Re: [PATCH v8 7/7] KVM: X86: Add user-space access interface for CET MSRs

From: Sean Christopherson
Date: Wed Dec 11 2019 - 11:27:06 EST


On Wed, Dec 11, 2019 at 10:19:51AM +0800, Yang Weijiang wrote:
> On Tue, Dec 10, 2019 at 01:58:59PM -0800, Sean Christopherson wrote:
> > On Fri, Nov 01, 2019 at 04:52:22PM +0800, Yang Weijiang wrote:
> > > There're two different places storing Guest CET states, states
> > > managed with XSAVES/XRSTORS, as restored/saved
> > > in previous patch, can be read/write directly from/to the MSRs.
> > > For those stored in VMCS fields, they're access via vmcs_read/
> > > vmcs_write.
> > >
> > >
> > > +#define CET_MSR_RSVD_BITS_1 0x3
> > > +#define CET_MSR_RSVD_BITS_2 (0xF << 6)
> > > +
> > > +static bool cet_msr_write_allowed(struct kvm_vcpu *vcpu, struct msr_data *msr)
> > > +{
> > > + u32 index = msr->index;
> > > + u64 data = msr->data;
> > > + u32 high_word = data >> 32;
> > > +
> > > + if ((index == MSR_IA32_U_CET || index == MSR_IA32_S_CET) &&
> > > + (data & CET_MSR_RSVD_BITS_2))
> > > + return false;
> > > +
> > > + if (is_64_bit_mode(vcpu)) {
> > > + if (is_noncanonical_address(data & PAGE_MASK, vcpu))
> >
> > I don't think this is correct. MSRs that contain an address usually only
> > fault on a non-canonical value and do the non-canonical check regardless
> > of mode. E.g. VM-Enter's consistency checks on SYSENTER_E{I,S}P only care
> > about a canonical address and are not dependent on mode, and SYSENTER
> > itself states that bits 63:32 are ignored in 32-bit mode. I assume the
> > same is true here.
> The spec. reads like this: Must be machine canonical when written on parts
> that support 64 bit mode. On parts that do not support 64 bit mode, the bits
> 63:32 are reserved and must be 0.

Yes, that agrees with me. The key word is "support", i.e. "on parts that
support 64 bit mode" means "on parts with CPUID.0x80000001.EDX.LM=1."

The reason the architecture works this way is that unless hardware clears
the MSRs on transition from 64->32, bits 63:32 need to be ignored on the
way out instead of being validated on the way in, e.g. software writes a
64-bit value to the MSR and then transitions to 32-bit mode. Clearing the
MSRs would be painful, slow and error prone, so it's easier for hardware
to simply ignore bits 63:32 in 32-bit mode.

> > If that is indeed the case, what about adding these to the common canonical
> > check in __kvm_set_msr()? That'd cut down on the boilerplate here and
> > might make it easier to audit KVM's canonical checks.
> >
> > > + return false;
> > > + else if ((index == MSR_IA32_PL0_SSP ||
> > > + index == MSR_IA32_PL1_SSP ||
> > > + index == MSR_IA32_PL2_SSP ||
> > > + index == MSR_IA32_PL3_SSP) &&
> > > + (data & CET_MSR_RSVD_BITS_1))
> > > + return false;
> > > + } else {
> > > + if (msr->index == MSR_IA32_INT_SSP_TAB)
> > > + return false;
> > > + else if ((index == MSR_IA32_U_CET ||
> > > + index == MSR_IA32_S_CET ||
> > > + index == MSR_IA32_PL0_SSP ||
> > > + index == MSR_IA32_PL1_SSP ||
> > > + index == MSR_IA32_PL2_SSP ||
> > > + index == MSR_IA32_PL3_SSP) &&
> > > + (high_word & ~0ul))
> > > + return false;
> > > + }
> > > +
> > > + return true;
> > > +}
> >
> > This helper seems like overkill, e.g. it's filled with index-specific
> > checks, but is called from code that has already switched on the index.
> > Open coding the individual checks is likely more readable and would require
> > less code, especially if the canonical checks are cleaned up.
> >
> I'm afraid if the checks are not wrapped in a helper, there're many
> repeat checking-code, that's why I'm using a wrapper.

But you're adding almost as much, if not more, code to re-split the checks
in this helper.

> > > +
> > > +static bool cet_msr_access_allowed(struct kvm_vcpu *vcpu, struct msr_data *msr)
> > > +{
> > > + u64 kvm_xss;
> > > + u32 index = msr->index;
> > > +
> > > + if (is_guest_mode(vcpu))
> > > + return false;
> >
> > I may have missed this in an earlier discussion, does CET not support
> > nesting?
> >
> I don't want to make CET avaible to nested guest at time being, first to
> make it available to L1 guest first. So I need to avoid exposing any CET
> CPUID/MSRs to a nested guest.