Re: [PATCH v3 03/37] KVM: nVMX: Invalidate all EPTP contexts when emulating INVEPT for L1

[Sean Christopherson]

On Mon, Mar 23, 2020 at 04:24:05PM +0100, Vitaly Kuznetsov wrote:
> Sean Christopherson <sean.j.christopherson@xxxxxxxxx> writes:
> 
> > Free all L2 (guest_mmu) roots when emulating INVEPT for L1.  Outstanding
> > changes to the EPT tables managed by L1 need to be recognized, and
> > relying on KVM to always flush L2's EPTP context on nested VM-Enter is
> > dangerous.
> >
> > Similar to handle_invpcid(), rely on kvm_mmu_free_roots() to do a remote
> > TLB flush if necessary, e.g. if L1 has never entered L2 then there is
> > nothing to be done.
> >
> > Nuking all L2 roots is overkill for the single-context variant, but it's
> > the safe and easy bet.  A more precise zap mechanism will be added in
> > the future.  Add a TODO to call out that KVM only needs to invalidate
> > affected contexts.
> >
> > Fixes: b119019847fbc ("kvm: nVMX: Remove unnecessary sync_roots from handle_invept")
> > Reported-by: Jim Mattson <jmattson@xxxxxxxxxx>
> > Signed-off-by: Sean Christopherson <sean.j.christopherson@xxxxxxxxx>
> > ---
> >  arch/x86/kvm/vmx/nested.c | 8 ++++----
> >  1 file changed, 4 insertions(+), 4 deletions(-)
> >
> > diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
> > index f3774cef4fd4..9624cea4ed9f 100644
> > --- a/arch/x86/kvm/vmx/nested.c
> > +++ b/arch/x86/kvm/vmx/nested.c
> > @@ -5160,12 +5160,12 @@ static int handle_invept(struct kvm_vcpu *vcpu)
> >  		if (!nested_vmx_check_eptp(vcpu, operand.eptp))
> >  			return nested_vmx_failValid(vcpu,
> >  				VMXERR_INVALID_OPERAND_TO_INVEPT_INVVPID);
> > +
> > +		/* TODO: sync only the target EPTP context. */
> >  		fallthrough;
> >  	case VMX_EPT_EXTENT_GLOBAL:
> > -	/*
> > -	 * TODO: Sync the necessary shadow EPT roots here, rather than
> > -	 * at the next emulated VM-entry.
> > -	 */
> > +		kvm_mmu_free_roots(vcpu, &vcpu->arch.guest_mmu,
> > +				   KVM_MMU_ROOTS_ALL);
> >  		break;
> 
> An ignorant reader may wonder "and how do we know that L1 actaully uses
> EPT" as he may find out that guest_mmu is not being used otherwise. The
> answer to the question will likely be "if L1 doesn't use EPT for some of
> its guests than there's nothing we should do here as we will be
> resetting root_mmu when switching to/from them". Hope the ignorant
> reviewer typing this is not very wrong :-)

A different way to put it would be:

  KVM never uses root_mmu to hold nested EPT roots.

Invalidating too much is functionally ok, though sub-optimal for performance.
Invalidating too little is what we really care about.

FWIW, VMX currently uses guest_mmu iff nested EPT is enabled.  In theory,
KVM could be enhanced to also used guest_mmu when nested-TDP is disabled,
e.g. to enable VMX to preserve L1's root_mmu when emulating INVVPID.  That
would likely be a decent performance boost for nested VMX+VPID without
nested EPT, but I'm guessing that the cross-section of users that care
about nested performance and don't use nested EPT is quite small.

> >  	default:
> >  		BUG_ON(1);
> 
> Reviewed-by: Vitaly Kuznetsov <vkuznets@xxxxxxxxxx>
> 
> -- 
> Vitaly
>