Re: [ANNOUNCE] PUCK Notes - 2024.04.03 - TDX Upstreaming Strategy

[Xiaoyao Li <xiaoyao.li@intel.com>]

On 4/9/2024 10:01 PM, Sean Christopherson wrote:
> On Tue, Apr 09, 2024, Xiaoyao Li wrote:
>> On 4/9/2024 12:20 AM, Sean Christopherson wrote:
>>> On Sun, Apr 07, 2024, Xiaoyao Li wrote:
>>>> On 4/6/2024 12:58 AM, Sean Christopherson wrote:
>>>>>     - For guest MAXPHYADDR vs. GPAW, rely on KVM_GET_SUPPORTED_CPUID to enumerate
>>>>>       the usable MAXPHYADDR[2], and simply refuse to enable TDX if the TDX Module
>>>>>       isn't compatible.  Specifically, if MAXPHYADDR=52, 5-level paging is enabled,
>>>>>       but the TDX-Module only allows GPAW=0, i.e. only supports 4-level paging.
>>>>
>>>> So userspace can get supported GPAW from usable MAXPHYADDR, i.e.,
>>>> CPUID(0X8000_0008).eaxx[23:16] of KVM_GET_SUPPORTED_CPUID:
>>>>    - if usable MAXPHYADDR == 52, supported GPAW is 0 and 1.
>>>>    - if usable MAXPHYADDR <= 48, supported GPAW is only 0.
>>>>
>>>> There is another thing needs to be discussed. How does userspace configure
>>>> GPAW for TD guest?
>>>>
>>>> Currently, KVM uses CPUID(0x8000_0008).EAX[7:0] in struct
>>>> kvm_tdx_init_vm::cpuid.entries[] of IOCTL(KVM_TDX_INIT_VM) to deduce the
>>>> GPAW:
>>>>
>>>> 	int maxpa = 36;
>>>> 	entry = kvm_find_cpuid_entry2(cpuid->entries, cpuid->nent, 0x80000008, 0);
>>>> 	if (entry)
>>>> 		max_pa = entry->eax & 0xff;
>>>>
>>>> 	...
>>>> 	if (!cpu_has_vmx_ept_5levels() && max_pa > 48)
>>>> 		return -EINVAL;
>>>> 	if (cpu_has_vmx_ept_5levels() && max_pa > 48) {
>>>> 		td_params->eptp_controls |= VMX_EPTP_PWL_5;
>>>> 		td_params->exec_controls |= TDX_EXEC_CONTROL_MAX_GPAW;
>>>> 	} else {
>>>> 		td_params->eptp_controls |= VMX_EPTP_PWL_4;
>>>> 	}
>>>>
>>>> The code implies that KVM allows the provided CPUID(0x8000_0008).EAX[7:0] to
>>>> be any value (when 5level ept is supported). when it > 48, configure GPAW of
>>>> TD to 1, otherwise to 0.
>>>>
>>>> However, the virtual value of CPUID(0x8000_0008).EAX[7:0] inside TD is
>>>> always the native value of hardware (for current TDX).
>>>>
>>>> So if we want to keep this behavior, we need to document it somewhere that
>>>> CPUID(0x8000_0008).EAX[7:0] in struct kvm_tdx_init_vm::cpuid.entries[] of
>>>> IOCTL(KVM_TDX_INIT_VM) is only for configuring GPAW, not for userspace to
>>>> configure virtual CPUID value for TD VMs.
>>>>
>>>> Another option is that, KVM doesn't allow userspace to configure
>>>> CPUID(0x8000_0008).EAX[7:0]. Instead, it provides a gpaw field in struct
>>>> kvm_tdx_init_vm for userspace to configure directly.
>>>>
>>>> What do you prefer?
>>>
>>> Hmm, neither.  I think the best approach is to build on Gerd's series to have KVM
>>> select 4-level vs. 5-level based on the enumerated guest.MAXPHYADDR, not on
>>> host.MAXPHYADDR.
>>
>> I see no difference between using guest.MAXPHYADDR (EAX[23:16]) and using
>> host.MAXPHYADDR (EAX[7:0]) to determine the GPAW (and EPT level) for TD
>> guest. The case for TDX diverges from what for non TDX VMs. The value of
>> them passed from userspace can only be used to configure GPAW and EPT level
>> for TD, but won't be reflected in CPUID inside TD.
> 
> But the TDX module will emulate EAX[7:0] to match hardware, no?  Whenever possible,
> the CPUID entries passed to KVM should match the CPUID values that are observed
> by the guest.  E.g. if host.MAXPHYADDR=52, but the CPU only supports 4-level
> paging, then KVM should get host.MAXPHYADDR=52, guest.MAXPHYADDR=48.

side topic: Do we expect KVM to check the input of EAX[7:0] to match 
with hardware value? or a zero value? or both are allowed?

> As I said in my response to Rick:
> 
>   : > An alternative would be to have the KVM API peak at the value, and then
>   : > discard it (not pass the leaf value to the TDX module). Not ideal.
>   :
>   : Heh, I typed up this idea before reading ahead.  This has my vote.  Unless I'm
>   : misreading where things are headed, using guest.MAXPHYADDR to communicate what
>   : is essentially GPAW to the guest is about to become the de facto standard.
>   :
>   : At that point, KVM can basically treat the current TDX module behavior as an
>   : erratum, i.e. discarding guest.MAXPHYADDR becomes a workaround for a "CPU" bug,
>   : not some goofy KVM quirk.

yes, bit [23:16] fits better.