Implement wasi-tls

[badeend]

wasi-tls has recently been accepted as a phase 1 proposal.

We'd like to start implementing this in wasmtime. There already exists some prior efforts:

• @dicej created a standalone WASI host with TLS support at: https://github.com/dicej/dotnet-wasi-tls
  • This implements a minimal subset of the draft spec.
  • Uses the native-tls crate.
• @jsturtevant modified the .NET runtime to use Joel's implementation at: dotnet/runtime@main...jsturtevant:runtime:wasi-tls-2
  • This successfully runs their SqlClient.
• Based on Joel's work, I (@badeend) started integrating it into wasmtime itself: https://github.com/badeend/wasmtime/tree/wasi-tls
  • This too is just a minimal subset of the draft spec. Although the WIT in this branch has evolved slightly.
  • Uses the rustls crate.
  • The changes are part of the wasi-sockets interface, which is not the right place.
  • And some boilerplate:
    • threading through a new tls CLI flag
    • Change rustls from a wasi-http-only dependency to a workspace dependency.

---

My suggestion is to add a new standalone wasi-tls crate:

• under the /crates folder, similar to the other proposals.
• that uses the WIT file from this PR
• that targets WASI v0.2 for the time being. Work on a v0.3 interface can happen in parallel in the future.
• that is completely experimental and therefore behind a tls flag.
• uses rustls for its implementation? (up for discussion, see next comment)

Thoughts?

---

CC @dicej @jsturtevant

[badeend]

I did an inventory of the two most popular rust TLS crates to see how suitable they are to implement the draft spec:

rustls:

• doesn't support post-handshake authentication.
• supports programmatic client certificate selection, but the callback isn't async.
• supports programmatic peer certificate verification, but the callback isn't async.

native-tls:

• does not support TLS1.3 (yet)
• uses MacOS Secure Transport, which is deprecated and will likely never receive TLS1.3 support.
• does not support programmatic client certificate selection
• does not support programmatic peer certificate verification
• can't inspect the peer's certificate chain, only the leaf cert
• can't inspect ClientHello nor select server certificate & other configurations based on SNI
• can't read SNI submitted server name
• can't read negotiated TLS version
• can't read negotiated cipher suite
• can't configure ALPN ids on the server side
• can't configure cipher suites
• can't request mid-handshake client certificate on the server side
• doesn't support post-handshake authentication

From wasmtime's POV, rustls seems an obvious choice:

• it is already used by wasi-http,
• it is the most feature complete,
• allows us to quickly iterate on the WIT definitions. Most additions are simply hooking it up to the corresponding rustls methods.

Despite its shortcomings, from a Standards POV, native-tls has one important leg up:
Instant validation of portability goals against three different industry-standard back-ends (OpenSSL, SChannel, SecureTransport).
AFAIK, most of these shortcomings are of the native-tls crate itself and not of the underlying providers. Case in point: .NET's SSLStream is also built on top of SChannel & OpenSSL, yet does support the desired features.

---

In the current stage the interface is still simple enough that it doesn't really matter which one we choose. I just wanted to throw it out there before we start sinking too much time into the integration of either option.

[alexcrichton]

How unreasonable do you think it would be to support both rustls and native-tls? For example via compile-time Cargo features? It seems reasonable to have rustls as the default given its breadth of features but being able to showcase both in the same codebase would be a nice example for others looking to implement the proposal as well.

[badeend]

Given the current simplicity of the WASI interface, it should be doable to have both backends. Looking into the future, I can't vouch for how tenable the situation will be.

The current plan is that @jsturtevant will work on the initial implementation. I haven't checked in with him on this specific point yet, but I suspect that we'll start with just a single backend. We could add the secondary backend in a follow-up PR.

@jsturtevant Does this sound about right to you?

[jsturtevant]

Sounds good to me. My biggest concern would be the lack of features on the native-tls side but the initial interface is pretty minimal so shouldn't be an issue initially. We can adjust as we go.

[alexcrichton]

Also to be clear if you the implementor @jsturtevant would prefer to pick one or the other I think that's totally ok too. I wouldn't consider it a requirement to support both at the beginning at all. Given how things are leaning I think it would make sense to start with rustls and once that's all working we could see if adding a native-tls backend would make sense?

[alexcrichton]

Ok sorry early morning strikes again, @badeend already said all that, disregard me.

[jsturtevant]

Just an update, I've gotten pretty far on this, debugging one last issue before opening up a PR.