Date: Thu, 12 Jul 2007 09:08:28 -0700 (PDT) From: Doug Ambrisko <ambrisko@ambrisko.com> To: Tom Judge <tom@tomjudge.com> Cc: Josh Paetzel <josh@tcbug.org>, freebsd-net@freebsd.org, Julian Elischer <julian@elischer.org>, Paul Schmehl <pauls@utdallas.edu> Subject: Re: Question about bce driver Message-ID: <200707121608.l6CG8Sur002750@ambrisko.com> In-Reply-To: <4695F4FD.2010102@tomjudge.com>
next in thread | previous in thread | raw e-mail | index | archive | help
Tom Judge writes: | Julian Elischer wrote: | > Tom Judge wrote: | >> Josh Paetzel wrote: | >>> On Wednesday 11 July 2007, Tom Judge wrote: | >>>> Hi Paul, | >>>> | >>>> From the testing that I have been doing for the last few months | >>>> the driver in 6.2 is stable if you are not using jumbo frames and | >>>> there is a light-moderate network load. | >>>> | >>>> However if you want to use Jumbo frames the driver is very | >>>> unstable. I posted a patch against 6.2 which should fix some load | >>>> based issues in the driver with standard frame sizes. | >>>> | >>>> Tom | >>> | >>> Paul, I was never able to solve the link up/link down problems with | >>> the driver....I was using the drivers from STABLE for a while, and | >>> without jumbo frames everything worked somewhat ok most of the | >>> time....the ultimate solution was to just get the intel PCI-X card | >>> and stop using the broadcoms. | >>> | >>> | >> | >> We have basically come to the same conclusion today, unfortunately | >> this is 35 machines, but if it makes them stable at least we can use | >> them. | > | > I'm not seeing any problems on our 2950s running 6.1 plus some backpatches. | | I am very surprised at that. The driver in 6.1 was un-usable in our | environment. 6.2 makes it usable with standard frames under moderate | load. However use jumbo frames and it all falls apart, and unfortunately | the network these systems are plugged into is GigE only with a 8192 | Jumbo mtu. There are several back-ports/patches in our 6.1 image that we are running! Also we don't do jumbo packets. It wouldn't surprize me if there were jumbo packets issues. Other drivers have had issues with that at certain sizes. What Julian is really saying, is that for our work-load and version of the bce driver things are pretty darn stable. None of our patches are private and are from various versions that have been in -current etc. (ie. pre-Serdes support). There is one possibility in that our IPMI support might be ahead of what is in -current. I forget. IPMI WRT to Broadcom can be "tricky" and cause issues at the PHY level. | I have been trying to get some help with the problem for over a month | now without luck. Firstly I was asked to enable some debugging in the | driver, this just uncovered what seems to be memory management bug in | the driver. A patch was suggest for this but it did not solve the | problem. Next I noticed the NetBSD guys had we written the offending | parts of code which I then ported to FreeBSD driver. Still no luck, | although the problem was not as bad as before. Then I took a look at | the OpenBSD driver, and it seems that they just completely disabled | jumbo frames as they are just so ropey. | | If a patch was released today that fixed the problem then I may be able | to fully test the driver and not have to replace the NIC's. However I am | running out of time with deployment deadlines and I need this kit in | production by the latest mid next week. Which is reasonable unless you have the time to try to really dig in and try to isolate the problem with the firmware or the driver (ie. hack on the driver). Unfortunately you are blazing into uncharted teritory with the bce driver. David has done some great work getting it to work under FreeBSD in what looks to me as his spare time. Things are improving all of the time. Doug A.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200707121608.l6CG8Sur002750>