Re: [PATCH 4/4] x86,module: Detect CRn and DRn manipulation

[Nadav Amit]

> On Apr 7, 2020, at 1:50 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> 
> On Tue, Apr 07, 2020 at 01:27:45PM -0700, Nadav Amit wrote:
>>> On Apr 7, 2020, at 12:38 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>>> 
>>> On Tue, Apr 07, 2020 at 11:55:21AM -0700, Nadav Amit wrote:
>>>>> On Apr 7, 2020, at 4:02 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>>>>> 
>>>>> Since we now have infrastructure to analyze module text, disallow
>>>>> modules that write to CRn and DRn registers.
>>>> 
>>>> Assuming the kernel is built without CONFIG_PARAVIRT, what is the right way
>>>> for out-of-tree modules to write to CRs? Letâs say CR2?
>>> 
>>> Most of them there is no real justification for ever writing to. CR2 I
>>> suppose we can have an exception for given a sane rationale for why
>>> you'd need to rewrite the fault address.
>> 
>> For the same reason that KVM writes to CR2 - to restore CR2 before entering
>> a guest, since CR2 not architecturally loaded from the VMCS. I suspect there
>> are additional use-cases which are not covered by the kernel interfaces.
> 
> So I'm not much of a virt guy (clearly), and *groan*, that's horrible.
> I'll go make an exception for CR2.

Clearly you are not a virt guy if you think that this is the horrible part
in x86 virtualization ;-)

Anyhow, I do not think it is the only use-case which is not covered by your
patches (even considering CRs/DRs alone). For example, there is no kernel
function to turn on CR4.VMXE, which is required to run hypervisors on x86.

I think a thorough analysis of existing software is needed to figure out
which use-cases are valid, and to exclude them during module scanning or to
provide alternative kernel interfaces to enable them. This may require a
transition phase in which module scanning would only issue warnings and
would not prevent the module from being loaded.