Re: [edk2-devel] [PATCH V12 42/47] OvmfPkg: Add TdxDxe driver

["Lendacky, Thomas" <thomas.lendacky@amd.com>]

On 4/16/22 19:56, Xu, Min M wrote:
> On April 16, 2022 11:09 PM, Lendacky, Thomas wrote:
>> On 4/15/22 20:57, Xu, Min M wrote:
>>> On April 16, 2022 4:52 AM, Lendacky, Thomas wrote:
>>>>
>>>> Unfortunately, this driver also breaks SEV-ES. I bypassed the TDX
>>>> code in the SEC library, but then hit an issue because this driver is
>>>> loaded before the AmdSevDxe driver. The AmdSevDxe driver performs a
>>>> MemEncryptSevClearMmioPageEncMask() call against the
>>>> PcdPciExpressBaseAddress range to mark it shared/unencrypted.
>>>> However, the TdxDxe driver is loaded before the AmdSevDxe driver, and
>>>> it appears the dependencies result in an MMIO being performed to an
>>>> address in the PcdPciExpressBaseAddress range. Since the range has
>>>> not been marked shared/unencrypted, the #VC handler terminates the
>>>> guest for trying to do MMIO to an encrypted region.
>>>>
>>> I carefully check the code TdxDxeEntryPoint@TdxDxe.c.
>>> If the working guest is NOT td guest, before it returns, it just does below:
>>> 1. check if the GuidHob exists
>>> 2. Set PcdOvmfHostBridgePciDevId with the information in the GuidHob
>>>
>>> SetMmioSharedBit() is called if the working guest is Td guest. So if it is sev
>> guest, SetMmioSharedBit will not be called.
>>>
>>> I don't have a SEV-ES in hand. Can you help to add some debug
>> information in TdxDxe to see what the last code before the exception is
>> triggered?
>>
>> I don't think it is anything in your code, I think it is another library that is
>> being loaded based on dependencies. I put a DEBUG statement at the start
>> of TdxDxeEntryPoint() and never see the output before the crash.
>>
> I check the libraries loaded by TdxDxe and AmdSev and find that they load different PciLib.
> TdxDxe load PciLib|OvmfPkg/Library/DxePciLibI440FxQ35/DxePciLibI440FxQ35.inf.
> AmdSev load PciLib|MdePkg/Library/BasePciLibCf8/BasePciLibCf8.inf.
> 
> PciLib is consumed by DxeAcpiTimerLib. In the AcpiTimerLibConstructor@DxeAcpiTimerLib there is below code:
> mAcpiTimerIoAddr = (PciRead32 (Pmba) & ~PMBA_RTE) + ACPI_TIMER_OFFSET;
> 
> I think this is the root cause of the exception.
> 
> There are 2 options to fix this issue.
> 1. Load AmdSev before TdxDxe
> 2. Make TdxDxe to import BasePciLibCf8.inf instead of DxePciLibI440FxQ35.inf (just like AmdSev)
> 
> I tried above 2 options in my Tdx guest and both work.
> Tom, Can you help to try above 2 options in your SEV guest to see whether they work?
> 
>>
>>>
>>> BTW, have you tried to load AmdSev.inf before TdxDxe.inf? I tried it in my
>> TDX guest and it works fine.
>>
>> Yes, moving AmdSevDxe.inf ahead of TdxDxe.inf does fix this issue. Do you
>> want to submit the patch or do you want me to?
>>
> If above option 2 works, I prefer this option to fix the issue. Because there is still potential issues in option 1. I will submit the patch.

I added the same library class override to TdxDxe and, yes, option 2 worked.

Thanks,
Tom

> 
> Thanks
> Min