sysfs topology for arm64 cluster_id

[ddutile@redhat.com (Don Dutile)]

On 01/13/2015 07:47 PM, Jon Masters wrote:
> Hi Folks,
>
> TLDR: I would like to consider the value of adding something like
> "cluster_siblings" or similar in sysfs to describe ARM topology.
>
> A quick question on intended data representation in /sysfs topology
> before I ask the team on this end to go down the (wrong?) path. On ARM
> systems today, we have a hierarchical CPU topology:
>
>                   Socket ---- Coherent Interonnect ---- Socket
>                     |                                    |
>           Cluster0 ... ClusterN                Cluster0 ... ClusterN
>              |             |                      |             |
>        Core0...CoreN  Core0...CoreN        Core0...CoreN  Core0...CoreN
>          |       |      |        |           |       |      |       |
>       T0..TN  T0..Tn  T0..TN  T0..TN       T0..TN T0..TN  T0..TN  T0..TN
>
> Where we might (or might not) have threads in individual cores (a la SMT
> - it's allowed in the architecture at any rate) and we group cores
> together into units of clusters usually 2-4 cores in size (though this
> varies between implementations, some of which have different but similar
> concepts, such as AppliedMicro Potenza PMDs CPU complexes of dual
> cores). There are multiple clusters per "socket", and there might be an
> arbitrary number of sockets. We'll start to enable NUMA soon.
>
> The existing ARM architectural code understands expressing topology in
> terms of the above, but it doesn't quite map these concepts directly in
> sysfs (does not expose cluster_ids as an example). Currently, a cpu-map
> in DeviceTree can expose hierarchies (included nested clusters) and this
> is parsed at boot time to populate scheduler information, as well as the
> topology files in sysfs (if that is provided - none of the reference
> devicetrees upstream do this today, but some exist). But the cluster
> information itself isn't quite exposed (whereas other whacky
> architectural concepts such as s390 books are exposed already today).
>
> Anyway. We have a small problem with tools such as those in util-linux
> (lscpu) getting confused as a result of translating x86-isms to ARM. For
> example, the lscpu utility calculates the number of sockets using the
> following computation:
>
> nsockets = desc->ncpus / nthreads / ncores
>
> (number of sockets = total number of online processing elements /
> threads within a single core / cores within a single socket)
>
> If you're not careful, you can end up with something like:
>
> # lscpu
> Architecture:          aarch64
> Byte Order:            Little Endian
> CPU(s):                8
> On-line CPU(s) list:   0-7
> Thread(s) per core:    1
> Core(s) per socket:    2
> Socket(s):             4
>
Basically, in the top-most diagram, lscpu (& hwloc) are equating Cluster<N>
as socket<N>.  I'm curious what the sysfs numa info will be interpreted
as when/if that is turned on for arm64.

> Now we can argue that the system in question needs an updated cpu-map
> (it'll actually be something ACPI but I'm keeping this discussion to DT
> to avoid that piece further in discussion, and you can assume I'm
> booting any test boxes in further work on this using DeviceTree prior to
> switching the result over to ACPI) but either way, util-linux is
> thinking in an x86-centric sense of what these files mean. And I think
> the existing topology/cpu-map stuff in arm64 is doing the same.
>
The above values are extracted from the MPIDR:Affx fields and is currently
independent of DT & ACPI.
The Aff1 field is the 'cluster-id' and is being used to associated cpu's (via cpu masks)
to siblings. lscpu & hwloc associate cpu-nums & siblings to sockets via the above
calculation, which doesn't quite show how siblings enter the equation
       ncores = CPU_COUNT_S(setsize, core_siblings) / nthreads;

Note: in the arm(32) tree, what was 'socket-id' is 'cluster-id' in arm64;
       I believe this 'mapping' (backporting/association) is one root problem
       in the arch/arm64/kernel/topology.c code.

Now, a simple, yet requiring lots of fun, cross-architecture testing, would
be to change lscpu to use the sysfs physical_package_id to get Socket correct.  Yet,
that won't fix the above 'Core(s) per socket' because that's being created
via the sibling masks, which are generated from the cluster-id.
This change would require arm(64) to implement DT & ACPI methods to
extract pcpu's to sockets (missing at the moment).

And modifying the cluster-id and/or the siblings masks creates non-topology
(non-lscpu, non-hwloc) issues like breaking gic init code paths which use
the cluster-id information as well. ... some 'empirical data' to note
if anyone thinks it's just a topology-presentation issue.

> Is it not a good idea to expose the cluster details directly in sysfs
> and have these utilities understand the possible extra level in the
> calculation? Or do we want to just fudge the numbers (as seems to be the
> case in some systems I am seeing) to make the x86 model add up?
>
Short-term, I'm trying to develop a reasonable 'fudge' for lscpu & hwloc,
that doesn't impact the (proper) operation of the gic code.
I haven't dug deep enough yet, but this also requires a check on how
the scheduler associates cpu-cache-sibling associativity when selecting
optimal cpu to schedule threads on.

> Let me know the preferred course...
>
> Jon.
>
> _______________________________________________
> linux-arm-kernel mailing list
> linux-arm-kernel at lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
>