separate OVMF binary for TDX? [was: OvmfPkg: Reserve the Secrets and Cpuid page for the SEV-SNP guest]


Laszlo Ersek
 

On 04/09/21 15:44, Yao, Jiewen wrote:
Hi Laszlo
Thanks.

We did provide a separate binary in the beginning - see https://github.com/tianocore/edk2-staging/tree/TDVF, with same goal - easy to maintain and develop. A clean solution, definitely.

However, we got requirement to deliver one binary solution together with 1) normal OVMF, 2) AMD-SEV, 3) Intel-TDX.
Now, we are struggling to merge them......

For DXE, we hope to isolate TDX driver whenever it is possible.
But we only have one reset vector here. Sigh...
Can we please pry a little bit at that "one binary" requirement?

Ultimately the "guest bundle" is going to be composed by much
higher-level code, I expect (such as some userspace code, written in
python or similar); selecting a firmware binary in such an environment
is surely easier than handling this "polymorphism" in the most
restrictive software environment imaginable (reset vector assembly code
in the guest)?

Thanks
Laszlo


Dr. David Alan Gilbert
 

* Laszlo Ersek (lersek@redhat.com) wrote:
On 04/09/21 15:44, Yao, Jiewen wrote:
Hi Laszlo
Thanks.

We did provide a separate binary in the beginning - see https://github.com/tianocore/edk2-staging/tree/TDVF, with same goal - easy to maintain and develop. A clean solution, definitely.

However, we got requirement to deliver one binary solution together with 1) normal OVMF, 2) AMD-SEV, 3) Intel-TDX.
Now, we are struggling to merge them......

For DXE, we hope to isolate TDX driver whenever it is possible.
But we only have one reset vector here. Sigh...
Can we please pry a little bit at that "one binary" requirement?

Ultimately the "guest bundle" is going to be composed by much
higher-level code, I expect (such as some userspace code, written in
python or similar); selecting a firmware binary in such an environment
is surely easier than handling this "polymorphism" in the most
restrictive software environment imaginable (reset vector assembly code
in the guest)?
I think also there's a security argument here; some people like to
measure security in kloc's; so having your secure boot image as small
as possible for the environment you're actually running does make some
sense, which favours the 2 image idea.

Dave

Thanks
Laszlo
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK


Yao, Jiewen
 

I totally agree with you that from security perspective, the best idea to isolate AMD SEV/Intel TDX from standard OVMF.

Do you want to propose move AMD SEV support to another SEC?

-----Original Message-----
From: devel@edk2.groups.io <devel@edk2.groups.io> On Behalf Of Dr. David
Alan Gilbert
Sent: Monday, April 12, 2021 4:35 PM
To: Laszlo Ersek <lersek@redhat.com>
Cc: Yao, Jiewen <jiewen.yao@intel.com>; Xu, Min M <min.m.xu@intel.com>;
devel@edk2.groups.io; thomas.lendacky@amd.com; jejb@linux.ibm.com;
Brijesh Singh <brijesh.singh@amd.com>; Justen, Jordan L
<jordan.l.justen@intel.com>; Ard Biesheuvel <ardb+tianocore@kernel.org>;
Paolo Bonzini <pbonzini@redhat.com>; Nathaniel McCallum
<npmccallum@redhat.com>
Subject: Re: [edk2-devel] separate OVMF binary for TDX? [was: OvmfPkg:
Reserve the Secrets and Cpuid page for the SEV-SNP guest]

* Laszlo Ersek (lersek@redhat.com) wrote:
On 04/09/21 15:44, Yao, Jiewen wrote:
Hi Laszlo
Thanks.

We did provide a separate binary in the beginning - see
https://github.com/tianocore/edk2-staging/tree/TDVF, with same goal - easy to
maintain and develop. A clean solution, definitely.

However, we got requirement to deliver one binary solution together with 1)
normal OVMF, 2) AMD-SEV, 3) Intel-TDX.
Now, we are struggling to merge them......

For DXE, we hope to isolate TDX driver whenever it is possible.
But we only have one reset vector here. Sigh...
Can we please pry a little bit at that "one binary" requirement?

Ultimately the "guest bundle" is going to be composed by much
higher-level code, I expect (such as some userspace code, written in
python or similar); selecting a firmware binary in such an environment
is surely easier than handling this "polymorphism" in the most
restrictive software environment imaginable (reset vector assembly code
in the guest)?
I think also there's a security argument here; some people like to
measure security in kloc's; so having your secure boot image as small
as possible for the environment you're actually running does make some
sense, which favours the 2 image idea.

