Re: [PATCH] direct IO breaks root filesystem

[Tachino Nobuhiro <tachino@open.nm.fujitsu.co.jp>]

Hello,

At Tue, 11 Dec 2001 09:37:37 -0800 (PST),
Linus Torvalds wrote:
> 
> 
> On Tue, 11 Dec 2001, GOTO Masanori wrote:
> >
> > Accessing with inode size unit (== 4096 byte) is ok, but if I accessed
> > with block size unit, generic_direct_IO() returns error.  The reason
> > is that blocksize is designated as inode->i_blkbits, and its value is
> > not disk minimal block size (512), but inode's unit size (4096).
> 
> Actually, I now found the _real_ bug that explains both the "inode->i_dev"
> _and_ the block size problem.
> 
> We have the wrong inode.
> 
> We use the on-disk inode, which is NOT the same as the "mapping" inode for
> actually doing the IO.
> 
> This simple (and completely untested) patch should fix it.
> 
> This two-liner should also actually make the previous patch completely
> unnecessary, because now "inode->i_dev" should be automatically correct. I
> should have realized that inode->i_dev should always be right, and have
> thought more about the fact that it wasn't.

  Your patch will break Goto's fix because mapping->host does not have
correct i_mode. The patch below is for 2.4.17, I think it works on 2.5.1.


diff -Nur linux-2.4.17-pre7.org/fs/block_dev.c linux-2.4.17-pre7/fs/block_dev.c
--- linux-2.4.17-pre7.org/fs/block_dev.c	Mon Dec 10 17:34:52 2001
+++ linux-2.4.17-pre7/fs/block_dev.c	Mon Dec 10 17:48:21 2001
@@ -329,6 +329,7 @@
 			inode->i_bdev = new_bdev;
 			inode->i_data.a_ops = &def_blk_aops;
 			inode->i_data.gfp_mask = GFP_USER;
+			inode->i_mode = S_IFBLK;
 			spin_lock(&bdev_lock);
 			bdev = bdfind(dev, head);
 			if (!bdev) {