It recently dawned on me that the way we handle PMU traps when the PMU is disabled is plain wrong. We consider that handling the registers as RAZ/WI is a fine thing to do, while the ARMv8 ARM is pretty clear that that's not OK and that such registers should UNDEF when FEAT_PMUv3 doesn't exist. I went all the way back to the first public version of the spec, and it turns out we were *always* wrong. It probably comes from the fact that we used not to trap the ID registers, and thus were unable to advertise the lack of PMU, but that's hardly an excuse. So let's fix the damned thing. This series adds an extra check in the helpers that check for the validity of the PMU access (most of the registers have to checked against some enable flags and/or the accessing exception level), and rids us of the RAZ/WI behaviour. This enables us to make additional cleanups, to the point where we can remove the PMU "ready" state that always had very bizarre semantics. All in all, a negative diffstat, and spec compliant behaviours. What's not to like? I've run a few guests with and without PMUs as well as KUT, and nothing caught fire. The patches are on top of kvmarm/queue. Marc Zyngier (8): KVM: arm64: Add kvm_vcpu_has_pmu() helper KVM: arm64: Set ID_AA64DFR0_EL1.PMUVer to 0 when no PMU support KVM: arm64: Refuse illegal KVM_ARM_VCPU_PMU_V3 at reset time KVM: arm64: Inject UNDEF on PMU access when no PMU configured KVM: arm64: Remove PMU RAZ/WI handling KVM: arm64: Remove dead PMU sysreg decoding code KVM: arm64: Gate kvm_pmu_update_state() on the PMU feature KVM: arm64: Get rid of the PMU ready state arch/arm64/include/asm/kvm_host.h | 3 ++ arch/arm64/kvm/pmu-emul.c | 11 +++---- arch/arm64/kvm/reset.c | 4 +++ arch/arm64/kvm/sys_regs.c | 51 ++++++++----------------------- include/kvm/arm_pmu.h | 3 -- 5 files changed, 24 insertions(+), 48 deletions(-) -- 2.28.0
