Re: [PATCH 1/2] KVM: x86 emulator: don't update vcpu state if instruction is restarted.

[Gleb Natapov <gleb@redhat.com>]

On Sun, Aug 01, 2010 at 01:00:11PM +0300, Avi Kivity wrote:
>  On 08/01/2010 12:24 PM, Gleb Natapov wrote:
> >On Sun, Aug 01, 2010 at 12:14:40PM +0300, Avi Kivity wrote:
> >>  On 08/01/2010 12:01 PM, Gleb Natapov wrote:
> >>>>>It uses rip _during_ pio. And pio emulation changes rip
> >>>>>only at the end of emulation.
> >>>>But non-emulated pio does a skip_emulated_instruction() immediately
> >>>>(or so the code in kvm-tpr-opt.c assumes:
> >>>>
> >>>Indeed, this is bug in non-emulated pio.
> >>But userspace depends on this bug.
> >We can fix that, or make it smarter. Look for io instruction at
> >rip/rip-2 and use rip accordingly for instance.
> 
> That requires everyone to update, or suffer major breakage.
> 
They will suffer major breakage when they update to a kvm that calls to
kvm-tpr-opt.c from emulator anyway.

> >>>But the patch does not change
> >>>rip behaviour for emulated pio. vcpu->arch.emulate_ctxt.eip is updated
> >>>only at the end of emulation.
> >>That will lead to failures if the emulator is used for the
> >>kvm-tpr-opt pio (which may happen with big real mode).
> >>
> >IIRC it was always this way in emulator. I'd rather fix userspace than
> >break emulator.
> 
> It wasn't a problem because the emulator wasn't (and still isn't)
> used for this.  But it has the potential to break badly once we make
> emulate_invalid_guest_state=1 the default.
> 
So what can we do about it?

--
			Gleb.