Re: [PATCH 2/2] KVM: x86: Fix INIT signal handling in various CPU states

[Liran Alon <liran.alon@oracle.com>]

> On 27 Aug 2019, at 0:17, Jim Mattson <jmattson@google.com> wrote:
> 
> On Mon, Aug 26, 2019 at 3:25 AM Liran Alon <liran.alon@oracle.com> wrote:
>> 
>> Commit cd7764fe9f73 ("KVM: x86: latch INITs while in system management mode")
>> changed code to latch INIT while vCPU is in SMM and process latched INIT
>> when leaving SMM. It left a subtle remark in commit message that similar
>> treatment should also be done while vCPU is in VMX non-root-mode.
>> 
>> However, INIT signals should actually be latched in various vCPU states:
>> (*) For both Intel and AMD, INIT signals should be latched while vCPU
>> is in SMM.
>> (*) For Intel, INIT should also be latched while vCPU is in VMX
>> operation and later processed when vCPU leaves VMX operation by
>> executing VMXOFF.
>> (*) For AMD, INIT should also be latched while vCPU runs with GIF=0
>> or in guest-mode with intercept defined on INIT signal.
>> 
>> To fix this:
>> 1) Add kvm_x86_ops->apic_init_signal_blocked() such that each CPU vendor
>> can define the various CPU states in which INIT signals should be
>> blocked and modify kvm_apic_accept_events() to use it.
>> 2) Modify vmx_check_nested_events() to check for pending INIT signal
>> while vCPU in guest-mode. If so, emualte vmexit on
>> EXIT_REASON_INIT_SIGNAL. Note that nSVM should have similar behaviour
>> but is currently left as a TODO comment to implement in the future
>> because nSVM don't yet implement svm_check_nested_events().
>> 
>> Note: Currently KVM nVMX implementation don't support VMX wait-for-SIPI
>> activity state as specified in MSR_IA32_VMX_MISC bits 6:8 exposed to
>> guest (See nested_vmx_setup_ctls_msrs()).
>> If and when support for this activity state will be implemented,
>> kvm_check_nested_events() would need to avoid emulating vmexit on
>> INIT signal in case activity-state is wait-for-SIPI. In addition,
>> kvm_apic_accept_events() would need to be modified to avoid discarding
>> SIPI in case VMX activity-state is wait-for-SIPI but instead delay
>> SIPI processing to vmx_check_nested_events() that would clear
>> pending APIC events and emulate vmexit on SIPI.
>> 
>> Reviewed-by: Joao Martins <joao.m.martins@oracle.com>
>> Co-developed-by: Nikita Leshenko <nikita.leshchenko@oracle.com>
>> Signed-off-by: Nikita Leshenko <nikita.leshchenko@oracle.com>
>> Signed-off-by: Liran Alon <liran.alon@oracle.com>
> 
> Thanks for doing this! I asked Marc to take a look at it earlier this
> month, but he's been swamped.

No problem. BTW Nikita is also preparing a patch on top of this to support wait-for-SIPI activity-state.
In case that was in your TODO list as-well.

