Dual opteron PIT timing problem?

Dual opteron PIT timing problem?

[John Clemens]

I have a dual Opteron 250: 2 socket, single core.. HP xw9300
workstation. It is having the problem described here:

http://bugzilla.xensource.com/bugzilla/show_bug.cgi?id=195

Even non-SMP, while I don't get the "Time went backwards!" messages, I
still get funky timing.. cursors flash way too fast, time runs 2-6x
faster then it should, keyboard repeat rate is off, etc. 

This happens with 3.0.2 and 3.0.3-testing.

My speculation is that this motherboard does -not- seem to have an HPET
timer, and the PIT timer Xen falls back to does not seem to work
correctly...?  

This is a modern NVIDIA chipset server motherboard that may (or may not)
need the acpi_skip_timer_override parameter.  

When I boot a real (native) kernel, it uses the PM Timer...but even when
I force it to use the PIT ('nopmtimer' kernel option) it still doesn't
exhibit the weird timing behavior shown when running a xen kernel.

Are the any other things I could try? I've tried all sorts of
combinations of [no]apic, [no]smp, [no]tsc, [no]lapic, [no]apm, etc.. Is
anyone running an SMP xen system without an HPET timer?  

thanks,
john.c

-- 
John Clemens           Revive Systems, www.revivesystems.com

Re: Dual opteron PIT timing problem?

[Keir Fraser]

On 6/10/06 10:00 pm, "John Clemens" <jclemens@revivesystems.com> wrote:

> Are the any other things I could try? I've tried all sorts of
> combinations of [no]apic, [no]smp, [no]tsc, [no]lapic, [no]apm, etc.. Is
> anyone running an SMP xen system without an HPET timer?

Most people, I should think. Many BIOSes do not advertise the presence of an
HPET, even if the chipset has one.

I'm not sure what's going on with your system. We had time problems in the
past but they are all believed fixed now. If things are behaving as weirdly
as you say, perhaps we can add some tracing for you to work out what's going
wrong and get this fixed for 3.0.4.

 -- Keir

Re: Dual opteron PIT timing problem?

[Daniel P. Berrange]

On Fri, Oct 06, 2006 at 11:47:44PM +0100, Keir Fraser wrote:
> On 6/10/06 10:00 pm, "John Clemens" <jclemens@revivesystems.com> wrote:
> 
> > Are the any other things I could try? I've tried all sorts of
> > combinations of [no]apic, [no]smp, [no]tsc, [no]lapic, [no]apm, etc.. Is
> > anyone running an SMP xen system without an HPET timer?
> 
> Most people, I should think. Many BIOSes do not advertise the presence of an
> HPET, even if the chipset has one.

Actually there's some pretty common machines which don't include a HPET. IIRC
HP's Opteron based systems such as the DL-{3,5}85 don't typically have a HPET. The
RHEL kernels would typically fallback to using the TSC or the ACPI PMTimer for
these systems (PMTimer is reliable but high overhead, while TSC is low overhead
but pretty tricky to get the code correct to maintain SMP synchronisation).

Regards,
Dan.
-- 
|=- Red Hat, Engineering, Emerging Technologies, Boston.  +1 978 392 2496 -=|
|=-           Perl modules: http://search.cpan.org/~danberr/              -=|
|=-               Projects: http://freshmeat.net/~danielpb/               -=|
|=-  GnuPG: 7D3B9505   F3C9 553F A1DA 4AC2 5648 23C1 B3DF F742 7D3B 9505  -=| 

Re: Dual opteron PIT timing problem?

[John Clemens]

On Fri, 2006-10-06 at 23:47 +0100, Keir Fraser wrote:

> Most people, I should think. Many BIOSes do not advertise the presence of an
> HPET, even if the chipset has one.
> 
> I'm not sure what's going on with your system. We had time problems in the
> past but they are all believed fixed now. If things are behaving as weirdly
> as you say, perhaps we can add some tracing for you to work out what's going
> wrong and get this fixed for 3.0.4.
> 
>  -- Keir

They are, and I'm apparently not alone as the last two entries in bug
number 195 are seeing the same problems, and the last entry is even the
exact same machine type (HP xw9300), posted in September. 

I would gladly try some instrumentation if you have any I can turn on. I
will also try the xen-unstable tree instead of xen-3.0.3-testing.

I have a gut feeling that there's something about the way the timer is
wired on this motherboard, as there seem to be workarounds in Linux
kernels for nvidia chipset systems and timers.  But I have no actual
facts to back that up.

john.c

-- 
John Clemens           Revive Systems, www.revivesystems.com