
3011  * the buffer_head.
3012  */
3013 int __sync_dirty_buffer(struct buffer_head *bh, int rw)
3014 {
3015         int ret = 0;
3016
3017         WARN_ON(atomic_read(&bh->b_count) < 1);
3018         lock_buffer(bh);
3019         if (test_clear_buffer_dirty(bh)) {
3020                 get_bh(bh);
3021                 bh->b_end_io = end_buffer_write_sync;
3022                 ret = submit_bh(rw, bh);
3023                 wait_on_buffer(bh);
3024                 if (!ret && !buffer_uptodate(bh))
3025                         ret = -EIO;
3026         } else {
3027                 unlock_buffer(bh);
3028         }
3029         return ret;
3030 }
3031 EXPORT_SYMBOL(__sync_dirty_buffer);
3032
3033 int sync_dirty_buffer(struct buffer_head *bh)
3034 {
3035         return __sync_dirty_buffer(bh, WRITE_SYNC);
3036 }
3037 EXPORT_SYMBOL(sync_dirty_buffer);
3038
3039 /*
3040  * try_to_free_buffers() checks if all the buffers on this particular page
3041  * are unused, and releases them if so.
3042  *
3043  * Exclusion against try_to_free_buffers may be obtained by either
3044  * locking the page or by holding its mapping's private_lock.
3045  *
3046  * If the page is dirty but all the buffers are clean then we need to
3047  * be sure to mark the page clean as well.  This is because the page
3048  * may be against a block device, and a later reattachment of buffers
3049  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3050  * filesystem data on the same device.
3051  *
3052  * The same applies to regular filesystem pages: if all the buffers are
3053  * clean then we set the page clean and proceed.  To do that, we require
3054  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3055  * private_lock.
3056  *
3057  * try_to_free_buffers() is non-blocking.
3058  */
3059 static inline int buffer_busy(struct buffer_head *bh)
3060 {
3061         return atomic_read(&bh->b_count) |
3062                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3063 }
3064
3065 static int
3066 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3067 {
3068         struct buffer_head *head = page_buffers(page);
3069         struct buffer_head *bh;
3070
3071         bh = head;
3072         do {
3073                 if (buffer_write_io_error(bh) && page->mapping)
3074                         set_bit(AS_EIO, &page->mapping->flags);
3075                 if (buffer_busy(bh))
3076                         goto failed;
3077                 bh = bh->b_this_page;
3078         } while (bh != head);
3079
3080         do {
3081                 struct buffer_head *next = bh->b_this_page;
3082
3083                 if (bh->b_assoc_map)
3084                         __remove_assoc_queue(bh);
3085                 bh = next;
3086         } while (bh != head);
3087         *buffers_to_free = head;
3088         __clear_page_buffers(page);
3089         return 1;
3090 failed:
3091         return 0;
3092 }
3093
3094 int try_to_free_buffers(struct page *page)
3095 {
3096         struct address_space * const mapping = page->mapping;
3097         struct buffer_head *buffers_to_free = NULL;
3098         int ret = 0;
3099
3100         BUG_ON(!PageLocked(page));
3101         if (PageWriteback(page))
3102                 return 0;
3103
3104         if (mapping == NULL) {          /* can this still happen? */
3105                 ret = drop_buffers(page, &buffers_to_free);
3106                 goto out;
3107         }
3108
3109         spin_lock(&mapping->private_lock);
3110         ret = drop_buffers(page, &buffers_to_free);
3111
3112         /*
3113          * If the filesystem writes its buffers by hand (eg ext3)
3114          * then we can have clean buffers against a dirty page.  We
3115          * clean the page here; otherwise the VM will never notice
3116          * that the filesystem did any IO at all.
3117          *
3118          * Also, during truncate, discard_buffer will have marked all
3119          * the page's buffers clean.  We discover that here and clean
3120          * the page also.
3121          *
3122          * private_lock must be held over this entire operation in order
3123          * to synchronise against __set_page_dirty_buffers and prevent the
3124          * dirty bit from being lost.
3125          */
3126         if (ret)
3127                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3128         spin_unlock(&mapping->private_lock);
3129 out:
3130         if (buffers_to_free) {
3131                 struct buffer_head *bh = buffers_to_free;
3132
3133                 do {
3134                         struct buffer_head *next = bh->b_this_page;
3135                         free_buffer_head(bh);
3136                         bh = next;
3137                 } while (bh != buffers_to_free);
3138         }
3139         return ret;
3140 }
3141 EXPORT_SYMBOL(try_to_free_buffers);
3142
3143 void block_sync_page(struct page *page)
3144 {
3145         struct address_space *mapping;
3146
3147         smp_mb();
3148         mapping = page_mapping(page);
3149         if (mapping)
3150                 blk_run_backing_dev(mapping->backing_dev_info, page);
3151 }
3152 EXPORT_SYMBOL(block_sync_page);
3153
3154 /*
3155  * There are no bdflush tunables left.  But distributions are
3156  * still running obsolete flush daemons, so we terminate them here.
