On Wed, Apr 22, 2020 at 02:06:49PM -0700, Sean Christopherson wrote:
> On Mon, Apr 13, 2020 at 05:09:45PM -0700, Jim Mattson wrote:
> > Fixes: f4124500c2c13 ("KVM: nVMX: Fully emulate preemption timer")
> > Signed-off-by: Jim Mattson <jmattson@xxxxxxxxxx>
> > Reviewed-by: Oliver Upton <oupton@xxxxxxxxxx>
> > Reviewed-by: Peter Shier <pshier@xxxxxxxxxx>
>
> ...
>
> > diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> > index 83050977490c..aae01253bfba 100644
> > --- a/arch/x86/kvm/vmx/vmx.c
> > +++ b/arch/x86/kvm/vmx/vmx.c
> > @@ -4682,7 +4682,7 @@ static int handle_exception_nmi(struct kvm_vcpu *vcpu)
> > if (is_icebp(intr_info))
> > WARN_ON(!skip_emulated_instruction(vcpu));
> >
> > - kvm_queue_exception(vcpu, DB_VECTOR);
> > + kvm_requeue_exception(vcpu, DB_VECTOR);
>
> This isn't wrong per se, but it's effectively papering over an underlying
> bug, e.g. the same missed preemption timer bug can manifest if the timer
> expires while in KVM context (because the hr timer is left running) and KVM
> queues an exception for _any_ reason.
I just reread your changelog and realized this patch was intended to fix a
different symptom than what I observed, i.e. the above probably doesn't
make a whole lot of sense. I just so happened that this change also
resolved my "missing timer" bug because directly injecting the #DB would
cause vmx_check_nested_events() to return -EBUSY on the preemption timer.
That being said, I'm 99% certain that the behavior you observed is fixed by
correctly handling priority of non-exiting events vs. exiting events, i.e.
slightly different justification, same net result.
