Re: AMD64 X2 lost ticks on PM timer

[Abdulla Bubshait <darkray@ic3man.com>]

On Tuesday 28 February 2006 01:47, Jason Baron wrote:
> On Mon, 27 Feb 2006, Bill Rugolsky Jr. wrote:
> > On Tue, Feb 28, 2006 at 12:22:40AM +0300, bubshait wrote:
> > > 	Losing some ticks... checking if CPU frequency changed.
> > > 	warning: many lost ticks.
> > > 	Your time source seems to be instable or some driver is hogging
> > > interupts rip __do_softirq+0x47/0xd1
> > >
> > > adding report_lost_ticks only prints repeating messages like
> > >
> > > 	Lost 3 timer tick(s)! rip __do_softirq+0x47/0xd1
> >
> > I'm seeing tons of these on a Tyan 2895 (Nvidia CKO4) running FC4 with
> > kernel-2.6.15-1.1830 (2.6.15.2) SMP:
> >
> > time.c: Lost 1 timer tick(s)! rip default_idle+0x37/0x7a)
> > time.c: Lost 2 timer tick(s)! rip __do_softirq+0x55/0xd4)
> >
> > [I've seen the same thing with earlier FC 2.6.14 kernels.]
> >
> > On our systems the __do_softirq messages are strongly correlated with
> > sata_nv interrupts, especially during our nightly tripwire-like fs
> > checksum job.  Unfortunately, the log messages are not very informative.
> > I'm not sure what ever happened to the following patch,
> >
> > http://www.kernel.org/pub/linux/kernel/people/akpm/patches/2.5/2.5.64/2.5
> >.64-mm3/broken-out/report-lost-ticks.patch
> >
> > but it was dropped.
> >
> > Unfortunately, I need to spend tomorrow patching kernels in search of a
> > fix or workaround, as I have to start using these boxes in production,
> > and they need to keep time.
>
> passing 'nohpet' and/or 'nopmtimer' will force the use of a different
> timer...but this is certainly a workaround, if it helps...

Unfortunately, I can't seem to find a way to force it to use hpet. Passing 
'notsc' and 'nopmtimer' I end up using PIT/TSC based timekeeping. TSC is 
already known to have problems with dual core. But I will sit with it for a 
while to see if it fairs better than the pm timer.

Bill, What timer do you use, and do these lost ticks persist after sata_nv 
interrupts stop?