Re: Questing regarding KVM Guest PMU

[Gleb Natapov <gleb@redhat.com>]

On Thu, Apr 05, 2012 at 04:48:57PM +0300, Gleb Natapov wrote:
> On Thu, Apr 05, 2012 at 04:28:48PM +0300, Avi Kivity wrote:
> > On 04/05/2012 04:26 PM, Gleb Natapov wrote:
> > > On Thu, Apr 05, 2012 at 03:48:51PM +0300, Avi Kivity wrote:
> > > > On 04/05/2012 03:37 PM, Gleb Natapov wrote:
> > > > > On Thu, Apr 05, 2012 at 03:27:18PM +0300, Avi Kivity wrote:
> > > > > > On 04/04/2012 01:29 PM, Gleb Natapov wrote:
> > > > > > > > >
> > > > > > > > ok. seems to be. will move over to perf as its working fine inside guest.
> > > > > > > > 
> > > > > > > Good riddance IMO. I managed to run it on a guest (but not on my
> > > > > > > host!). The thing is buggy. It does not use global ctrl MSR to enable
> > > > > > > counters and kvm has all of them disabled by default. I didn't find what
> > > > > > > value this MSR should have after reset, so this may be either kvm bug or
> > > > > > > real BIOSes enable all counters in global ctrl MSR for PMUv1
> > > > > > > compatibility. Doing "wrmsr 0x38f 0x70000000f" solves this problem. The
> > > > > > > second problem is that oprofile reprogram PMU counters without
> > > > > > > disabling them first and this is explicitly prohibited by Intel SDM.
> > > > > > > The patch below solve that, but oprofile is the one who should be fixed.
> > > > > > 
> > > > > > Both should be fixed, there may be other profilers affected.
> > > > > > 
> > > > > Global ctrl msr issue I need to investigate further, but second one is
> > > > > bug that should be fixed in oprofile. Intel spec clearly says:
> > > > >
> > > > >  EN (Enable Counters) Flag (bit 22) — When set, performance counting
> > > > >  is enabled in the corresponding performance-monitoring counter; when
> > > > >  clear, the corresponding counter is disabled. The event logic unit
> > > > >  for a UMASK must be disabled by setting IA32_PERFEVTSELx[bit 22] = 0,
> > > > >  before writing to IA32_PMCx.
> > > > >
> > > > > I suspect that on real HW they got wrong result too. It simply subtly
> > > > > wrong, so it is not as noticeable as with kvm.
> > > > >
> > > > 
> > > > First, there is the standard Linus rant to never break userspace (and
> > > We broke nothing. Oprofile is broken and never could have worked
> > > according to Intel spec. (It didn't manage to get any result from it on
> > > real CPU, but may be this is unrelated).
> > 
> > But it did work sometime in the past, I've used it.
> > 
> May be it used NMI based profiling. We should ask oprofile developers.
> As I said I am almost sure my inability to run it on a host is probably
> PEBKAC, although I ran the same script exactly on the host and the
> guest (the script is from the first email of this thread)
> 
After upgrading the kernel to latest git from whatever it was there the
same script works and counts CPU_CLK_UNHALT events.

--
			Gleb.