Re: [PATCH 0/9] perf: Adding better precise_ip field handling

[Peter Zijlstra <peterz@infradead.org>]

On Fri, May 10, 2013 at 12:18:23PM +0200, Ingo Molnar wrote:
> 
> * Peter Zijlstra <peterz@infradead.org> wrote:
> 
> > > so you mean just detect that by opening events with increasing precise 
> > > and see how far we could get.. could be I guess, though the 'precise' 
> > > sysfs attribute seems more fit to me
> > 
> > The other way around, start at ppp end at !p, then use the one that 
> > worked.
> 
> Really, instead of this silly 'probing until it works' notion, how about 
> the radical idea that we pass to the kernel our request, and the kernel 
> fulfills our wish to the best of its ability?
> 
> This could be done as a new PERF_COUNT_HW_CPU_CYCLES_PRECISE event, to 
> which tooling could migrate, without changing existing semantics.
> 
> The problem with the complex probing is that it's totally unnecessary 
> complexity that results from lack of passing the right information to the 
> kernel. Forcing that will only hinder user-space adoption of our precise 
> profiling facilities.

The part I have trouble with is that its a vague request and you'll get a vague
answer.

Its very similar to: 'do something' and getting a string of random bytes back.
Yes it is 'something', request fulfilled.

By doing the probing, userspace knows exactly what it asked and what it'll get.