Re: [PATCH 1/2] sched/topology: Allow archs to override cpu_smt_mask

From: peterz
Date: Thu Aug 06 2020 - 13:28:55 EST

On Thu, Aug 06, 2020 at 10:25:12PM +1000, Michael Ellerman wrote:
> peterz@xxxxxxxxxxxxx writes:
> > On Thu, Aug 06, 2020 at 03:32:25PM +1000, Michael Ellerman wrote:
> >
> >> That brings with it a bunch of problems, such as existing software that
> >> has been developed/configured for Power8 and expects to see SMT8.
> >>
> >> We also allow LPARs to be live migrated from Power8 to Power9 (and back), so
> >> maintaining the illusion of SMT8 is considered a requirement to make that work.
> >
> > So how does that work if the kernel booted on P9 and demuxed the SMT8
> > into 2xSMT4? If you migrate that state onto a P8 with actual SMT8 you're
> > toast again.
> The SMT mask would be inaccurate on the P8, rather than the current case
> where it's inaccurate on the P9.
> Which would be our preference, because the backward migration case is
> not common AIUI.
> Or am I missing a reason we'd be even more toast than that?

Well, the scheduler might do a wee bit funny. We just had a patch that
increase load-balancing opportunities between SMT siblings because they
all share L1 anyway.

But yeah, nothing terminal.

> Under PowerVM the kernel does know it's being migrated, so we could
> actually update the mask, but I'm not sure if that's really feasible.

As long as you get a notification, rebuilding the sched domains isn't
terribly hard to do, there's more code that does that.

> >> Yeah I agree the naming is confusing.
> >>
> >> Let's call them "SMT4 cores" and "SMT8 cores"?
> >
> > Works for me, thanks!
> >
> >> The problem is we are already lying to userspace, because firmware lies to us.
> >>
> >> ie. the firmware on these systems shows us an SMT8 core, and so current kernels
> >> show SMT8 to userspace. I don't think we can realistically change that fact now,
> >> as these systems are already out in the field.
> >>
> >> What this patch tries to do is undo some of the mess, and at least give the
> >> scheduler the right information.
> >
> > What a mess... I think it depends on what you do with that P9 to P8
> > migration case. Does it make sense to have a "p8_compat" boot arg for
> > the case where you want LPAR migration back onto P8 systems -- in which
> > case it simply takes the firmware's word as gospel and doesn't untangle
> > things, because it can actually land on a P8.
> We already get told by firmware that we're running in "p8 compat" mode,
> because we have to pretend to userspace that it's running on a P8. So we
> could use that as a signal to leave things alone.
> But my understanding is most LPARs don't get migrated back and forth,
> they'll start life on a P8 and only get migrated to a P9 once when the
> customer gets a P9. They might then run for a long time (months to
> years) on the P9 in P8 compat mode, not because they ever want to
> migrate back to a real P8, but because the software in the LPAR is still
> expecting to be on a P8.
> I'm not a real expert on all the Enterprisey stuff though, so someone
> else might be able to give us a better picture.
> But the point of mentioning the migration stuff was mainly just to
> explain why we feel we need to present SMT8 to userspace even on P9.

OK, fair enough. The patch wasn't particularly onerous, I was just
wondering why etc..

The case of starting on a P8 and being migrated to a P9 makes sense to
me; in that case you'd like to rebuild your sched domains, but can't go
about changing user visible topolofy information.

I suppose:

Acked-by; Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>

An updated Changelog that recaps some of this discussion might also be