Re: [PATCH 03/15] KVM: SVM: Inject #UD on RDTSCP when it should be disabled in the guest

From: Jim Mattson <jmattson@google.com>
To: Sean Christopherson <seanjc@google.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>,
	Vitaly Kuznetsov <vkuznets@redhat.com>,
	Wanpeng Li <wanpengli@tencent.com>,
	Joerg Roedel <joro@8bytes.org>, kvm list <kvm@vger.kernel.org>,
	LKML <linux-kernel@vger.kernel.org>,
	Xiaoyao Li <xiaoyao.li@intel.com>,
	Reiji Watanabe <reijiw@google.com>
Subject: Re: [PATCH 03/15] KVM: SVM: Inject #UD on RDTSCP when it should be disabled in the guest
Date: Tue, 4 May 2021 15:24:04 -0700	[thread overview]
Message-ID: <CALMp9eT9HMKs_JcQHLsyc9MxFLFaAt9Ve8ev=inH-+8NeHtayw@mail.gmail.com> (raw)
In-Reply-To: <YJHGQgEE3mqUhbAc@google.com>

On Tue, May 4, 2021 at 3:10 PM Sean Christopherson <seanjc@google.com> wrote:
>
> On Tue, May 04, 2021, Jim Mattson wrote:
> > On Tue, May 4, 2021 at 2:53 PM Sean Christopherson <seanjc@google.com> wrote:
> > >
> > > On Tue, May 04, 2021, Jim Mattson wrote:
> > > > On Tue, May 4, 2021 at 10:17 AM Sean Christopherson <seanjc@google.com> wrote:
> > > > >
> > > > > Intercept RDTSCP to inject #UD if RDTSC is disabled in the guest.
> > > > >
> > > > > Note, SVM does not support intercepting RDPID.  Unlike VMX's
> > > > > ENABLE_RDTSCP control, RDTSCP interception does not apply to RDPID.  This
> > > > > is a benign virtualization hole as the host kernel (incorrectly) sets
> > > > > MSR_TSC_AUX if RDTSCP is supported, and KVM loads the guest's MSR_TSC_AUX
> > > > > into hardware if RDTSCP is supported in the host, i.e. KVM will not leak
> > > > > the host's MSR_TSC_AUX to the guest.
> > > > >
> > > > > But, when the kernel bug is fixed, KVM will start leaking the host's
> > > > > MSR_TSC_AUX if RDPID is supported in hardware, but RDTSCP isn't available
> > > > > for whatever reason.  This leak will be remedied in a future commit.
> > > > >
> > > > > Fixes: 46896c73c1a4 ("KVM: svm: add support for RDTSCP")
> > > > > Cc: stable@vger.kernel.org
> > > > > Signed-off-by: Sean Christopherson <seanjc@google.com>
> > > > > ---
> > > > ...
> > > > > @@ -4007,8 +4017,7 @@ static void svm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
> > > > >         svm->nrips_enabled = kvm_cpu_cap_has(X86_FEATURE_NRIPS) &&
> > > > >                              guest_cpuid_has(vcpu, X86_FEATURE_NRIPS);
> > > > >
> > > > > -       /* Check again if INVPCID interception if required */
> > > > > -       svm_check_invpcid(svm);
> > > > > +       svm_recalc_instruction_intercepts(vcpu, svm);
> > > >
> > > > Does the right thing happen here if the vCPU is in guest mode when
> > > > userspace decides to toggle the CPUID.80000001H:EDX.RDTSCP bit on or
> > > > off?
> > >
> > > I hate our terminology.  By "guest mode", do you mean running the vCPU, or do
> > > you specifically mean running in L2?
> >
> > I mean is_guest_mode(vcpu) is true (i.e. running L2).
>
> No, it will not do the right thing, whatever "right thing" even means in this
> context.  That's a pre-existing issue, e.g. INVCPID handling is also wrong.
> I highly doubt VMX does, or even can, do the right thing either.
>
> I'm pretty sure I lobbied in the past to disallow KVM_SET_CPUID* if the vCPU is
> in guest mode since it's impossible to do the right thing without forcing an
> exit to L1, e.g. changing MAXPHYSADDR will allow running L2 with an illegal
> CR3, ditto for various CR4 bits.

With that caveat understood,

Reviewed-by: Jim Mattson <jmattson@google.com>