ACAI: Extending Arm Confidential Computing Architecture Protection from CPUs to Accelerators

Trusted execution environments in several existing and upcoming CPUs demonstrate the success of confidential computing, with the caveat that tenants cannot use accelerators such as GPUs and FPGAs. If the accelerators have TEE support, the user-code executing on the CPU in a confidential VM has to rely on software-based encryption to facilitate communication between VMs and accelerators. Even after hardware changes to enable TEEs on both sides and software changes to adopt existing code to leverage these features, it results in redundant data copies and hardware encryption at the bus-level and on the accelerator thus degrading the performance and defeating the purpose of using accelerators. In this paper, we reconsider the Arm Confidential Computing Architecture (CCA) design-an upcoming TEE feature in Arm v9-to address this gap. We observe that CCA offers the right abstraction and mechanisms to allow confidential VM to use accelerators as a first class abstraction, while relying on the hardware-based memory protection to preserve security. We build Acai, a CCA-based solution, to demonstrate the feasibility of our approach without changes to hardware or software on the CPU and the accelerator. Our experimental results on GPU and FPGA show that Acai can achieve strong security guarantees with low performance overheads.