Dave

Thanks
Laszlo
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK





James Bottomley
 

On Mon, 2021-04-12 at 11:54 +0000, Yao, Jiewen wrote:
I totally agree with you that from security perspective, the best
idea to isolate AMD SEV/Intel TDX from standard OVMF.
There's a big difference between building tuned binaries and separating
the subsystems entirely. Ideally we don't want customers running
images to have to build them differently for Intel or AMD (a bit like
how QEMU/KVM hide the VM differences from users), and Confidential
Computing shares a huge amount of interface similarity, so we wouldn't
want that separated. I think the rule should be that if both Intel and
AMD expose a feature in different ways, OVMF tries to expose a uniform
API for that feature over two differing implementations.

Do you want to propose move AMD SEV support to another SEC?
You mean have an entirely separate SEC for AMD, OVMF and Intel? I
really wouldn't do that: much that's in the SEC: page table setup,
memory mapping and decompression is common to all of them. This all
follows for a lot of the components.

To build separate binaries, we just need separate dsc and fdf files.
Then I think the goal would be to share as much as possible to avoid
duplicating the maintenance and possibly diverging the user API.

James


Erdem Aktas
 

Hi all,

Can we please pry a little bit at that "one binary" requirement?
I think when we call it a "one binary" requirement, it sounds like we
are asking something new but what we are asking is pretty much
captured by James Bottomley.
We do not want to generate different binaries for AMD, Intel, Intel
with TDX, AMD with SEV/SNP etc therefore we were expecting the TDX
changes to be part of the upstream code.
Of course there can be tuning or customization for specific use cases
but those are all future and product specific questions. I just do not
see the reason why the upstreamed code should have a limitation of not
being able to generate a single binary for the same architecture.

-Erdem

On Mon, Apr 12, 2021 at 7:34 AM James Bottomley <jejb@linux.ibm.com> wrote:

On Mon, 2021-04-12 at 11:54 +0000, Yao, Jiewen wrote:
I totally agree with you that from security perspective, the best
idea to isolate AMD SEV/Intel TDX from standard OVMF.
There's a big difference between building tuned binaries and separating
the subsystems entirely. Ideally we don't want customers running
images to have to build them differently for Intel or AMD (a bit like
how QEMU/KVM hide the VM differences from users), and Confidential
Computing shares a huge amount of interface similarity, so we wouldn't
want that separated. I think the rule should be that if both Intel and
AMD expose a feature in different ways, OVMF tries to expose a uniform
API for that feature over two differing implementations.

Do you want to propose move AMD SEV support to another SEC?
You mean have an entirely separate SEC for AMD, OVMF and Intel? I
really wouldn't do that: much that's in the SEC: page table setup,
memory mapping and decompression is common to all of them. This all
follows for a lot of the components.

To build separate binaries, we just need separate dsc and fdf files.
Then I think the goal would be to share as much as possible to avoid
duplicating the maintenance and possibly diverging the user API.

James







Paolo Bonzini
 

On 15/04/21 01:34, Erdem Aktas wrote:
We do not want to generate different binaries for AMD, Intel, Intel
with TDX, AMD with SEV/SNP etc
My question is why the user would want a single binary for VMs with and without TDX/SNP. I know there is attestation, but why would you even want the _possibility_ that your guest starts running without TDX or SNP protection, and only find out later via attestation?

For a similar reason, OVMF already supports shipping a binary that fails to boot if SMM is not available to the firmware, because then secure boot would be trivially circumvented.

I can understand having a single binary for both TDX or SNP. That's not a problem since you can set up the SEV startup VMSA to 32-bit protected mode just like TDX wants.

therefore we were expecting the TDX
changes to be part of the upstream code.
Having 1 or more binaries should be unrelated to the changes being upstream (or more likely, I am misunderstanding you).

Thanks,

Paolo


Erdem Aktas
 

Thanks Paolo.

On Thu, Apr 15, 2021 at 12:59 AM Paolo Bonzini <pbonzini@redhat.com> wrote:

On 15/04/21 01:34, Erdem Aktas wrote:
We do not want to generate different binaries for AMD, Intel, Intel
with TDX, AMD with SEV/SNP etc
My question is why the user would want a single binary for VMs with and
without TDX/SNP. I know there is attestation, but why would you even
want the _possibility_ that your guest starts running without TDX or SNP
protection, and only find out later via attestation?
There might be multiple reasons why customers want it but we need this
requirement for a couple of other reasons too.

