On 07/29/2011 08:25 PM, Andreas Dilger wrote:

On 2011-07-29, at 11:44 AM, Bernd Schubert wrote:

quoted

thanks a lot for sending those patches! From a previous review I still remember ext2fs_mmp_read(), see below.
I will look through the other parts later on.

quoted

+errcode_t ext2fs_mmp_read(ext2_filsys fs, blk64_t mmp_blk, void *buf)
+{
+	/* ext2fs_open reserves fd0,1,2 to avoid stdio collision */
+	if (fs->mmp_fd<= 0) {
+		fs->mmp_fd = open(fs->device_name, O_RDWR | O_DIRECT);
+		if (fs->mmp_fd<   0) {
+			retval = EXT2_ET_MMP_OPEN_DIRECT;
+			goto out;
+		}
+	}
+
+	if (ext2fs_llseek(fs->mmp_fd, mmp_blk * fs->blocksize, SEEK_SET) !=
+	    mmp_blk * fs->blocksize) {
+		retval = EXT2_ET_LLSEEK_FAILED;
+		goto out;
+	}
+
+	if (read(fs->mmp_fd, fs->mmp_cmp, fs->blocksize) != fs->blocksize) {
+		retval = EXT2_ET_SHORT_READ;
+		goto out;
+	}

While ext2fs_llseek() and read() works fine, I still wonder why the code does not use io_channel_read_blk64() here, similarly as io_channel_write_blk64() write in ext2fs_mmp_write().

One of the reasons for doing a direct open+seek+read here is because previous
versions of e2fsprogs did not allow O_DIRECT reads via io_channel_read_blk64(),
which is critical for the correct MMP functioning.  It cannot allow read from
a cached block, either in libext2fs or in the VM, though a cached block on the
disk is OK as long as it is visible to all hosts that are accessing it.

Yeah, I understand that it had to be O_DIRECT, but I didn't know about 
(previous) limitation of io_channel_read_blk64(). Maybe you could add a 
comment about it to the code about that, if you want to keep it? 
Something like
"/* We don't use io_channel_read_blk64() here to make sure blocks are 
really not cached */"?

Thanks,
Bernd