Re: [RFC v2 2/2] KVM: VMX: Enable bus lock VM exit

[Xiaoyao Li]

On 9/3/2020 6:44 AM, Sean Christopherson wrote:
> On Tue, Sep 01, 2020 at 10:43:12AM +0200, Vitaly Kuznetsov wrote:
> > > @@ -6809,6 +6824,19 @@ static fastpath_t vmx_vcpu_run(struct kvm_vcpu *vcpu)
> > >   	if (unlikely(vmx->exit_reason.failed_vmentry))
> > >   		return EXIT_FASTPATH_NONE;
> > >   
> > > +	/*
> > > +	 * check the exit_reason to see if there is a bus lock
> > > +	 * happened in guest.
> > > +	 */
> > > +	if (kvm_bus_lock_exit_enabled(vmx->vcpu.kvm)) {
> > > +		if (vmx->exit_reason.bus_lock_detected) {
> > > +			vcpu->stat.bus_locks++;
>
> Why bother with stats?  Every bus lock exits to userspace, having quick
> stats doesn't seem all that interesting.

OK. We will remove it.

> > > +			vcpu->arch.bus_lock_detected = true;
> > > +		} else {
> > > +			vcpu->arch.bus_lock_detected = false;
> >
> > This is a fast path so I'm wondering if we can move bus_lock_detected
> > clearing somewhere else.
>
> Why even snapshot vmx->exit_reason.bus_lock_detected?  I don't see any
> reason why vcpu_enter_guest() needs to handle the exit to userspace, e.g.
> it's just as easily handled in VMX code.

Because we want to handle the exit to userspace only in one place, i.e., 
after kvm_x86_ops.handle_exit(vcpu, exit_fastpath). Otherwise, we would 
have to check vmx->exit_reason.bus_lock_detected in every other handler, 
at least in those can preempt the bus lock VM-exit theoretically.

> > > +		}
> > > +	}
> > > +
> > >   	vmx->loaded_vmcs->launched = 1;
> > >   	vmx->idt_vectoring_info = vmcs_read32(IDT_VECTORING_INFO_FIELD);
> > >   
> > > @@ -8060,6 +8088,9 @@ static __init int hardware_setup(void)
> > >   		kvm_tsc_scaling_ratio_frac_bits = 48;
> > >   	}
> > >   
> > > +	if (cpu_has_vmx_bus_lock_detection())
> > > +		kvm_has_bus_lock_exit = true;
> > > +
> > >   	set_bit(0, vmx_vpid_bitmap); /* 0 is reserved for host */
> > >   
> > >   	if (enable_ept)
>
> ...
>
> > > @@ -4990,6 +4996,12 @@ int kvm_vm_ioctl_enable_cap(struct kvm *kvm,
> > >   		kvm->arch.exception_payload_enabled = cap->args[0];
> > >   		r = 0;
> > >   		break;
> > > +	case KVM_CAP_X86_BUS_LOCK_EXIT:
> > > +		if (!kvm_has_bus_lock_exit)
> > > +			return -EINVAL;
> >
> > ... because userspace can check for -EINVAL when enabling the cap. Or we
> > can return e.g. -EOPNOTSUPP here. I don't have a strong opinion on the matter..
> >
> > > +		kvm->arch.bus_lock_exit = cap->args[0];
>
> Assuming we even want to make this per-VM, I think it'd make sense to make
> args[0] a bit mask, e.g. to provide "off" and "exit" (this behavior) while
> allowing for future modes, e.g. log-only.

Good idea, will do it in next version.

> > > +		r = 0;
> > > +		break;
> > >   	default:
> > >   		r = -EINVAL;
> > >   		break;
> > > @@ -7732,12 +7744,23 @@ static void post_kvm_run_save(struct kvm_vcpu *vcpu)