[0/3] KVM: x86: Revert dirty tracking for GPRs

Message

Sean Christopherson Jan. 22, 2021, 11:50 p.m. UTC

This is effectively belated feedback on the SEV-ES series.  My primary
interest is to revert the GPR dirty/available tracking, as it's pure
overhead for non-SEV-ES VMs, and even for SEV-ES I suspect the dirty
tracking is at best lost in the noise, and possibly even a net negative.

My original plan was to submit patches 1+3 as patch 1, taking a few
creative liberties with the GHCB spec to justify writing the GHCB GPRs
after every VMGEXIT.  But, since KVM is effectively writing the GHCB GPRs
on every VMRUN, I feel confident in saying that my interpretation of the
spec has already been proven correct.

The SEV-ES changes are effectively compile tested only, but unless I've
overlooked a code path, patch 1 is a nop.  Patch 3 definitely needs
testing.

Paolo, I'd really like to get patches 1 and 2 into 5.11, the code cost of
the dirty/available tracking is not trivial.

Sean Christopherson (3):
  KVM: SVM: Unconditionally sync GPRs to GHCB on VMRUN of SEV-ES guest
  KVM: x86: Revert "KVM: x86: Mark GPRs dirty when written"
  KVM: SVM: Sync GPRs to the GHCB only after VMGEXIT

 arch/x86/kvm/kvm_cache_regs.h | 51 +++++++++++++++++------------------
 arch/x86/kvm/svm/sev.c        | 14 +++++-----
 arch/x86/kvm/svm/svm.h        |  1 +
 3 files changed, 34 insertions(+), 32 deletions(-)

Comments

Paolo Bonzini Jan. 25, 2021, 5:22 p.m. UTC | #1

On 23/01/21 00:50, Sean Christopherson wrote:
> This is effectively belated feedback on the SEV-ES series.  My primary
> interest is to revert the GPR dirty/available tracking, as it's pure
> overhead for non-SEV-ES VMs, and even for SEV-ES I suspect the dirty
> tracking is at best lost in the noise, and possibly even a net negative.
> 
> My original plan was to submit patches 1+3 as patch 1, taking a few
> creative liberties with the GHCB spec to justify writing the GHCB GPRs
> after every VMGEXIT.  But, since KVM is effectively writing the GHCB GPRs
> on every VMRUN, I feel confident in saying that my interpretation of the
> spec has already been proven correct.
> 
> The SEV-ES changes are effectively compile tested only, but unless I've
> overlooked a code path, patch 1 is a nop.  Patch 3 definitely needs
> testing.
> 
> Paolo, I'd really like to get patches 1 and 2 into 5.11, the code cost of
> the dirty/available tracking is not trivial.
> 
> Sean Christopherson (3):
>    KVM: SVM: Unconditionally sync GPRs to GHCB on VMRUN of SEV-ES guest
>    KVM: x86: Revert "KVM: x86: Mark GPRs dirty when written"
>    KVM: SVM: Sync GPRs to the GHCB only after VMGEXIT
> 
>   arch/x86/kvm/kvm_cache_regs.h | 51 +++++++++++++++++------------------
>   arch/x86/kvm/svm/sev.c        | 14 +++++-----
>   arch/x86/kvm/svm/svm.h        |  1 +
>   3 files changed, 34 insertions(+), 32 deletions(-)
> 

Queued 1-2, thanks!  Yes, these should be in 5.11.

Paolo