Re: [PATCH] direct IO breaks root filesystem

[GOTO Masanori]

Hi,

Kernel both 2.4.16-pre7 and 2.5.1-pre8 have serious problem 
when we use direct IO for the block device.
These kernel breaks my root filesystem which has /dev 
(block device entry) when an application like dd write data
into /dev/sdXX (= not root filesystem's device).

My environment is:
	/dev/hda	root filesystem with /dev/ dir.
	/dev/sda	target direct io device
I lost my root filesystem with:
	fd = open("/dev/sda", O_DIRECT);
	write(fd, data, large_size);

The reason is that when kernel accesses /dev/sda with O_DIRECT,
blkdev_direct_IO() is called. But, it calls generic_direct_IO(),
and generic_direct_IO() calls brw_kiovec(..., inode->i_dev, ...).
If inode->i_dev is root filesystem, not target direct io device,
then inode->i_dev is used for root filesystem overwriting (= then
I lost my root file system...).

Here is patch, it was based some parts in 2.4.10 (and some 2.4.1x-aa).
In my test with this patch, it works well and
kernel does not break my root partition.

Linus and Marcelo, please apply this patch.

-- gotom

--- fs/block_dev.c.vanilla	Sat Dec  8 22:46:15 2001
+++ fs/block_dev.c	Mon Dec 10 10:45:20 2001
@@ -115,7 +115,25 @@
 
 static int blkdev_direct_IO(int rw, struct inode * inode, struct kiobuf * iobuf, unsigned long blocknr, int blocksize)
 {
-	return generic_direct_IO(rw, inode, iobuf, blocknr, blocksize, blkdev_get_block);
+	int i, nr_blocks, retval;
+	unsigned long * blocks = iobuf->blocks;
+
+	nr_blocks = iobuf->length / blocksize;
+	/* build the blocklist */
+	for (i = 0; i < nr_blocks; i++, blocknr++) {
+		struct buffer_head bh;
+
+		retval = blkdev_get_block(inode, blocknr, &bh, 0);
+		if (retval)
+			goto out;
+
+			blocks[i] = bh.b_blocknr;
+	}
+
+	retval = brw_kiovec(rw, 1, &iobuf, inode->i_rdev, iobuf->blocks, blocksize);
+
+ out:
+	return retval;
 }
 
 static int blkdev_writepage(struct page * page)