README.md

Cryptographic signatures for unsafe code

Functions containing unsafe code demand extra scrutiny, because they can break Rust's memory safety guarantees. Some projects may desire a formal process for auditing unsafe code whenever it is added or modified. This compiler plugin supports a workflow where audit status is tracked in the source code, and the history of audits is part of each file's version control history.

It works by attaching a cryptographic signature to every unsafe fn, as well as every fn that contains an unsafe block.

#![feature(plugin)]

#[no_link]
#[plugin(public_key="examples/pubkey")]
extern crate launch_code;

#[launch_code="⠐⡛⢾⣯⢓⢵⢖⡆⣈⠇⠸⣼⢁⢦⢰⢷⡫⢙⠻⠺⢗⢻⣷⠋⣸⡐⣂⡜⠇⡍⢁⢗⢜⠢⡢⣵⠩⠲⡈⢈⢂⡑⣷⣩⢲⢖⢃⡓⠄⣴⠩⡹⡸⠥⢱⢭⡼⠡⣻⡥⢜⢔⡌⠅"]
fn totally_fine() -> u64 {
    unsafe {
        *std::ptr::null()
    }
}

The "launch code" is an Ed25519 signature of the whole fn ... { ... } block, plus the function's attributes, other than launch_code itself. This is computed from the original source files, so it authenticates comments and formatting as well as the code's behavior. The Ed25519 implementation is libsodium via Sodium Oxide. The signature is encoded using Unicode's braille characters.

This is a proof of concept only. My only security guarantee is that I guarantee there's a rather obvious backdoor in the library, to deter people from using it for realsises. (Also the name is kind of silly.)

I'd be interested in collaborating on a production-quality version of this idea, but it needs a lot more thought. A full-featured version would support multiple keys, with project-specific policy on which keys can authorize which language features and APIs.

Hypothesized workflow

• The auditor generates a keypair and gives the public key to developers.

• Developers use allow(unauthorized_unsafe), as do automated tests that block merging to the dev branch. The default setting of warn(wrong_launch_code) alerts the developer to modifications in audited functions. If the modification is intentional, they simply delete the launch_code attribute, committing an unaudited version.

• The auditor uses forbid(unauthorized_unsafe). This causes the compiler to print a list of blocks that need auditing, in the form of error messages. Once the auditor is satisfied with a function, they add a bare #[launch_code] attribute, equivalent to #[launch_code=""]. Because the auditor provides a secret_key file, the warning for an incorrect launch code will include the correct code. The auditor commits and pushes these codes at their own pace.

• Release builds use forbid(unauthorized_unsafe). Missing or incorrect launch codes are a release-blocking issue, even if they don't block landing patches day-to-day.

You can generate a keypair with

cargo run --bin gen_keypair pubkey seckey