On Wed, Apr 20, 2022, Paolo Bonzini wrote:
> On 4/20/22 18:15, Sean Christopherson wrote:
> > > > Let's just require X86_FEATURE_NRIPS, either in general or just to
> > > > enable nested virtualiazation
> > > 👍
> > Hmm, so requiring NRIPS for nested doesn't actually buy us anything. KVM still
> > has to deal with userspace hiding NRIPS from L1, so unless I'm overlooking something,
> > the only change would be:
> >
> > diff --git a/arch/x86/kvm/svm/nested.c b/arch/x86/kvm/svm/nested.c
> > index bdf8375a718b..7bed4e05aaea 100644
> > --- a/arch/x86/kvm/svm/nested.c
> > +++ b/arch/x86/kvm/svm/nested.c
> > @@ -686,7 +686,7 @@ static void nested_vmcb02_prepare_control(struct vcpu_svm *svm,
> > */
> > if (svm->nrips_enabled)
> > vmcb02->control.next_rip = svm->nested.ctl.next_rip;
> > - else if (boot_cpu_has(X86_FEATURE_NRIPS))
> > + else
> > vmcb02->control.next_rip = vmcb12_rip;
> >
> > if (is_evtinj_soft(vmcb02->control.event_inj)) {
> >
> > And sadly, because SVM doesn't provide the instruction length if an exit occurs
> > while vectoring a software interrupt/exception, making NRIPS mandatory doesn't buy
> > us much either.
> >
> > I believe the below diff is the total savings (plus the above nested thing) against
> > this series if NRIPS is mandatory (ignoring the setup code, which is a wash). It
> > does eliminate the rewind in svm_complete_soft_interrupt() and the funky logic in
> > svm_update_soft_interrupt_rip(), but that's it AFAICT. The most obnoxious code of
> > having to unwind EMULTYPE_SKIP when retrieving the next RIP for software int/except
> > injection doesn't go away:-(
> >
> > I'm not totally opposed to requiring NRIPS, but I'm not in favor of it either.
>
> Yeah, you're right. However:
>
> * the rewind might already be worth it;
FWIW, I don't actually care about supporting nrips=false, it's the ability to test
EMULTYPE_SKIP that I find valuable. I also find the extra perspective of how RIP
interacts with software interrupts/exceptions to be very helpful/educational, though
that's of debatable value going forward.
> * if we require NRIPS for nested, we can also assume that the SVM save state
> data has a valid next_rip; even if !svm->nrips_enabled. There's the pesky
> issue of restoring from an old system that did not have NRIPS, but let's
> assume for now that NRIPS was set on the source as well.
How about I throw a Not-signed-off-by patch on the end of the series to make NRIPS
mandatory, that way we (in theory) have a fully functional snapshot for nrips=false
if we want to go back in time. And we probably need to give a deprecation grace
period, i.e. wait a release or two before disappearing nrips?
