On Thu, Jun 01, 2023, Jim Mattson wrote:
> On Thu, Jun 1, 2023 at 12:28 PM Jon Kohler <jon@xxxxxxxxxxx> wrote:
> > > diff --git a/arch/x86/kvm/x86.h b/arch/x86/kvm/x86.h
> > > index c544602d07a3..454bcbf5b543 100644
> > > --- a/arch/x86/kvm/x86.h
> > > +++ b/arch/x86/kvm/x86.h
> > > @@ -492,7 +492,31 @@ static inline void kvm_machine_check(void)
> > >
> > > void kvm_load_guest_xsave_state(struct kvm_vcpu *vcpu);
> > > void kvm_load_host_xsave_state(struct kvm_vcpu *vcpu);
> > > +
> > > int kvm_spec_ctrl_test_value(u64 value);
> > > +
> > > +static inline bool kvm_account_msr_spec_ctrl_write(struct kvm_vcpu *vcpu)
> > > +{
> > > + if ((vcpu->stat.exits - vcpu->arch.spec_ctrl_nr_exits_snapshot) < 20)
>
> I think you mean 200 here. If it's bad to have more than 1
> WRMSR(IA32_SPEC_CTRL) VM-exit in 20 VM-exits, then more than 10 such
> VM-exits in 200 VM-exits represents sustained badness.
No? The snapshot is updated on every write, i.e. this check is whether or not
the last wrmsr(SPEC_CTRL) was less than 20 cycles ago.
if ((vcpu->stat.exits - vcpu->arch.spec_ctrl_nr_exits_snapshot) < 20)
vcpu->arch.nr_quick_spec_ctrl_writes++;
else
vcpu->arch.nr_quick_spec_ctrl_writes = 0;
vcpu->arch.spec_ctrl_nr_exits_snapshot = vcpu->stat.exits; <= new snapshot
return vcpu->arch.nr_quick_spec_ctrl_writes >= 10;
> (Although, as Sean noted, these numbers are just placeholders.)
And the logic is very off-the-cuff, e.g. it may be better to have a rolling 200-exit
window instead of 10 somewhat independent 20-exit windows.
