Re: [RFC PATCH 25/35] KVM: x86: Update __get_sregs() / __set_sregs() to support SEV-ES

[Sean Christopherson]

On Tue, Sep 15, 2020 at 03:37:21PM -0500, Tom Lendacky wrote:
> On 9/15/20 11:33 AM, Sean Christopherson wrote:
> > On Tue, Sep 15, 2020 at 09:19:46AM -0500, Tom Lendacky wrote:
> >> On 9/14/20 4:37 PM, Sean Christopherson wrote:
> >>> On Mon, Sep 14, 2020 at 03:15:39PM -0500, Tom Lendacky wrote:
> >>>> From: Tom Lendacky <thomas.lendacky@xxxxxxx>
> >>>>
> >>>> Since many of the registers used by the SEV-ES are encrypted and cannot
> >>>> be read or written, adjust the __get_sregs() / __set_sregs() to only get
> >>>> or set the registers being tracked (efer, cr0, cr4 and cr8) once the VMSA
> >>>> is encrypted.
> >>>
> >>> Is there an actual use case for writing said registers after the VMSA is
> >>> encrypted?  Assuming there's a separate "debug mode" and live migration has
> >>> special logic, can KVM simply reject the ioctl() if guest state is protected?
> >>
> >> Yeah, I originally had it that way but one of the folks looking at live
> >> migration for SEV-ES thought it would be easier given the way Qemu does
> >> things. But I think it's easy enough to batch the tracking registers into
> >> the VMSA state that is being transferred during live migration. Let me
> >> check that out and likely the SET ioctl() could just skip all the regs.
> > 
> > Hmm, that would be ideal.  How are the tracked registers validated when they're
> > loaded at the destination?  It seems odd/dangerous that KVM would have full
> > control over efer/cr0/cr4/cr8.  I.e. why is KVM even responsibile for migrating
> > that information, e.g. as opposed to migrating an opaque blob that contains
> > encrypted versions of those registers?
> > 
> 
> KVM doesn't have control of them. They are part of the guest's encrypted
> state and that is what the guest uses. KVM can't alter the value that the
> guest is using for them once the VMSA is encrypted. However, KVM makes
> some decisions based on the values it thinks it knows.  For example, early
> on I remember the async PF support failing because the CR0 that KVM
> thought the guest had didn't have the PE bit set, even though the guest
> was in protected mode. So KVM didn't include the error code in the
> exception it injected (is_protmode() was false) and things failed. Without
> syncing these values after live migration, things also fail (probably for
> the same reason). So the idea is to just keep KVM apprised of the values
> that the guest has.

Ah, gotcha.  Migrating tracked state through the VMSA would probably be ideal.
The semantics of __set_sregs() kinda setting state but not reaaaally setting
state would be weird.