I'm asking Ken why we don't just close and invalidate conn->fd when killing a connection. It seems deliberate given how the code was introduced, but I can't think of a reason. Depending on the answer, I'll apply your patch or adjust kill_conn().