Re: [PATCH 00/14] nSVM fixes and optional features

[Paolo Bonzini]

On 14/09/21 17:48, Maxim Levitsky wrote:
> Those are few patches I was working on lately, all somewhat related
> to the two CVEs that I found recently.
>
> First 7 patches fix various minor bugs that relate to these CVEs.
>
> The rest of the patches implement various optional SVM features,
> some of which the guest could enable anyway due to incorrect
> checking of virt_ext field.
>
> Last patch is somewhat an RFC, I would like to hear your opinion
> on that.
>
> I also implemented nested TSC scaling while at it.
>
> As for other optional SVM features here is my summary of few features
> I took a look at:
>
> X86_FEATURE_DECODEASSISTS:
>     this feature should make it easier
>     for the L1 to emulate an instruction on MMIO access, by not
>     needing to read the guest memory but rather using the instruction
>     bytes that the CPU already fetched.
>
>     The challenge of implementing this is that we sometimes inject
>     #PF and #NPT syntenically and in those cases we must be sure
>     we set the correct instruction bytes.
>
>     Also this feature adds assists for MOV CR/DR, INTn, and INVLPG,
>     which aren't that interesting but must be supported as well to
>     expose this feature to the nested guest.
>
> X86_FEATURE_VGIF
>     Might allow the L2 to run the L3 a bit faster, but due to crazy complex
>     logic we already have around int_ctl and vgif probably not worth it.
>
> X86_FEATURE_VMCBCLEAN
>     Should just be enabled, because otherwise L1 doesn't even attempt
>     to set the clean bits. But we need to know if we can take an
>     advantage of these bits first.
>
> X86_FEATURE_FLUSHBYASID
> X86_FEATURE_AVIC
>     These two features would be very good to enable, but that
>     would require lots of work, and will be done eventually.
>
> There are few more nested SVM features that I didn't yet had a
> chance to take a look at.
>
> Best regards,
> 	Maxim Levitsky
>
> Maxim Levitsky (14):
>    KVM: x86: nSVM: restore int_vector in svm_clear_vintr
>    KVM: x86: selftests: test simultaneous uses of V_IRQ from L1 and L0
>    KVM: x86: nSVM: test eax for 4K alignment for GP errata workaround
>    KVM: x86: nSVM: don't copy pause related settings
>    KVM: x86: nSVM: don't copy virt_ext from vmcb12
>    KVM: x86: SVM: don't set VMLOAD/VMSAVE intercepts on vCPU reset
>    KVM: x86: SVM: add warning for CVE-2021-3656
>    KVM: x86: SVM: add module param to control LBR virtualization
>    KVM: x86: nSVM: correctly virtualize LBR msrs when L2 is running
>    KVM: x86: nSVM: implement nested LBR virtualization
>    KVM: x86: nSVM: implement nested VMLOAD/VMSAVE
>    KVM: x86: SVM: add module param to control TSC scaling
>    KVM: x86: nSVM: implement nested TSC scaling
>    KVM: x86: nSVM: support PAUSE filter threshold and count
>
>   arch/x86/kvm/svm/nested.c                     | 105 +++++++--
>   arch/x86/kvm/svm/svm.c                        | 218 +++++++++++++++---
>   arch/x86/kvm/svm/svm.h                        |  20 +-
>   arch/x86/kvm/vmx/vmx.c                        |   1 +
>   arch/x86/kvm/x86.c                            |   1 +
>   tools/testing/selftests/kvm/.gitignore        |   1 +
>   tools/testing/selftests/kvm/Makefile          |   1 +
>   .../selftests/kvm/x86_64/svm_int_ctl_test.c   | 128 ++++++++++
>   8 files changed, 427 insertions(+), 48 deletions(-)
>   create mode 100644 tools/testing/selftests/kvm/x86_64/svm_int_ctl_test.c

Queued more patches, with 9-10-11-14 left now.

Paolo