[vortex-bug] transfer errors with 3C905C-TX-M

Donald Becker becker@scyld.com
Mon, 4 Sep 2000 11:19:30 -0400 (EDT)


On Mon, 4 Sep 2000, Bogdan Costescu wrote:

> > their number is random and differs from transfer to transfer also for
> > the same file. This happens with all drivers available to me: 3Com's
> > 3c90x versions 1.0.0.c (standard for RH6.2) and 1.0.0.i (the latest
> > version from the 3Com's web site), and 3c59x driver V.0.99H. Any ideas?
> 
> If you're using RH 6.2's default kernel (2.2.14-5), the 3c59x driver does
> not support properly the 905C cards. You should upgrade to 2.2.16-3.
> It might be a media selection problem, but the rate of failure (if correct
> computed) is IMHO too low. Anyway, 3Com's drivers should get the right
> media settings.

This couldn't possibly be a media selection problem.
Any data correction due to media-level errors would be caught by the CRC,
unless your physical error rate is unusably high.

Note: You should always track down CRC error sources on your network.  The
Ethernet CRC is very good, providing almost perfect protection against
single and double bit flips.  But with enough media-level errors, it's
possible that an error will finally slip through.  But that's not what is
happening here.

> Subject: Re: [vortex-bug] transfer errors with 3C905C-TX-M

The 3c905C includes hardware IP checksumming.  Thus the software TCP/UDP/IP
checksumming is not used.  This adds the possibility of data corruption
when transferred from the Ethernet adapter to main memory.

You can disable this checksumming by changing line 1826 (approximately)
from
		skb->ip_summed = CHECKSUM_UNNECESSARY;
to
		skb->ip_summed = 0;

My guess is that you will still have data corruption.  If so, you have a
problem with your memory or disk controller.


Donald Becker				becker@scyld.com
Scyld Computing Corporation		http://www.scyld.com
410 Severn Ave. Suite 210		Beowulf-II Cluster Distribution
Annapolis MD 21403