Hi all,
In the current implementation, every secure function has an associated veneer function. Therefore, there are so many veneer functions in ‘tfm_veneers.c’, which have a similar prototype.
This would lead to:
* Waste of the veneer and secure function size – these APIs have a similar prototype, and they have unified NS dispatcher already. * More secure functions lead to more veneers and potential re-configuration of the non-secure callable area.
This patch tries to unify the service entry so that:
* Similar function codes do not need to be duplicated. * Keeping almost the same performance.
This is also an experiment patch to start the journey to the SFN model Andrew proposed. Let’s see the feedbacks and decide what to do in the next step. https://review.trustedfirmware.org/c/trusted-firmware-m/+/4115
Please do feedback, especially the library model users – please check what kinds of the inconvenience it brings so that we can discuss the correct shape.
Here are some details in the patch:
Prototype of the unified veneer function: psa_status_t tfm_sfc_call(uint32_t ctrl, psa_invec *in_vec, psa_outvec *out_vec); where: the uint32_t type parameter ‘ctrl’ is a pack of parameters - psa invec length, psa outvec length, and function identifier. [8 bits for inlen][8 bits for outlen][16 bits for function identifier]
This is to avoid the condition that 5 parameters will cause re-wrapping of parameters.
Time cost and code size measurement: github-tracking
Use the unified veneer cost of a veneer call is 1264 cost of an interrupt is 941 veneer used 832B, region size 832B, 100%
cost of a veneer call is 1274 cost of an interrupt is 941 veneer used 64B, region size 832B, 7.69%
Thanks, Mingyang