Parsing dumps shows that 2.2 moves snd_nxt
beyond right edge of window... :-( The result is that acks from it
are never accepted by another side and it is the finish, certainly.
I still do not know why it occurs. It smells like a race condition.
It is natural to suspect that immediate reason is full-duplex io in rsync,
2.2 is pretty dirty with mutual locking of multiple processes
accessing the socket...
Stephen, please, try 2.3 so aggressively as you are able. If it will also
lockup, it is not a race and it is easier to investigate there.