Re: [PATCH 2/2] arm64: errata: Add Cortex-A520 speculative unprivileged load workaround

[Marc Zyngier]

On Tue, 19 Sep 2023 13:29:07 +0100,
Rob Herring <robh@xxxxxxxxxx> wrote:
> 
> On Mon, Sep 18, 2023 at 5:18 AM Marc Zyngier <maz@xxxxxxxxxxxxxxx> wrote:
> >
> > On 2023-09-18 11:01, Will Deacon wrote:
> > > On Tue, Sep 12, 2023 at 07:11:15AM -0500, Rob Herring wrote:
> > >> Implement the workaround for ARM Cortex-A520 erratum 2966298. On an
> > >> affected Cortex-A520 core, a speculatively executed unprivileged load
> > >> might leak data from a privileged level via a cache side channel.
> > >>
> > >> The workaround is to execute a TLBI before returning to EL0. A
> > >> non-shareable TLBI to any address is sufficient.
> > >
> > > Can you elaborate at all on how this works, please? A TLBI addressing a
> > > cache side channel feels weird (or is "cache" referring to some TLB
> > > structures rather than e.g. the data cache here?).
> > >
> > > Assuming there's some vulnerable window between the speculative
> > > unprivileged load and the completion of the TLBI, what prevents another
> > > CPU from observing the side-channel during that time? Also, does the
> > > TLBI need to be using the same ASID as the unprivileged load? If so,
> > > then
> > > a context-switch could widen the vulnerable window quite significantly.
> >
> > Another 'interesting' case is the KVM world switch. If EL0 is
> > affected, what about EL1? Can such a data leak exist cross-EL1,
> > or from EL2 to El1? Asking for a friend...
> 
> I'm checking for a definitive answer, but page table isolation also
> avoids the issue. Wouldn't these scenarios all be similar to page
> table isolation in that the EL2 or prior EL1 context is unmapped?

No, EL2 is always mapped, and we don't have anything like KPTI there.

Maybe the saving grace is that EL2 and EL2&0 are different translation
regimes from EL1&0, but there's nothing in the commit message that
indicates it. As for EL1-to-EL1 leaks, it again completely depends on
how the TLBs are tagged.

You'd hope that having different VMIDs would save the bacon, but if
you can leak EL1 translations into EL0, it means that the associated
permission and/or tags do not contain all the required information...

	M.

-- 
Without deviation from the norm, progress is not possible.