> 
>> ---
>> arch/x86/include/asm/kvm_host.h |  2 ++
>> arch/x86/kvm/lapic.c            | 11 +++++++----
>> arch/x86/kvm/svm.c              | 20 ++++++++++++++++++++
>> arch/x86/kvm/vmx/nested.c       | 14 ++++++++++++++
>> arch/x86/kvm/vmx/vmx.c          |  6 ++++++
>> 5 files changed, 49 insertions(+), 4 deletions(-)
>> 
>> diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
>> index 74e88e5edd9c..158483538181 100644
>> --- a/arch/x86/include/asm/kvm_host.h
>> +++ b/arch/x86/include/asm/kvm_host.h
>> @@ -1209,6 +1209,8 @@ struct kvm_x86_ops {
>>        uint16_t (*nested_get_evmcs_version)(struct kvm_vcpu *vcpu);
>> 
>>        bool (*need_emulation_on_page_fault)(struct kvm_vcpu *vcpu);
>> +
>> +       bool (*apic_init_signal_blocked)(struct kvm_vcpu *vcpu);
>> };
>> 
>> struct kvm_arch_async_pf {
>> diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
>> index 685d17c11461..9620fe5ce8d1 100644
>> --- a/arch/x86/kvm/lapic.c
>> +++ b/arch/x86/kvm/lapic.c
>> @@ -2702,11 +2702,14 @@ void kvm_apic_accept_events(struct kvm_vcpu *vcpu)
>>                return;
>> 
>>        /*
>> -        * INITs are latched while in SMM.  Because an SMM CPU cannot
>> -        * be in KVM_MP_STATE_INIT_RECEIVED state, just eat SIPIs
>> -        * and delay processing of INIT until the next RSM.
>> +        * INITs are latched while CPU is in specific states.
>> +        * Because a CPU cannot be in these states immediately
>> +        * after it have processed an INIT signal (and thus in
>> +        * KVM_MP_STATE_INIT_RECEIVED state), just eat SIPIs
>> +        * and delay processing of INIT until CPU leaves
>> +        * the state which latch INIT signal.
>>         */
>> -       if (is_smm(vcpu)) {
>> +       if (kvm_x86_ops->apic_init_signal_blocked(vcpu)) {
>>                WARN_ON_ONCE(vcpu->arch.mp_state == KVM_MP_STATE_INIT_RECEIVED);
>>                if (test_bit(KVM_APIC_SIPI, &apic->pending_events))
>>                        clear_bit(KVM_APIC_SIPI, &apic->pending_events);
>> diff --git a/arch/x86/kvm/svm.c b/arch/x86/kvm/svm.c
>> index 6462c386015d..0e43acf7bea4 100644
>> --- a/arch/x86/kvm/svm.c
>> +++ b/arch/x86/kvm/svm.c
>> @@ -7205,6 +7205,24 @@ static bool svm_need_emulation_on_page_fault(struct kvm_vcpu *vcpu)
>>        return false;
>> }
>> 
>> +static bool svm_apic_init_signal_blocked(struct kvm_vcpu *vcpu)
>> +{
>> +       struct vcpu_svm *svm = to_svm(vcpu);
>> +
>> +       /*
>> +        * TODO: Last condition latch INIT signals on vCPU when
>> +        * vCPU is in guest-mode and vmcb12 defines intercept on INIT.
>> +        * To properly emulate INIT intercept, SVM should also implement
>> +        * kvm_x86_ops->check_nested_events() and process pending INIT
>> +        * signal to cause nested_svm_vmexit(). As SVM currently don't
>> +        * implement check_nested_events(), this work is delayed
>> +        * for future improvement.
>> +        */
>> +       return is_smm(vcpu) ||
>> +                  !gif_set(svm) ||
>> +                  (svm->vmcb->control.intercept & (1ULL << INTERCEPT_INIT));
>> +}
>> +
>> static struct kvm_x86_ops svm_x86_ops __ro_after_init = {
>>        .cpu_has_kvm_support = has_svm,
>>        .disabled_by_bios = is_disabled,
>> @@ -7341,6 +7359,8 @@ static struct kvm_x86_ops svm_x86_ops __ro_after_init = {
>>        .nested_get_evmcs_version = nested_get_evmcs_version,
>> 
>>        .need_emulation_on_page_fault = svm_need_emulation_on_page_fault,
>> +
>> +       .apic_init_signal_blocked = svm_apic_init_signal_blocked,
>> };
>> 
>> static int __init svm_init(void)
>> diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
>> index ced9fba32598..d655fcd04c01 100644
>> --- a/arch/x86/kvm/vmx/nested.c
>> +++ b/arch/x86/kvm/vmx/nested.c
>> @@ -3401,6 +3401,15 @@ static int vmx_check_nested_events(struct kvm_vcpu *vcpu, bool external_intr)
>>        unsigned long exit_qual;
>>        bool block_nested_events =
>>            vmx->nested.nested_run_pending || kvm_event_needs_reinjection(vcpu);
>> +       struct kvm_lapic *apic = vcpu->arch.apic;
>> +
>> +       if (lapic_in_kernel(vcpu) &&
>> +               test_bit(KVM_APIC_INIT, &apic->pending_events)) {
>> +               if (block_nested_events)
>> +                       return -EBUSY;
>> +               nested_vmx_vmexit(vcpu, EXIT_REASON_INIT_SIGNAL, 0, 0);
>> +               return 0;
>> +       }
> 
> Suppose that L0 just finished emulating an L2 instruction with
> EFLAGS.TF set. So, we've just queued up a #DB trap in
> vcpu->arch.exception. On this emulated VM-exit from L2 to L1, the
> guest pending debug exceptions field in the vmcs12 should get the
> value of vcpu->arch.exception.payload, and the queued #DB should be
> squashed.

