[RFC PATCH v6 03/35] KVM: arm64: Add CONFIG_KVM_ARM_SPE Kconfig option

[Alexandru Elisei]

Hi James,

On Mon, Jan 12, 2026 at 03:18:53PM +0000, Alexandru Elisei wrote:
> Hi James,
> 
> On Mon, Jan 12, 2026 at 12:09:19PM +0000, James Clark wrote:
> > 
> > 
> > On 12/01/2026 11:26 am, Alexandru Elisei wrote:
> > > Hi James,
> > > 
> > > On Fri, Jan 09, 2026 at 04:29:50PM +0000, James Clark wrote:
> > > > 
> > > > 
> > > > On 14/11/2025 4:06 pm, Alexandru Elisei wrote:
> > > > > Add a new configuration option that will be used for KVM SPE emulation.
> > > > > CONFIG_KVM_ARM_SPE depends on the SPE driver being builtin because:
> > > > > 
> > > > > 1. The SPE driver maintains a cpumask of physical CPUs that support SPE,
> > > > >      and that will be used by KVM to emulate SPE on heterogeneous systems.
> > > > > 
> > > > > 2. KVM will rely on the SPE driver enabling the SPE interrupt at the GIC
> > > > >      level.
> > > > > 
> > > > > Signed-off-by: Alexandru Elisei <alexandru.elisei at arm.com>
> > > > > ---
> > > > >    arch/arm64/kvm/Kconfig | 8 ++++++++
> > > > >    1 file changed, 8 insertions(+)
> > > > > 
> > > > > diff --git a/arch/arm64/kvm/Kconfig b/arch/arm64/kvm/Kconfig
> > > > > index 4f803fd1c99a..31388b5b2655 100644
> > > > > --- a/arch/arm64/kvm/Kconfig
> > > > > +++ b/arch/arm64/kvm/Kconfig
> > > > > @@ -83,4 +83,12 @@ config PTDUMP_STAGE2_DEBUGFS
> > > > >    	  If in doubt, say N.
> > > > > +config KVM_ARM_SPE
> > > > > +	bool
> > > > > +	depends on KVM && ARM_SPE_PMU=y
> > > > 
> > > > I think the most common configuration is module, so requiring built-in isn't
> > > > great. If there's any way of avoiding it, even if it costs a little bit of
> > > > pain, it would be good for adoption.
> > > 
> > > I'm not sure how that could be done. You need the buffer maintenance interrupt
> > > to do a VM exit so KVM can inject the virtual interrupt back into the guest,
> > > otherwise there's a possibility of a large blackout window when the buffer is
> > > full.
> > 
> > Without expanding more I don't see how injecting an interrupt is strongly
> > related to the way it's compiled.
> 
> It's the SPE driver that enables the hardware interrupt in the GIC.
> 
> > 
> > > 
> > > The code also relies on the SPE driver to probe all the SPUs in the system.
> > > 
> > 
> > I don't think that's a hard requirement either, the KVM code could fairly
> > easily do this itself. Inline the common code and put it in a shared header
> > etc. I'm sure there are lots of creative ways to flip the dependencies that
> > might not be "proper" software engineering. But like I said a bit of pain
> > here might be beneficial overall.
> 
> Userspace gets the maximum buffer size and the SPE instance id from the sysfs
> files.
> 
> So KVM would have to parse the DTB, read PMBIDR_EL1 and PMSIDR_EL1 on each
> physical CPU in the system, create the sysfs files (at least 'type', 'cpumask'
> and 'max_buffer_size' or whatever that ends up being named) and install a dummy
> IRQ handler before the SPE driver is loaded.
> 
> Then when the SPE driver is loaded, the driver would have to remove the IRQ
> handler and install its own, and create the rest of the sysfs files.
> 
> That's assuming that the SPE driver is loaded after KVM. I think it's possible
> for the SPE driver to be loaded first. So both KVM and SPE would have be able to
> communicate with one other in case the other probed first.
> 
> Is having to change the symbol ARM_SPE_PMU from 'm' to 'y' worth this? I am not
> convinced and I don't see much use for going down this rabbit hole at this stage
> of the series.
> 
> Also, since KVM would partially populate the sysfs directory for each SPE
> instance, would it be reasonable for software to assume that the SPE driver is
> loaded based on this and start using it? For example, how does perf detect the
> presence of the arm_spe event?

Coming back to this, I do understand and agree with your concerns, so after
talking in private I came up with exporting the symbol kvm_host_spe_init() -
that's the function which adds the SPE instance to the list of SPUs.

With this change the SPE driver can be built as a module, but SPE for a VM is
available only as long as the driver is loaded.

I hope that improves the adoption story :)

Thanks,
Alex