We do not only have hardware based confidential VMs. We might have
some other solutions which measure the initial image before boot.
Ultimately we might want to use a common attestation interface where
customers might be running different kinds of VMs. Using a single
binary will make it easier to manage/verify measurements for both of
us and the customers. I am not a PM so I cannot give more context on
customer use cases.

Another reason is how we deploy and manage guest firmware. We have a
lot of optimization and customization to speed up firmware loading
time and also reduce the time to deploy new builds on the whole fleet
uniformly. Adding a new firmware binary is a big challenge for us to
enable these features. On the top of integration challenges, it will
create maintainability issues in the long run for us when we provide
tools to verify/reproduce the hashes in the attestation report.

want the _possibility_ that your guest starts running without TDX or SNP
protection, and only find out later via attestation?
I am missing the point here. Customers should rely on only the
attestation report to establish the trust.
-If firmware does not support TDX and TDX is enabled, that firmware
will crash at some point.
-If firmware is generic firmware that supports TDX and SNP and others,
and TDX is enabled or not, still the customer needs to verify the TDX
enablement through attestation.
-If firmware is a customized binary compiled to support TDX,
irrelevant of TDX being enabled or not, still the customer needs to
verify the TDX enablement through attestation.


For a similar reason, OVMF already supports shipping a binary that fails
to boot if SMM is not available to the firmware, because then secure
boot would be trivially circumvented.

I can understand having a single binary for both TDX or SNP. That's not
a problem since you can set up the SEV startup VMSA to 32-bit protected
mode just like TDX wants.
I agree that this is doable but I am not sure if we need to also
modify the reset vector for AMD SNP in that case. Also it will not
solve our problem. If we start to generate a new firmware for every
feature , it will not end well for us, I think. Both TDX and SNP are
still new features in the same architecture, and it seems to me that
they are sharing a lot of common/similar code. AMD has already made
some of their patches in (SEV and SEV-ES) which works very nicely for
our use case and integration. Looks like Intel just has an issue on
how to fix their reset vector problem. Once they solve it and upstream
accepts the changes, I do not see any other big blocker. OVMF was
doing a great job on abstracting differences and providing a common
interface without creating multiple binaries. I do not see why it
should not do the same thing here.

therefore we were expecting the TDX
changes to be part of the upstream code.
Having 1 or more binaries should be unrelated to the changes being
upstream (or more likely, I am misunderstanding you).
You are right, it is my bad for not clarifying it. What I mean is we
want it to be part of the upstream so it can be easier for us to pull
the changes and they are compatible with the changes that SNP is doing
but we also do not want to use different configuration files to
generate different binaries for each use case.


Thanks,

Paolo


Yao, Jiewen
 

Hello
Do we have some conclusion on this topic?

Do we agree the one-binary solution in OVMF or we need more discussion?


Thank you
Yao Jiewen

-----Original Message-----
From: Erdem Aktas <erdemaktas@google.com>
Sent: Friday, April 16, 2021 3:43 AM
To: Paolo Bonzini <pbonzini@redhat.com>
Cc: devel@edk2.groups.io; jejb@linux.ibm.com; Yao, Jiewen
<jiewen.yao@intel.com>; dgilbert@redhat.com; Laszlo Ersek
<lersek@redhat.com>; Xu, Min M <min.m.xu@intel.com>;
thomas.lendacky@amd.com; Brijesh Singh <brijesh.singh@amd.com>; Justen,
Jordan L <jordan.l.justen@intel.com>; Ard Biesheuvel
<ardb+tianocore@kernel.org>; Nathaniel McCallum
<npmccallum@redhat.com>; Ning Yang <ningyang@google.com>
Subject: Re: [edk2-devel] separate OVMF binary for TDX? [was: OvmfPkg:
Reserve the Secrets and Cpuid page for the SEV-SNP guest]

Thanks Paolo.

On Thu, Apr 15, 2021 at 12:59 AM Paolo Bonzini <pbonzini@redhat.com>
wrote:

On 15/04/21 01:34, Erdem Aktas wrote:
We do not want to generate different binaries for AMD, Intel, Intel
with TDX, AMD with SEV/SNP etc
My question is why the user would want a single binary for VMs with and
without TDX/SNP. I know there is attestation, but why would you even
want the _possibility_ that your guest starts running without TDX or SNP
protection, and only find out later via attestation?
There might be multiple reasons why customers want it but we need this
requirement for a couple of other reasons too.

