Re: [PATCH] KVM: VMX: expose the host's ARCH_CAPABILITIES MSR to userspace

["Radim Krčmář" <rkrcmar@redhat.com>]

2018-03-07 12:53+0100, Paolo Bonzini:
> On 02/03/2018 22:42, Radim Krčmář wrote:
> > Ok, sounds good.  I've deferred it to rc5 as I think we'll want to use
> > this to replace the auto setting:  I would not bet that it is going to
> > be safe to expose future bits, so having the userspace always sanitize
> > the capabilities would be safer (and more in line with what we do with
> > other MSRs).  i.e. this patch would also
> > 
> > diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
> > index 051dab74e4e9..86ea4a83af1f 100644
> > --- a/arch/x86/kvm/vmx.c
> > +++ b/arch/x86/kvm/vmx.c
> > @@ -5740,9 +5740,6 @@ static void vmx_vcpu_setup(struct vcpu_vmx *vmx)
> >  		++vmx->nmsrs;
> >  	}
> >  
> > -	if (boot_cpu_has(X86_FEATURE_ARCH_CAPABILITIES))
> > -		rdmsrl(MSR_IA32_ARCH_CAPABILITIES, vmx->arch_capabilities);
> > -
> >  	vm_exit_controls_init(vmx, vmcs_config.vmexit_ctrl);
> >  
> >  	/* 22.2.1, 20.8.1 */
> 
> I don't know. There are good reasons for both behaviors, and especially
> the following two for _not_ removing the rdmsr:
> 
> 1) so far you could just pass the result of KVM_GET_SUPPORTED_CPUID to
> KVM_SET_CPUID2, and expect the result to be "as close as possible to the
> host";
>
> 2) having different behavior for VMX and ARCH_CAPABILITIES MSRs would be
> confusing.

Right, we can't just stop setting them by default ...

(I am in favor of forcing the userspace to configure everything and I'd
 accept this exception as a mistake of the past.)

> I think I'm leaning more towards the following direction: whitelist
> ARCH_CAPABILITIES, like we do for the AMD LFENCE MSR already, and
> default the AMD LFENCE MSR to the host value.

The whitelisting is a good idea and I'm ok with just that, thanks.

The MSR_F10H_DECFG default is questionable -- MSR_F10H_DECFG is an
architectural MSR, so we'd be changing the guest under the sight of
existing userspaces.
A potential security risk if they migrate the guest to a CPU that
doesn't serialize LFENCE.  ARCH_CAPABILITIES are at least hidden by a
new CPUID bit.

The feature MSR defaults are going to be a mess anyway: we have
MSR_IA32_UCODE_REV that is tightly coupled with CPUID.  Not a good
candidate for passing by default and currently also has a default value.