[PATCH 0/4] Expose the isa-string via the AT_BASE_PLATFORM aux vector

[Evan Green]

On Mon, May 8, 2023 at 10:33 AM Philipp Tomsich
<philipp.tomsich at vrull.eu> wrote:
>
> Evan,
>
> On Mon, 8 May 2023 at 18:50, Evan Green <evan at rivosinc.com> wrote:
>>
>> On Wed, May 3, 2023 at 3:31 AM Heiko Stübner <heiko at sntech.de> wrote:
>> >
>> > Hi,
>> >
>> > Am Dienstag, 2. Mai 2023, 19:15:29 CEST schrieb Palmer Dabbelt:
>> > > On Tue, 02 May 2023 02:13:10 PDT (-0700), philipp.tomsich at vrull.eu wrote:
>> > > > On Tue, 2 May 2023 at 09:58, Björn Töpel <bjorn at kernel.org> wrote:
>> > > >>
>> > > >> Philipp Tomsich <philipp.tomsich at vrull.eu> writes:
>> > > >>
>> > > >> > It is a pity that the current interface was designed without involving
>> > > >> > RVI (and that I had to ask my team to put together a patch set for
>> > > >> > further discussion, given that none of the other major vendors in RVI
>> > > >> > stepped forward).  I guarantee that plenty of reviewers would have
>> > > >> > highlighted that a central registry (even if it is just a kernel
>> > > >> > header) should be avoided.
>> > > >>
>> > > >> Are you claiming that the hwprobe work was not done in the open, but
>> > > >> secretly merged? That is not only incorrect, but rude to upstream RISC-V
>> > > >> Linux developers. I suggest you review how you interact with upstream
>> > > >> kernel work.
>> > > >
>> > > > Please don't put words into my mouth...
>> > > >
>> > > > I was merely pointing out that there was no engagement by the RVI
>> > > > member companies (in regard to this mechanism) within RVI, which would
>> > > > have prevented Jessica's issue.
>> > > > This would have also helped to address the concerns on vendor-defined
>> > > > extensions.
>> > > >
>> > > > Also who do you refer to when you say "how _you_ interact"?  If it is
>> > > > RVI that you refer to: it doesn't interact with upstream work
>> > > > directly, as it doesn't own any engineering resources.
>> > > > RVI provides a forum for member companies to come to an
>> > > > understanding/design and then have the member companies perform the
>> > > > work and take it upstream.
>> > >
>> > > I'm not even sure what you're looking for here: if RVI doesn't want to
>> > > work upstream, then complaining that RVI isn't part of upstream
>> > > discussions is pretty pointless.
>> > >
>> > > >> Why didn't RVI get involved in the review of the series? The expectation
>> > > >> cannot be that all open source projects go to RVI, but rather the other
>> > > >> way around.
>> > > >
>> > > > That is exactly the point I was making and which you seem to miss: RVI
>> > > > does not own any engineering resources and depends solely on its
>> > > > member companies to project into open source projects.
>> > > >
>> > > >> Take a look at commit ea3de9ce8aa2 ("RISC-V: Add a syscall for HW
>> > > >> probing"). Your team was very much involved in the review.
>> > > >
>> > > > I am aware, as I had reviewed and commented on these are well.
>> > > > And my only request (was and) is that we need to figure out a way to
>> > > > efficiently deal with vendor-defined extensions.
>> > >
>> > > Maybe you should go talk to you team, then?  Handling vendor extensions
>> > > via hwprobe has been discussed, sounds like you're confused again.
>> >
>> > I too have this vague memory of us talking about vendor extensions,
>> > but my memory is really bad for stuff like this, so I spent the morning
>> > combing through all the hwprobe iterations looking for it, but so far
>> > have only found
>> >
>> > https://lore.kernel.org/lkml/CALs-HstoeoTWjTEZrLWouCgwq0t3tDB6uL=tB68RJDs1ub4Frw@mail.gmail.com/
>> >
>> > I'm most likely just blind, but does someone have another pointer?
>>
>> Hello! That's probably the only pointer.
>
>
> Thanks for following up, as we were debating internally if and what discussions we had missed.
>
>>
>> Couldn't handling vendor extensions within the hwprobe framework be as
>> straightforward as explicitly carving out a region for them? Say
>> 0x8000000000000000+ belongs to the vendor? The layout of keys within
>> the vendor hwprobe region then depends on the value in mvendorid (and
>> if the vendor so chooses, archid and impid as well). Then vendors can
>> expose keys to their hearts (avoiding the dumb pun there) content.
>>
>> We can probably skip caching the vendor keys in the vDSO for now. If
>> it really needs to be done we can add it later.
>>
>> This would enforce that there's only one "vendor" at a time for a
>> given hart, but I think that's reasonable. Let me know what you think.
>
>
> We generally try to treat vendor-extensions as "vendor-defined" and not "vendor-specific".  In other words, an implementor would be free to implement XVentanaCondOp and XTHeadVDot.  While we could simply alias things into the implementor's "vendor"-space, I see some benefits to having a unique id for every vendor-defined property…
>
> Could we use ( 0x8000000000000000 | vendor-id << VENDOR_SHIFT | key-in-vendor-space )?
> If so, do we have vendor-ids allocated today that we could use for this purpose?

I can sort of see why you'd make that choice architecturally, and the
translation of that idea to this implementation is reasonable. But I
think there are too many dragons there for not enough benefit. Some
thoughts below:

 * I'm a little skeptical of the idea that vendors will actually
implement each other's extensions. Have there been a lot of instances
of this so far we can point to? However, even if they did, I think the
chances that they'll do it identically, such that software can simply
look at the feature bit without also considering the vendor, are
virtually nil. So doing all of these acrobatics to allow usermode to
query a single bit are pointless if the implementations come out
different (or even might come out different) and a (correctly)
defensive usermode filters it by mvendorid anyway.

 * Outside of parsing the ISA string, there's no sane implementation
for a vendor to populate/probe their vendor-specific feature set on
another vendor's hardware. So all the implementations are going to
start with if (mvendorid != mine) return -1, which again makes
exposing all vendors keys on all chips a bit silly.

 * If we did have that rare vendor extension that the vendors all
loved and implemented perfectly faithfully, it would be easy enough to
lower it into the general hwprobe keyspace. I think this will be rare
enough that we won't be tripping over our own merge conflicts.

 * As others have documented, there are real downsides to slicing up
the keyspace this finely: we'd be imposing lower ceilings on the
number of vendors and the number of keys per vendor, and we now have
to become the arbiter of a vendor ID to enum mapping. And since this
is ABI, once that keyspace is allocated to some vendor, it's gone
forever.

One thing we could do if we're really concerned we're making the wrong
choice is to make a smaller slice for the (single) vendor space, like
size 0x1000000. Giving away smaller keyspace acreage now for vendor
extensions gives us the flexibility to either expand it later if it's
working-but-too-small (since everything up there is currently
reserved), or abandon that chunk of keys and do the other thing with
the remaining reserved area.

-Evan

>
> Thanks,
> Philipp.
>
>>
>>
>> -Evan