Re: Thread+network crashes in 2.0/2.1

Gabriel Paubert (
Mon, 2 Mar 1998 14:41:40 +0100 (MET)

On Sat, 28 Feb 1998, Larry McVoy wrote:

> : No, the problems we had with the 21140 was that our routers reported
> : excessive number of "runt" packets - there seemed to be something fishy
> : with the collision handling of the cards or similar. The Linux machines
> : themselves never reported anything really wrong, they just showed bad
> : performance.
> :
> : It seemed to be mainly an issue when the network was under heavy load
> : anyway (the cards seemed fine when load was low), and then it just got
> : really bad rather quickly. We haven't seen the same kinds of problems
> : with the eepro100 cards.
> I can add some more data to this. I believe that the tulip (21140)
> cards are very sensitive to cable quality and also that the cards vary in
> quality even with in the same vendor/production run. Why, I have no idea.

Actually the problem might not be with the tulip chip itself. When using
it at 100MB/s, you have to go through another one or two chips (for
example NS DP83840 + DP83223) and the magnetics. The interface from the
dec chip runs at 25MHz (on a 4 or 5 bits wide bus) with very reasonable
timings. Than come the PCS (physical coding sublayer), scrambler and
parallel to serial conversion. The first two can be enabled directly in
the 21140, but the serial<->parallel conversion is always performed in an
external chip. If the problems arise more often with longer cables, their
origin is most likely in the analog part. To continue the example, the
DP83840 performs PCS, scrambling and serial<->parallel conversion and
interfaces with the D3223 on 125MHz differential ECL signal pairs.

The 83223 in turn is a twisted pair transceiver that directly drives the
isolation transformer. If you want to have both 10Mb and 100Mb/s, you have
to connect several pins (differential transmission lines to be exact) in
parallel to the transformer (at least with the DP83840). A bad PCB
routing can easily ruin the design.


To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to