We do not only have hardware based confidential VMs. We might have
some other solutions which measure the initial image before boot.
Ultimately we might want to use a common attestation interface where
customers might be running different kinds of VMs. Using a single
binary will make it easier to manage/verify measurements for both of
us and the customers. I am not a PM so I cannot give more context on
customer use cases.

Another reason is how we deploy and manage guest firmware. We have a
lot of optimization and customization to speed up firmware loading
time and also reduce the time to deploy new builds on the whole fleet
uniformly. Adding a new firmware binary is a big challenge for us to
enable these features. On the top of integration challenges, it will
create maintainability issues in the long run for us when we provide
tools to verify/reproduce the hashes in the attestation report.

want the _possibility_ that your guest starts running without TDX or SNP
protection, and only find out later via attestation?
I am missing the point here. Customers should rely on only the
attestation report to establish the trust.
-If firmware does not support TDX and TDX is enabled, that firmware
will crash at some point.
-If firmware is generic firmware that supports TDX and SNP and others,
and TDX is enabled or not, still the customer needs to verify the TDX
enablement through attestation.
-If firmware is a customized binary compiled to support TDX,
irrelevant of TDX being enabled or not, still the customer needs to
verify the TDX enablement through attestation.


For a similar reason, OVMF already supports shipping a binary that fails
to boot if SMM is not available to the firmware, because then secure
boot would be trivially circumvented.

I can understand having a single binary for both TDX or SNP. That's not
a problem since you can set up the SEV startup VMSA to 32-bit protected
mode just like TDX wants.
I agree that this is doable but I am not sure if we need to also
modify the reset vector for AMD SNP in that case. Also it will not
solve our problem. If we start to generate a new firmware for every
feature , it will not end well for us, I think. Both TDX and SNP are
still new features in the same architecture, and it seems to me that
they are sharing a lot of common/similar code. AMD has already made
some of their patches in (SEV and SEV-ES) which works very nicely for
our use case and integration. Looks like Intel just has an issue on
how to fix their reset vector problem. Once they solve it and upstream
accepts the changes, I do not see any other big blocker. OVMF was
doing a great job on abstracting differences and providing a common
interface without creating multiple binaries. I do not see why it
should not do the same thing here.

therefore we were expecting the TDX
changes to be part of the upstream code.
Having 1 or more binaries should be unrelated to the changes being
upstream (or more likely, I am misunderstanding you).
You are right, it is my bad for not clarifying it. What I mean is we
want it to be part of the upstream so it can be easier for us to pull
the changes and they are compatible with the changes that SNP is doing
but we also do not want to use different configuration files to
generate different binaries for each use case.


Thanks,

Paolo


Laszlo Ersek
 

On 04/21/21 02:38, Yao, Jiewen wrote:
Hello
Do we have some conclusion on this topic?

Do we agree the one-binary solution in OVMF or we need more discussion?
Well it's not technically impossible to do, just very ugly and brittle.
And I'm doubtful that this is a unique problem ("just fix the reset
vector") the likes of which will supposedly never return during the
integration of SEV and TDX. Once we make this promise ("one firmware
binary at all costs"), the hacks we accept for its sake will only
accumulate over time, and we'll have more and more precedent to justify
the next hack. Technical debt is not exactly what we don't have enough
of, in edk2.

I won't make a secret out of the fact that I'm slightly annoyed that
this approach is being dictated by Google (as far as I understand, at
this point, anyway). I don't see or recall a lot of Google contributions
in the edk2 history or the bug tracker. I'm not enthusiastic about
complexity without explicit commitment / investment on the beneficiary's
side.

I won't nack the approach personally, but I'm quite unhappy about it.
Can Google at least propose a designated reviewer ("R") for the
"OvmfPkg: Confidential Computing" section of "Maintainers.txt", in a patch?

Thanks
Laszlo



Thank you
Yao Jiewen

-----Original Message-----
From: Erdem Aktas <erdemaktas@google.com>
Sent: Friday, April 16, 2021 3:43 AM
To: Paolo Bonzini <pbonzini@redhat.com>
Cc: devel@edk2.groups.io; jejb@linux.ibm.com; Yao, Jiewen
<jiewen.yao@intel.com>; dgilbert@redhat.com; Laszlo Ersek
<lersek@redhat.com>; Xu, Min M <min.m.xu@intel.com>;
thomas.lendacky@amd.com; Brijesh Singh <brijesh.singh@amd.com>; Justen,
Jordan L <jordan.l.justen@intel.com>; Ard Biesheuvel
<ardb+tianocore@kernel.org>; Nathaniel McCallum
<npmccallum@redhat.com>; Ning Yang <ningyang@google.com>
Subject: Re: [edk2-devel] separate OVMF binary for TDX? [was: OvmfPkg:
Reserve the Secrets and Cpuid page for the SEV-SNP guest]

Thanks Paolo.

