System hang with heavy network traffic using rtl8139c

Hi -

We have a multi-threaded load test program that will cause any x86 system configuration upon which we've tried that that contains at least one rtl8139c to hard lock the system. The exact same OS image and test program has been shown to run on several exact same configurations by simply replacing the rtl8139c device(s) with intel 82559ER's or Digital 21143TD's. Please read on for more info.

I am involved in reliability testing of our linux-based network device with which we plan to use realtek 8139c parts as its three main network interfaces. The system is architecturally very much like an x86 PC and runs a standard linux 2.4.16 kernel with some well-used community patches and/or driver modules.

We created a multi-threaded test program that generates bidirectional traffic over all network interfaces simultaneously as fast as possible. Using this test, the system will lock up hard, taking from 90 seconds to 1.5 weeks to do so. Most failures occur, however, within the 2-5 hour range. The lock-up is complete, in that not even sysrq works. More typical network traffic patterns do not cause lockups. In general, the test configurations have between 4 and 8 network devices installed in the system, with most tests running with 8 devices. No test has used more than 4 devices of any single chipset.

We have ran our test on various hardware configurations to narrow the problem boundaries, and the results are always the same: if the configuration includes a RealTek 8139C part it will lock up. If it contains *no* 8139C it will not.

We can duplicate this behaviour on our hardware using either Geode GX1 processors or Transmeta Crusoe TM5800 processors. We can duplicate it on desktop Pentium MMX-based systems that contain none of our hardware. Realtek configurations lock up regardless of which driver we are using (rtl8139 v1.18 / pci-scan v1.08 or 8139too v0.9.22). With any configuration, replacing the rtl parts with intel (82559ER) or digital (21143TD) and the system will not fail. The reverse has also been true. Mixed-part environments fail even with a single rtl part installed, but will not if there are no rtl parts.

I'm hoping to get some feedback from some of you out there who have worked with this part more than I. I'll be happy to post specifics as requested. I also appreciate thoughts on how to debug this problem given that it fully locks up the system.

I have posted this message to LKML and to the realtek ML, so reply accordingly.

All the best,
R. Steve McKown
Titanium Mirror, Inc.

try the scyld/Donald Becker driver...

http://www.scyld.com/network/rtl8139.html

The realtek 8139 chip has, a probably well deserved reputation of being unsuitable for high performance network applications so it may the case that they are simply unsuitable for your application.

joelja

-------------------------------------------------------------------------- 
Joel Jaeggli	      Academic User Services   [email protected]    
--    PGP Key Fingerprint: 1DE9 8FCA 51FB 4195 B42A 9C32 A30D 121E      --
  In Dr. Johnson's famous dictionary patriotism is defined as the last
  resort of the scoundrel.  With all due respect to an enlightened but
  inferior lexicographer I beg to submit that it is the first.
	   	            -- Ambrose Bierce, "The Devil's Dictionary"