Re: [PATCH 2/2] perf_counter: optimize context switch between identical inherited contexts

[Ingo Molnar <mingo@elte.hu>]

* Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:

> On Fri, 2009-05-22 at 19:56 +1000, Paul Mackerras wrote:
> > Peter Zijlstra writes:
> > 
> > > On Fri, 2009-05-22 at 14:27 +1000, Paul Mackerras wrote:
> > > > Since we don't have individual fds for the counters in a cloned
> > > > context, the only thing that can make two clones of a given parent
> > > > different after they have been cloned is enabling or disabling all
> > > > counters with prctl.  To account for this, we keep a count of the
> > > > number of enabled counters in each context.  Two contexts must have
> > > > the same number of enabled counters to be considered equivalent.
> > > 
> > > Curious point that.. so prctl() can disable counters it doesn't own.
> > > 
> > > Shouldn't we instead fix that?
> > 
> > Well, prctl enables/disables the counters that are counting on the
> > current process, regardless of who or what created them.  I always
> > thought that was a little strange; maybe it is useful to be able to
> > disable all the counters that any other process might have put on to
> > you, but I can't think of a scenario where you'd really need to do
> > that, particularly since the disable is a one-shot operation, and
> > doesn't prevent new (enabled) counters being attached to you.
> > 
> > On the other hand, what does "all the counters I own" mean?  
> > Does it mean all the ones that I have fds open for?  Or does it 
> > mean all the ones that I created?  Either way we don't have a 
> > good way to enumerate them.
> 
> I'm for all counters you created (ie have a fd for). Being able to 
> disable counters others created on you just sounds wrong.
> 
> If we can settle on a semantic, I'm sure we can implement it :-)
> 
> Ingo, Corey, any opinions?

It indeed doesnt sound correct that we can disable counters others 
created on us - especially if they are in a different (higher 
privileged) security context than us.

OTOH, enabling/disabling counters in specific functions of a library 
might be a valid use-case. So perhaps make this an attribute: 
.transparent or so, with perf stat defaulting on it to be 
transparent (i.e. not child context disable-able).

	Ingo