[PATCH v3 9/9] RISC-V: KVM: Upgrade the supported SBI version to 3.0

[Andrew Jones]

On Thu, May 29, 2025 at 11:44:38AM -0700, Atish Patra wrote:
> 
> On 5/29/25 3:24 AM, Radim Krčmář wrote:
> > I originally gave up on the idea, but I feel kinda bad for Drew now, so
> > trying again:
> 
> I am sorry if some of my replies came across in the wrong way. That was
> never
> the intention.

Not at all. Radim only meant that I was defending his patches, even though
he wasn't :-)

> 
> 
> > 2025-05-28T12:21:59-07:00, Atish Patra <atish.patra at linux.dev>:
> > > On 5/28/25 8:09 AM, Andrew Jones wrote:
> > > > On Wed, May 28, 2025 at 07:16:11AM -0700, Atish Patra wrote:
> > > > > On 5/26/25 4:13 AM, Andrew Jones wrote:
> > > > > > On Mon, May 26, 2025 at 11:00:30AM +0200, Radim Krčmář wrote:
> > > > > > > 2025-05-23T10:16:11-07:00, Atish Patra <atish.patra at linux.dev>:
> > > > > > > > On 5/23/25 6:31 AM, Radim Krčmář wrote:
> > > > > > > > > 2025-05-22T12:03:43-07:00, Atish Patra <atishp at rivosinc.com>:
> > > > > > > > > > Upgrade the SBI version to v3.0 so that corresponding features
> > > > > > > > > > can be enabled in the guest.
> > > > > > > > > > 
> > > > > > > > > > Signed-off-by: Atish Patra <atishp at rivosinc.com>
> > > > > > > > > > ---
> > > > > > > > > > diff --git a/arch/riscv/include/asm/kvm_vcpu_sbi.h b/arch/riscv/include/asm/kvm_vcpu_sbi.h
> > > > > > > > > > -#define KVM_SBI_VERSION_MAJOR 2
> > > > > > > > > > +#define KVM_SBI_VERSION_MAJOR 3
> > > > > > > > > I think it's time to add versioning to KVM SBI implementation.
> > > > > > > > > Userspace should be able to select the desired SBI version and KVM would
> > > > > > > > > tell the guest that newer features are not supported.
> > > > > > We need new code for this, but it's a good idea.
> > > > > > 
> > > > > > > > We can achieve that through onereg interface by disabling individual SBI
> > > > > > > > extensions.
> > > > > > > > We can extend the existing onereg interface to disable a specific SBI
> > > > > > > > version directly
> > > > > > > > instead of individual ones to save those IOCTL as well.
> > > > > > > Yes, I am all in favor of letting userspace provide all values in the
> > > > > > > BASE extension.
> > > > > We already support vendorid/archid/impid through one reg. I think we just
> > > > > need to add the SBI version support to that so that user space can set it.
> > > > > 
> > > > > > This is covered by your recent patch that provides userspace_sbi.
> > > > > Why do we need to invent new IOCTL for this ? Once the user space sets the
> > > > > SBI version, KVM can enforce it.
> > > > If an SBI spec version provides an extension that can be emulated by
> > > > userspace, then userspace could choose to advertise that spec version,
> > > > implement a BASE probe function that advertises the extension, and
> > > > implement the extension, even if the KVM version running is older
> > > > and unaware of it. But, in order to do that, we need KVM to exit to
> > > > userspace for all unknown SBI calls and to allow BASE to be overridden
> > > You mean only the version field in BASE - Correct ?
> > No, "BASE probe function" is the sbi_probe_extension() ecall.
> > 
> > > > by userspace. The new KVM CAP ioctl allows opting into that new behavior.
> > > But why we need a new IOCTL for that ? We can achieve that with existing
> > > one reg interface with improvements.
> > It's an existing IOCTL with a new data payload, but I can easily use
> > ONE_REG if you want to do everything through that.
> > 
> > KVM doesn't really need any other IOCTL than ONE_REGs, it's just
> > sometimes more reasonable to use a different IOCTL, like ENABLE_CAP.
> > 
> > > > The old KVM with new VMM configuration isn't totally far-fetched. While
> > > > host kernels tend to get updated regularly to include security fixes,
> > > > enterprise kernels tend to stop adding features at some point in order
> > > > to maximize stability. While enterprise VMMs would also eventually stop
> > > > adding features, enterprise consumers are always free to use their own
> > > > VMMs (at their own risk). So, there's a real chance we could have
> > > I think we are years away from that happening (if it happens). My
> > > suggestion was not to
> > > try to build a world where no body lives ;). When we get to that
> > > scenario, the default KVM
> > > shipped will have many extension implemented. So there won't be much
> > > advantage to
> > > reimplement them in the user space. We can also take an informed
> > > decision at that time
> > > if the current selective forwarding approach is better
> > Please don't repeat the design of SUSP/SRST/DBCN.
> > Seeing them is one of the reasons why I proposed the new interface.
> > 
> > "Blindly" forwarding DBCN to userspace is even a minor optimization. :)
> > 
> > >                                                         or we need to
> > > blindly forward any
> > > unknown SBI calls to the user space.
> > Yes, KVM has to do what userpace configures it to do.
> > 
> > I don't think that implementing unsupported SBI extensions in KVM is
> > important -- they should not be a hot path.
> > 
> > > > deployments with older, stable KVM where users want to enable later SBI
> > > > extensions, and, in some cases, that should be possible by just updating
> > > > the VMM -- but only if KVM is only acting as an SBI implementation
> > > > accelerator and not as a userspace SBI implementation gatekeeper.
> > > But some of the SBI extensions are so fundamental that it must be
> > > implemented in KVM
> > > for various reasons pointed by Anup on other thread.
> > No, SBI does not have to be implemented in KVM at all.
> > 
> > We do have a deep disagreement on what is virtualization and the role of
> > KVM in it.  I think that userspace wants a generic ISA accelerator.
> 
> I think the disagreement is the role of SBI in KVM virtualization rather
> than
> a generic virtualization and the role of KVM in it. I completely agree that
> KVM should act as an accelerator and defer the control to the user space in
> most of the cases
> such e.g I/O operations or system related functionalities. However, SBI
> specification solves
> much wider problems than those. Broadly we can categorize SBI
> functionalities into the following
> areas
> 
> 1. Bridging ISA GAP
> 2. Higher Privilege Assistance
> 3. Virtualization
> 4. Platform abstraction
> 5. Confidential computing
> 
> For #1, #3 and #5, I believe user space shouldn't be involved in
> implementation
> some of them are in hot path as well.

