Re: [PATCH] arm64/numa: Add more vetting in numa_set_distance()

[John Garry]

> > > >  I think we should either factor out the sanity check
> > > > > into a core helper or make the core code robust to these funny configurations.
> > > >
> > > > OK, so to me it would make sense to factor out a sanity check into a core
> > > > helper.
> > >
> > > That, or have the OF code perform the same validation that slit_valid() is
> > > doing for ACPI. I'm just trying to avoid other architectures running into
> > > this problem down the line.
> >
> > Right, OF code should do this validation job if ACPI is doing it (especially since the DT bindings actually specify the distance rules), and not rely on the arch NUMA code to accept/reject numa_set_distance() combinations.
>
> I would say this particular condition checking still falls under arch NUMA init
> code sanity check like other basic tests what numa_set_distance() currently does
> already but it should not be a necessity for the OF driver to check these.

The checks in the arch NUMA code mean that invalid inter-node distance 
combinations are ignored.

However, if any entries in the table are invalid, then the whole table 
can be discarded as none of it can be believed, i.e. it's better to 
validate the table.

It can
> choose to check but arch NUMA should check basic things like two different NUMA
> nodes should not have LOCAL_DISTANCE as distance like in this case.
>
> 	(from == to && distance != LOCAL_DISTANCE) ||
> 		(from != to && distance == LOCAL_DISTANCE))
>
>
> > And, in addition to this, I'd say OF should disable NUMA if given an invalid table (like ACPI does).
>
> Taking a decision to disable NUMA should be with kernel (arch NUMA) once kernel
> starts booting. Platform should have sent right values, OF driver trying to
> adjust stuff what platform has sent with FDT once the kernel starts booting is
> not right. For example "Kernel NUMA wont like the distance factors lets clean
> then up before passing on to MM".

Sorry, but I don't know who was advocating this.

Disabling NUMA is one such major decision which
> should be with arch NUMA code not with OF driver.

I meant parsing the table would fail, so arch NUMA would fall back on 
dummy NUMA.

>

Thanks,
John