Re: [PATCH v2 2/2] perf tools: Make Power7 events available for perf

From: Ingo Molnar <mingo@kernel.org>
To: Michael Ellerman <michael@ellerman.id.au>
Cc: Peter Zijlstra <peterz@infradead.org>,
	Vince Weaver <vincent.weaver@maine.edu>,
	Vince Weaver <vince@deater.net>,
	Runzhen Wang <runzhen@linux.vnet.ibm.com>,
	linuxppc-dev@lists.ozlabs.org, linux-kernel@vger.kernel.org,
	paulus@samba.org, acme@redhat.com,
	Stephane Eranian <eranian@google.com>,
	sukadev@linux.vnet.ibm.com, xiaoguangrong@linux.vnet.ibm.com
Subject: Re: [PATCH v2 2/2] perf tools: Make Power7 events available for perf
Date: Wed, 10 Jul 2013 10:34:35 +0200	[thread overview]
Message-ID: <20130710083435.GA24070@gmail.com> (raw)
In-Reply-To: <20130710030926.GD7491@concordia>

* Michael Ellerman <michael@ellerman.id.au> wrote:

> On Tue, Jul 09, 2013 at 10:14:34AM +0200, Peter Zijlstra wrote:
> > On Mon, Jul 08, 2013 at 10:24:34PM -0400, Vince Weaver wrote:
> > > 
> > > So something like they have on ARM?
> > > 
> > > vince@pandaboard:/sys/bus/event_source/devices$ ls -l
> > > lrwxrwxrwx 1 root root 0 Jul  8 21:57 ARMv7 Cortex-A9 -> ../../../devices/ARMv7 Cortex-A9
> > > lrwxrwxrwx 1 root root 0 Jul  8 21:57 breakpoint -> ../../../devices/breakpoint
> > > lrwxrwxrwx 1 root root 0 Jul  8 21:57 software -> ../../../devices/software
> > > lrwxrwxrwx 1 root root 0 Jul  8 21:57 tracepoint -> ../../../devices/tracepoint
> > 
> > Right so what I remember of the ARM case is that their /proc/cpuinfo isn't
> > sufficient to identify their PMU. And they don't have a cpuid like instruction
> > at all.
> > 
> > > > For the cpu you can obviously just detect what processor you're on with
> > > > cpuid or whatever, but it's a bit of a hack. And that really doesn't
> > > > work for non-cpu PMUs.
> > > 
> > > why is it a hack to use cpuid?
> > 
> > I agree, for x86 cpuid is perfectly fine, as would /proc/cpuinfo be, I suspect
> > that just the model number is sufficient in most cases, even for uncore stuff.
>  
> What about things on PCI? Other strange buses?
> 
> As long as everything's in /sys then it should be _possible_ for 
> userspace to work out what's what, but it's going to end up with a bunch 
> of detection logic and heuristics in the library.
> 
> At which point you've just rewritten libpfm4.

Exactly - PMUs enumerated in /sys should be self-identifying, it's a 
hardware topology after all ...

Anytime userspace is forced to look into /proc, or into weird places in 
/sys it's a FAIL really.

perf ABIs want to be self-identifying and self-sufficient, anytime 
userspace is forced to look elsewhere it adds another source of fragility.

And duplication with something that is 'already in /proc' is not a problem 
_at all_, these are computers that provide us different views into the 
same physical reality with dozens of different abstractions, so 
duplication of information is natural and _good_.

Thanks,

	Ingo