Re: [PATCH 5/8] KVM: SVM: Re-inject INT3/INTO instead of retrying the instruction

[Maciej S. Szmigiero]

On 4.04.2022 21:33, Sean Christopherson wrote:
> On Mon, Apr 04, 2022, Maciej S. Szmigiero wrote:
> > > > > index 47e7427d0395..a770a1c7ddd2 100644
> > > > > --- a/arch/x86/kvm/svm/svm.h
> > > > > +++ b/arch/x86/kvm/svm/svm.h
> > > > > @@ -230,8 +230,8 @@ struct vcpu_svm {
> > > > >    	bool nmi_singlestep;
> > > > >    	u64 nmi_singlestep_guest_rflags;
> > > > > -	unsigned int3_injected;
> > > > > -	unsigned long int3_rip;
> > > > > +	unsigned soft_int_injected;
> > > > > +	unsigned long soft_int_linear_rip;
> > > > >    	/* optional nested SVM features that are enabled for this guest  */
> > > > >    	bool nrips_enabled                : 1;
> > > >
> > > >
> > > > I mostly agree with this patch, but think that it doesn't address the
> > > > original issue that Maciej wanted to address:
> > > >
> > > > Suppose that there is *no* instruction in L2 code which caused the software
> > > > exception, but rather L1 set arbitrary next_rip, and set EVENTINJ to software
> > > > exception with some vector, and that injection got interrupted.
> > > >
> > > > I don't think that this code will support this.
> > >
> > > Argh, you're right.  Maciej's selftest injects without an instruction, but it doesn't
> > > configure the scenario where that injection fails due to an exception+VM-Exit that
> > > isn't intercepted by L1 and is handled by L0.  The event_inj test gets the coverage
> > > for the latter, but always has a backing instruction.
> >
> > Still reviewing the whole patch set, but want to clear this point quickly:
> > The selftest does have an implicit intervening NPF (handled by L0) while
> > injecting the first L1 -> L2 event.
>
> I'll do some debug to figure out why the test passes for me.  I'm guessing I either
> got lucky, e.g. IDT was faulted in already, or I screwed up and the test doesn't
> actually pass.

Try applying Maxim's proposed change to this test (adding two additional ud2s at the
start of L2 guest code and adjusting just the next_rip field accordingly).

We might have been lucky in the patched KVM code and have next_rip = rip at this
re-injection case now (the next_rip field was zero in the original KVM code).

Thanks,
Maciej