On Fri, Sep 27, 2019 at 9:05 AM Xiaoyao Li <xiaoyao.li@xxxxxxxxx> wrote: > > On Fri, 2019-09-27 at 17:46 +0200, Vitaly Kuznetsov wrote: > > > > > > > Is this something known already or should I investigate? > > > > > > > > > > > > No, I didn't know about it, it works here. > > > > > > > > > > > > > > > > Ok, this is a bit weird :-) '194' is 'MSR_ARCH_PERFMON_EVENTSEL0 + > > > > > 14'. In intel_pmu_refresh() nr_arch_gp_counters is set to '8', however, > > > > > rdmsr_safe() for this MSR passes in kvm_init_msr_list() (but it fails > > > > > for 0x18e..0x193!) so it stay in the list. get_gp_pmc(), however, checks > > > > > it against nr_arch_gp_counters and returns a failure. > > > > > > > > Huh, 194h apparently is a "FLEX_RATIO" MSR. I agree that PMU MSRs need > > > > to be checked against CPUID before allowing them. > > > > > > My vote would be to programmatically generate the MSRs using CPUID and the > > > base MSR, as opposed to dumping them into the list and cross-referencing > > > them against CPUID. E.g. there should also be some form of check that the > > > architectural PMUs are even supported. > > > > Yes. The problem appears to be that msrs_to_save[] and emulated_msrs[] > > are global and for the MSRs in question we check > > kvm_find_cpuid_entry(vcpu, 0xa, ) to find out how many of them are > > available so this can be different for different VMs (and even vCPUs :-) > > However, > > > > "KVM_GET_MSR_INDEX_LIST returns the guest msrs that are supported. The list > > varies by kvm version and host processor, but does not change otherwise." > > > > Indeed, "KVM_GET_MSR_INDEX_LIST" returns the guest msrs that KVM supports and > they are free from different guest configuration since they're initialized when > kvm module is loaded. > > Even though some MSRs are not exposed to guest by clear their related cpuid > bits, they are still saved/restored by QEMU in the same fashion. > > I wonder should we change "KVM_GET_MSR_INDEX_LIST" per VM? Yes! Quoting Paolo from a few days ago, "If there's a complex or really weird behavior that userspace would most definitely get wrong, we should design the API to simplify its job." > > So it seems that PMU MSRs just can't be there. Revert? > > >
