Re: [PATCH 1/6] KVM: In kernel pit model

["Yang, Sheng" <sheng.yang@intel.com>]

On Thursday 06 March 2008 17:41:03 Yang, Sheng wrote:
> On Thursday 06 March 2008 16:43:18 Yang, Sheng wrote:
> > On Thursday 06 March 2008 16:06:51 Avi Kivity wrote:
> > > Yang, Sheng wrote:
> > > > Here is the updated patch. I kept 0xff because I think it's OK for
> > > > understand easily. :)
> > >
> > > Any news on the regression with older Linux guests?  That's the only
> > > thing keeping my from applying the patchset.
> >
> > Not much. PIT interrupts injection is all right, 1000 per second. Just
> > found two clock source in guest got problem: PM timer and TSC. Seems both
> > due to compensate for lost ticks. Get rid of something like "jiffies_64
> > += lost -1" get the time ok.
> >
> > And I think it's a exist bug. As you see, in most condition, userspace
> > pit + in kernel irqchip resulted in time flow slowly, due to the lost of
> > interrupts. But RHEL4 runs even faster than host...
> >
> > I will do more investigate.
>
> Get some clues.
>
> It seems like for the kernel which is active to inject lost interrupt, when
> some PIT interrupts were pending, the TSC/other clocksource got the wrong
> impression that some PIT interrupt lost, then using itself's counter to
> adjust the jiffies. So the problem occurs.
>
> Xen adjust TSC to fit this mode. It pull TSC backward to get the correct
> value when injecting one PIT interrupt. But this would causing trouble on
> some Windows. Then, it got the "time mode" concept...

Found more complex for KVM. Xen pulled pm timer down to kernel part, and used 
the guest TSC as source. So only adjust TSC is OK for it. But we are still 
using pm timer in QEmu, which using host time as source. So even we pull back 
TSC, the problem still exists, for 2.6.9 prefer to pm timer by default.

-- 
Thanks
Yang, Sheng

-------------------------------------------------------------------------
This SF.net email is sponsored by: Microsoft
Defy all challenges. Microsoft(R) Visual Studio 2008.
http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/