Re: Too many collision errors with Full Duplex auto negotiation

On Fri, 22 Aug 2003, Cavelero, Mark wrote:

> system.  The user interface and data acquisition/collection box is a
> stock Red Hat 7.2 system running 2.4.7-10smp kernel.  This kernel still
> uses eepro100 v1.09j-t:rev1.36 for both of its Intel ethernet interfaces
...
> To facilitate unrestricted bandwidth between the acquisition box (linux)
> and the hardware control box (win98), I use a custom cross-over cable
> connecting the two boxes.

I always suspect locally-made cables.  Most people do not get the
pairing correct.

> When both boxes are configured to auto
> negotiate, the results are random.  Using mii-diag on the linux box
> always gives the negotiated speed as 100baseTX-FD, however it has become
> apparent that communication speeds are often much slower, shown most
> easily by file transfer.  Here is the output from ifconfig after typical
> data collection:
..
>           collisions:61207 txqueuelen:100 

You clearly are not set to full duplex.
The eepro100 chip is very reliable at detecting full duplex.  In the
default mode the hardware just uses FDX when it is negotiated, the
driver doesn't even have to set the configuration.

> In all instances, using mii-tool -R corrects the problem for a
> non-specific amount of time.

That's a hint that something is going wrong with the physical link that
causes re-negotiation and failure.

> I obviously have tried the latest driver set available from scyld
> (v1.27), however it segmentation faults when loading:

The current version is v1.28, which fixes the bug you are seeing.
I must update the netdrivers.tgz file where you found the old
version... done.

-- 
Donald Becker				[email protected]
Scyld Computing Corporation		http://www.scyld.com
914 Bay Ridge Road, Suite 220		Scyld Beowulf cluster system
Annapolis MD 21403			410-990-9993