Hi All,

Following is a summary of the situation.

FF-A 1.1 does not use more than 8 registers for parameter/return values.

FF-A 1.2 introduced Direct Req/Resp 2 which utilises up to 18 registers for parameter/return values.

SMCCC 1.2 permitted the use of X4-X17 as return registers and X8-X17 as argument registers.

See ‘Appendix F: Changelog’, in the SMCCC specification, DEN0028, version 1.6G

(https://developer.arm.com/documentation/den0028/latest/)

The section ‘2.4 Conduits’, in the FF-A 1.0 specification specifies that the SMC conduit as described in the SMCCC 1.2 specification should be used, see https://documentation-service.arm.com/static/5fb7e8a6ca04df4095c1d65e?token=

Considering the above one can infer that FF-A 1.0+ requires SMCCC 1.2 which allows usage of X1-X17 as parameter and X0-X17 as return values when using SMC64/HVC64 argument passing. See SMCCC 1.2 specification, Section 2.7 SMC64/HVC64 argument passing.

To support StandaloneMM (StMM), TF-A provides a SPMC+SPMD@EL3 which currently supports FF-A 1.1 with partial support for Direct Req/Resp2 (i.e. only 8 registers supported as parameter/return values). This was done to have a minimal software stack for supporting FF-A with StMM in the upstream, i.e. the FVP software stack with FF-A support has UEFI + TF-A (with SPMC+SPMD@EL3) + StMM.

Apparently, the SPMC+SPMD@EL3 support in TF-A is deprecated and no further updates are planned. Therefore, to support a minimal software stack we were looking at a build option in edk2 that supports configuring the use of 8 registers.

However, considering the fact that FF-A 1.0+ requires compliance with SMCCC 1.2, I think we should just extend the SPMC+SPMD@EL3 support in TF-A to support 18 registers. That way even though the SPMC+SPMD@EL3 support in TF-A is deprecated, it is at least fully compliant with the SMCCC 1.2 requirements.

Levi has submitted a change to TF-A at https://review.trustedfirmware.org/c/TF-A/trusted-firmware-a/+/36018. If this TF-A change is approved and merged, the ArmCallSmc() implementation in edk2 can just support 18 registers by default.

Regards,

Sami Mujawar

From: Leif Lindholm <leif.lindholm@oss.qualcomm.com>
Date: Thursday, 6 March 2025 at 22:58
To: Kun Qin <kuqin12@gmail.com>
Cc: Sami Mujawar <Sami.Mujawar@arm.com>, Olivier Deprez <Olivier.Deprez@arm.com>, Yeo Reum Yun <YeoReum.Yun@arm.com>, edk2-devel-groups-io <devel@edk2.groups.io>, Ard Biesheuvel <ardb+tianocore@kernel.org>
Subject: Re: [edk2-devel] 18 register support for SMC/SVC on AARCH64

Hi Kun,

My point is this has nothing to do with FF-A or likeliness. It is
architecturally broken just from an SMCCC standpoint.
But yes, I would like to hear more from Arm about the specific concern.

Regards,

Leif

On Thu, 6 Mar 2025 at 20:42, Kun Qin <kuqin12@gmail.com> wrote:
>
> Hi Leif,
>
> Thanks for the input. I agree that platforms supporting FF-A v1.2+ will rely on SMCCC v1.1+, and thus platforms not supporting whole 18 register usage are not complying with the spec.
>
> I think Sami, Levi, or Olivier could chime in for better insights on their concerns about SPMC at EL3. As far as the setup we are using (Hafnium as SPMC), the 18-register usage is good across all firmware entities.
>
> Regards,
> Kun
>
> On Thu, Mar 6, 2025 at 1:42 AM Leif Lindholm <leif.lindholm@oss.qualcomm.com> wrote:
>>
>> Hi Kun,
>>
>> On Thu, 6 Mar 2025 at 06:13, Kun Qin <kuqin12@gmail.com> wrote:
>> >
>> > Hi ARM enthusiasts,
>> >
>> > I recently filed a PR to allow 18 register support for SMC/SVC calls between UEFI and secure partition components: https://github.com/tianocore/edk2/pull/10685/files.
>> >
>> > The main purpose of this change is to allow more registers to hold values while doing FF-A transactions. In FF-A spec v1.2 and onward, the section "FFA_MSG_SEND_DIRECT_REQ2" mentions that up to 18 general-purpose registers can be used for such calls. However, the current SMC/SVC implementation in EDK2 only supports up to 8 registers.
>> >
>> > There were some differing opinions on how to support this more properly. Could you please review the PR and chime in on the email thread about how to proceed with it?
>> >
>> > TL;DR:
>> >
>> > In conversations with ARM stakeholders, they revealed concerns about using 18 registers all along because some older firmware components on the secure side do not support full 18 register usage, and the returned values may not be sane. Therefore, there is a need for a build flag that controls how many registers are used during SMC calls to be backwards compatible, which is the PcdSxcUse18Registers approach I went with in the PR.
>>
>> I'm not sure I follow this one (and this is very much the reason I
>> asked for email thread breakout - thank you).
>> Code that relies on the 18 registers is relying on SMCCC >= 1.1.
>> If code is relying on SMCCC >= 1.1, then it must verify that the
>> secure side supports that
>> by making an SMCCC_VERSION call.
>> If that returns NOT_SUPPORTED, or that the version is 1.0, then the
>> fewer-registers calling
>> conventions MUST be used. Otherwise, the 18-register variant is safe.
>> Am I missing something?
>>
>> If we're talking about supporting secure sides that don't comply with
>> the spec, then I think
>> that should be very much a "deal with broken secure firmware quirk"
>> and not a different
>> library.
>> And in that case, it seems to me platform ports that felt the need to
>> deal with broken
>> secure sides should opt into that, with special handling in the single library.
>>
>> If we're talking about supporting edk2 code that doesn't sanity check
>> the version, then
>> I'd suggest we fix the buggy edk2 code instead.
>>
>> Best Regards,
>>
>> Leif
>>
>> > The original approach of using the PCD was to make it a feature flag so that all header files, assembly files, and C files will not even compile the code that supports more than 8 registers if not needed. But that would involve the PCDs getting pre-processed by the build framework, and all components using the ArmSmcLib would thus have to add the PCD in their inf files. So instead, we went with the runtime code evaluation.
>> >
>> > On the PR, Sami suggested creating a new interface that supports SMC with 18 registers and making the PCD control which function to call. For FF-A functions that only involve 8 registers or under, the caller should just use the legacy interfaces. But the issue is, once Standalone MM hands off the control using an 8 register SMC call, it will only be able to process 8 register incoming requests, which will not work if it is woken up by an FFA_MSG_SEND_DIRECT_REQ2 call using 18 registers.
>> >
>> > Any input is appreciated.
>> >
>> > Regards,
>> > Kun