Hi Congkai, Thanks for respinning.
On Thu, Jul 02, 2026 at 07:04:18PM +0000, Congkai Tan wrote: > Today when the perf tool runs in a guest on cores with PMUv3p4, it fails > to parse the default metrics with "Failure to read '#slots'", since perf > can only read 0 from sysfs caps/slots, which is backed by PMMIR_EL1.SLOTS > that KVM traps as RAZ/WI. > > Taking into account backward compatibility and heterogeneous systems, the > exposure of PMMIR_EL1.SLOTS is gated behind a new vCPU feature flag: > > - Patch 1 adds the new flag KVM_ARM_VCPU_PMU_V3_STRICT. When set, KVM does > not create a default PMU during vCPU init, and the VMM must select one > explicitly via KVM_ARM_VCPU_PMU_V3_SET_PMU before the first KVM_RUN. > - Patch 2 exposes PMMIR_EL1.SLOTS of the selected PMU under the flag, and > adds userspace get/set for PMMIR_EL1 so that SLOTS can be reset to 0 > for backward compatibility. > - Patch 3 stops masking STALL_SLOT* in PMCEID1 under the flag. I think the series is starting to shape up. Few more things to address: - Didn't mention it in v1, but writes to PMCR_EL0.N should be ignored when V3_STRICT is set. We now have a vCPU attribute for configuring event counters and the register-based thing is just broken :) - The vCPU feature flag needs a corresponding KVM_CAP so userspace can detect it - In terms of patch ordering, the vCPU flag / KVM_CAP exposure should come last after all the behavior changes are implemented (and flag-guarded) - Move enforcement of a non-NULL arm_pmu to kvm_arm_pmu_v3_init() since userspace must call KVM_ARM_VCPU_PMU_V3_INIT before KVM_RUN - Prevent the PMU event filter from being configured until a hardware PMU has been selected I've addressed all of this locally and pushed to my tree [*]. Untested, as always :) Would you be able to give it a spin? Also, do you have VMM patches for using the new feature flag? [*]: https://git.kernel.org/pub/scm/linux/kernel/git/oupton/linux.git/log/?h=kvm-arm64/pmu-7.3 Thanks, Oliver
