Re: [PATCH v3] topology: make core_mask include at least cluster_siblings

[Darren Hart <darren@os.amperecomputing.com>]

On Tue, Mar 08, 2022 at 05:03:07PM +0100, Dietmar Eggemann wrote:
> On 08/03/2022 12:04, Vincent Guittot wrote:
> > On Tue, 8 Mar 2022 at 11:30, Will Deacon <will@kernel.org> wrote:
> 
> [...]
> 
> >>> ---
> >>> v1: Drop MC level if coregroup weight == 1
> >>> v2: New sd topo in arch/arm64/kernel/smp.c
> >>> v3: No new topo, extend core_mask to cluster_siblings
> >>>
> >>>  drivers/base/arch_topology.c | 8 ++++++++
> >>>  1 file changed, 8 insertions(+)
> >>>
> >>> diff --git a/drivers/base/arch_topology.c b/drivers/base/arch_topology.c
> >>> index 976154140f0b..a96f45db928b 100644
> >>> --- a/drivers/base/arch_topology.c
> >>> +++ b/drivers/base/arch_topology.c
> >>> @@ -628,6 +628,14 @@ const struct cpumask *cpu_coregroup_mask(int cpu)
> >>>                       core_mask = &cpu_topology[cpu].llc_sibling;
> >>>       }
> >>>
> >>> +     /*
> >>> +      * For systems with no shared cpu-side LLC but with clusters defined,
> >>> +      * extend core_mask to cluster_siblings. The sched domain builder will
> >>> +      * then remove MC as redundant with CLS if SCHED_CLUSTER is enabled.
> 
> IMHO, if core_mask weight is 1, MC will be removed/degenerated anyway.
> 
> This is what I get on my Ampere Altra (I guess I don't have the ACPI
> changes which would let to a CLS sched domain):
> 
> # cat /sys/kernel/debug/sched/domains/cpu0/domain*/name
> DIE
> NUMA
> root@oss-altra01:~# zcat /proc/config.gz | grep SCHED_CLUSTER
> CONFIG_SCHED_CLUSTER=y

I'd like to follow up on this. Would you share your dmidecode BIOS
Information section?

Which kernel version?

> >>> +      */
> >>> +     if (cpumask_subset(core_mask, &cpu_topology[cpu].cluster_sibling))
> >>> +             core_mask = &cpu_topology[cpu].cluster_sibling;
> >>> +
> >>
> >> Sudeep, Vincent, are you happy with this now?
> > 
> > I would not say that I'm happy because this solution skews the core
> > cpu mask in order to abuse the scheduler so that it will remove a
> > wrong but useless level when it will build its domains.
> > But this works so as long as the maintainer are happy, I'm fine

I did explore the other options and they added considerably more
complexity without much benefit in my view. I prefer this option which
maintains the cpu_topology as described by the platform, and maps it
into something that suits the current scheduler abstraction. I agree
there is more work to be done here and intend to continue with it.

> I do not have any better idea than this tweak here either in case the
> platform can't provide a cleaner setup.

I'd argue The platform is describing itself accurately in ACPI PPTT
terms. The topology doesn't fit nicely within the kernel abstractions
today. This is an area where I hope to continue to improve things going
forward.

> Maybe the following is easier to read but then we use
> '&cpu_topology[cpu].llc_sibling' in cpu_coregroup_mask() already ...
> 
> @@ -617,6 +617,7 @@ EXPORT_SYMBOL_GPL(cpu_topology);
>  const struct cpumask *cpu_coregroup_mask(int cpu)
>  {
>         const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));
> +       const cpumask_t *cluster_mask = cpu_clustergroup_mask(cpu);
> 
>         /* Find the smaller of NUMA, core or LLC siblings */
>         if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
> @@ -628,6 +629,9 @@ const struct cpumask *cpu_coregroup_mask(int cpu)
>                         core_mask = &cpu_topology[cpu].llc_sibling;
>         }
> 
> +       if (cpumask_subset(core_mask, cluster_mask))
> +               core_mask = cluster_mask;
> +

Either works for me. I felt the version I sent was parallel to the
existing implementation, but have no preference either way.

>         return core_mask;
>  }
> 
> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
> 

Thanks for the review Dietmar.

-- 
Darren Hart
Ampere Computing / OS and Kernel