On Thu, Apr 15, 2021 at 12:59 AM Paolo Bonzini <pbonzini@redhat.com>
wrote:

On 15/04/21 01:34, Erdem Aktas wrote:
We do not want to generate different binaries for AMD, Intel, Intel
with TDX, AMD with SEV/SNP etc
My question is why the user would want a single binary for VMs with and
without TDX/SNP. I know there is attestation, but why would you even
want the _possibility_ that your guest starts running without TDX or SNP
protection, and only find out later via attestation?
There might be multiple reasons why customers want it but we need this
requirement for a couple of other reasons too.

We do not only have hardware based confidential VMs. We might have
some other solutions which measure the initial image before boot.
Ultimately we might want to use a common attestation interface where
customers might be running different kinds of VMs. Using a single
binary will make it easier to manage/verify measurements for both of
us and the customers. I am not a PM so I cannot give more context on
customer use cases.

Another reason is how we deploy and manage guest firmware. We have a
lot of optimization and customization to speed up firmware loading
time and also reduce the time to deploy new builds on the whole fleet
uniformly. Adding a new firmware binary is a big challenge for us to
enable these features. On the top of integration challenges, it will
create maintainability issues in the long run for us when we provide
tools to verify/reproduce the hashes in the attestation report.

want the _possibility_ that your guest starts running without TDX or SNP
protection, and only find out later via attestation?
I am missing the point here. Customers should rely on only the
attestation report to establish the trust.
-If firmware does not support TDX and TDX is enabled, that firmware
will crash at some point.
-If firmware is generic firmware that supports TDX and SNP and others,
and TDX is enabled or not, still the customer needs to verify the TDX
enablement through attestation.
-If firmware is a customized binary compiled to support TDX,
irrelevant of TDX being enabled or not, still the customer needs to
verify the TDX enablement through attestation.


For a similar reason, OVMF already supports shipping a binary that fails
to boot if SMM is not available to the firmware, because then secure
boot would be trivially circumvented.

I can understand having a single binary for both TDX or SNP. That's not
a problem since you can set up the SEV startup VMSA to 32-bit protected
mode just like TDX wants.
I agree that this is doable but I am not sure if we need to also
modify the reset vector for AMD SNP in that case. Also it will not
solve our problem. If we start to generate a new firmware for every
feature , it will not end well for us, I think. Both TDX and SNP are
still new features in the same architecture, and it seems to me that
they are sharing a lot of common/similar code. AMD has already made
some of their patches in (SEV and SEV-ES) which works very nicely for
our use case and integration. Looks like Intel just has an issue on
how to fix their reset vector problem. Once they solve it and upstream
accepts the changes, I do not see any other big blocker. OVMF was
doing a great job on abstracting differences and providing a common
interface without creating multiple binaries. I do not see why it
should not do the same thing here.

therefore we were expecting the TDX
changes to be part of the upstream code.
Having 1 or more binaries should be unrelated to the changes being
upstream (or more likely, I am misunderstanding you).
You are right, it is my bad for not clarifying it. What I mean is we
want it to be part of the upstream so it can be easier for us to pull
the changes and they are compatible with the changes that SNP is doing
but we also do not want to use different configuration files to
generate different binaries for each use case.


Thanks,

Paolo


Erdem Aktas
 

Hi Laszlo,

I am sorry to hear that it sounded like we are dictating a certain approach. Although I can see why it sounded that way, it certainly was not my intention.
We want to work with the EDK2 community to have a solution that is beneficial for everyone and we appreciate the inputs that we got from you and Paolo.  Code quality is always a high priority for us. Therefore, if, at some point, things get too hacky to impact the quality/maintainability of the upstream code, we will consider making adjustments on our side.

With the current discussion, I was just trying to describe our use case and the importance of having a single binary where there is no absolute need for architectural differences. As far as I know, the only problematic area is modifying the reset vector to be compatible with TDX and it seems like Intel has a solution for it without impacting the overall quality of the upstream code. I just want to reiterate that we are open for discussion and what we ask should not be considered "at all cost" and please let us know that if edk2 community/maintainers are still thinking that what Intel is proposing is not feasible. 

