Re: [RFC PATCH 0/5] target/ppc: initial SMT support in TCG

["Nicholas Piggin" <npiggin@gmail.com>]

On Thu Jun 1, 2023 at 5:56 PM AEST, Cédric Le Goater wrote:
> Hello Nick,
>
> On 5/31/23 03:23, Nicholas Piggin wrote:
> > Hi,
> > 
> > I'm posting this now just to get some first thoughts. I wouldn't say
> > it's ready but it does actually work with some basic tests including
> > pseries booting a Linux distro. I have powernv booting too, it just
> > requires some more SPRs converted, nothing fundamentally different so
> > for the purpose of this RFC I leave it out.
> > 
> > A couple of things, I don't know the object model well enough to do
> > something nice with topology. Iterating siblings I would have thought
> > should be going to parent core then iterating its children CPUs. Should
> > that be done with the object model, or is it better to add direct
> > pointers in CPUs to core and core to CPUs? It is (semi) important for> performance so maybe that is better than object iterators. If we go that
> > way, the PnvCore and SpaprCore have pointers to the SMT threads already,
> > should those be abstracted go in the CPUCore?
>
> You should be able to move the thread array into the CPUCore. If you do
> that, please check that migration compat is not impacted by the state
> change. However, I am not sure you can use the CPUCore model under the
> insn modeling. Something to check.

Okay.

> Anyhow, the way you implemented the loop on the siblings is sufficiently
> fast for a small numbers of CPU and safe, w.r.t to CPU hotplug. So
> I would leave that part for now, if it runs decently with 4*4 vCPUs in
> TCG it should be fine.

Yeah you're right I'm overly paranoid about it but we don't do hundreds
of CPUs in TCG so it should be fine. Maybe I will defer it for now
then and just do the CPU iteration.

Thanks,
Nick