Message ID | 20240523121828.808-2-ravi.bangoria@amd.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | KVM: SEV-ES: Fix KVM_{GET|SET}_MSRS and LBRV handling | expand |
On 5/23/24 14:18, Ravi Bangoria wrote: > From: Nikunj A Dadhania <nikunj@amd.com> > > KVM currently allows userspace to read/write MSRs even after the VMSA is > encrypted. This can cause unintentional issues if MSR access has side- > effects. For ex, while migrating a guest, userspace could attempt to > migrate MSR_IA32_DEBUGCTLMSR and end up unintentionally disabling LBRV on > the target. Fix this by preventing access to those MSRs which are context > switched via the VMSA, once the VMSA is encrypted. > > Suggested-by: Sean Christopherson <seanjc@google.com> > Signed-off-by: Nikunj A Dadhania <nikunj@amd.com> > Signed-off-by: Ravi Bangoria <ravi.bangoria@amd.com> > --- > arch/x86/kvm/svm/svm.c | 18 ++++++++++++++++++ > 1 file changed, 18 insertions(+) > > diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c > index 3d0549ca246f..489b0183f37d 100644 > --- a/arch/x86/kvm/svm/svm.c > +++ b/arch/x86/kvm/svm/svm.c > @@ -2834,10 +2834,24 @@ static int svm_get_msr_feature(struct kvm_msr_entry *msr) > return 0; > } > > +static bool > +sev_es_prevent_msr_access(struct kvm_vcpu *vcpu, struct msr_data *msr_info) > +{ > + return sev_es_guest(vcpu->kvm) && > + vcpu->arch.guest_state_protected && > + svm_msrpm_offset(msr_info->index) != MSR_INVALID && > + !msr_write_intercepted(vcpu, msr_info->index); > +} > + > static int svm_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info) > { > struct vcpu_svm *svm = to_svm(vcpu); > > + if (sev_es_prevent_msr_access(vcpu, msr_info)) { > + msr_info->data = 0; > + return 0; This should return -EINVAL, not 0. Likewise below in svm_set_msr(). Paolo > + } > + > switch (msr_info->index) { > case MSR_AMD64_TSC_RATIO: > if (!msr_info->host_initiated && > @@ -2988,6 +3002,10 @@ static int svm_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr) > > u32 ecx = msr->index; > u64 data = msr->data; > + > + if (sev_es_prevent_msr_access(vcpu, msr)) > + return 0; > + > switch (ecx) { > case MSR_AMD64_TSC_RATIO: >
On 5/28/2024 10:01 PM, Paolo Bonzini wrote: > On 5/23/24 14:18, Ravi Bangoria wrote: >> From: Nikunj A Dadhania <nikunj@amd.com> >> >> KVM currently allows userspace to read/write MSRs even after the VMSA is >> encrypted. This can cause unintentional issues if MSR access has side- >> effects. For ex, while migrating a guest, userspace could attempt to >> migrate MSR_IA32_DEBUGCTLMSR and end up unintentionally disabling LBRV on >> the target. Fix this by preventing access to those MSRs which are context >> switched via the VMSA, once the VMSA is encrypted. >> >> Suggested-by: Sean Christopherson <seanjc@google.com> >> Signed-off-by: Nikunj A Dadhania <nikunj@amd.com> >> Signed-off-by: Ravi Bangoria <ravi.bangoria@amd.com> >> --- >> arch/x86/kvm/svm/svm.c | 18 ++++++++++++++++++ >> 1 file changed, 18 insertions(+) >> >> diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c >> index 3d0549ca246f..489b0183f37d 100644 >> --- a/arch/x86/kvm/svm/svm.c >> +++ b/arch/x86/kvm/svm/svm.c >> @@ -2834,10 +2834,24 @@ static int svm_get_msr_feature(struct kvm_msr_entry *msr) >> return 0; >> } >> +static bool >> +sev_es_prevent_msr_access(struct kvm_vcpu *vcpu, struct msr_data *msr_info) >> +{ >> + return sev_es_guest(vcpu->kvm) && >> + vcpu->arch.guest_state_protected && >> + svm_msrpm_offset(msr_info->index) != MSR_INVALID && >> + !msr_write_intercepted(vcpu, msr_info->index); >> +} >> + >> static int svm_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info) >> { >> struct vcpu_svm *svm = to_svm(vcpu); >> + if (sev_es_prevent_msr_access(vcpu, msr_info)) { >> + msr_info->data = 0; >> + return 0; > > This should return -EINVAL, not 0. Likewise below in svm_set_msr(). Sure. Thanks, Ravi