>>Can Google at least propose a designated reviewer ("R") for the
>>"OvmfPkg: Confidential Computing" section of "Maintainers.txt", in a patch?
Sure I would be happy too. 

-Erdem

On Wed, Apr 21, 2021 at 3:44 AM Laszlo Ersek <lersek@...> wrote:
On 04/21/21 02:38, Yao, Jiewen wrote:
> Hello
> Do we have some conclusion on this topic?
>
> Do we agree the one-binary solution in OVMF or we need more discussion?

Well it's not technically impossible to do, just very ugly and brittle.
And I'm doubtful that this is a unique problem ("just fix the reset
vector") the likes of which will supposedly never return during the
integration of SEV and TDX. Once we make this promise ("one firmware
binary at all costs"), the hacks we accept for its sake will only
accumulate over time, and we'll have more and more precedent to justify
the next hack. Technical debt is not exactly what we don't have enough
of, in edk2.

I won't make a secret out of the fact that I'm slightly annoyed that
this approach is being dictated by Google (as far as I understand, at
this point, anyway). I don't see or recall a lot of Google contributions
in the edk2 history or the bug tracker. I'm not enthusiastic about
complexity without explicit commitment / investment on the beneficiary's
side.

I won't nack the approach personally, but I'm quite unhappy about it.
Can Google at least propose a designated reviewer ("R") for the
"OvmfPkg: Confidential Computing" section of "Maintainers.txt", in a patch?

Thanks
Laszlo

>
>
> Thank you
> Yao Jiewen
>
>> -----Original Message-----
>> From: Erdem Aktas <erdemaktas@...>
>> Sent: Friday, April 16, 2021 3:43 AM
>> To: Paolo Bonzini <pbonzini@...>
>> Cc: devel@edk2.groups.io; jejb@...; Yao, Jiewen
>> <jiewen.yao@...>; dgilbert@...; Laszlo Ersek
>> <lersek@...>; Xu, Min M <min.m.xu@...>;
>> thomas.lendacky@...; Brijesh Singh <brijesh.singh@...>; Justen,
>> Jordan L <jordan.l.justen@...>; Ard Biesheuvel
>> <ardb+tianocore@...>; Nathaniel McCallum
>> <npmccallum@...>; Ning Yang <ningyang@...>
>> Subject: Re: [edk2-devel] separate OVMF binary for TDX? [was: OvmfPkg:
>> Reserve the Secrets and Cpuid page for the SEV-SNP guest]
>>
>> Thanks Paolo.
>>
>> On Thu, Apr 15, 2021 at 12:59 AM Paolo Bonzini <pbonzini@...>
>> wrote:
>>>
>>> On 15/04/21 01:34, Erdem Aktas wrote:
>>>> We do not want to generate different binaries for AMD, Intel, Intel
>>>> with TDX, AMD with SEV/SNP etc
>>>
>>> My question is why the user would want a single binary for VMs with and
>>> without TDX/SNP.  I know there is attestation, but why would you even
>>> want the _possibility_ that your guest starts running without TDX or SNP
>>> protection, and only find out later via attestation?
>>
>> There might be multiple reasons why customers want it but we need this
>> requirement for a couple of other reasons too.
>>
>> We do not only have hardware based confidential VMs. We might have
>> some other solutions which measure the initial image before boot.
>> Ultimately we might want to use a common attestation interface where
>> customers might be running different kinds of VMs. Using a single
>> binary will make it easier to manage/verify measurements for both of
>> us and the customers. I am not a PM so I cannot give more context on
>> customer use cases.
>>
>> Another reason is how we deploy and manage guest firmware. We have a
>> lot of optimization and customization to speed up firmware loading
>> time and also reduce the time to deploy new builds on the whole fleet
>> uniformly.  Adding a new firmware binary is a big challenge for us to
>> enable these features. On the top of integration challenges, it will
>> create maintainability issues in the long run for us when we provide
>> tools to verify/reproduce the hashes in the attestation report.
>>
>>> want the _possibility_ that your guest starts running without TDX or SNP
>>> protection, and only find out later via attestation?
>>
>> I am missing the point here. Customers should rely on only the
>> attestation report to establish the trust.
>> -If firmware does not support TDX and TDX is enabled, that firmware
>> will crash at some point.
>> -If firmware is generic firmware that supports TDX and SNP and others,
>> and TDX is enabled or not,  still the customer needs to verify the TDX
>> enablement through attestation.
>> -If firmware is a customized binary compiled to support TDX,
>> irrelevant of TDX being enabled or not,  still the customer needs to
>> verify the TDX enablement through attestation.
>>
>>
>>> For a similar reason, OVMF already supports shipping a binary that fails
>>> to boot if SMM is not available to the firmware, because then secure
>>> boot would be trivially circumvented.
>>>
>>> I can understand having a single binary for both TDX or SNP.  That's not
>>> a problem since you can set up the SEV startup VMSA to 32-bit protected
>>> mode just like TDX wants.
>>
>> I agree that this is doable but I am not sure if we need to also
>> modify the reset vector for AMD SNP in that case. Also it will not
>> solve our problem. If we start to generate a new firmware for every
>> feature , it will not end well for us, I think. Both TDX and SNP are
>> still new features in the same architecture, and it seems to me that
>> they are sharing a lot of common/similar code. AMD has already made
>> some of their patches in (SEV and SEV-ES) which works very nicely for
>> our use case and integration. Looks like Intel just has an issue on
>> how to fix their reset vector problem. Once they solve it and upstream
>> accepts the changes, I do not see any other big blocker. OVMF was
>> doing a great job on abstracting differences and providing a common
>> interface without creating multiple binaries. I do not see why it
>> should not do the same thing here.
>>
>>>> therefore we were expecting the TDX
>>>> changes to be part of the upstream code.
>>>
>>> Having 1 or more binaries should be unrelated to the changes being
>>> upstream (or more likely, I am misunderstanding you).
>>
>> You are right, it is my bad for not clarifying it. What I mean is we
>> want it to be part of the upstream so it can be easier for us to pull
>> the changes and they are compatible with the changes that SNP is doing
>> but we also do not want to use different configuration files to
>> generate different binaries for each use case.
>>
>>
>>> Thanks,
>>>
>>> Paolo
>>>







