Try bumping VR_RX_LIST_CNT in if_vrreg.h in our driver from 64
to 128. The FBsd driver has some flow control stuff in it too but
that's only for 6105+ chips, and yours is a 6102 or 6103. There
is a RX pause feature commented out in the FBsd driver for 6102+
chips (but its commented out so...).
DragonFly's NFS implementation runs RPCs fully asynchronously so it
can seriously load down the network. Both FBsd and DragonFly have
variable-sized TCP buffers so scp-style transfers should work about
the same, but you can check by looking at the backlog in the sockbuf
on the sender for the connection ('netstat -p tcp' on the sending side)
while it is going full-bore.
-Matt
Matthew Dillon
<[email protected]>
