> Ah, yes, GATEWAY option is enabled.  So, datagram passes thru
> ipflow_fastforward() rathole.  End note through-put is slower.
> The driver does not use Tx interrupts.  It takes "shoot-away" strateray
> to send.  (need to beat a Linksys)
>

I haven't heard of this "shoot-away" strategy.  How much performance is
gained by not taking Tx interrupts?  How do you know when the Tx mbufs can
be freed?  It would be interesting to see some "ttcp" or "netperf"
benchmarks so I could compare against the ep93xx enet driver/hardware. 
Here are some of the ep93xx numbers:

... TTCP from NetBSD-current i386 ex0 to ep93xx epe0 iface
# ttcp -n 8192 -t -s 10.0.0.10
ttcp-t: buflen=8192, nbuf=8192, align=16384/0, port=5001  tcp  -> 10.0.0.10
ttcp-t: socket
ttcp-t: connect
ttcp-t: 67108864 bytes in 19.43 real seconds = 3373.39 KB/sec +++
ttcp-t: 8192 I/O calls, msec/call = 2.43, calls/sec = 421.67
ttcp-t: -1.9user 0.2sys 0:19real 1% 0i+0d 0maxrss 0+6248pf 24544+17csw

... Same run as above, but as recorded on ep93xx
# ttcp -r -s
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001  tcp
ttcp-r: socket
ttcp-r: accept from 10.0.0.1
ttcp-r: 67108864 bytes in 19.44 real seconds = 3371.08 KB/sec +++
ttcp-r: 8204 I/O calls, msec/call = 2.43, calls/sec = 422.00
ttcp-r: 0.3user 8.5sys 0:19real 45% 0i+0d 0maxrss 0+2pf 20+244csw

//Jesse Off