Security and Correctness in Wasmtime (opens in new tab)

> WebAssembly programs are sandboxed and isolated from one another and from the host, so they can’t read or write external regions of memory, transfer control to arbitrary code in the process, or freely access the network and filesystem. This makes it safe to run untrusted WebAssembly programs: they cannot escape the sandbox to steal private data from elsewhere on your laptop or run a botnet on your servers.

As if users will not concede every requested permission to the first Monero miner that asks.

This seems like a pretty comprehensive strategy to take wasm security and correctness seriously. It pretty much covers everything I would want to see if I were relying on this system, including auditing, fuzzing, formal correctness, spectre, and even a clear-eyed organizational stance toward reported security vulnerabilities.

The post mentions using `cargo vet` to organize audits of third party crates, discussed here a few months ago [0]. I'm more familiar with cargo-crev which does something similar, how do these auditing tools compare? The audit format [1] seems somewhat reasonable, but it doesn't include the review date and there's no mechanism to validate the authenticity of the auditors.

[0]: https://news.ycombinator.com/item?id=31719532

[1]: https://mozilla.github.io/cargo-vet/recording-audits.html

I'm certainly curious to see how in-process sandboxing plays out with Spectre. Even the process boundary sometimes doesn't feel like enough, heavy handed as it may be. I wonder if there's a way to prove the absence of side channels by encoding side effects more directly and ensuring that those side effects never propagate across a boundary. The problem would probably be enumerating them... and then idk, everything has side effects to some degree, "the value was read, which caused a L cache line to flush" I guess it's probably not tractable.

I kind of, vaguely loosely, feel like running multiple 'workers' within a single process is just not a reasonable goal. Ultimately if you have a multi-tenant requirement you should be using separate processes and pinning them to separate physical CPUs, and hope that that is enough. Not to discourage this, I can't wait to look back in a decade and see how this all has changed.

edit: Also, there are other use cases. Like, maybe I'm a single tenant and I'm deploying multiple workers to a single VM. I trust myself, but it would still be nice to have it be hard for those boundaries to be violated - driving up the cost is sane.

It also sort of reminds me of the Sysiphean task of removing ROP gadgets from the Linux kernel.

What is the overhead of the spectre mitigations? I read a PhD dissertation showing that v8 suffers a 20% overhead in practice. People can sugar those numbers, but I wouldn't expect another userspace-emulator-like program to behave differently. Is this number in the ballpark for wasmtime?