Re: [RFC 2/3] arm_pmu: acpi: spe: Add initial MADT/SPE probing

[Jeremy Linton]

Hi,

On 2/15/19 9:00 AM, Will Deacon wrote:
> On Thu, Feb 14, 2019 at 12:03:57PM -0600, Jeremy Linton wrote:
> > On 2/14/19 11:11 AM, Will Deacon wrote:
> > > On Fri, Feb 08, 2019 at 06:47:17PM -0600, Jeremy Linton wrote:
> > > > +/*
> > > > + * For lack of a better place, hook the normal PMU MADT walk
> > > > + * and create a SPE device if we detect a recent MADT with
> > > > + * a homogeneous PPI mapping.
> > > > + */
> > > > +static int arm_spe_acpi_parse_irqs(void)
> > > > +{
> > > > +	int cpu, ret, irq;
> > > > +	u16 gsi = 0;
> > > > +	bool first = true;
> > > > +
> > > > +	struct acpi_madt_generic_interrupt *gicc;
> > > > +
> > > > +	/*
> > > > +	 * sanity check all the GICC tables for the same interrupt number
> > > > +	 * for now we only support homogeneous ACPI/SPE machines.
> > > > +	 */
> > > > +	for_each_possible_cpu(cpu) {
> > > > +		gicc = acpi_cpu_get_madt_gicc(cpu);
> > > > +
> > > > +		if (gicc->header.length < ACPI_MADT_GICC_SPE)
> > > > +			return -ENODEV;
> > > > +
> > > > +		if (first) {
> > > > +			gsi = gicc->spe_overflow_interrupt;
> > > > +			if (!gsi)
> > > > +				return -ENODEV;
> > > > +			first = false;
> > > > +		} else if (gsi != gicc->spe_overflow_interrupt) {
> > > > +			pr_warn("ACPI: SPE must have homogeneous interrupts\n");
> > > > +			return -EINVAL;
> > > > +		}
> > >
> > > Unfortunately, I don't think this is sufficient to detect a homogeneous
> > > system: we'll have to check the MIDRs instead, which is nasty. I would
> > > personally be in favour of enforcing homogeneity for ACPI systems when we
> > > bring up secondary CPUs, but I suspect others would disagree.
> >
> > Given that all the SPE capable machines i'm aware of at the moment are
> > homogeneous, are we ok with just doing an online CPU MIDR check for now, and
> > cleaning that up if/when someone builds a machine and complains?
>
> Yeah, I think we added a new bit to the PPTT to tell you that the machine is
> homogenous, so just check that first and bail if it's not set.

Yes of course, 100% better plan. Although its probably going to have to 
be more of a case of walking all the possible cores and assuring they 
have the same flag level (similar to how the socket flag is handled). Of 
course that information is useful enough it should probably just be done 
as part of the normal cpu topology walk. Then the people who have to 
back port these patches end up with a big dependent set... <chuckle>