Re: [PATCH 25/25] KVM: x86: Add CPUID bits missing from KVM_GET_SUPPORTED_CPUID

[Chao Gao <chao.gao@intel.com>]

On Wed, Aug 14, 2024 at 02:45:50AM +0800, Edgecombe, Rick P wrote:
>On Tue, 2024-08-13 at 19:34 +0800, Chao Gao wrote:
>> Mandating that all fixed-1 bits be supported by KVM would be a burden for both
>> KVM and the TDX module: the TDX module couldn't add any fixed-1 bits until KVM
>> supports them, and 
>
>> KVM shouldn't drop any feature that was ever a fixed-1 bit
>> in any TDX module.
>
>Honest question...can/does this happen for normal VMs? KVM dropping support for
>features? I think I recall even MPX getting limped along for backward
>compatibility reasons.
>
>>  I don't think this is a good idea. TDX module support for a
>> feature will likely be ready earlier than KVM's, as TDX module is smaller and
>> is developed inside Intel. Requiring the TDX module to avoid adding fixed-1
>> bits doesn't make much sense, as making all features configurable would
>> increase its complexity.
>> 
>> I think adding new fixed-1 bits is fine as long as they don't break KVM, i.e.,
>> KVM shouldn't need to take any action for the new fixed-1 bits, like
>> saving/restoring more host CPU states across TD-enter/exit or emulating
>> CPUID/MSR accesses from guests
>
>If these would only be simple features, then I'd wonder how much complexity
>making them configurable would really add to the TDX module.
>
>I think there are more concerns than just TDX module breaking KVM. (my 2 cents
>would be that it should just be considered a TDX module bug) But KVM should also
>want to avoid getting boxed into some ABI. For example a a new userspace
>developed against a new TDX module, but old KVM could start using some new
>feature that KVM would want to handle differently. As you point out KVM
>implementation could happen later, at which point userspace could already expect
>a certain behavior. Then KVM would have to have some other opt in for it's
>preferred behavior.

I don't fully understand "getting boxed into some ABI". But filtering out
unsupported bits could also cause ABI breakage if those bits later become
supported and are no longer filtered, but userspace may still expect them to be
cleared.

It seems that KVM would have to refuse to work with the TDX module if it
detects some fixed-1/native bits are unsupported/unknown.

But if we do that, IIUC, disabling certain features using the "clearcpuid="
kernel cmdline on the host may cause KVM to be incompatible with the TDX
module. Anyway, this is probably a minor issue.

>
>Now, that is comparing *sometimes* KVM needing to have an opt-in, with TDX
>module *always* needing an opt-in. But I don't see how never having fixed bits
>is more complex for KVM.