Laszlo Ersek
 

On 04/21/21 19:07, Erdem Aktas wrote:
Hi Laszlo,

I am sorry to hear that it sounded like we are dictating a certain
approach. Although I can see why it sounded that way, it certainly was not
my intention.
We want to work with the EDK2 community to have a solution that is
beneficial for everyone and we appreciate the inputs that we got from you
and Paolo. Code quality is always a high priority for us. Therefore, if,
at some point, things get too hacky to impact the
quality/maintainability of the upstream code, we will consider making
adjustments on our side.

With the current discussion, I was just trying to describe our use case and
the importance of having a single binary where there is no absolute need
for architectural differences. As far as I know, the only problematic area
is modifying the reset vector to be compatible with TDX and it seems like
Intel has a solution for it without impacting the overall quality of the
upstream code. I just want to reiterate that we are open for discussion and
what we ask should not be considered "at all cost" and please let us know
that if edk2 community/maintainers are still thinking that what Intel is
proposing is not feasible.
OK.

It's not lost on me that we're talking about ~3 instructions.

Let's keep a close eye on further "polymorphisms" that would require hacks.


Can Google at least propose a designated reviewer ("R") for the
"OvmfPkg: Confidential Computing" section of "Maintainers.txt", in a
patch?
Sure I would be happy too.
Yes, please do that. It can be included in the TDX patch set from Min Xu
that modifies the beginning of reset vector as discussed above.

Thanks!
Laszlo


-Erdem

On Wed, Apr 21, 2021 at 3:44 AM Laszlo Ersek <lersek@redhat.com> wrote:

On 04/21/21 02:38, Yao, Jiewen wrote:
Hello
Do we have some conclusion on this topic?

Do we agree the one-binary solution in OVMF or we need more discussion?
Well it's not technically impossible to do, just very ugly and brittle.
And I'm doubtful that this is a unique problem ("just fix the reset
vector") the likes of which will supposedly never return during the
integration of SEV and TDX. Once we make this promise ("one firmware
binary at all costs"), the hacks we accept for its sake will only
accumulate over time, and we'll have more and more precedent to justify
the next hack. Technical debt is not exactly what we don't have enough
of, in edk2.

I won't make a secret out of the fact that I'm slightly annoyed that
this approach is being dictated by Google (as far as I understand, at
this point, anyway). I don't see or recall a lot of Google contributions
in the edk2 history or the bug tracker. I'm not enthusiastic about
complexity without explicit commitment / investment on the beneficiary's
side.

