On Fri, 2022-02-25 at 16:40 +0000, Sean Christopherson wrote: > On Fri, Feb 25, 2022, David Woodhouse wrote: > > On Fri, 2022-02-25 at 01:39 +0000, Sean Christopherson wrote: > > > @@ -11160,7 +11162,7 @@ int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu) > > > vcpu->arch.msr_platform_info = MSR_PLATFORM_INFO_CPUID_FAULT; > > > kvm_vcpu_mtrr_init(vcpu); > > > vcpu_load(vcpu); > > > - kvm_set_tsc_khz(vcpu, max_tsc_khz); > > > + kvm_set_tsc_khz(vcpu, max_tsc_khz ? : tsc_khz); > > > kvm_vcpu_reset(vcpu, false); > > > kvm_init_mmu(vcpu); > > > vcpu_put(vcpu); > > > > > > > Hm, now if you hit that race you end up potentially giving *different* > > frequencies to different vCPUs in a single guest, depending on when > > they were created. > > Yep. Though the race is much harder to hit (userspace vs TSC refinement). The > existing race being hit is essentially do_initcalls() vs. TSC refinement. > > > How about this... (and as noted, I think I want to add an explicit KVM > > ioctl to set kvm->arch.default_tsc_khz for subsequently created vCPUs). > > This wouldn't necessarily help. E.g. assuming userspace knows the actual TSC > frequency, creating a vCPU before refinement completes might put the vCPU in > "always catchup" purgatory. Right. But at least they'd be *consistent*. I was actually making that change anyway, for the benefit of VMs where we are intentionally scaling to a known, different, TSC frequency — which is currently completely hosed when all the vCPUs set it for themselves because the TSC sync then fails. > To really fix the race, KVM needs a notification that refinement completed (or > failed). KVM could simply refuse to create vCPUs until it got the notification. > In the non-constant case, KVM would also need to refresh max_tsc_khz. Hm, would the world be a better place if we knew that the delta between the unrefined and refined TSC values was always within the tolerance of tsc_tolerance_ppm for which we wouldn't bother scaling anyway?
Attachment:
smime.p7s
Description: S/MIME cryptographic signature
