Re: [PATCH v5 1/4] dt-bindings: EDAC: Add Amazon's Annapurna Labs L1 EDAC

[Rob Herring <robh@kernel.org>]

On Fri, Aug 30, 2019 at 7:45 AM James Morse <james.morse@arm.com> wrote:
>
> Hi guys,
>
> On 27/08/2019 14:49, Rob Herring wrote:
> > On Mon, Aug 26, 2019 at 9:49 AM Hawa, Hanna <hhhawa@amazon.com> wrote:
> >> On 8/21/2019 10:17 PM, Rob Herring wrote:
> >>> Why is this even in DT? AFAICT, this is all just CortexA57 core features
> >>> (i.e. nothing Amazon specific). The core type and the ECC capabilities
> >>> are discoverable.
> >>
> >> Added to the DT in order to easily enable/disable the driver.
> >
> > That alone is not reason enough to put it in DT. From a DT
> > perspective, I have no idea what the whims of a OS maintainer are
> > regarding whether they want all this to be 1 driver or 2 drivers.
> > (IMO, it should be 1 as this is ECC for an A57. For a core and memory
> > controller, then 2 seems appropriate.)
> >
> >> You are
> >> correct that they are CortexA57 core features and nothing Amazon
> >> specific, but it's IMPLEMENTATION DEFINED, meaning that in different
> >> cortex revisions (e.g. A57) the register bitmap may change. Because of
> >> that we added an Amazon compatible which corresponds to the specific
> >> core we are using.
>
> I think its that the instruction encoding is in the imp-def space that is important.
>
> CPU-implementers can add whatever registers they find useful here. A57 and A72 both
> implemented some ECC registers here. (They are not guaranteed to be the same, but I can't
> find any differences).

Two cores potentially being the same only furthers my argument that
this shouldn't be an Amazon driver.

> We need some information from DT because the TRM doesn't say what happens when you read
> from these registers on an A57 that doesn't have the 'optional ECC protection'. It could
> take an exception due to an unimplemented system register.

My read of the TRM is that L2 ECC is always there and L1 ECC/parity is
optional. Furthermore, bit 22 of L2CTRL_EL1 indicates if L1 ECC/parity
is supported or not and there's other non-ECC stuff like cache RAM
timing values in that register.

> The imp-def instruction space may also be trapped by a higher exception level. KVM does
> this, and emulates these registers as if they were all undefined.

So KVM provides a semi-CortexA57? Code that runs on real h/w won't as a guest.

However, if we do need DT to indicate ECC support in a core or not,
then we already have an A57 node and should put that info there.

Rob