Re: [PATCH v2 0/4] NUMA: Apply socket-NUMA-node boundary for aarch64 and RiscV machines

[Gavin Shan]

Hi Drew,

On 2/23/23 11:25 PM, Andrew Jones wrote:
> On Thu, Feb 23, 2023 at 04:13:57PM +0800, Gavin Shan wrote:
> > For arm64 and RiscV architecture, the driver (/base/arch_topology.c) is
> > used to populate the CPU topology in the Linux guest. It's required that
> > the CPUs in one socket can't span mutiple NUMA nodes. Otherwise, the Linux
> > scheduling domain can't be sorted out, as the following warning message
> > indicates. To avoid the unexpected confusion, this series attempts to
> > rejects such kind of insane configurations.
>
> I'm not sure 'insane' is the right word, maybe 'irregular'. I've never
> seen, and don't really expect to find, specifications stating that arm
> and/or riscv must have 1:1 mappings for numa nodes and sockets. But, as
> QEMU machine types can choose certain configurations to support, if
> everyone is in agreement to restrict sbsa-ref and arm/riscv virt to 1:1
> mappings, then why not. However, I suggest stating that this is simply a
> choice being made for these boards, since the expectation of needing
> "irregular" configurations is low. We should not be trying to use Linux
> assumptions as rationale, since other OSes may cope better with irregular
> configurations.
>
> Also, I suggest adding a comment to the boards, which adopt this
> restriction, which explains that it's only a platform choice, not
> something specified by the architecture.

[...]

Thanks for your comments. Yes, "irregular" is definitely better words. I failed
to find the statement that ARM64 has the requirement for 1:1 mapping between
socket and NUMA node, which makes it hard to tell if the restriction is required
from architectural level. However, the linux driver 
(drivers/base/arch_topology.c)
considers both CPU-to-NUMA-association and CPU toplogy from ACPI PPTT when the
CPU set in the core group is calculated. It means the CPU-to-NUMA-association
can affect the CPU topology seen by Linux scheduler.

With the restriction of 1:1 mapping between socket and NUMA node, or between
cluster and NUMA node as I mentioned in the reply to Daniel P. Berrangé 
<berrange@redhat.com>,
the existing configuration for other OSes can become invalid.

Yes, a comment to mention it's a platform choice, not architectural requirement,
definitely helps to understand the context.

Thanks,
Gavin