Re: [PATCH 3/3] perf, tools: Add support for guest/host-only profiling

["Roedel, Joerg" <Joerg.Roedel@amd.com>]

On Tue, May 10, 2011 at 11:25:14AM -0400, Peter Zijlstra wrote:
> On Tue, 2011-05-10 at 17:08 +0200, Roedel, Joerg wrote:
> > On Tue, May 10, 2011 at 10:50:08AM -0400, Peter Zijlstra wrote:
> > > On Tue, 2011-05-10 at 16:35 +0200, Joerg Roedel wrote:
> > > > @@ -740,6 +740,12 @@ parse_event_modifier(const char **strp, struct perf_event_attr *attr)
> > > >                         if (!exclude)
> > > >                                 exclude = eu = ek = eh = 1;
> > > >                         eh = 0;
> > > > +               } else if (*str == 'G') {
> > > > +                       eg   = 0;
> > > > +                       ehst = 1;
> > > > +               } else if (*str == 'H') {
> > > > +                       eg   = 1;
> > > > +                       ehst = 0; 
> > > 
> > > This doesn't match the existing exclude logic, also eH and eG come to
> > > mind.
> > 
> > OK, eH and eG seems like a better choice. Regarding the logic I
> > explictly decided to do it this way. The reason is that guest/host
> > counting is orthogonal to user/kernel/hv counting. You can decide to
> > only count guest-kernel for example. And if a user just specifies
> > -e cycles:G this would automatically exlucde user and kernel counting.
> > This didn't make sense to me so I decided to keep the logic seperate for
> > guest/host exclusions.
> 
> OK, so the changelog lacked that bit of information ;-)

Okay, so I add this information to the changelog.

> How about you do something like:
> 
> 
> 
> +               } else if (*str == 'G') {
> +			if (!excl_GH)
> +				excl_GH = eH = eG = 1;
> +                       eG   = 0;
> +               } else if (*str == 'H') {
> +			if (!excl_GH)
> +				excl_GH = eH = eG = 1;
> +                       eH   = 0;
> 
> Which mirrors the existing logic but keeps it orthogonal?

Right, it would better fit to the u/k/hv logic. It is not strictly
needed because there are only two excludes here but it makes the code
look more consistent. I'll change it.

> Hmm,. does this nicely integrate with exclude_hv? that seems to want to
> be grouped with G/H.

I intended to re-use the exlude_hv bit originally, but then I looked
into how this bit is used. On PPC it looks like this bit is set when
Linux itself runs as a guest to exclude the hypervisor code being
profiled. The meaning here is different (beacause Linux itself is the
hypervisor) and I wanted to avoid different semantics for this bit
across architectures. So I introduces seperate bits.
The exclude_hv bit can be used when we have some kind of perf-ctr
support for KVM guests.

Regards,

	Joerg