RE: [tip: sched/core] sched: Add cluster scheduler level for x86

[Song Bao Hua (Barry Song)]

> -----Original Message-----
> From: Peter Zijlstra [mailto:peterz@xxxxxxxxxxxxx]
> Sent: Friday, October 22, 2021 2:23 AM
> To: Barry Song <21cnbao@xxxxxxxxx>
> Cc: Tom Lendacky <thomas.lendacky@xxxxxxx>; LKML
> <linux-kernel@xxxxxxxxxxxxxxx>; linux-tip-commits@xxxxxxxxxxxxxxx; Tim Chen
> <tim.c.chen@xxxxxxxxxxxxxxx>; Song Bao Hua (Barry Song)
> <song.bao.hua@xxxxxxxxxxxxx>; x86 <x86@xxxxxxxxxx>
> Subject: Re: [tip: sched/core] sched: Add cluster scheduler level for x86
> 
> On Thu, Oct 21, 2021 at 11:32:36PM +1300, Barry Song wrote:
> > On Thu, Oct 21, 2021 at 9:43 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > >
> > > On Wed, Oct 20, 2021 at 10:36:19PM +0200, Peter Zijlstra wrote:
> > >
> > > > OK, I think I see what's happening.
> > > >
> > > > AFAICT cacheinfo.c does *NOT* set l2c_id on AMD/Hygon hardware, this
> > > > means it's set to BAD_APICID.
> > > >
> > > > This then results in match_l2c() to never match. And as a direct
> > > > consequence set_cpu_sibling_map() will generate cpu_l2c_shared_mask with
> > > > just the one CPU set.
> > > >
> > > > And we have the above result and things come unstuck if we assume:
> > > >   SMT <= L2 <= LLC
> > > >
> > > > Now, the big question, how to fix this... Does AMD have means of
> > > > actually setting l2c_id or should we fall back to using match_smt() for
> > > > l2c_id == BAD_APICID ?
> > >
> > > The latter looks something like the below and ought to make EPYC at
> > > least function as it did before.
> > >
> > >
> > > ---
> > >  arch/x86/kernel/smpboot.c | 2 +-
> > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > >
> > > diff --git a/arch/x86/kernel/smpboot.c b/arch/x86/kernel/smpboot.c
> > > index 849159797101..c2671b2333d1 100644
> > > --- a/arch/x86/kernel/smpboot.c
> > > +++ b/arch/x86/kernel/smpboot.c
> > > @@ -472,7 +472,7 @@ static bool match_l2c(struct cpuinfo_x86 *c, struct
> cpuinfo_x86 *o)
> > >
> > >         /* Do not match if we do not have a valid APICID for cpu: */
> > >         if (per_cpu(cpu_l2c_id, cpu1) == BAD_APICID)
> > > -               return false;
> > > +               return match_smt(c, o); /* assume at least SMT shares L2 */
> >
> > Rather than making a fake cluster_cpus and cluster_cpus_list which
> > will expose to userspace
> > through /sys/devices/cpus/cpux/topology, could we just fix the
> > sched_domain mask by the
> > below?
> 
> I don't think it's fake; SMT fundamentally has to share all cache
> levels. And having the sched domains differ in setup from the reported
> (nonsensical) topology also isn't appealing.

Fair enough. I was actually inspired by cpu_coregroup_mask() which
is a combination of a couple of cpumask set:
drivers/base/arch_topology.c

const struct cpumask *cpu_coregroup_mask(int cpu)
{
	const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));

	/* Find the smaller of NUMA, core or LLC siblings */
	if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
		/* not numa in package, lets use the package siblings */
		core_mask = &cpu_topology[cpu].core_sibling;
	}
	if (cpu_topology[cpu].llc_id != -1) {
		if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask))
			core_mask = &cpu_topology[cpu].llc_sibling;
	}

	return core_mask;
}

Thanks
Barry