Re: [PATCH] cpu-topology: warn if NUMA configurations conflicts with lower layer

[Morten Rasmussen]

On Sun, Jan 12, 2020 at 01:22:02PM +0000, Valentin Schneider wrote:
> On 09/01/2020 10:52, Morten Rasmussen wrote:
> >> AFAIA what matters here is memory controllers, less so LLCs. Cores within
> >> a single die could have private LLCs and separate memory controllers, or
> >> shared LLC and separate memory controllers.
> > 
> > Don't confuse cache boundaries, packages and nodes :-)
> > 
> > core_siblings are cpus in the same package and doesn't say anything
> > about cache boundaries. It is not given that there is sched_domain that
> > matches the core_sibling span.
> > 
> > The MC sched_domain is supposed to match the LLC span which might
> > different for core_siblings. So the about example should be valid for a
> > NUMA-in-package system with one package containing two nodes.
> > 
> 
> Right, the point I was trying to make is that node boundaries can be pretty
> much anything, so nodes can span over LLCs, or LLCs can span over nodes,
> which is why we need checks such as the one in arch_topology() that lets us
> build up a usable domain hierarchy (which cares about LLCs, at least at some
> level).

Indeed. The topology masks can't always be used as is to define the
sched_domain hierarchy.