On Wed, May 29, 2024 at 04:14:10PM +0530, Ravi Bangoria wrote: > On 5/28/2024 10:01 PM, Paolo Bonzini wrote: > > On 5/23/24 14:18, Ravi Bangoria wrote: > >> From: Nikunj A Dadhania <nikunj@amd.com> > >> > >> KVM currently allows userspace to read/write MSRs even after the VMSA is > >> encrypted. This can cause unintentional issues if MSR access has side- > >> effects. For ex, while migrating a guest, userspace could attempt to > >> migrate MSR_IA32_DEBUGCTLMSR and end up unintentionally disabling LBRV on > >> the target. Fix this by preventing access to those MSRs which are context > >> switched via the VMSA, once the VMSA is encrypted. > >> > >> Suggested-by: Sean Christopherson <seanjc@google.com> > >> Signed-off-by: Nikunj A Dadhania <nikunj@amd.com> > >> Signed-off-by: Ravi Bangoria <ravi.bangoria@amd.com> > >> --- > >> arch/x86/kvm/svm/svm.c | 18 ++++++++++++++++++ > >> 1 file changed, 18 insertions(+) > >> > >> diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c > >> index 3d0549ca246f..489b0183f37d 100644 > >> --- a/arch/x86/kvm/svm/svm.c > >> +++ b/arch/x86/kvm/svm/svm.c > >> @@ -2834,10 +2834,24 @@ static int svm_get_msr_feature(struct kvm_msr_entry *msr) > >> return 0; > >> } > >> +static bool > >> +sev_es_prevent_msr_access(struct kvm_vcpu *vcpu, struct msr_data *msr_info) > >> +{ > >> + return sev_es_guest(vcpu->kvm) && > >> + vcpu->arch.guest_state_protected && > >> + svm_msrpm_offset(msr_info->index) != MSR_INVALID && > >> + !msr_write_intercepted(vcpu, msr_info->index); > >> +} > >> + > >> static int svm_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info) > >> { > >> struct vcpu_svm *svm = to_svm(vcpu); > >> + if (sev_es_prevent_msr_access(vcpu, msr_info)) { > >> + msr_info->data = 0; > >> + return 0; > > > > This should return -EINVAL, not 0. Likewise below in svm_set_msr(). > > Sure. One consequence of this change is that older VMMs that might still call svm_get_msr()/svm_set_msr() for SEV-ES guests. Newer VMMs that are aware of KVM_SEV_INIT2 however are already aware of the stricter limitations of what vCPU state can be sync'd during guest run-time, so newer QEMU for instance will work both for legacy KVM_SEV_ES_INIT interface as well as KVM_SEV_INIT2. So when using KVM_SEV_INIT2 it's okay to assume userspace can deal with -EINVAL, whereas for legacy KVM_SEV_ES_INIT we sort of have to assume the VMM does not have the necessary changes to deal with -EINVAL, so in that case it's probably more appropriate to return 0 and just silently noop. We had a similar situations with stricter limitations on fpstate sync'ing for KVM_SEV_INIT2 and that was the approach taken there: https://lore.kernel.org/kvm/ZfRhu0GVjWeAAJMB@google.com/ so I'll submit a patch that takes the same approach. -Mike > > Thanks, > Ravi
diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c index 3d0549ca246f..489b0183f37d 100644 --- a/arch/x86/kvm/svm/svm.c +++ b/arch/x86/kvm/svm/svm.c @@ -2834,10 +2834,24 @@ static int svm_get_msr_feature(struct kvm_msr_entry *msr) return 0; } +static bool +sev_es_prevent_msr_access(struct kvm_vcpu *vcpu, struct msr_data *msr_info) +{ + return sev_es_guest(vcpu->kvm) && + vcpu->arch.guest_state_protected && + svm_msrpm_offset(msr_info->index) != MSR_INVALID && + !msr_write_intercepted(vcpu, msr_info->index); +} + static int svm_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info) { struct vcpu_svm *svm = to_svm(vcpu); + if (sev_es_prevent_msr_access(vcpu, msr_info)) { + msr_info->data = 0; + return 0; + } + switch (msr_info->index) { case MSR_AMD64_TSC_RATIO: if (!msr_info->host_initiated && @@ -2988,6 +3002,10 @@ static int svm_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr) u32 ecx = msr->index; u64 data = msr->data; + + if (sev_es_prevent_msr_access(vcpu, msr)) + return 0; + switch (ecx) { case MSR_AMD64_TSC_RATIO: