[PATCH 1/2] arm64: smp: Fix pseudo NMI issues w/ broken Mediatek FW

[Mark Rutland]

On Wed, Oct 04, 2023 at 07:04:12AM -0700, Doug Anderson wrote:
> On Wed, Oct 4, 2023 at 3:15 AM Marc Zyngier <maz at kernel.org> wrote:
> >
> > On Wed, 04 Oct 2023 10:59:50 +0100,
> > Mark Rutland <mark.rutland at arm.com> wrote:
> > >
> > > Given you haven't seen any issues, I suspect those are getting reset to fixed
> > > values that happens to work out for us, but it is a bit worrisome more
> > > generally (e.g. the LPI case above).
> >
> > It is likely that these SoCs don't even have an ITS.
> 
> Right. That was what we decided [1] when Marc pointed this out earlier.
> 
> Overall: we know that this firmware behavior is not good but we're
> stuck with it. :( At the very least, any new devices coming out will
> have this fixed. Presumably if old devices are working OK enough today
> (as long as you don't enable pseudo-NMI) then they can be made to keep
> working?
> 
> So circling back: what patch should we actually land?

For now I'd prefer we took the patch I sent in:

  https://lore.kernel.org/linux-arm-kernel/ZRr8r7XMoyDKaitd@FVFF77S0Q05N.cambridge.arm.com/

... as that leaves us no worse than before this series, and it's pretty simple.

> As of right now only pseudo-NMI is broken, but it would be good to make sure
> that if the kernel later adds other features that would be broken on this
> hardware that it gets handled properly...

Going further than the above, I think there are three options here:

1) Complete fix: depend on a working firmware, and throw this workaround away.

   IIUC from the above, that's not something you can commit to.

2) Partial fix: have the kernel save/restore everything.

   IIUC this is unpalatable.

3) Partial fix: make the ARM64_HAS_GIC_PRIO_MASKING cpucap depend on the
   absence of a "mediatek,broken-save-restore-fw" property in the DT. I believe
   we can check that in early_enable_pseudo_nmi() or can_use_gic_priorities().

   That'll avoid potential issues if/when we change the priorities used for
   pNMI (which is something I've been looking at).

I'm happy with (3) if Marc is.

Mark.