Re: [PATCH 1/4] perf_counter: Default to higher paranoia level

[Ingo Molnar <mingo@elte.hu>]

* Peter Zijlstra <peterz@infradead.org> wrote:

> On Fri, 2009-08-21 at 16:21 +0200, Ingo Molnar wrote:
> > * Peter Zijlstra <peterz@infradead.org> wrote:
> 
> > > diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h
> > > index 9ba1822..2b0528f 100644
> > > --- a/include/linux/perf_counter.h
> > > +++ b/include/linux/perf_counter.h
> > > @@ -439,6 +439,7 @@ enum perf_callchain_context {
> > >  struct perf_callchain_entry {
> > >  	__u64				nr;
> > >  	__u64				ip[PERF_MAX_STACK_DEPTH];
> > > +	int				restricted;
> > >  };
> > 
> > i'd love to have something more specific here - i.e. a context type 
> > ID that identifies these basic types:
> > 
> >  - process
> >  - softirq
> >  - hardirq
> >  - NMI
> > 
> > and then let it be up to upper layers to decide what they do with a 
> > restricted entry, and how to further process this information.
> > 
> > And it's not just security: for example it would be interesting to 
> > sample pure, non-irq overhead - as IRQ overhead is often unrelated 
> > to the process being measured.
> 
> Yes it is, this is purely about not showing some data. If you 
> don't want to sample IRQ stuff that's something else, we'd have to 
> grow that capability in hardware (like the OS/USR bits) or put 
> perf enable/disable hooks into the irq entry/exit hooks (which 
> doesn't sound all too hot an idea to me).
> 
> Simply not showing the call-trace is something all-together 
> different from not profiling it.

Well, it's not about not profiling it - it's about being able to 
_separate out_ the samples from the various contexts.

Right now we already have context separators for call-chains:

        PERF_CONTEXT_HV                 = (__u64)-32,
        PERF_CONTEXT_KERNEL             = (__u64)-128,
        PERF_CONTEXT_USER               = (__u64)-512,

        PERF_CONTEXT_GUEST              = (__u64)-2048,
        PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
        PERF_CONTEXT_GUEST_USER         = (__u64)-2560,

All i'm suggesting is to also have these context separators:

	PERF_CONTEXT_KERNEL_HARDIRQ
	PERF_CONTEXT_KERNEL_SOFTIRQ
	PERF_CONTEXT_KERNEL              /* syscall level */

So that if user-space wants to visualize just a portion of it, it 
can do it.

	Ingo