A channel could be opened with the 'max packet size' set to 0.  The behaviour is not defined in the ssh rfc:s, and openssh does not handle that case in some special way besides looping.

The introduced solution just "keep quiet" but removes the eternal loop that caused Erlang to core dump.

The ssh acceptor process listens for connections and
spawns a process to handle each new connection that it accepts.
The ownership of the accepted socket will be transfered to the new process
that will handle the incomming messages. Before the ownership is
transfered the socket should be in {active, false} or the acceptor
process may receive data aimed for the connection. As the accept socket
inherits the listen options we set {active, false} there.

A queue is the behaviour that we want, so this makes the code
easier to understand and more effective.

The error report was assumed to only happen if our code was wrongly
implemented "internal error". However it would also occur when
bad input was recived from the peer, and could hence cause extensive
logging on DoS attacks.