On Wed, Mar 29, 2023, Manali Shukla wrote:
> On 3/25/2023 1:25 AM, Sean Christopherson wrote:
> > On Mon, Feb 06, 2023, Manali Shukla wrote:
> >> - if (sev_es_guest(vcpu->kvm))
> >> + if (sev_es_guest(vcpu->kvm)) {
> >> + bool ibs_fetch_active, ibs_op_active;
> >> + u64 ibs_fetch_ctl, ibs_op_ctl;
> >> +
> >> + if (svm->prevent_hostibs_enabled) {
> >> + /*
> >> + * With PreventHostIBS enabled, IBS profiling cannot
> >> + * be active when VMRUN is executed. Disable IBS before
> >> + * executing VMRUN and, because of a race condition,
> >> + * enable the PreventHostIBS window if IBS profiling was
> >> + * active.
> >
> > And the race can't be fixed because...?
>
> Race can not be fixed because VALID and ENABLE bit for IBS_FETCH_CTL and IBS_OP_CTL
> are contained in their same resepective MSRs. Due to this reason following scenario can
> be generated:
> Read IBS_FETCH_CTL (IbsFetchEn bit is 1 and IBSFetchVal bit is 0)
> Write IBS_FETCH_CTL (IbsFetchEn is 0 now)
> Imagine in between Read and Write, IBSFetchVal changes to 1. Write to IBS_FETCH_CTL will
> clear the IBSFetchVal bit. When STGI is executed after VMEXIT, the NMI is taken and check for
> valid mask will fail and generate Dazed and Confused NMI messages.
> Please refer to cover letter for more details.
I understand the race, I'm asking why this series doesn't fix the race. Effectively
suppressing potentially unexpected NMIs because PreventHostIBS was enable is ugly.
> >> + */
> >> + ibs_fetch_active =
> >> + amd_disable_ibs_fetch(&ibs_fetch_ctl);
> >> + ibs_op_active =
> >> + amd_disable_ibs_op(&ibs_op_ctl);
> >> +
> >> + amd_prevent_hostibs_window(ibs_fetch_active ||
> >> + ibs_op_active);
> >> + }
> >> +
> >> __svm_sev_es_vcpu_run(svm, spec_ctrl_intercepted);
> >> - else
> >> +
> >> + if (svm->prevent_hostibs_enabled) {
> >> + if (ibs_fetch_active)
> >> + amd_restore_ibs_fetch(ibs_fetch_ctl);
> >> +
> >> + if (ibs_op_active)
> >> + amd_restore_ibs_op(ibs_op_ctl);
> >
> > IIUC, this adds up to 2 RDMSRs and 4 WRMSRs to the VMRUN path. Blech. There's
> > gotta be a better way to implement this.
>
> I will try to find a better way to implement this.
>
> > Like PeterZ said, this is basically
> > exclude_guest.
>
> As I mentioned before, exclude_guest lets the profiler decide whether it wants to trace the guest
> data or not, whereas PreventHostIBS lets the owner of the guest decide whether host can trace guest's
> data or not.
PreventHostIBS is purely an enforcement, it does not actually do anything to
disable tracing of the guest. What PeterZ and I are complaining about is that
instead of integrating this feature with exclude_guest, e.g. finding a way to
make guest tracing mutually exclusive with KVM_RUN so that PreventHostIBS can be
contexted switched according, this series instead backdoors into perf to forcefully
disable tracing.
In other words, please try to create a sane contract between userspace, perf, and
KVM, e.g. disallow tracing a guest with PreventHostIBS at some level instead of
silently toggling tracing around VMRUN.
