Re: [PATCH V2 1/2] riscv: Add RISC-V svpbmt extension

From: Anup Patel
Date: Mon Sep 27 2021 - 23:50:40 EST

On Tue, Sep 28, 2021 at 6:32 AM Nick Kossifidis <mick@xxxxxxxxxxxx> wrote:
> Στις 2021-09-27 23:13, Atish Patra έγραψε:
> >> We need to decide whether we should support the upstream kernel for
> >> D1. Few things to consider.
> >> – Can it be considered as an errata ?
> It's one thing to follow the spec and have an error in the
> implementation, and another to not follow the spec.
> >> – Does it set a bad precedent and open can of worms in future ?
> IMHO yes, I'm thinking of Kendryte 210 devs for example coming up and
> asking for MMU support, they 've also shipped many chips already. I can
> also imagine other vendors in the future coming up with implementations
> that violate the spec in which case handling the standard stuff will
> become messy and complex, and hurt performance/security. We'll end up
> filling the code with exceptions and tweaks all over the place. We need
> to be strict about what is "riscv" and what's "draft riscv" or "riscv
> inspired", and what we are willing to support upstream. I can understand
> supporting vendor extensions upstream but they need to fit within the
> standard spec, we can't have for example extensions that use encoding
> space/csrs/fields etc reserved for standard use, they may only use
> what's reserved for custom/vendor use. At least let's agree on that.

Totally agree with Nick here. It's a slippery slope.

Including D1 PTE bits (or Kendryte K210 MMU) part of the Linux RISC-V
means future hardware which intentionally violates specs will also have to
be merged and the RISC-V patch acceptance policy will have no significance.

> >> – Can we just ignore D1 given the mass volume ?
> >>
> IMHO no, we need to find a way to support it upstream but I believe
> there is another question to answer:
> Do we also guarantee "one image to rule them all" approach, required by
> binary distros, for implementations that violate the spec ? Are we ok
> for example to support Allwinner D1 upstream but require a custom
> configuration/build instead of supporting it with the "generic" image ?
> In one case we need to handle the violation at runtime and introduce
> overhead for everyone (like looking up __riscv_svpbmt every time we set
> a PTE in this case), in the other it's an #ifdef.

At least, we should not have hardware violating specs as part of the
unified kernel image instead have these intentional deviations/violations
under separate kconfig which will not be enabled by default. This means
vendors (of such hardware) and distros will have to explicitly enable
support for such violations/deviations.