Re: [RFC PATCH 1/1] proc: introduce /proc/<pid>/lbr_stack

[Peter Zijlstra <peterz@infradead.org>]

On Fri, Feb 27, 2015 at 03:57:02PM -0800, Andi Kleen wrote:
> On Fri, Feb 27, 2015 at 11:05:45PM +0100, Peter Zijlstra wrote:
> > On Fri, Feb 27, 2015 at 09:54:34AM -0800, Andi Kleen wrote:
> > > > > perf record doesn't show where you're currently blocked.
> > > > 
> > > > Of course it does; look at perf inject -s.
> > > 
> > > Trace points don't support the LBR stack.
> > 
> > Yes, indeed. But would it not make much more sense to squirrel the LBR
> > state into sched:sched_switch and teach that inject -s thing to dtrt,
> > than to make a proc file that's available on all archs but will only
> > work on 1-2 x86 uarchs and only if you're also running the right magic
> > perf record at the same time?
> 
> Yes. It would be nice to capture the whole PMU state in trace points.
> There are use models for this where it can work better than
> sampling.
> 
> But that would be a lot bigger project than this simple file,
> which is already quite useful with minimal effort.

Its also the most horrible hack of an interface ever, so no go.