Re: [edk2-devel] [PATCH RFC v2 03/28] MdePkg: Define the GHCB GPA structure

["Brijesh Singh" <brijesh.singh@amd.com>]

On 5/3/21 7:19 AM, Laszlo Ersek wrote:
> On 05/03/21 12:24, Laszlo Ersek wrote:
>> On 04/30/21 13:51, Brijesh Singh wrote:
>>> BZ: https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fbugzilla.tianocore.org%2Fshow_bug.cgi%3Fid%3D3275&data=04%7C01%7Cbrijesh.singh%40amd.com%7C9eac9a93753d403dcc4d08d90e2dbcb5%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637556411874265560%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=eNafEGfhCMOkOboQOJnxq8Rw%2BOTuvAUGIziDuELV8%2Bk%3D&reserved=0
>>>
>>> An SEV-SNP guest is required to perform the GHCB GPA registration. See
>>> the GHCB specification for further details.
>>>
>>> Cc: James Bottomley <jejb@linux.ibm.com>
>>> Cc: Min Xu <min.m.xu@intel.com>
>>> Cc: Jiewen Yao <jiewen.yao@intel.com>
>>> Cc: Tom Lendacky <thomas.lendacky@amd.com>
>>> Cc: Jordan Justen <jordan.l.justen@intel.com>
>>> Cc: Ard Biesheuvel <ardb+tianocore@kernel.org>
>>> Cc: Laszlo Ersek <lersek@redhat.com>
>>> Cc: Erdem Aktas <erdemaktas@google.com>
>>> Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
>>> ---
>>>  MdePkg/Include/Register/Amd/Fam17Msr.h | 7 +++++++
>>>  1 file changed, 7 insertions(+)
>>>
>>> diff --git a/MdePkg/Include/Register/Amd/Fam17Msr.h b/MdePkg/Include/Register/Amd/Fam17Msr.h
>>> index a65d51ab12..e19bd04b6c 100644
>>> --- a/MdePkg/Include/Register/Amd/Fam17Msr.h
>>> +++ b/MdePkg/Include/Register/Amd/Fam17Msr.h
>>> @@ -53,6 +53,11 @@ typedef union {
>>>      UINT64  Features:52;
>>>    } GhcbHypervisorFeatures;
>>>  
>>> +  struct {
>>> +    UINT64  Function:12;
>>> +    UINT64  GuestFrameNumber:52;
>>> +  } GhcbGpaRegister;
>>> +
>>>    VOID    *Ghcb;
>>>  
>>>    UINT64  GhcbPhysicalAddress;
>>> @@ -62,6 +67,8 @@ typedef union {
>>>  #define GHCB_INFO_SEV_INFO_GET             2
>>>  #define GHCB_INFO_CPUID_REQUEST            4
>>>  #define GHCB_INFO_CPUID_RESPONSE           5
>>> +#define GHCB_INFO_GHCB_GPA_REGISTER_REQUEST   18
>>> +#define GHCB_INFO_GHCB_GPA_REGISTER_RESPONSE  19
>>>  #define GHCB_HYPERVISOR_FEATURES_REQUEST   128
>>>  #define GHCB_HYPERVISOR_FEATURES_RESPONSE  129
>>>  #define GHCB_INFO_TERMINATE_REQUEST        256
>>>
>> The number match the spec (2.0), but I have some remarks / questions.
>>
>> (1) Patch #2 (SVM_EXIT_HYPERVISOR_FEATURES) and this patch
>> (GHCB_INFO_GHCB_GPA_REGISTER_REQUEST) break the nice alignments of the
>> macro values (replacement texts) in both header files. Can you prepend a
>> whitespace-only patch that simply moves the affected "columns" to the
>> right far enough?

Sure, do you want me to the post after all the new VMGEXIT's are defined ?


>>
>> (2) I've checked section 2.3.2 "GHCB GPA Registration" in the spec
>> (2.0). What is the specific risk of allowing a guest to switch from one
>> GHCB address to another?

The GHCB is a shared page, there is no risk to switch from one page to
another. This feature is designed to simplify some of the hypervisor
implementation. Since the GHCB is accessed on every vmgexit, a
hypervisor may prefer to create a map during the registration and refer
the map instead of creating a new mapping on every vmgexit.


>>
>> (3) It seems strange to expect that a guest stick with a particular GHCB
>> address for its entire lifetime (including firmware and OS) -- in fact
>> OVMF already uses multiple GHCB addresses. The spec does not explain how
>> the guest can "unlock" (de-register) a registered GHCB address.
>> Furthermore, if a guest can do that *at all* (which I think it must --
>> we're already using different GHCB addresses between SEC and DXE, for
>> example), then what protection does the *temporary* locking of the GHCB
>> address provide?

The spec does not force that GHCB should *never* change once registered.
It says that before switching to new GHCB page, the guest must register
the page. As you rightly said that OVMF uses multiple GHCBs from SEC to
DXE. There is no unregister, registering a new GHCB is a hint to
hypervisor that it should drop the old GHCB mapping. The GHCB
registration is not a PSP function, and are not designed to mitigate a
security exploits. It is purely a hypevisor virtualized feature.


>> I'll stop reviewing here, because I think I need to understand your
>> answers. I'd like to have a rudimentary mental basis for reviewing the rest.
> ... interestingly, with reference to my question (2) under patch "RFC v2
> 02/28", the GHCB GPA registration function is one that can *only* be
> performed with the GHCB MSR protocol, and not through the GHCB page.
>
> So that shows that the MSR protocol's functions cannot be considered a
> pure subset of the GHCB page's functions. If
> SVM_EXIT_HYPERVISOR_FEATURES didn't exist (and the same function would
> only be accessible via GHCB_HYPERVISOR_FEATURES_REQUEST), then no
> "larger principle" would be damaged.

That is correct, not every exit have both MSR and non MSR protocol based
vmgexit. It seems that during the spec review no other HV vendor saw the
need for non-MSR based exit. Certainly, I don't see a need for it in KVM
and can't comment on other HV ;)


>
> Thanks
> Laszlo
>