Broadcom NetXtremeII Packet Loss/PE1950

[Matt Saladna]

Hi,
	Has anyone noticed sporadic packet loss with the BCM5708 NICs on Linux?
 I have been attempting to track down the root cause, but everything is
turning up empty.

The symptoms are once every few hundred HTTP requests or so, there will
be a timeout; same goes for DNS resolution.  It looks like any TCP or
UDP request has a 1 in 30 or so chance of abruptly timing out on this
card.  I have been testing by running a lookup query to 4.2.2.1 and
averaging out trip times with the following script:

( set -o pipefail ; let COUNT=0 ; RET=0 ; until [ $RET -ne 0 -o $COUNT
-gt 1000 ] ; do dig +short +trace @4.2.2.1 google.com | grep 'from
server' | awk '{print $7}' ; RET=$? ; ((COUNT++)) ; done ) | awk '{ SUM
+= $1; COUNT += 1; } END { print SUM/COUNT, COUNT }' ; clock

Three of the servers running CentOS 4 with the BCM5708 chipset die
within the first 30 queries:

dig: couldn't get address for 'H.ROOT-SERVERS.NET': not found
722.948 77

Two of the three servers are running the official 2.6.22.1 kernel, while
another is running on 2.6.24-rc6.  There is a fourth server, a PE SC1435
with the Broadcom BCM5721 chipset, which performs the lookups
flawlessly.  All four servers are on the same switch in the facility
effectively ruling out network issues on the uplink.  I removed the TOE
from one of the troubled PE1950s last weekend, but that hasn't affected
packet loss.  Packet loss occurs independent of any offload settings
made via ethtool.  I have turned all of the parameters off and on with
no success.

Finally, just for kicks, I decided to run the DNS lookup test on a
development machine with a mirror image of the filesystem.  I'm running
2.6.22.2 with Via's VT6102 (Rhine-II) embedded NIC.  This works without
a problem either.

Thanks!
   Matt Saladna
   Apis Networks