Re: [PATCH v3 16/21] KVM:x86: Save/Restore GUEST_SSP to/from SMM state save area

From: Sean Christopherson
Date: Mon Jun 26 2023 - 17:20:28 EST


On Mon, Jun 26, 2023, Weijiang Yang wrote:
>
> On 6/24/2023 6:30 AM, Sean Christopherson wrote:
> > On Thu, May 11, 2023, Yang Weijiang wrote:
> > > Save GUEST_SSP to SMM state save area when guest exits to SMM
> > > due to SMI and restore it VMCS field when guest exits SMM.
> > This fails to answer "Why does KVM need to do this?"
>
> How about this:
>
> Guest SMM mode execution is out of guest kernel, to avoid GUEST_SSP
> corruption,
>
> KVM needs to save current normal mode GUEST_SSP to SMRAM area so that it can
> restore original GUEST_SSP at the end of SMM.

The key point I am looking for is a call out that KVM is emulating architectural
behavior, i.e. that smram->ssp is defined in the SDM and that the documented
behavior of Intel CPUs is that the CPU's current SSP is saved on SMI and loaded
on RSM. And I specifically say "loaded" and not "restored", because the field
is writable.

> > > Signed-off-by: Yang Weijiang <weijiang.yang@xxxxxxxxx>
> > > ---
> > > arch/x86/kvm/smm.c | 20 ++++++++++++++++++++
> > > 1 file changed, 20 insertions(+)
> > >
> > > diff --git a/arch/x86/kvm/smm.c b/arch/x86/kvm/smm.c
> > > index b42111a24cc2..c54d3eb2b7e4 100644
> > > --- a/arch/x86/kvm/smm.c
> > > +++ b/arch/x86/kvm/smm.c
> > > @@ -275,6 +275,16 @@ static void enter_smm_save_state_64(struct kvm_vcpu *vcpu,
> > > enter_smm_save_seg_64(vcpu, &smram->gs, VCPU_SREG_GS);
> > > smram->int_shadow = static_call(kvm_x86_get_interrupt_shadow)(vcpu);
> > > +
> > > + if (kvm_cet_user_supported()) {
> > This is wrong, KVM should not save/restore state that doesn't exist from the guest's
> > perspective, i.e. this needs to check guest_cpuid_has().
>
> Yes, the check missed the case that user space disables SHSTK. Will change
> it, thanks!
>
> >
> > On a related topic, I would love feedback on my series that adds a framework for
> > features like this, where KVM needs to check guest CPUID as well as host support.
> >
> > https://lore.kernel.org/all/20230217231022.816138-1-seanjc@xxxxxxxxxx
>
> The framework looks good, will it be merged in kvm_x86?

Yes, I would like to merge it at some point.

> > > @@ -565,6 +575,16 @@ static int rsm_load_state_64(struct x86_emulate_ctxt *ctxt,
> > > static_call(kvm_x86_set_interrupt_shadow)(vcpu, 0);
> > > ctxt->interruptibility = (u8)smstate->int_shadow;
> > > + if (kvm_cet_user_supported()) {
> > > + struct msr_data msr;
> > > +
> > > + msr.index = MSR_KVM_GUEST_SSP;
> > > + msr.host_initiated = true;
> > > + msr.data = smstate->ssp;
> > > + /* Mimic host_initiated access to bypass ssp access check. */
> > No, masquerading as a host access is all kinds of wrong. I have no idea what
> > check you're trying to bypass, but whatever it is, it's wrong. Per the SDM, the
> > SSP field in SMRAM is writable, which means that KVM needs to correctly handle
> > the scenario where SSP holds garbage, e.g. a non-canonical address.
>
> MSR_KVM_GUEST_SSP is only accessible to user space, e.g., during LM, it's not
> accessible to VM itself. So in kvm_cet_is_msr_accessible(), I added a check to
> tell whether the access is initiated from user space or not, I tried to bypass
> that check. Yes, I will add necessary checks here.
>
> >
> > Why can't this use kvm_get_msr() and kvm_set_msr()?
>
> If my above assumption is correct, these helpers are passed by
> host_initiated=false and cannot meet the requirments.

Sorry, I don't follow. These writes are NOT initiated from the host, i.e.
kvm_get_msr() and kvm_set_msr() do the right thing, unless I'm missing something.