Re: Virtualization of the CPU Performance Monitoring Unit

[Dietmar Hahn <dietmar.hahn@ts.fujitsu.com>]

Am Montag 16 April 2012, 13:39:36 schrieb Konrad Rzeszutek Wilk:
> On Mon, Apr 16, 2012 at 04:16:07PM +0800, Lin Ming wrote:
> > On Wed, Apr 11, 2012 at 10:30 PM, Lin Ming <mlin@ss.pku.edu.cn> wrote:
> > 
> > [....]
> > 
> > >> That isn't actually true. If you run it, you will see it working
> > >> in the guest - it just that it does not use the performence counters
> > >> but instead uses the timer to sample data.
> > >
> > > Right, I mean "hardware event" does not work.
> > >
> > > Hardware event, for example, perf top -e cycles, does not work.
> > 
> > Just found that vpmu is disabled by default.
> > You need to pass xen boot parameter "vpmu" to make hardware event work.
> 
> Oh, I wonder why it was disabled by default? Wei, would you know
> by any chance?

This had to do with a problem in the intel nehalem processors which could cause
endless interrupt loops in the hypervisor if a hvm guest uses the performance
counters so Keir proposed to add the vpmu boot flag.

http://lists.xen.org/archives/html/xen-devel/2009-10/msg01460.html
and
http://lists.xen.org/archives/html/xen-devel/2009-11/msg00088.html

Dietmar

> 
> > 
> > > Software event, for example, perf top -e cpu-clock, works.
> > 
> > So both hardware and software event work in DomU.
> > Great!
> 
> Excellent!
> > 
> > >
> > >>
> > >> > Run "perf top", but no data was collected.
> > >>
> > >> Hm, I am able to collect data using Fedora Core 16 PV guest.
> > >> For dom0 or domU? For dom0 there is a bug somewhere where
> > >
> > > For domU HVM guest.
> > > I have problem to run domU PV guest. Still looking at it.
> > >
> > >> the machine crashes after 30 seconds or so - hadn't actually
> > >> gotten to the bottom of it. There was an email thread:
> > >> https://lkml.org/lkml/2012/2/12/74 about this.
> > >>
> > >> Patches are most welcome!
> > 
> > Here are the patches.
> > https://lkml.org/lkml/2012/4/15/12
> 
> Let me play with them a bit. At first glance they look ok - but I recall
> Peter Z saying something about not implementing the IRQ WORKER, but I can't
> recall the reasons.
> > 
> > Regards,
> > Lin Ming
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xen.org
> http://lists.xen.org/xen-devel
> 
-- 
Company details: http://ts.fujitsu.com/imprint.html