18. July 2022

[dimakuv]

Agenda

(please write your proposed agenda items in comments under this discussion)

• Enrico Borello from Leonardo Cyber Security: their application and questions about Gramine (30 minutes slot).
• Grant proposal from Chia-Che and Don (10 minutes slot).
• Sockets rewrite: cannot connect to UNIX domain socket's path.

Enrico Borello's questions

Enrico and his team are at the phase of investigation (how to run Linux applications in SGX enclaves) and they found Gramine. The main question was about the general process of SGX remote attestation: "can custom, non-Intel-based attestation schemes be used?". Mona and Dmitrii explained the EPID and DCAP attestation schemes (the first one requires internet connectivity to verify the SGX quote via the Intel IAS web service, the latter has a "caching service" that may alleviate the need for constant internet connectivity but still relies on the Intel PCS web service).

Grant proposal from Chia-Che and Don

There is a proposal for NSF (National Science Foundation) funding, in the "Secure and Trustworthy Cyberspace: transition to practive" program. 4 years of funding, with different possible tasks around Gramine open source development:

• Set of tasks 1: implementing new syscalls, pseudo-files, etc. Also implementing syscall unit tests and module-level unit tests. Developing analysis and logging tools.
• Set of tasks 2: Surveys for user feedback, to improve usability. Manifest editor tool. Upstreaming hooks to Glibc and Musl. AMD SEV PAL. Windows SGX PAL. Tasks around containers and Kubernetes.
• Set of tasks 3: Fuzz testing, symbolic execution, data flow analysis. Improving the IPC layer, e.g., rewriting it in Rust with compile-time protocol checks.

Request for feedback: should Gramine work on boot camps or webinars? Should Gramine offer open hours?

• Feedback from Eric:
  • User experience becomes very important at this point of Gramine life (and generally in CC solutions).
  • Packaging becomes very important: ideally, there should be a way to abstract away different HW TEEs, so that a user chooses one package to install, and the distro installer picks a corresponding runtime + config to install on user's particular machine (with a particular HW TEE).

Sockets rewrite and UNIX domain sockets

Issue: how can two Gramine instances (i.e., Gramine processes that were launched separately, e.g., in two Docker containers) communicate over UNIX Domain Sockets (UDS)?

The problem is that UDS communication is transparently TLS-encrypted by Gramine (think of UDS as "better version of UNIX pipes"). So if two processes want to communicate over UDS, they must possess the same master encryption key (because Gramine uses TLS-PSK for pipes and UDS). But two Gramine instances will not share any encryption keys, so they can't establish a secure TLS session over UDS.

Transparent encryption of UDS communication is a core design choice. We will not change this logic.

Currently, the only two solutions to this issue are:

1. Replace UDS with explicitly-TLS-protected TCP sockets at the application level. (Requires modifications to the app; some apps support such a switch via configuration options.)
2. Start Gramine with a "wrapper" process/script that will spawn child processes -- they will all share the same master encryption key, and UDS communication will be possible. (Requires modifications to how the apps are started.)

Vijay mentioned that they have a similar scenario: two independent SGX enclaves are launched separately, and they would like to communicate over UDS. With current Gramine design, this is impossible.

Mona proposed an idea to implement something similar to RA-TLS (used for SGX remote attestation), but in a local fashion -- LA-TLS (used for SGX local attestation). The core concept will be similar to RA-TLS: during the TLS handshake, the two enclaves exchange SGX evidence (TARGETINFO and SGXREPORT), attest each other, and establish a secure channel (skipping the "normal PKI" verification of X.509 certificates). By verifying MRSIGNER and possibly MRENCLAVE, it will be possible to connect two independently launched SGX enclaves, but from the same "package".

Dmitrii believes it is viable, but LA-TLS will require a lot of effort/time to implement.

[donporter]

I would like to present a summary of a grant proposal Chia-Che and I are working on, and get feedback/suggestions on the proposed Gramine work from the team. I expect we would need no more than 10 minutes.

[dimakuv]

Added to the agenda.

[dimakuv]

We should discuss the problem of not being able to connect to the UNIX domain socket's path in the child process. Example from Nginx + PHP:

[P5:T5:nginx] trace: ---- connect(9, {family=UNIX,path=/var/run/php/php7.4-fpm.sock}, 110) ...
[P5:T5:nginx] trace: ---- return from connect(...) = -2

UPDATE: This was wrong. I guess I have misunderstood the problem. The root cause is actually that two independent Gramine instances (one Nginx, one PHP) wanted to communicate over the UNIX domain socket, and this is unsupported.