libext2fs: fix Direct I/O fallback code so it implements RMW correctly

There is a bug in Unix I/O manager where if an aligned I/O is
required, it does not correctly do the read-modify-write cycle
correctly.  Specifically, it was not doing an lseek between the read
and the write calls, so the update was going to block N+1 instead of
block N.  Oops.

Fortunately in practice we almost never use this fallback path, so
file systems weren't getting horribly corrupted, because (a) we almost
never use Direct I/O in e2fsprogs, at least not by default, and (b)
when we do the buffers end up being aligned anyway, so it's OK.

We only noticed this because the new Undo I/O manager in e2fsprogs
1.43 was doing unaligned I/O and FreeBSD requires that I/O requests be
aligned even if you are not doing Direct I/O, and the e2undo
regression tests were all failing as a result.

Signed-off-by: Theodore Ts'o <tytso@mit.edu>

1 files changed, 4 insertions, 0 deletions

diff --git a/lib/ext2fs/unix_io.c b/lib/ext2fs/unix_io.c
index 23f22e30f..4403aa0e6 100644
--- a/lib/ext2fs/unix_io.c
+++ b/lib/ext2fs/unix_io.c
@@ -282,6 +282,10 @@ static errcode_t raw_write_blk(io_channel channel,
 		if (size > channel->block_size)
 			actual = channel->block_size;
 		memcpy(data->bounce, buf, actual);
+		if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
+			retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
+			goto error_out;
+		}
 		actual = write(data->dev, data->bounce, channel->block_size);
 		if (actual != channel->block_size)
 			goto short_write;