Re: [PATCH 2/4] kvm: svm: Enable UMIP feature on AMD

[Jim Mattson]

On Fri, Nov 1, 2019 at 1:04 PM Moger, Babu <Babu.Moger@xxxxxxx> wrote:
>
>
>
> On 11/1/19 2:24 PM, Andy Lutomirski wrote:
> > On Fri, Nov 1, 2019 at 12:20 PM Moger, Babu <Babu.Moger@xxxxxxx> wrote:
> >>
> >>
> >>
> >> On 11/1/19 1:29 PM, Jim Mattson wrote:
> >>> On Fri, Nov 1, 2019 at 10:33 AM Moger, Babu <Babu.Moger@xxxxxxx> wrote:
> >>>>
> >>>> AMD 2nd generation EPYC processors support UMIP (User-Mode Instruction
> >>>> Prevention) feature. The UMIP feature prevents the execution of certain
> >>>> instructions if the Current Privilege Level (CPL) is greater than 0.
> >>>> If any of these instructions are executed with CPL > 0 and UMIP
> >>>> is enabled, then kernel reports a #GP exception.
> >>>>
> >>>> The idea is taken from articles:
> >>>> https://lwn.net/Articles/738209/
> >>>> https://lwn.net/Articles/694385/
> >>>>
> >>>> Enable the feature if supported on bare metal and emulate instructions
> >>>> to return dummy values for certain cases.
> >>>>
> >>>> Signed-off-by: Babu Moger <babu.moger@xxxxxxx>
> >>>> ---
> >>>>  arch/x86/kvm/svm.c |   21 ++++++++++++++++-----
> >>>>  1 file changed, 16 insertions(+), 5 deletions(-)
> >>>>
> >>>> diff --git a/arch/x86/kvm/svm.c b/arch/x86/kvm/svm.c
> >>>> index 4153ca8cddb7..79abbdeca148 100644
> >>>> --- a/arch/x86/kvm/svm.c
> >>>> +++ b/arch/x86/kvm/svm.c
> >>>> @@ -2533,6 +2533,11 @@ static void svm_decache_cr4_guest_bits(struct kvm_vcpu *vcpu)
> >>>>  {
> >>>>  }
> >>>>
> >>>> +static bool svm_umip_emulated(void)
> >>>> +{
> >>>> +       return boot_cpu_has(X86_FEATURE_UMIP);
> >>>> +}
> >>>
> >>> This makes no sense to me. If the hardware actually supports UMIP,
> >>> then it doesn't have to be emulated.
> >> My understanding..
> >>
> >> If the hardware supports the UMIP, it will generate the #GP fault when
> >> these instructions are executed at CPL > 0. Purpose of the emulation is to
> >> trap the GP and return a dummy value. Seems like this required in certain
> >> legacy OSes running in protected and virtual-8086 modes. In long mode no
> >> need to emulate. Here is the bit explanation https://lwn.net/Articles/738209/
> >>
> >
> > Indeed.  Again, what does this have to do with your patch?
> >
> >>
> >>>
> >>> To the extent that kvm emulates UMIP on Intel CPUs without hardware
> >>> UMIP (i.e. smsw is still allowed at CPL>0), we can always do the same
> >>> emulation on AMD, because SVM has always offered intercepts of sgdt,
> >>> sidt, sldt, and str. So, if you really want to offer this emulation on
> >>> pre-EPYC 2 CPUs, this function should just return true. But, I have to
> >>> ask, "why?"
> >>
> >>
> >> Trying to support UMIP feature only on EPYC 2 hardware. No intention to
> >> support pre-EPYC 2.
> >>
> >
> > I think you need to totally rewrite your changelog to explain what you
> > are doing.
> >
> > As I understand it, there are a couple of things KVM can do:
> >
> > 1. If the underlying hardware supports UMIP, KVM can expose UMIP to
> > the guest.  SEV should be irrelevant here.
> >
> > 2. Regardless of whether the underlying hardware supports UMIP, KVM
> > can try to emulate UMIP in the guest.  This may be impossible if SEV
> > is enabled.
> >
> > Which of these are you doing?
> >
> My intention was to do 1.  Let me go back and think about this again.

(1) already works.