Re: [PATCH 2/2] KVM: VMX: Inject #GP, not #UD, if SGX2 ENCLS leafs are unsupported

[Huang, Kai]

On Thu, 2023-04-06 at 11:00 -0700, Sean Christopherson wrote:
> On Thu, Apr 06, 2023, Huang, Kai wrote:
> > On Wed, 2023-04-05 at 16:45 -0700, Sean Christopherson wrote:
> > > Per Intel's SDM, unsupported ENCLS leafs result in a #GP, not a #UD.
> > > SGX1 is a special snowflake as the SGX1 flag is used by the CPU as a
> > > "soft" disable, e.g. if software disables machine check reporting, i.e.
> > > having SGX but not SGX1 is effectively "SGX completely unsupported" and
> > > and thus #UDs.
> > 
> > If I recall correctly, this is an erratum which can clear SGX1 in CPUID while
> > the SGX flag is still in CPUID?
> 
> Nope, not an erratum, architectural behavior.

I found the relevant section in SDM:

All supported IA32_MCi_CTL bits for all the machine check banks must be set for
Intel SGX to be available
(CPUID.SGX_Leaf.0:EAX[SGX1] == 1). Any act of clearing bits from '1 to '0 in any
of the IA32_MCi_CTL register
may disable Intel SGX (set CPUID.SGX_Leaf.0:EAX[SGX1] to 0) until the next
reset.

Looking at the code, it seems currently KVM won't clear SGX1 bit in CPUID when
guest disables IA32_MCi_CTL (writing 0).  Should we do that?

> 
> > But I am not sure whether this part is relevant to this patch?  Because SDM
> > already says ENCLS causes #UD if SGX1 isn't present.  This patch changes
> > "unsupported leaf" from causing #UD to causing #GP, which is also documented in
> > SDM.
> 
> I wanted to capture why SGX1 is different and given special treatment in the SDM.
> I.e. to explain why SGX1 leafs are an exception to the "#GP if leaf unsupported"
> clause.

OK.

> 
> > > Fixes: 9798adbc04cf ("KVM: VMX: Frame in ENCLS handler for SGX virtualization")
> > > Cc: Binbin Wu <binbin.wu@xxxxxxxxxxxxxxx>
> > > Cc: Kai Huang <kai.huang@xxxxxxxxx>
> > > Signed-off-by: Sean Christopherson <seanjc@xxxxxxxxxx>
> > > ---
> > >  arch/x86/kvm/vmx/sgx.c | 15 +++++++++------
> > >  1 file changed, 9 insertions(+), 6 deletions(-)
> > > 
> > > diff --git a/arch/x86/kvm/vmx/sgx.c b/arch/x86/kvm/vmx/sgx.c
> > > index f881f6ff6408..1c092ab89c33 100644
> > > --- a/arch/x86/kvm/vmx/sgx.c
> > > +++ b/arch/x86/kvm/vmx/sgx.c
> > > @@ -350,11 +350,12 @@ static int handle_encls_einit(struct kvm_vcpu *vcpu)
> > >  
> > >  static inline bool encls_leaf_enabled_in_guest(struct kvm_vcpu *vcpu, u32 leaf)
> > >  {
> > > -	if (!enable_sgx || !guest_cpuid_has(vcpu, X86_FEATURE_SGX))
> > > -		return false;
> > > -
> > > +	/*
> > > +	 * ENCLS #UDs if SGX1 isn't supported, i.e. this point will be reached
> > 
> > Why #UDs instead of #UD?  Is #UD a verb?
> 
> Heh, it is now ;-)  I can reword to something like
> 
> 	/*
> 	 * ENCLS generates a #UD if SGX1 isn't supported ...
> 	 */
> 
> if my made up grammar is confusing.
> 
> > > +	 * if and only if the SGX1 leafs are enabled.
> > > +	 */
> > 
> > Is it better to move "ENCLS #UDs if SGX1 isn't supported" part to ...
> > 
> > >  	if (leaf >= ECREATE && leaf <= ETRACK)
> > > -		return guest_cpuid_has(vcpu, X86_FEATURE_SGX1);
> > > +		return true;
> > >  
> > >  	if (leaf >= EAUG && leaf <= EMODT)
> > >  		return guest_cpuid_has(vcpu, X86_FEATURE_SGX2);
> > > @@ -373,9 +374,11 @@ int handle_encls(struct kvm_vcpu *vcpu)
> > >  {
> > >  	u32 leaf = (u32)kvm_rax_read(vcpu);
> > >  
> > > -	if (!encls_leaf_enabled_in_guest(vcpu, leaf)) {
> > > +	if (!enable_sgx || !guest_cpuid_has(vcpu, X86_FEATURE_SGX) ||
> > > +	    !guest_cpuid_has(vcpu, X86_FEATURE_SGX1)) {
> > >  		kvm_queue_exception(vcpu, UD_VECTOR);
> > 
> > ... above here, where the actual code reside?
> 
> My goal was to document why encls_leaf_enabled_in_guest() unconditionally returns
> true for SGX1 leafs, i.e. why it doesn't query X86_FEATURE_SGX1.  I'm definitely
> not opposed to also adding a comment here, but I do want to keep the comment in
> encls_leaf_enabled_in_guest().

Sure.

Anyway,

Reviewed-by: Kai Huang <kai.huang@xxxxxxxxx>