Re: [PATCH v2 05/16] KVM: VMX: Drop initialization of IA32_FEATURE_CONTROL MSR

From: Paolo Bonzini
Date: Fri Nov 15 2019 - 05:10:21 EST


On 14/11/19 19:34, Sean Christopherson wrote:
> On Tue, Oct 22, 2019 at 08:16:22AM -0700, Sean Christopherson wrote:
>> On Tue, Oct 22, 2019 at 12:51:01PM +0200, Paolo Bonzini wrote:
>>> On 22/10/19 02:08, Sean Christopherson wrote:
>>>> Remove the code to initialize IA32_FEATURE_CONTROL MSR when KVM is
>>>> loaded now that the MSR is initialized during boot on all CPUs that
>>>> support VMX, i.e. can possibly load kvm_intel.
>>>>
>>>> Reviewed-by: Jim Mattson <jmattson@xxxxxxxxxx>
>>>> Signed-off-by: Sean Christopherson <sean.j.christopherson@xxxxxxxxx>
>>>> ---
>>>> arch/x86/kvm/vmx/vmx.c | 48 +++++++++++++++++-------------------------
>>>> 1 file changed, 19 insertions(+), 29 deletions(-)
>>>
>>> I am still not sure about this... Enabling VMX is adding a possible
>>> attack vector for the kernel, we should not do it unless we plan to do a
>>> VMXON.
>>
>> An attacker would need arbitrary cpl0 access to toggle CR4.VMXE and do
>> VMXON (and VMLAUNCH), would an extra WRMSR really slow them down?
>>
>> And practically speaking, how often do you encounter systems whose
>> firmware leaves IA32_FEATURE_CONTROL unlocked?

I honestly don't know... always on nested virtualization probably
doesn't count as an answer. :)

>From a totally abstract point of view I like the idea of KVM being an
independent driver and thus the only place that touches VMX stuff
(including the relevant bit in IA32_FEATURE_CONTROL). But I understand
that this doesn't really make any concrete difference, so I guess you
can go ahead with this.

>>> Why is it so important to operate with locked
>>> IA32_FEATURE_CONTROL (so that KVM can enable VMX and the kernel can
>>> still enable SGX if desired).
>>
>> For simplicity. The alternative that comes to mind is to compute the
>> desired MSR value and write/lock the MSR on demand, e.g. add a sequence
>> similar to KVM's hardware_enable_all() for SGX, but that's a fair amount
>> of complexity for marginal benefit (IMO).
>>
>> If a user really doesn't want VMX enabled, they can clear the feature bit
>> via the clearcpuid kernel param.
>>
>> That being said, enabling VMX in IA32_FEATURE_CONTROL if and only if
>> IS_ENABLED(CONFIG_KVM) is true would be an easy enhancement.
>
> Paolo, any follow up thoughts on this approach?

Yes, that would be a simple enhancement, useful at least for
documentation purpose.

Paolo