Re: [PATCH RFC 0/7] perf pmu-events: Support event aliasing for system PMUs

From: Mark Rutland
Date: Tue Feb 18 2020 - 13:13:43 EST


On Tue, Feb 18, 2020 at 05:58:46PM +0000, John Garry wrote:
> On 18/02/2020 17:08, Mark Rutland wrote:
> > > > I also don't understand how a SoC ID makes things
> > > > any easier in this regard.
> > > It's doesn't necessarily make things easier in this regard. But using a SoC
> > > ID is an alternative to checking the SMMU_ID or the kernel driver having to
> > > know that it was a MMU-600 at all.
> > Using SOC_ID means that going forward, userspace needs to learn about
> > the integration details of each SoC in order to identify a component. As
> > you said:
> >
> > | As constantly checking what the SoC ID means throughout system components
> > | does not scale.
> >
> > ... and I think that equally applies to userspace in this case. Who knows how
> > many SoCs are going to have MMU-600?
> >
> > I also know that SOC_ID is going to be optional, and I think it's near-certain
> > that someone will end up producing two SoCs exposing the same ID.
>
> Wouldn't different SoCs having same SMC SOC_ID and revision be a (fixable)
> mistake in the SMC FW?
>
> And if it's not implemented, then no PMU events aliasing in perf tool for
> those uncore PMUs - nothing gets broken though and no regression. But I do
> understand your concern here.
>
> > For system PMUs, I'd rather the system PMU driver exposed some sort of
> > implementation ID. e.g. the SMMU_ID for SMMU. We can give that a generic name,
> > and mandate that where a driver exposes it, the format/meaning is defined in
> > the documentation for the driver.
>
> Then doesn't that per-PMU ID qualify as brittle and non-standard also?

Not in my mind; any instances of the same IP can have the same ID,
regardless of which SoC they're in. Once userspace learns about
device-foo-4000, it knows about it on all SoCs. That also means you can
support heterogeneous instances in the same SoC.

If a device varies so much on a SoC-by-SoC basis and or the driver has
no idea what to expose, it could be legitimate for the PMU driver to
expose the SoC ID as its PMU-specific ID, but I don't think we should
make that the common/only case.

> At least the SMC SoC ID is according to some standard.
>
> And typically most PMU HW would have no ID reg, so where to even get this
> identification info? Joakim Zhang seems to have this problem for the imx8
> DDRC PMU driver.

For imx8, the DT compat string or additional properties on the DDRC node
could be used to imply the id.

> > That can be namespace by driver, so e.g. keys would be smmu_sysfs_name/<id> and
> > ddrc_sysfs_name/<id>.
> >
> > > > > So even if it is solvable here, the kernel driver(s) will need to be
> > > > > reworked. And that is just solving one case in many.
> > > > PMU drivers will need to expose more information to userspace so that they
> > > > can be identified more precisely, yes. I wouldn't say they would need to be
> > > > "reworked".
> > > OK, so some combination of changes would still be required for the SMMU
> > > PMCG, IORT, and SMMUv3 drivers.
> > To expose the SMMU ID, surely that's just the driver?
>
> This case is complicated, like others I anticipate.
>
> So the SMMU PMCG HW has no ID register itself, and this idea relies on using
> the associated SMMUv3 IIDR in lieu. For that, we need to involve the IORT,
> SMMUv3, and SMMU PMCG drivers to create this linkage, and even then I still
> have my doubts on whether this is even proper.

Ok, I hadn't appreciated that the PMCG did not have an ID register
itself.

I think that the relationship between the SMMU and PMCG is a stronger
argument against using the SOC_ID. If the PMCGs in a system are
heterogeneous, then you must know the type of the specific instance.

> Please see https://lore.kernel.org/linux-iommu/1569854031-237636-1-git-send-email-john.garry@xxxxxxxxxx/
> for reference.
>
> Or are there
> > implementations where the ID register is bogus and have to be overridden?
> >
>
> I will also note that perf tool PMU events framework relies today on
> generating a table of events aliases per CPU and matching based on that. If
> you want to totally disassociate a CPU or any SoC ID mapping, then this will
> require big perf tool rework.

I think that might be necessary, as otherwise we're going to back
ourselves into a corner by building what's simple now.

Thanks,
Mark.