Re: [PATCH 1/3] perf, core: Introduce attrs to count in either host or guest mode

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Tue, 2011-05-10 at 17:38 +0200, Roedel, Joerg wrote:
> On Tue, May 10, 2011 at 11:18:29AM -0400, Peter Zijlstra wrote:
> > On Tue, 2011-05-10 at 16:59 +0200, Roedel, Joerg wrote:
> > > On Tue, May 10, 2011 at 10:45:46AM -0400, Peter Zijlstra wrote:
> > > > On Tue, 2011-05-10 at 16:35 +0200, Joerg Roedel wrote:
> > > > > +       /* Can't exclude counting in guest and in host mode */
> > > > > +       if (attr.exclude_host && attr.exclude_guest)
> > > > > +               return -EINVAL;
> > > > 
> > > > Why not?
> > > 
> > > By definition the counter won't count at all. The hardware just ignores
> > > the bits if they are both set. My rationale here was that it does not
> > > makes sense to setup a counter and exclude guest and host mode.
> > 
> > I would expect it to 'work' but simply return 0. If that isn't what the
> > AMD hardware does you need to fix that in the AMD driver.
> 
> By 'work' you mean that userspace can set it up but it doesn't count at
> all in this situation? 

Right.

> This would certainly be consistent behavior but I
> can't imagine any use-case for it so that this code assumes that such a
> situation is most likely a bug.
> I can certainly change that if wanted, but I think its better to inform
> userspace if we get weird values?

The eternal how much rope to give and what knots to teach argument I
guess. As it is, I think we allow people to exclude both user- and
kernel-space, giving a similar situation, so allowing to exclude both
host and guest is consistent.