Advice needed on SMP regression after cpu_core_mask change

[Srikar Dronamraju]

* Daniel Henrique Barboza <danielhb413 at gmail.com> [2021-03-17 10:00:34]:

> Hello,
> 
> Patch 4bce545903fa ("powerpc/topology: Update topology_core_cpumask") introduced
> a regression in both upstream and RHEL downstream kernels [1]. The assumption made
> in the commit:
> 
> "Further analysis shows that cpu_core_mask and cpu_cpu_mask for any CPU would be
> equal on Power"
> 
> Doesn't seem to be true. After this commit, QEMU is now unable to set single NUMA
> node SMP topologies such as:
> 
> -smp 8,maxcpus=8,cores=2,threads=2,sockets=2

What does it mean for a NUMA to have more than one sockets?
If they are all part of the same node, there are at local distance to each
other. cache is per core. So what resources are shared by the Sockets that
are part of the same NUMA. And how does Userspace/ application make use of
the same.

Please don't mistake this as attempt to downplay your report but a honest
attempt to better understand the situation.

For example, if the socket denotes the hemisphere logic in P10, then can we
see if the coregroup feature can be used. "Coregroup" is suppose to mean a
set of cores within a NUMA that have some characteristics and there can be
multiple coregroups within a NUMA. We add that mostly to mimic hemisphere in
P10. However the number of coregroups in a NUMA is not exported to userspace
at this time.

However if each Socket is associated with a memory and node distance, then
should they be NUMA?

Can you provide me with the unique ibm,chip-ids in your 2 NUMA, 4 node case?
Does this cause an performance issues with the guest/application?

Till your report, I was under the impression that NUMAs == Sockets.

> 
> lscpu will give the following output in this case:
> 
> # lscpu
> Architecture:        ppc64le
> Byte Order:          Little Endian
> CPU(s):              8
> On-line CPU(s) list: 0-7
> Thread(s) per core:  2
> Core(s) per socket:  4
> Socket(s):           1
> NUMA node(s):        1
> Model:               2.2 (pvr 004e 1202)
> Model name:          POWER9 (architected), altivec supported
> Hypervisor vendor:   KVM
> Virtualization type: para
> L1d cache:           32K
> L1i cache:           32K
> NUMA node0 CPU(s):   0-7
> 
> 
> This is happening because the macro cpu_cpu_mask(cpu) expands to
> cpumask_of_node(cpu_to_node(cpu)), which in turn expands to node_to_cpumask_map[node].
> node_to_cpumask_map is a NUMA array that maps CPUs to NUMA nodes (Aneesh is on CC to
> correct me if I'm wrong). We're now associating sockets to NUMA nodes directly.
> 
> If I add a second NUMA node then I can get the intended smp topology:
> 
> -smp 8,maxcpus=8,cores=2,threads=2,sockets=2
> -numa node,memdev=mem0,cpus=0-3,nodeid=0 \
> -numa node,memdev=mem1,cpus=4-7,nodeid=1 \
> 
> # lscpu
> Architecture:        ppc64le
> Byte Order:          Little Endian
> CPU(s):              8
> On-line CPU(s) list: 0-7
> Thread(s) per core:  2
> Core(s) per socket:  2
> Socket(s):           2
> NUMA node(s):        2
> Model:               2.2 (pvr 004e 1202)
> Model name:          POWER9 (architected), altivec supported
> Hypervisor vendor:   KVM
> Virtualization type: para
> L1d cache:           32K
> L1i cache:           32K
> NUMA node0 CPU(s):   0-3
> NUMA node1 CPU(s):   4-7
> 
> 
> However, if I try a single socket with multiple NUMA nodes topology, which is the case
> of Power10, e.g.:
> 
> 
> -smp 8,maxcpus=8,cores=4,threads=2,sockets=1
> -numa node,memdev=mem0,cpus=0-3,nodeid=0 \
> -numa node,memdev=mem1,cpus=4-7,nodeid=1 \
> 
> 
> This is the result:
> 
> # lscpu
> Architecture:        ppc64le
> Byte Order:          Little Endian
> CPU(s):              8
> On-line CPU(s) list: 0-7
> Thread(s) per core:  2
> Core(s) per socket:  2
> Socket(s):           2
> NUMA node(s):        2
> Model:               2.2 (pvr 004e 1202)
> Model name:          POWER9 (architected), altivec supported
> Hypervisor vendor:   KVM
> Virtualization type: para
> L1d cache:           32K
> L1i cache:           32K
> NUMA node0 CPU(s):   0-3
> NUMA node1 CPU(s):   4-7
> 
> 
> This confirms my suspicions that, at this moment, we're making sockets == NUMA nodes.
> 
> 
> Cedric, the reason I'm CCing you is because this is related to ibm,chip-id. The commit
> after the one that caused the regression, 4ca234a9cbd7c3a65 ("powerpc/smp: Stop updating
> cpu_core_mask"), is erasing the code that calculated cpu_core_mask. cpu_core_mask, despite
> its shortcomings that caused its removal, was giving a precise SMP topology. And it was
> using physical_package_id/'ibm,chip-id' for that.
> 
> Checking in QEMU I can say that the ibm,chip-id calculation is the only place in the code
> that cares about cores per socket information. The kernel is now ignoring that, starting
> on 4bce545903fa, and now QEMU is unable to provide this info to the guest.
> 
> If we're not going to use ibm,chip-id any longer, which seems sensible given that PAPR does
> not declare it, we need another way of letting the guest know how much cores per socket
> we want.
> 
> 
> 
> [1] https://bugzilla.redhat.com/1934421
> 
> 
> 
> Thanks,
> 
> 
> DHB

-- 
Thanks and Regards
Srikar Dronamraju