Add staged ClientHello acceptor for ServerConfig selection

[iadev09]

Closes #2024.

Adds a staged ClientHello accept path so a server can inspect the rustls ClientHello before finalizing the Quinn ServerConfig.

Commit Structure

1. rustls 0.24 upgrade path

Kept as one compatibility commit. This is where the temporary rustls git dependencies live, including the rustls-platform-verifier 0.8 branch.

• Update rustls dependencies to upstream 0.24

Related rustls 0.24 PR outside Quinn: rustls/rustls-platform-verifier#233.

2. staged ClientHello acceptor primitive

• Refactor rustls transport error mapping
• Start server sessions from rustls Accepted
• Extract shared server accept finalization
• Make connection assembler available to endpoint
• Add staged rustls ClientHello acceptor
• Support multi-packet staged ClientHello
• Create connection from rustls Accepted
• Expose staged acceptor on Incoming
• Test staged rustls acceptor paths

3. selector helper

Adds a ServerConfigSelector helper trait on top of the staged acceptor API.

• Add ServerConfigSelector helper

pub trait ServerConfigSelector {
    /// Error returned by the selector.
    type Error;

    /// Select a server config for the provided ClientHello.
    fn select_server_config<'a>(
        &'a self,
        client_hello: rustls::server::ClientHello<'a>,
    ) -> impl Future<Output = Result<Option<Arc<ServerConfig>>, Self::Error>> + 'a;
}

Tests

cargo test -p quinn incoming_acceptor \
  --features runtime-tokio,rustls-aws-lc-rs \
  --no-default-features \
  --locked \
  -- --nocapture

cargo test -p quinn incoming_acceptor \
  --features runtime-tokio,rustls-ring \
  --no-default-features \
  --locked \
  -- --nocapture

Bench

https://github.com/iadev09/quinn-acceptor-bench

In local 1000-sample smoke runs, the selected path stayed in the same range as continuing with the default ServerConfig.

Comparison

Compared with the TCP/rustls shape, this exposes the staged primitive directly.

For TCP, config selection usually needs an intermediate acceptor/future shape:

let acceptor =
    LazyConfigAcceptor::new(rustls::server::Acceptor::default(), tcp_stream);
tokio::pin!(acceptor);
let start = acceptor.as_mut().await?;

For Quinn, the proposed primitive lets the caller inspect the ClientHello and choose the quinn::ServerConfig directly.

Because the Endpoint already owns the Initial/CRYPTO handling, it does not need a LazyConfigAcceptor-style stream wrapper.

let accepted = incoming.acceptor()?.await?;
let hello = accepted.client_hello();
let config = app_state.select_server_config(hello).await?;
let connection = accepted.accept_with(config)?.await?;

Same scenario on TCP/rustls LazyConfigAcceptor:

https://github.com/iadev09/lazy-config-acceptor-bench

This only measures the selector/no-selector difference; TCP connect and kernel transport work are not included.

[djc]

I think this feature refactoring can be done separately, earlier (suggest in a separate PR). rustls 0.24 requires that callers provide a crypto provider explicitly. Since rustls is a public dependency for Quinn anyway, I suggest that we drop the provider-specific features in favor of requiring an explicit provider, too.

@Ralith do you agree?

If so, I'd suggest that the explicit provider passing can also become a separate commit that can be merged ahead of this one. (May obviate the need for a __rustls feature?)

[Ralith]

I suggest that we drop the provider-specific features in favor of requiring an explicit provider, too.

Yep, that should be much more composable. We may still want to spend some effort having a reasonable default, but we should try to avoid needing to know about other providers.

[djc]

I guess the challenge here is that we still have the raw primitives in use for packet protection. Maybe the first objective is finishing main...ctz:quinn:jbp-retry-calculation-uses-rustls...

[djc]

So I think if you want to make progress on getting this merged, this needs to be done as a separate PR:

I guess the challenge here is that we still have the raw primitives in use for packet protection. Maybe the first objective is finishing main...ctz:quinn:jbp-retry-calculation-uses-rustls...

I think the upgrade to 0.24 should also be a separate PR -- not sure which of these two makes sense to merge first.

[iadev09]

I split the rustls 0.24 upgrade out into a separate branch first. #2701

After that, I can look at the retry packet protection cleanup as another small PR before rebasing the staged ClientHello acceptor work on top.