Message-ID: <mhng-d64da1be-bacd-4885-aaf2-fea3c763418c@palmerdabbelt-glaptop>
Date:   Fri, 09 Apr 2021 11:58:19 -0700 (PDT)
From:   Palmer Dabbelt <palmer@...belt.com>
To:     pbonzini@...hat.com
CC:     anup@...infault.org, Anup Patel <Anup.Patel@....com>,
        Paul Walmsley <paul.walmsley@...ive.com>,
        aou@...s.berkeley.edu, graf@...zon.com,
        Atish Patra <Atish.Patra@....com>,
        Alistair Francis <Alistair.Francis@....com>,
        Damien Le Moal <Damien.LeMoal@....com>, kvm@...r.kernel.org,
        kvm-riscv@...ts.infradead.org, linux-riscv@...ts.infradead.org,
        linux-kernel@...r.kernel.org
Subject:     Re: [PATCH v16 00/17] KVM RISC-V Support

On Wed, 31 Mar 2021 02:21:58 PDT (-0700), pbonzini@...hat.com wrote:
> On 30/03/21 07:48, Anup Patel wrote:
>>
>> It seems Andrew does not want to freeze H-extension until we have virtualization
>> aware interrupt controller (such as RISC-V AIA specification) and IOMMU. Lot
>> of us feel that these things can be done independently because RISC-V
>> H-extension already has provisions for external interrupt controller with
>> virtualization support.

Sorry to hear that.  It's really gotten to a point where I'm just 
embarrassed with how the RISC-V foundation is being run -- not sure if 
these other ones bled into Linux land, but this is the third ISA 
extension that's blown up over the last few weeks.  We had a lot of 
discussion about this on the binutils/GCC side of things and I've 
managed to convince myself that coupling the software stack to the 
specification process isn't viable -- we made that decision under the 
assumption that specifications would actually progress through the 
process, but in practice that's just not happening.

My goal with the RISC-V stuff has always been getting us to a place 
where we have real shipping products running a software stack that is as 
close as possible to the upstream codebases.  I see that as the only way 
to get the software stack to a point where it can be sustainably 
maintained.  The "only frozen extensions" policy was meant to help this 
by steering vendors towards a common base we could support, but in 
practice it's just not working out.  The specification process is just 
so unreliable that in practice everything that gets built ends up 
relying on some non-standard behavior: whether it's a draft extension, 
some vendor-specific extension, or just some implementation quirks.  
There's always going to be some degree of that going on, but over the 
last year or so we've just stopped progressing.

My worry with accepting the draft extensions is that we have no 
guarantee of compatibility between various drafts, which makes 
supporting multiple versions much more difficult.  I've always really 
only been worried about supporting what gets implemented in a chip I can 
actually run code on, as I can at least guarantee that doesn't change.  
In practice that really has nothing to do with the specification freeze: 
even ratified specifications change in ways that break compatibility so 
we need to support multiple versions anyway.  That's why we've got 
things like the K210 support (which doesn't quite follow the ratified 
specs) and are going to take the errata stuff.  I hadn't been all that 
worried about the H support because there was a plan to get is to 
hardware, but with the change I'm not really sure how that's going to 
happen.

> Yes, frankly that's pretty ridiculous as it's perfectly possible to
> emulate the interrupt controller in software (and an IOMMU is not needed
> at all if you are okay with emulated or paravirtualized devices---which
> is almost always the case except for partitioning hypervisors).

There's certainly some risk to freezing the H extension before we have 
all flavors of systems up and running.  I spent a lot of time arguing 
that case years ago before we started telling people that the H 
extension just needed implementation, but that's not the decision we 
made.  I don't really do RISC-V foundation stuff any more so I don't 
know why this changed, but it's just too late.  It would be wonderful to 
have an implementation of everything we need to build out one of these 
complex systems, but I just just don't see how the current plan gets 
there: that's a huge amount of work and I don't see why anyone would 
commit to that when they can't count on it being supported when it's 
released.

There are clearly some systems that can be built with this as it stands.  
They're not going to satisfy every use case, but at least we'll get 
people to start seriously using the spec.  That's the only way I can see 
to move forward with this.  It's pretty clear that sitting around and 
waiting doesn't work, we've tried that.

> Palmer, are you okay with merging RISC-V KVM?  Or should we place it in
> drivers/staging/riscv/kvm?

I'm certainly ready to drop my objections to merging the code based on 
it targeting a draft extension, but at a bare minimum I want to get a 
new policy in place that everyone can agree to for merging code.  I've 
tried to draft up a new policy a handful of times this week, but I'm not 
really quite sure how to go about this: ultimately trying to build 
stable interfaces around an unstable ISA is just a losing battle.  I've 
got a bunch of stuff going on right now, but I'll try to find some time 
to actually sit down and finish one.

I know it might seem odd to complain about how slowly things are going 
and then throw up another roadblock, but I really do think this is a 
very important thing to get right.  I'm just not sure how we're going to 
get anywhere with RISC-V without someone providing stability, so I want 
to make sure that whatever we do here can be done reliably.  If we don't 
I'm worried the vendors are just going to go off and do their own 
software stacks, which will make getting everyone back on the same page 
very difficult.

> Either way, the best way to do it would be like this:
>
> 1) you apply patch 1 in a topic branch
>
> 2) you merge the topic branch in the risc-v tree
>
> 3) Anup merges the topic branch too and sends me a pull request.
>
> Paolo

Powered by blists - more mailing lists