Re: [PATCH v3 1/4] KVM: VMX: Flush CPU buffers as needed if L1D cache flush is skipped

[Pawan Gupta <pawan.kumar.gupta@linux.intel.com>]

On Mon, Oct 27, 2025 at 04:58:10PM -0700, Jim Mattson wrote:
> On Mon, Oct 27, 2025 at 4:17 PM Pawan Gupta
> <pawan.kumar.gupta@linux.intel.com> wrote:
> >
> > On Mon, Oct 27, 2025 at 03:03:23PM -0700, Jim Mattson wrote:
> > > On Tue, Oct 21, 2025 at 6:20 PM Pawan Gupta
> > > <pawan.kumar.gupta@linux.intel.com> wrote:
> > > >
> > > > ...
> > > > Thinking more on this, the software sequence is only invoked when the
> > > > system doesn't have the L1D flushing feature added by a microcode update.
> > > > In such a case system is not expected to have a flushing VERW either, which
> > > > was introduced after L1TF. Also, the admin needs to have a very good reason
> > > > for not updating the microcode for 5+ years :-)
> > >
> > > KVM started reporting MD_CLEAR to userspace in Linux v5.2, but it
> > > didn't report L1D_FLUSH to userspace until Linux v6.4, so there are
> > > plenty of virtual CPUs with a flushing VERW that don't have the L1D
> > > flushing feature.
> >
> > Shouldn't only the L0 hypervisor be doing the L1D_FLUSH?
> >
> > kvm_get_arch_capabilities()
> > {
> > ...
> >         /*
> >          * If we're doing cache flushes (either "always" or "cond")
> >          * we will do one whenever the guest does a vmlaunch/vmresume.
> >          * If an outer hypervisor is doing the cache flush for us
> >          * (ARCH_CAP_SKIP_VMENTRY_L1DFLUSH), we can safely pass that
> >          * capability to the guest too, and if EPT is disabled we're not
> >          * vulnerable.  Overall, only VMENTER_L1D_FLUSH_NEVER will
> >          * require a nested hypervisor to do a flush of its own.
> >          */
> >         if (l1tf_vmx_mitigation != VMENTER_L1D_FLUSH_NEVER)
> >                 data |= ARCH_CAP_SKIP_VMENTRY_L1DFLUSH;
> >
> 
> Unless L0 has chosen L1D_FLUSH_NEVER. :)
> 
> On GCE's L1TF-vulnerable hosts, we actually do an L1D flush at ASI
> entry rather than VM-entry. ASI entries are two orders of magnitude
> less frequent than VM-entries, so we get comparable protection to
> L1D_FLUSH_ALWAYS at a fraction of the cost.
> 
> At the moment, we still do an L1D flush on emulated VM-entry, but
> that's just because we have historically advertised
> IA32_ARCH_CAPABILITIES.SKIP_L1DFL_VMENTRY to L1.

Thanks for the background.

I still don't see the problem, CPUs that are vulnerable to L1TF are also
vulnerable to MDS. So, they don't set mmio_stale_data_clear, instead they
set X86_FEATURE_CLEAR_CPU_BUF and execute VERW in __vmx_vcpu_run()
regardless of whether L1D_FLUSH was done.

But, I agree it is best to decouple L1D flush and MMIO Stale Data to be
avoid any confusion.