Re: Issues with eepro100 and full duplex

See notes below.

On Thu, 2003-08-14 at 15:23, Donald Becker wrote:

> That's actually a modified driver from my perspective.  I'm guessing
> what you mean is "the driver that is in the kernel.org 2.4.21-pre source
> tree, not one munged by a distribution vendor".

You are correct.  

> > Using mii-tool, I configure the device for autonegotiate.  Mii-tool
> > reports autonegotiate at 100 Full duplex.  The NetGear switch also
> > reports 100 Full duplex.  So far so good.
> 
> Yes, Good.
> 
> > Then I use mii-tool to configure the card to disable autonegotiate and
> > run at 100 full duplex.  Mii-tool reports that the card is running at
> > 100 full duplex, but the switch now reports that the link is at 100 HALF
> > duplex.
> 
> That's exactly what is expected.  You have just created a duplex
> mismatch, which as a flawed configuration.

Does this mean that if I hard-coded both ends to run at 100 full duplex,
that I would end up with 100 half duplex being the end result?  

>   Running some more tests reveals similar results, where
> > regardless of setting the card to half or full duplex, the switch always
> > reports the link at half duplex.
> 
> That's exactly the expected behavior:
>    Forcing the speed must turn off autonegotiation
> Forcing the duplex results in a non-standard-conforming link.
> Full duplex on twisted pair is only standard-conforming when it has been
> autonegotiated. 

Is the problem that the duplex is only being forced at one end of the
link?  Or is it that full duplex has been forced at all?  If so, why
would there be a setting to set the link to full duplex, if there is no
way to actually get this result?

> > I ran the same test against a Cisco managed switch, and am seeing the
> > same results.  In either case, it always runs half-duplex, unless I set
> > the card to autonegotiate.
> ...
> > Is there something in the Ethernet spec that would make this expected
> > behavior?   
> 
> Yes.  If autonegotiation does not take place, and it won't with a forced
> link, the link speed is sensed based on the link beat.
> Ethernet _is_ CSMA/CD (informally "half duplex"), so you end up with a
> half duplex connection.

>From what it sounds like, everything I am seeing is expected behavior,
which is fine by me.  I am just trying to provide a mechanism that
allows my customers to configure the product to work with their
sometimes buggy switching equipment.  If this is how it works, then so
be it.

Thanks for all your help,

-- 
Devin Heitmueller
Senior Software Engineer
Netilla Networks Inc.