IMO, userspace should still be in control of whether or not it's involved
in #1, #3, and #5. It may make little sense for it to be involved, but the
choice should still be its.

> For #4 and #2, there are some
> opportunities which
> can be implemented in user space depending on the exact need. I am still not
> clear what is the exact
> motivation /right now/ to pursue such a path. May be I missed something.
> As per my understanding from our discussion threads, there are two use cases
> possible
> 
> 1. userspace wants to update more states in HSM. What are the states user
> space should care about scounteren (fixed already in usptream) ?
> 2. VMM vs KVM version difference - this may be true in the future depending
> on the speed of RISC-V virtualization adoption in the industry.
> But we are definitely not there yet. Please let me know if I misunderstood
> any use cases.

That's what I'm aware of as well, but I see giving userspace back full
control of what gets accelerated by KVM, and what doesn't, as a fix, which
is why I wouldn't want to delay it any longer.

> 
> > Even if userspace wants SBI for the M-mode interface, security minded
> This is probably a 3rd one ? Why we want M-mode interface in the user space
> ?
> > userspace aims for as little kernel code as possible.
> 
> We trust VMM code more than KVM code ?

We should be skeptical of both, which is why we'd rather put as much code
in userspace as possible. Insecure/faulty userspace will hopefully have
exploits/bugs contained to the single process. An insecure/faulty KVM
means the host is compromised/crashed. On x86, Google put a lot of effort
into moving instruction emulation out of KVM for security concerns[1]. In
general, if it's not a hot path and there's a way to do it in userspace,
then it should be done in userspace (or at least there should be an
option to use userspace -- each use case can choose what's best for
itself).

[1] https://www.linux-kvm.org/images/3/3d/01x02-Steve_Rutherford-Performant_Security_Hardening_of_KVM.pdf

Thanks,
drew

> 
> > Userspace might want to accelerate some SBI extension in KVM, but it
> > should not be KVM who decides what userspace wants.