Re: [PATCH v2 05/19] riscv: add ISA extension parsing for vector crypto extensions

[Evan Green]

On Thu, Oct 19, 2023 at 8:33 AM Conor Dooley <conor@xxxxxxxxxx> wrote:
>
> On Thu, Oct 19, 2023 at 11:35:59AM +0200, Clément Léger wrote:
> >
> >
> > On 18/10/2023 19:26, Evan Green wrote:
> > > On Wed, Oct 18, 2023 at 5:53 AM Clément Léger <cleger@xxxxxxxxxxxx> wrote:
> > >>
> > >>
> > >>
> > >> On 18/10/2023 03:45, Jerry Shih wrote:
> > >>> On Oct 17, 2023, at 21:14, Clément Léger <cleger@xxxxxxxxxxxx> wrote:
> > >>>> @@ -221,6 +261,22 @@ const struct riscv_isa_ext_data riscv_isa_ext[] = {
> > >>>>      __RISCV_ISA_EXT_DATA(zkt, RISCV_ISA_EXT_ZKT),
> > >>>>      __RISCV_ISA_EXT_DATA(zksed, RISCV_ISA_EXT_ZKSED),
> > >>>>      __RISCV_ISA_EXT_DATA(zksh, RISCV_ISA_EXT_ZKSH),
> > >>>> +    __RISCV_ISA_EXT_DATA(zvbb, RISCV_ISA_EXT_ZVBB),
> > >>>> +    __RISCV_ISA_EXT_DATA(zvbc, RISCV_ISA_EXT_ZVBC),
> > >>>> +    __RISCV_ISA_EXT_DATA(zvkb, RISCV_ISA_EXT_ZVKB),
> > >>>
> > >>> The `Zvkb` is the subset of `Zvbb`[1]. So, the `Zvkb` should be bundled with `Zvbb`.
> > >>
> > >> Hi Jerry,
> > >>
> > >> Thanks for catching this, I think some other extensions will fall in
> > >> this category as well then (Zvknha/Zvknhb). I will verify that.
> > >
> > > The bundling mechanism works well when an extension is a pure lasso
> > > around other extensions. We'd have to tweak that code if we wanted to
> > > support cases like this, where the extension is a superset of others,
> > > but also contains loose change not present anywhere else (and
> > > therefore also needs to stand as a separate bit).
> >
> > For Zvbb and Zvknhb, I used the following code:
> >
> > static const unsigned int riscv_zvbb_bundled_exts[] = {
> >       RISCV_ISA_EXT_ZVKB,
> >       RISCV_ISA_EXT_ZVBB
> > };
> >
> > static const unsigned int riscv_zvknhb_bundled_exts[] = {
> >       RISCV_ISA_EXT_ZVKNHA,
> >       RISCV_ISA_EXT_ZVKNHB
> > };
> >
> > Which correctly results in both extension (superset + base set) being
> > enabled when only one is set. Is there something that I'm missing ?
> >
> > >
> > > IMO, decomposing "pure" bundles makes sense since otherwise usermode
> > > would have to query multiple distinct bitmaps that meant the same
> > > thing (eg check the Zk bit, or maybe check the Zkn/Zkr/Zkt bits, or
> > > maybe check the Zbkb/Zbkc... bits, and they're all equivalent). But
> > > when an extension is a superset that also contains loose change, there
> > > really aren't two equivalent bitmasks, each bit adds something new.
> >
> > Agreed but if a system only report ZVBB for instance and the user wants
> > ZVKB, then it is clear that ZVKB should be reported as well I guess. So
> > in the end, it works much like "bundle" extension, just that the bundle
> > is actually a "real" ISA extension by itself.
> >
> > Clément
> >
> > >
> > > There's an argument to be made for still turning on the containing
> > > extensions to cover for silly ISA strings (eg ISA strings that
> > > advertise the superset but fail to advertise the containing
> > > extensions). We can decide if we want to work that hard to cover
> > > hypothetical broken ISA strings now, or wait until they show up.
> > > Personally I would wait until something broken shows up. But others
> > > may feel differently.
>
> I'm not really sure that those are "silly" ISA strings. People are going
> to do it that way because it is much easier than spelling out 5 dozen
> sub-components, and it is pretty inevitable that subsets will be
> introduced in the future for extensions we currently have.
>
> IMO, it's perfectly valid to say you have the supersets and not spell
> out all the subcomponents.

Hm, ok. If ISA strings are likely to be written that way, then I agree
having the kernel flip on all the contained extensions is a good idea.
We can tweak patch 2 to support the parsing of struct
riscv_isa_ext_data with both .id and .bundle_size set (instead of only
one or the other as it is now). Looking back at that patch, it looks
quite doable. Alright!

-Evan