Re: Hard lockup with 2.6.24.7-rt26 on x86

[Clark Williams <williams@redhat.com>]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On Tue, 03 Feb 2009 16:44:10 +0100
Carsten Emde <Carsten.Emde@osadl.org> wrote:

> Clark,
> 
> >> On two different desktop PC (both Core 2 Duo) and on a notebook
> >> (Core Duo) the kernel (2.6.24.7-rt26, CONFIG_X86_32) locks up hard
> >> under a certain rt load. The failure can be reproduced reliably
> >> with the following command:
> >> sudo ./cyclictest -p99 -t10 -n -i250
> > My guess would be that since you're running 10 threads on two 
> > processors at the highest available priority, you're starving all the
> > hard IRQ threads (as well as soft irq and other kernel threads).
> > This is one of those power-tool moments where you can lock up the 
> > system with the wrong workload/priority combination.
> Hmm, I would agree immediately, if the tasks were using all of the CPU
> power. But this is not the case. In the present case, running
> 
> # cyclictest -p99 -t10 -n -i250
> 
> results in about 10% CPU load. So there is plenty of time in between to
> let the system respond. I can reproduce the problem here (2.6.24-rt,
> 2.6.26-rt, various configurations). The system does not crash every time
> when cyclictest is started, only once in about 5 trials or so. If it
> crashes, then it does so immediately after being started. Whenever
> cyclictest survives for several seconds, then it never crashes at a
> later time. I tend to believe that we have a bug here, not a regular
> behavior.
> 

That's what I get for firing off a response with out running it first :)

I don't have a 2.6.26 test box handy, but I'll set that up this
afternoon. I agree with your results that this is a bug, since I
haven't been able to kill my 2.6.24.7-based RT test system and if it
was just starvation I should be able to do that easily.

Clark
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2.0.10 (GNU/Linux)

iEYEARECAAYFAkmIcDMACgkQHyuj/+TTEp21DgCfUihH/6BUUS5o3SwlrnK57dJC
/HQAoMEZp8bYuOEBiS6wVtbGgA0vi0MS
=X1jB
-----END PGP SIGNATURE-----