On Tue, Aug 24, 2021, Xiaoyao Li wrote:
> On 8/24/2021 10:20 PM, Sean Christopherson wrote:
> > On Tue, Aug 24, 2021, Xiaoyao Li wrote:
> > > Per SDM, it triggers #GP for all the accessing of PT MSRs, if
> > > X86_FEATURE_INTEL_PT is not available.
> > >
> > > Signed-off-by: Xiaoyao Li <xiaoyao.li@xxxxxxxxx>
> > > ---
> > > arch/x86/kvm/vmx/vmx.c | 20 ++++++++++++++------
> > > 1 file changed, 14 insertions(+), 6 deletions(-)
> > >
> > > diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> > > index 4a70a6d2f442..1bbc4d84c623 100644
> > > --- a/arch/x86/kvm/vmx/vmx.c
> > > +++ b/arch/x86/kvm/vmx/vmx.c
> > > @@ -1010,9 +1010,16 @@ static unsigned long segment_base(u16 selector)
> > > static inline bool pt_can_write_msr(struct vcpu_vmx *vmx)
> > > {
> > > return vmx_pt_mode_is_host_guest() &&
> > > + guest_cpuid_has(&vmx->vcpu, X86_FEATURE_INTEL_PT) &&
> > > !(vmx->pt_desc.guest.ctl & RTIT_CTL_TRACEEN);
> > > }
> > > +static inline bool pt_can_read_msr(struct kvm_vcpu *vcpu)
> > > +{
> > > + return vmx_pt_mode_is_host_guest() &&
> > > + guest_cpuid_has(vcpu, X86_FEATURE_INTEL_PT);
> > > +}
> > > +
> > > static inline bool pt_output_base_valid(struct kvm_vcpu *vcpu, u64 base)
> > > {
> > > /* The base must be 128-byte aligned and a legal physical address. */
> > > @@ -1849,24 +1856,24 @@ static int vmx_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
> > > &msr_info->data);
> > > break;
> > > case MSR_IA32_RTIT_CTL:
> > > - if (!vmx_pt_mode_is_host_guest())
> > > + if (!pt_can_read_msr(vcpu))
> >
> > These all need to provide exemptions for accesses from the host. KVM allows
> > access to MSRs that are not exposed to the guest so long as all the other checks
> > pass.
>
> Not all the MSRs are allowed to be accessed from host regardless of whether
> it's exposed to guest. e.g., MSR_IA32_TSC_ADJUST, it checks guest CPUID
> first.
>
> For me, for those PT MSRs, I cannot think of any reason that host/userspace
> would access them without PT being exposed to guest.
Order of operations. Userspace is allowed to do KVM_GET/SET_MSR before
KVM_SET_CPUID2.
> On the other hand, since this patch indeed breaks the existing userspace VMM
> who accesses those MSRs without checking guest CPUID.
>
> So I will follow your advice to allow the host_initiated case in next
> version.
>
> > Same for the next patch.
>
> Sorry, I don't know how it matters next patch.
Me either. Ignore that comment. :-)
