quiche-h3

⚠️ Experimental / work in progress. This crate is pre-1.0 and under active development. The API is unstable and may change without notice, and it is not yet recommended for production use.

An h3::quic transport-adapter bridge that runs hyperium h3 (HTTP/3) over Cloudflare quiche, driven asynchronously by tokio-quiche.

It exposes a standalone acceptor/connector whose accept() / connect() yield a Connection that implements h3::quic::Connection<Bytes>, so it plugs directly into h3::client / h3::server. It is usable as the quiche backend for tonic-h3 (gRPC over HTTP/3).

Design

The full design lives in docs/design/quiche-h3-bridge.md. In short:

A single-task QuicheDriver worker (tokio_quiche::ApplicationOverQuic) is the sole toucher of quiche::Connection. It owns all cross-task state and never awaits inside its synchronous read/write callbacks (try_send/try_reserve).
The front end implements the h3::quic traits over bounded byte/accept channels, an unbounded control channel, and race-free sticky terminal cells, so a full data queue can never hide a stream/connection terminal, and every terminal reason is typed data carried over the channel — channel closure is never itself a semantic signal.
Error mapping is designed from quiche's synchronous error surface (not inherited from a lossy adapter): peer RESET_STREAM / STOP_SENDING codes, timeouts, and local vs. peer connection closes are all distinguished.

Pinned build

Depends on semver ranges (tokio-quiche 0.19, which pulls quiche 0.29; h3 0.0.8) and pins exact builds via the committed Cargo.lock. A CI compatibility test (tests/ci_compat.rs) constructs one value of every mapped h3 error variant and names every load-bearing quiche/tokio-quiche API, so a minor upstream bump that reshapes the surface fails the build rather than silently mismapping.

Usage

Server:

use quiche_h3::{H3QuicheAcceptor, H3QuicheServerConfig};

# async fn run() -> Result<(), quiche_h3::Error> {
let config = H3QuicheServerConfig {
    cert_path: "cert.pem".into(),
    key_path: "key.pem".into(),
    ..Default::default()
};
let socket = tokio::net::UdpSocket::bind("0.0.0.0:4433").await?;
let mut acceptor = H3QuicheAcceptor::bind([socket], &config)?.pop().unwrap();

while let Some(conn) = acceptor.accept().await? {
    tokio::spawn(async move {
        let mut h3 = h3::server::Connection::new(conn).await.unwrap();
        // ... accept and serve requests over `h3` ...
    });
}
# Ok(()) }

Graceful shutdown (server): obtain a cloneable H3QuicheEndpoint from the acceptor before serving, then close and wait for idle. accept() stops yielding new connections once close() is observed and returns None once the pending handshakes drain (already-established connections keep their own workers, which are torn down by the broadcast close). Await the serve task to completion to drop the acceptor, then drive wait_idle() from the retained handle.

use quiche_h3::{H3QuicheAcceptor, H3QuicheServerConfig};

# async fn run(config: H3QuicheServerConfig) -> Result<(), quiche_h3::Error> {
let socket = tokio::net::UdpSocket::bind("0.0.0.0:4433").await?;
let mut acceptor = H3QuicheAcceptor::bind([socket], &config)?.pop().unwrap();

// Take the shutdown handle before moving the acceptor into the serve loop.
let endpoint = acceptor.endpoint();
let server = tokio::spawn(async move {
    while let Ok(Some(_conn)) = acceptor.accept().await {
        // ... spawn a task to serve each connection ...
    }
});

// Later — graceful shutdown, in this order:
endpoint.close(h3::error::Code::H3_NO_ERROR, b"server shutting down");
server.await.unwrap();      // accept loop ends → the acceptor is dropped
endpoint.wait_idle().await; // resolves once every live connection worker ends
// The same UDP port is now rebindable (use a short bounded retry to absorb the
// tokio-quiche router-task residual — see `H3QuicheAcceptor::endpoint` docs).
# Ok(()) }

Client:

use quiche_h3::{H3QuicheConnector, H3QuicheClientConfig};

# async fn run(addr: std::net::SocketAddr) -> Result<(), quiche_h3::Error> {
let config = H3QuicheClientConfig::default();
let connector = H3QuicheConnector::new(addr, "example.com".into(), config)?;
let conn = connector.connect().await?;
let (mut driver, mut send_request) = h3::client::new(conn).await.unwrap();
// ... drive `driver` and issue requests via `send_request` ...
# Ok(()) }

tokio_quiche and quiche are re-exported (quiche_h3::tokio_quiche, quiche_h3::quiche) so downstreams can build QuicSettings, TLS material, and Hooks without a separate dependency.

Known limitation (§5.5)

At zero connection-level send capacity, a peer that opens a writable-only bidi stream (e.g. via STOP_SENDING) is undiscoverable through any public quiche 0.29 API — quiche's tx_cap == 0 guard precedes the stopped-stream branch. This pathological case is an explicitly documented gap in the otherwise drop-in adapter contract, pending an upstream stream-enumeration API.

Features

tracing — structured instrumentation.
gcongestion / rpk — pass-through toggles for tokio-quiche.

Testing

cargo test -p quiche-h3                       # unit + CI compatibility tests
cargo test -p quiche-h3 -- --include-ignored  # + loopback/e2e integration tests

The ignored suite includes a real end-to-end HTTP/3 GET round-trip (tests/h3_e2e.rs) through h3::client ↔ h3::server over the bridge.

License

MIT.

quiche-h3 0.0.2