Re: [PATCH] direct IO breaks root filesystem

[GOTO Masanori]

At Tue, 11 Dec 2001 04:13:06 -0500 (EST),
Alexander Viro <[e-mail]> wrote:
> > +	/*
> > +	 * If inode is BLK and it is not already mounted,
> > +	 * blocksize change hardsect_size.
> > +	 */
> > +	if (S_ISBLK(inode->i_mode) && !is_mounted(inode->i_rdev)) {
> 
> That's obviously wrong.  What's to stop it from getting mounted
> right after that check?

Ah! You're right. I was confused, I'm sorry.
New patch is here (and works well)... Ugh.

--- linux-2.4.17-pre8.vanilla/mm/filemap.c	Tue Dec 11 15:54:57 2001
+++ linux-2.4.17-pre8/mm/filemap.c	Tue Dec 11 19:36:09 2001
@@ -1508,8 +1508,21 @@
 		new_iobuf = 1;
 	}
 
-	blocksize = 1 << inode->i_blkbits;
-	blocksize_bits = inode->i_blkbits;
+	/*
+	 * If inode is block device, blocksize is set as hardsect_size.
+	 */
+	if (S_ISBLK(inode->i_mode)) {
+		int size;
+
+		size = get_hardsect_size(inode->i_rdev);
+		blocksize = size;
+		for (blocksize_bits = 0; !(size & 1); )
+			size >>= 1, blocksize_bits++;
+	} else {
+		blocksize = 1 << inode->i_blkbits;
+		blocksize_bits = inode->i_blkbits;
+	}
+
 	blocksize_mask = blocksize - 1;
 	chunk_size = KIO_MAX_ATOMIC_IO << 10;

-- gotom