Re: [PATCH v3 04/37] KVM: nVMX: Invalidate all roots when emulating INVVPID without EPT

[Sean Christopherson <sean.j.christopherson@intel.com>]

On Mon, Mar 23, 2020 at 05:33:08PM +0100, Vitaly Kuznetsov wrote:
> Sean Christopherson <sean.j.christopherson@intel.com> writes:
> 
> > On Mon, Mar 23, 2020 at 04:34:17PM +0100, Vitaly Kuznetsov wrote:
> >> Sean Christopherson <sean.j.christopherson@intel.com> writes:
> >> 
> >> > From: Junaid Shahid <junaids@google.com>
> >> >
> >> > Free all roots when emulating INVVPID for L1 and EPT is disabled, as
> >> > outstanding changes to the page tables managed by L1 need to be
> >> > recognized.  Because L1 and L2 share an MMU when EPT is disabled, and
> >> > because VPID is not tracked by the MMU role, all roots in the current
> >> > MMU (root_mmu) need to be freed, otherwise a future nested VM-Enter or
> >> > VM-Exit could do a fast CR3 switch (without a flush/sync) and consume
> >> > stale SPTEs.
> >> >
> >> > Fixes: 5c614b3583e7b ("KVM: nVMX: nested VPID emulation")
> >> > Signed-off-by: Junaid Shahid <junaids@google.com>
> >> > [sean: ported to upstream KVM, reworded the comment and changelog]
> >> > Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
> >> > ---
> >> >  arch/x86/kvm/vmx/nested.c | 14 ++++++++++++++
> >> >  1 file changed, 14 insertions(+)
> >> >
> >> > diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
> >> > index 9624cea4ed9f..bc74fbbf33c6 100644
> >> > --- a/arch/x86/kvm/vmx/nested.c
> >> > +++ b/arch/x86/kvm/vmx/nested.c
> >> > @@ -5250,6 +5250,20 @@ static int handle_invvpid(struct kvm_vcpu *vcpu)
> >> >  		return kvm_skip_emulated_instruction(vcpu);
> >> >  	}
> >> >  
> >> > +	/*
> >> > +	 * Sync the shadow page tables if EPT is disabled, L1 is invalidating
> >> > +	 * linear mappings for L2 (tagged with L2's VPID).  Free all roots as
> >> > +	 * VPIDs are not tracked in the MMU role.
> >> > +	 *
> >> > +	 * Note, this operates on root_mmu, not guest_mmu, as L1 and L2 share
> >> > +	 * an MMU when EPT is disabled.
> >> > +	 *
> >> > +	 * TODO: sync only the affected SPTEs for INVDIVIDUAL_ADDR.
> >> > +	 */
> >> > +	if (!enable_ept)
> >> > +		kvm_mmu_free_roots(vcpu, &vcpu->arch.root_mmu,
> >> > +				   KVM_MMU_ROOTS_ALL);
> >> > +
> >> 
> >> This is related to my remark on the previous patch; the comment above
> >> makes me think I'm missing something obvious, enlighten me please)
> >> 
> >> My understanding is that L1 and L2 will share arch.root_mmu not only
> >> when EPT is globally disabled, we seem to switch between
> >> root_mmu/guest_mmu only when nested_cpu_has_ept(vmcs12) but different L2
> >> guests may be different on this. Do we need to handle this somehow?
> >
> > guest_mmu is used iff nested EPT is enabled, which requires enable_ept=1.
> > enable_ept is global and cannot be changed without reloading kvm_intel.
> >
> > This most definitely over-invalidates, e.g. it blasts away L1's page
> > tables.  But, fixing that requires tracking VPID in mmu_role and/or adding
> > support for using guest_mmu when L1 isn't using TDP, i.e. nested EPT is
> > disabled.  Assuming the vast majority of nested deployments enable EPT in
> > L0, the cost of both options likely outweighs the benefits.
> >
> 
> Yes but my question rather was: what if global 'enable_ept' is true but
> nested EPT is not being used by L1, don't we still need to do
> kvm_mmu_free_roots(&vcpu->arch.root_mmu) here?

No, because L0 isn't shadowing the L1->L2 page tables, i.e. there can't be
unsync'd SPTEs for L2.  The vpid_sync_*() above flushes the TLB for L2's
effective VPID, which is all that's required.