Re: [PATCH] KVM: arm64: Expose PMMIR_EL1.SLOTS to guests

[Oliver Upton <oupton@kernel.org>]

On Mon, Jun 01, 2026 at 01:06:17PM -0700, Oliver Upton wrote:
> Hi Congkai,
> 
> On Mon, Jun 01, 2026 at 07:39:54PM +0000, Congkai Tan wrote:
> > Commit 46081078feb4 ("KVM: arm64: Upgrade PMU support to ARMv8.4")
> > trapped PMMIR_EL1 as RAZ/WI and masked STALL_SLOT* from PMCEID1 to
> > discourage guests from relying on a register they could not read.
> > 
> > Forward the SLOTS field of PMMIR_EL1 so that the perf userspace tool
> > can read the correct value from
> > /sys/bus/event_source/devices/armv8_pmuv3_0/caps/slots. Today, perf
> > stat fails with message "Failure to read '#slots'" because it can
> > only read 0x0 from the sysfs location, causing the parsing failure
> > of the default metrics.
> > 
> > Fix this by:
> > 1. Adding an access_pmmir() handler that reads arm_pmu->reg_pmmir
> >    and returns a masked value containing only the SLOTS field [7:0]
> >    to the guest. Other PMMIR_EL1 fields are kept as RAZ to limit
> >    the extra information that this change exposes; individual
> >    fields can be unmasked as KVM gains support for each feature.
> > 2. Removing the STALL_SLOT, STALL_SLOT_FRONTEND and STALL_SLOT_BACKEND
> >    mask in PMCEID1. The mask existed to hide these events under the
> >    sysfs events/ directory when PMMIR_EL1 was RAZ; with SLOTS now
> >    readable they should be correctly exposed.
> > 
> > Tested on Graviton 2 (Neoverse N1, pre-PMUv3p4), Graviton 3
> > (Neoverse V1) and Graviton 4 (Neoverse V2) metal hosts with QEMU:
> > caps/slots reads 0x00000008 in guests on Graviton 3/4 and 0x00000000
> > on Graviton 2 (correct for pre-PMUv3p4). perf stat correctly
> > evaluates the default metrics.
> > 
> > Fixes: 46081078feb4 ("KVM: arm64: Upgrade PMU support to ARMv8.4")
> > Cc: stable@vger.kernel.org
> 
> This is not stable-worthy, nor is it a bugfix. As KVM presently does not
> implement the STALL_SLOTS* events, PMMIR_EL0.SLOTS = 0 is a legal
> implementation of FEAT_PMUv3p4.
> 
> > ---
> >  arch/arm64/kvm/pmu-emul.c | 11 +----------
> >  arch/arm64/kvm/sys_regs.c | 26 ++++++++++++++++++++++++--
> >  2 files changed, 25 insertions(+), 12 deletions(-)
> > 
> > diff --git a/arch/arm64/kvm/pmu-emul.c b/arch/arm64/kvm/pmu-emul.c
> > index e1860acae641..bafd5a258927 100644
> > --- a/arch/arm64/kvm/pmu-emul.c
> > +++ b/arch/arm64/kvm/pmu-emul.c
> > @@ -864,16 +864,7 @@ static u64 compute_pmceid0(struct arm_pmu *pmu)
> >  
> >  static u64 compute_pmceid1(struct arm_pmu *pmu)
> >  {
> > -	u64 val = __compute_pmceid(pmu, 1);
> > -
> > -	/*
> > -	 * Don't advertise STALL_SLOT*, as PMMIR_EL0 is handled
> > -	 * as RAZ
> > -	 */
> > -	val &= ~(BIT_ULL(ARMV8_PMUV3_PERFCTR_STALL_SLOT - 32) |
> > -		 BIT_ULL(ARMV8_PMUV3_PERFCTR_STALL_SLOT_FRONTEND - 32) |
> > -		 BIT_ULL(ARMV8_PMUV3_PERFCTR_STALL_SLOT_BACKEND - 32));
> > -	return val;
> > +	return __compute_pmceid(pmu, 1);
> >  }
> 
> Make the change to PMCEID1 in a separate patch from PMMIR_EL1.
> 
> > +/*
> > + * Expose only PMMIR_EL1.SLOTS to the guest, which is consumed by perf in its
> > + * topdown default metric group. Other PMMIR_EL1 fields remain RAZ. Future
> > + * patches can extend the exposed mask incrementally as KVM gains support for
> > + * those features.
> > + */
> > +static bool access_pmmir(struct kvm_vcpu *vcpu, struct sys_reg_params *p,
> > +			 const struct sys_reg_desc *r)
> > +{
> > +	struct arm_pmu *cpu_pmu = vcpu->kvm->arch.arm_pmu;
> > +
> > +	if (p->is_write)
> > +		return write_to_read_only(vcpu, p, r);
> > +
> > +	if (check_pmu_access_disabled(vcpu, 0))
> > +		return false;

Also, we don't need this check. PMMIR_EL1 is UNDEF at EL0 so we should
never take a trap from userspace. Same goes for access_pminten() which
is where this probably came from.

> > +	p->regval = cpu_pmu->reg_pmmir & ARMV8_PMU_SLOTS;
> > +	return true;
> > +}
> > +
> 
> We can't change the value of PMMIR_EL1 unconditionally since older KVM
> treated this register as RAZ/WI. This also mixes poorly with the default
> PMU garbage that we have since as the value of the register can change
> based on where KVM_ARM_VCPU_INIT gets called...
> 
> Considering everything, I'd like to see this wired up where:
> 
>  - PMMIR_EL1.SLOTS takes the value of the underlying hardware PMU only
>    if the VMM explicitly selects a particular PMU implementation
> 
>  - KVM allows userspace to set PMMIR_EL1.SLOTS=0 for backwards
>    compatibility
> 
> Thanks,
> Oliver
>

Thanks,
Oliver