On Thu, 2021-05-27 at 15:08 +0200, Paolo Bonzini wrote:

On 27/05/21 10:33, Stamatis, Ilias wrote:

quoted

quoted

  #ifdef CONFIG_X86_64

@@ -10444,6 +10461,7 @@ void kvm_arch_vcpu_postcreate(struct kvm_vcpu *vcpu)
             return;
     vcpu_load(vcpu);
     kvm_synchronize_tsc(vcpu, 0);
+    kvm_vcpu_write_tsc_multiplier(vcpu, kvm_default_tsc_scaling_ratio);

Hmm, I'm actually thinking now that this might not be correct. For example in
case we hotplug a new vCPU but the other vCPUs don't use the default ratio.

It is correct, the TSC frequency can be set per CPU (which is useless
except possibly for debugging OS timekeeping, but still).  So, the
default kHz after hotplug is the host frequency.

It doesn't really matter because it only affects the fixed delta between
the hotplugged CPU and the others as soon as userspace sets the
frequency to the correct value.

Paolo

So this patch is wrong anyway. 

kvm_arch_vcpu_create() does a kvm_set_tsc_khz(vcpu, max_tsc_khz) when
initializing the vcpu. This wouldn't normally result in a VMWRITE, but now
(after applying patch 9) it does. The problem is that this write now happens too
early and it raises an exception. To fix this, that line needs to be moved to
kvm_arch_vcpu_postcreate() (like above) but before calling
kvm_synchronize_tsc(vcpu, 0).

I will re-submit this patch with the fix.

Best,
Ilias