I won't nack the approach personally, but I'm quite unhappy about it.
Can Google at least propose a designated reviewer ("R") for the
"OvmfPkg: Confidential Computing" section of "Maintainers.txt", in a patch?

Thanks
Laszlo



Thank you
Yao Jiewen

-----Original Message-----
From: Erdem Aktas <erdemaktas@google.com>
Sent: Friday, April 16, 2021 3:43 AM
To: Paolo Bonzini <pbonzini@redhat.com>
Cc: devel@edk2.groups.io; jejb@linux.ibm.com; Yao, Jiewen
<jiewen.yao@intel.com>; dgilbert@redhat.com; Laszlo Ersek
<lersek@redhat.com>; Xu, Min M <min.m.xu@intel.com>;
thomas.lendacky@amd.com; Brijesh Singh <brijesh.singh@amd.com>; Justen,
Jordan L <jordan.l.justen@intel.com>; Ard Biesheuvel
<ardb+tianocore@kernel.org>; Nathaniel McCallum
<npmccallum@redhat.com>; Ning Yang <ningyang@google.com>
Subject: Re: [edk2-devel] separate OVMF binary for TDX? [was: OvmfPkg:
Reserve the Secrets and Cpuid page for the SEV-SNP guest]

Thanks Paolo.

On Thu, Apr 15, 2021 at 12:59 AM Paolo Bonzini <pbonzini@redhat.com>
wrote:

On 15/04/21 01:34, Erdem Aktas wrote:
We do not want to generate different binaries for AMD, Intel, Intel
with TDX, AMD with SEV/SNP etc
My question is why the user would want a single binary for VMs with and
without TDX/SNP. I know there is attestation, but why would you even
want the _possibility_ that your guest starts running without TDX or
SNP
protection, and only find out later via attestation?
There might be multiple reasons why customers want it but we need this
requirement for a couple of other reasons too.

We do not only have hardware based confidential VMs. We might have
some other solutions which measure the initial image before boot.
Ultimately we might want to use a common attestation interface where
customers might be running different kinds of VMs. Using a single
binary will make it easier to manage/verify measurements for both of
us and the customers. I am not a PM so I cannot give more context on
customer use cases.

Another reason is how we deploy and manage guest firmware. We have a
lot of optimization and customization to speed up firmware loading
time and also reduce the time to deploy new builds on the whole fleet
uniformly. Adding a new firmware binary is a big challenge for us to
enable these features. On the top of integration challenges, it will
create maintainability issues in the long run for us when we provide
tools to verify/reproduce the hashes in the attestation report.

want the _possibility_ that your guest starts running without TDX or
SNP
protection, and only find out later via attestation?
I am missing the point here. Customers should rely on only the
attestation report to establish the trust.
-If firmware does not support TDX and TDX is enabled, that firmware
will crash at some point.
-If firmware is generic firmware that supports TDX and SNP and others,
and TDX is enabled or not, still the customer needs to verify the TDX
enablement through attestation.
-If firmware is a customized binary compiled to support TDX,
irrelevant of TDX being enabled or not, still the customer needs to
verify the TDX enablement through attestation.


For a similar reason, OVMF already supports shipping a binary that
fails
to boot if SMM is not available to the firmware, because then secure
boot would be trivially circumvented.

I can understand having a single binary for both TDX or SNP. That's
not
a problem since you can set up the SEV startup VMSA to 32-bit protected
mode just like TDX wants.
I agree that this is doable but I am not sure if we need to also
modify the reset vector for AMD SNP in that case. Also it will not
solve our problem. If we start to generate a new firmware for every
feature , it will not end well for us, I think. Both TDX and SNP are
still new features in the same architecture, and it seems to me that
they are sharing a lot of common/similar code. AMD has already made
some of their patches in (SEV and SEV-ES) which works very nicely for
our use case and integration. Looks like Intel just has an issue on
how to fix their reset vector problem. Once they solve it and upstream
accepts the changes, I do not see any other big blocker. OVMF was
doing a great job on abstracting differences and providing a common
interface without creating multiple binaries. I do not see why it
should not do the same thing here.

therefore we were expecting the TDX
changes to be part of the upstream code.
Having 1 or more binaries should be unrelated to the changes being
upstream (or more likely, I am misunderstanding you).
You are right, it is my bad for not clarifying it. What I mean is we
want it to be part of the upstream so it can be easier for us to pull
the changes and they are compatible with the changes that SNP is doing
but we also do not want to use different configuration files to
generate different binaries for each use case.


Thanks,

Paolo