Quoting "Stephen M. Rumble" <stephen.rumble@utoronto.ca>:

> The Set Engineering "GIO Fast Ethernet" board (based on the TI ThunderLAN
> chip)
> also attaches, however it appears to have some issues with transmits and
> isn't
> ready for public consumption. I hope to fix this eventually but may not
> have
> sufficient time before the semester begins.

Since I've yet to solve the problem and am not likely to be able to work on this
for several months, I'll document what I've found in case anybody else is
interested.

The only thing that appears wrong is that some tcp checksums fail. These errors
kill the transmission completely. That is, tcp never recovers.

George Harvey indicated that ttcp appears to work fine, and I've confirmed this,
though all I have been able to learn is that doing write(2)'s of certain sizes
has an effect. For instance, 8192 seems to not display the problem. ttcp's
alignment and other parameters don't appear to adversely change the behaviour.

It appears that nothing is wrong with the driver's transmission queue: I can run
ttcp in the background, reproduce the problem with ftp, and not cause ttcp to
hiccup. Nothing but ftp's tcp session appears to wedge. The tlan never issues an
adapter check or gives any indication of running improperly. Perhaps this is
something very GIO bus-specific as tl(4) boards appear to work fine in O2's. Yet
my Phobos G130 (also GIO, but tlp(4)) hasn't any issues at all..

For the record, SOSEND_NO_LOAN doesn't appear to help either.

Steve