Issue can be fixed by specifying

 option strict-readdir on

to the replicate volume.

Behavior is expected. Option "strict-readdir" fixes it. Perhaps strict-readdir should be the default?

(In reply to comment #3)
> Behavior is expected. Option "strict-readdir" fixes it. Perhaps strict-readdir
> should be the default?

Isn't keeping this on a performance bottleneck?

Setup is GlusterFS 3.0.4, replicate with two subvolumes, no performance translators.

The bug was seen when the following sequence of operations were done:

# cd /mnt/glusterfs/test1

Kill glusterfsd on server1

# mkdir test2
# cd test2
# echo "hi" > file

Start glusterfsd on server1

# ls
.: File descriptor in bad state

Also, self-heal was not triggered on the test2 directory until a "cd .." was done on the client.

TRACE logs for the client are attached.

(In reply to comment #4)
> (In reply to comment #3)
> > Behavior is expected. Option "strict-readdir" fixes it. Perhaps strict-readdir
> > should be the default?
> 
> Isn't keeping this on a performance bottleneck?

The whole point of using the replicate translator is redundancy....if the clients need to use readdir, it should be set by default.