> Let me know if you find any difference.

OK.  I am mainly uncomfortable because we're depending on some other
bookkeeping being done right, I think.  If we keep a state variable
specifically for this purpose we don't have to worry about someone
screwing us up later by making a tweak somewhere else.  But, I'm
still cogitating.

> Also, i prefer the bumping the cwnd up, calling tcp_output and
> then restoring it back method because it avoids adding code in
> tcp_output that would be executed for every outgoing packet if i
> understood it correctly.

Ah, nice point.  I'd love to hear other's opinions.

> This would work well unless the reason for less than four packet
> cwnd is the arrival of a ICMP source quench just before the Fast
> Retransmit. In which case the cwnd is reset to 1 segment.

SQs are fairly rare.  I think the *heuristic* I suggested would
still work OK.

> Yes, i thought a counter that can be displayed through 'netstat
> -s' could be useful.

Certainly.

Thanks,
allman