If I understand correctly you are discussing a case where L2 exited to L0 for
emulating some instruction when L2’s RFLAGS.TF is set. Therefore, after x86
emulator finished emulating L2 instruction, it queued a #DB exception.

Then before resuming L2 guest, some other vCPU send an INIT signal
to this vCPU. When L0 will reach vmx_check_nested_events() it will
see pending INIT signal and exit on EXIT_REASON_INIT_SIGNAL
but nested_vmx_vmexit() will basically drop pending #DB exception
in prepare_vmcs12() when it calls kvm_clear_exception_queue()
because vmcs12_save_pending_event() only saves injected exceptions.
(As changed by myself a long time ago)

I think you are right this is a bug.
I also think it could trivially be fixed by just making sure vmx_check_nested_events()
first evaluates pending exceptions and only then pending apic events.
However, we also want to make sure to request an “immediate-exit” in case
eval of pending exception require emulation of an exit from L2 to L1.

I will add this scenario to my kvm-unit-tests that I’m about to submit.

Thanks,
-Liran

> 
>>        if (vcpu->arch.exception.pending &&
>>                nested_vmx_check_exception(vcpu, &exit_qual)) {
>> @@ -4466,7 +4475,12 @@ static int handle_vmoff(struct kvm_vcpu *vcpu)
>> {
>>        if (!nested_vmx_check_permission(vcpu))
>>                return 1;
>> +
>>        free_nested(vcpu);
>> +
>> +       /* Process a latched INIT during time CPU was in VMX operation */
>> +       kvm_make_request(KVM_REQ_EVENT, vcpu);
>> +
>>        return nested_vmx_succeed(vcpu);
>> }
>> 
>> diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
>> index b5b5b2e5dac5..5a1aa0640f2a 100644
>> --- a/arch/x86/kvm/vmx/vmx.c
>> +++ b/arch/x86/kvm/vmx/vmx.c
>> @@ -7479,6 +7479,11 @@ static bool vmx_need_emulation_on_page_fault(struct kvm_vcpu *vcpu)
>>        return false;
>> }
>> 
>> +static bool vmx_apic_init_signal_blocked(struct kvm_vcpu *vcpu)
>> +{
>> +       return is_smm(vcpu) || to_vmx(vcpu)->nested.vmxon;
>> +}
>> +
>> static __init int hardware_setup(void)
>> {
>>        unsigned long host_bndcfgs;
>> @@ -7803,6 +7808,7 @@ static struct kvm_x86_ops vmx_x86_ops __ro_after_init = {
>>        .get_vmcs12_pages = NULL,
>>        .nested_enable_evmcs = NULL,
>>        .need_emulation_on_page_fault = vmx_need_emulation_on_page_fault,
>> +       .apic_init_signal_blocked = vmx_apic_init_signal_blocked,
>> };
>> 
>> static void vmx_cleanup_l1d_flush(void)
>> --
>> 2.20.1
>>