Subject: Re: x86, current-20031102, vr0.
To: None <firstname.lastname@example.org>
From: Thor Lancelot Simon <email@example.com>
Date: 11/07/2003 11:29:31
On Fri, Nov 07, 2003 at 08:25:17AM -0700, Herb Peyerl wrote:
> So I have two -current i386 machines in a rack next to one another on an
> Extreme Summit switch. One machine has an fxp0 and the other a vr0. Both
> are set to 100/full, the switch is on auto and acknowledges the current
> setting of 100/full on each of the two ports.
Don't hardwire either side. I think you'll be pleased with the results.
Most modern PHYs _really_ screw up if they never get a chance to do the
nway autonegotiation, which is exactly what happens with most of our
drivers: we reset the PHY, then bring it up with nway turned off. In
contrast, both Windows and OS X interpret a "hardwired" setting as
"do nway, but refuse to negotiate anything but..." The result is almost
always a duplex mismatch on the host end that is hard to diagnose
because ifconfig says the card is set "right". Remember, there's a bug
in the spec -- with a link partner who won't do nway at all, you *cannot*
be guaranteed to correctly negotiate the duplex!
The old advice to "solve" duplex mismatch problems by hardwiring one or
both ends may have been good advice at the time, but it's _really_ bad
advice now, IME. Autonegotiation works -- use it.