[PlanetCCRMA] Re: updated: cheesetracker, added: brutefir, more kernels on the edge

Fernando Pablo Lopez-Lezcano nando@ccrma.Stanford.EDU
Sat Apr 24 11:13:01 2004

> > > It causes crashes on certain smp systems, and even hyperthread
> > > processors:
> > > 
> > > http://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=121413
> > > http://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=109962
> > > 
> > > Whoever hits such a bug should simply use the 2.4.26 ccrma kernels. :)
> > 
> > Yuck. Do you have lowlat disabled in your kernels (if you have kernels
> > derived from the fedora ones)?
> Some of the bug reports were reporting success on my Red Hat derived
> kernels (which include the PlanetCCRMA patches) even before I turned
> LOLAT off, and I myself never had problems with LOLAT on HT smp
> machines. Perhaps some of the PlanetCCRMA patches cure part of the
> problems?
> But there were some ATrpms users still having problems on smp asking
> to turn off LOLAT for smp (only), and I did so. Just a few days later
> LOLAT was turned off in the Red Hat kernel altogether, and I simply
> copied this practice. The ll patches there are adjusted by RH, so
> there may be some distribution specific bug, that the CCRMA 2.4.26
> kernels will never see.
> > I guess I'll leave it enabled for now...  :-(
> I agree. We should start thinking about and investing in 2.6.x
> instead. :)

Yes, I have already started. There is currently a showstopper lowlat
patch missing, the one for mga/radeon/r128 latency hits is "illegal", it
is doing reschedules with a lock taken (and that causes a oops with the
current 2.6.x, an "on purpose" oops that detects those problems). Until
a kernel guru makes a proper patch, drm with those cards will take
12+Msec hits, not good at all. 

-- Fernando