/*
* There is a congestion window for outstanding rpcs maintained per mount
* point. The cwnd size is adjusted in roughly the way that:
* Van Jacobson, Congestion avoidance and Control, In "Proceedings of
* SIGCOMM '88". ACM, August 1988.
* describes for TCP. The cwnd size is chopped in half on a retransmit timeout
* and incremented by 1/cwnd when each rpc reply is received and a full cwnd
* of rpcs is in progress.
* (The sent count and cwnd are scaled for integer arith.)
* Variants of "slow start" were tried and were found to be too much of a
* performance hit (ave. rtt 3 times larger),
* I suspect due to the large rtt that nfs rpcs have.
*/