3157  *
3158  * Use of bdflush() is deprecated and will be removed in a future kernel.
3159  * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3160  */
3161 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3162 {
3163         static int msg_count;
3164
3165         if (!capable(CAP_SYS_ADMIN))
3166                 return -EPERM;
3167
3168         if (msg_count < 5) {
3169                 msg_count++;
3170                 printk(KERN_INFO
3171                         "warning: process `%s' used the obsolete bdflush"
3172                         " system call\n", current->comm);
3173                 printk(KERN_INFO "Fix your initscripts?\n");
3174         }
3175
3176         if (func == 1)
3177                 do_exit(0);
3178         return 0;
3179 }
3180
3181 /*
3182  * Buffer-head allocation
3183  */
3184 static struct kmem_cache *bh_cachep;
3185
3186 /*
3187  * Once the number of bh's in the machine exceeds this level, we start
3188  * stripping them in writeback.
3189  */
3190 static int max_buffer_heads;
3191
3192 int buffer_heads_over_limit;
3193
3194 struct bh_accounting {
3195         int nr;                 /* Number of live bh's */
3196         int ratelimit;          /* Limit cacheline bouncing */
3197 };
3198
3199 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3200
3201 static void recalc_bh_state(void)
3202 {
3203         int i;
3204         int tot = 0;
3205
3206         if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3207                 return;
3208         __get_cpu_var(bh_accounting).ratelimit = 0;
3209         for_each_online_cpu(i)
3210                 tot += per_cpu(bh_accounting, i).nr;
3211         buffer_heads_over_limit = (tot > max_buffer_heads);
3212 }
3213         
3214 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3215 {
3216         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3217         if (ret) {
3218                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3219                 get_cpu_var(bh_accounting).nr++;
3220                 recalc_bh_state();
3221                 put_cpu_var(bh_accounting);
3222         }
3223         return ret;
3224 }
3225 EXPORT_SYMBOL(alloc_buffer_head);
3226
3227 void free_buffer_head(struct buffer_head *bh)
3228 {
3229         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3230         kmem_cache_free(bh_cachep, bh);
3231         get_cpu_var(bh_accounting).nr--;
3232         recalc_bh_state();
3233         put_cpu_var(bh_accounting);
3234 }
3235 EXPORT_SYMBOL(free_buffer_head);
3236
3237 static void buffer_exit_cpu(int cpu)
3238 {
3239         int i;
3240         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3241
3242         for (i = 0; i < BH_LRU_SIZE; i++) {
3243                 brelse(b->bhs[i]);
3244                 b->bhs[i] = NULL;
3245         }
3246         get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3247         per_cpu(bh_accounting, cpu).nr = 0;
3248         put_cpu_var(bh_accounting);
3249 }
3250
3251 static int buffer_cpu_notify(struct notifier_block *self,
3252                               unsigned long action, void *hcpu)
3253 {
3254         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3255                 buffer_exit_cpu((unsigned long)hcpu);
3256         return NOTIFY_OK;
3257 }
3258
3259 /**
3260  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3261  * @bh: struct buffer_head
3262  *
3263  * Return true if the buffer is up-to-date and false,
3264  * with the buffer locked, if not.
3265  */
3266 int bh_uptodate_or_lock(struct buffer_head *bh)
3267 {
3268         if (!buffer_uptodate(bh)) {
3269                 lock_buffer(bh);
3270                 if (!buffer_uptodate(bh))
3271                         return 0;
3272                 unlock_buffer(bh);
3273         }
3274         return 1;
3275 }
3276 EXPORT_SYMBOL(bh_uptodate_or_lock);
3277
3278 /**
3279  * bh_submit_read - Submit a locked buffer for reading
3280  * @bh: struct buffer_head
3281  *
3282  * Returns zero on success and -EIO on error.
3283  */
3284 int bh_submit_read(struct buffer_head *bh)
3285 {
3286         BUG_ON(!buffer_locked(bh));
3287
3288         if (buffer_uptodate(bh)) {
3289                 unlock_buffer(bh);
3290                 return 0;
3291         }
3292
3293         get_bh(bh);
3294         bh->b_end_io = end_buffer_read_sync;
3295         submit_bh(READ, bh);
3296         wait_on_buffer(bh);
3297         if (buffer_uptodate(bh))
3298                 return 0;
3299         return -EIO;
3300 }
3301 EXPORT_SYMBOL(bh_submit_read);
3302
3303 void __init buffer_init(void)
3304 {
3305         int nrpages;
3306
3307         bh_cachep = kmem_cache_create("buffer_head",
3308                         sizeof(struct buffer_head), 0,
3309                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3310                                 SLAB_MEM_SPREAD),
3311                                 NULL);
3312
3313         /*
3314          * Limit the bh occupancy to 10% of ZONE_NORMAL
3315          */
3316         nrpages = (nr_free_buffer_pages() * 10) / 100;
3317         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3318         hotcpu_notifier(buffer_cpu_notify, 0);
3319 }