[Intel-gfx] [RFC v2 2/3] drm/i915/pmu: serve global events and support perf stat

[Rogozhkin, Dmitry V]

On Tue, 2017-08-29 at 11:28 +0200, Peter Zijlstra wrote:
> On Wed, Aug 23, 2017 at 11:38:43PM +0000, Rogozhkin, Dmitry V wrote:
> > On Wed, 2017-08-23 at 08:26 -0700, Dmitry Rogozhkin wrote:
> > > +static cpumask_t i915_pmu_cpumask = CPU_MASK_CPU0;
> > 
> > Peter, this hardcoding of cpumask to use CPU0 works, but should I
> > implement something smarter or this will be sufficient? I see that
> > cstate.c you have pointed me to tries to track CPUs going online/offline
> > and migrates PMU context to another CPU if selected one went offline.
> > Should I follow this way?
> 
> Yes.. x86 used to not allow hotplug of CPU0, but they 'fixed' that :/
> 
> And the perf core needs _a_ valid CPU to run things from, which leaves
> you having to track online/offline things.
> 
> Now, I suppose its all fairly similar for a lot of uncore PMUs, so maybe
> you can pull some of this into a library and avoid the endless
> duplication between all (most?) uncore driveres.
> 
> > If I should track CPUs going online/offline, then I have questions:
> > 1. How I should register tracking callbacks? I see that cstate.c
> > registers CPUHP_AP_PERF_X86_CSTATE_STARTING and
> > CPUHP_AP_PERF_X86_CSTATE_ONLINE, uncore.c registers
> > CPUHP_AP_PERF_X86_UNCORE_ONLINE. What I should use? I incline to UNCORE.
> 
> Egads, what a mess :/ Clearly I didn't pay too much attention there.
> 
> So ideally we'd not hate a state per PMU, and
> __cpuhp_setup_state_cpuslocked() has a .multi_instance argument that
> allows reuse of a state.
> 
> So yes, please use the PERF_X86_UNCORE ones if possible.
> 
> > 2. If I will register for, say UNCORE, then how double registrations
> > will be handled if both uncore.c and i915.c will register callbacks? Any
> > conflict here?
> 
> Should work with .multi_instance I _think_, I've not had the pleasure of
> using the new and improved CPU hotplug infrastructure much.
> 
> > 3. What I should pass as 2nd argument? Will "perf/x86/intel/i915:online"
> > be ok?
> 
> Yeah, whatever I think.. something unique. Someone or something will
> eventually yell if its no good I suppose ;-)


I figured out how to track cpus online/offline status in PMU. Here is a
question however. What is the reason for uncore PMUs (cstate.c for
example) to register for cpus other than cpu0? I see it registers for
first thread of each cpu, on my 8 logical-core systems it registers for
cpu0-3 it seems. If they register for few cpus then perf-stat will
aggregate counters which can be disabled with '-A, --no-aggr' option.
Ok... but they could register for just cpu0. Besides, it looks like on
Linux cpu0 can't go offline at all at least of x86 architecture. Peter,
could you, please, clarify the reasoning to register designated readers
of uncore PMU for few CPUs?