utp

A pure-Go implementation of the Micro Transport Protocol (uTP, BEP 29): a reliable, ordered byte stream over UDP with delay-based congestion control, as used by BitTorrent. It needs no cgo and builds with CGO_ENABLED=0.

go get github.com/arxodeus/utp

sock, err := utpnet.Listen(ctx, "udp", ":6881", nil)
if err != nil {
	return err
}
defer sock.Close()

// Accept incoming connections.
go func() {
	for {
		conn, err := sock.Accept()
		if err != nil {
			return
		}
		go handle(conn)
	}
}()

// Dial out from the same port.
conn, err := sock.DialContext(ctx, "udp", "203.0.113.7:6881")

sock is also a net.PacketConn: datagrams that are not uTP, such as DHT traffic sharing the port, are returned by ReadFrom, and WriteTo sends on the same port.

• Classic LEDBAT congestion control (RFC 6817) by default. LEDBAT++ and RACK loss detection (RFC 8985) are available per connection as experimental options.
• Selective acknowledgements, fast retransmission, a tail loss probe, and undoing of window reductions caused by reordering.
• Path MTU discovery, with don't-fragment probes where the platform supports it and optional ICMP input.
• Half-close in both directions (CloseWrite, CloseRead).
• Idle connections release their goroutine and buffers.
• Connection limits, a firewall callback for refusing peers before any state is created, and per-connection metrics.
• Batched socket reads (recvmmsg on Linux) and one acknowledgement per batch.
• Structured logging through log/slog.

The library speaks the uTP wire protocol of BEP 29. Every change is tested for interoperability with BitTorrent's own uTP implementation and with the pure-Go one in anacrolix/go-utp, and against anacrolix/torrent. The experimental options change only how this side paces and detects loss; they need nothing from the peer.

This is a 0.x release: the API may still change before 1.0.

• All measurements so far are on a single machine, over loopback and an emulated network. Nothing has yet been measured over wide-area paths.
• LEDBAT++ is experimental. A LEDBAT++ flow that joins a link where another LEDBAT++ flow is already running takes about 70% of the link while both run, rather than an even share.
• RACK loss detection is experimental.
• Classic LEDBAT does not yield fully to loss-based TCP traffic: it keeps about 69% of a shared bottleneck, by design of the protocol.
• When a connection's event queue is full, further packets for it are dropped and counted in Socket.PacketsDroppedFullConnQueue.
• The full test suite, the race detector and fuzzing run on Linux only; macOS and Windows run the short suite.
• The Go version in go.mod is the minimum, not a recommendation: older standard libraries have known vulnerabilities this code reaches. Build with a current Go release.

go test ./...          # full suite
go test -short ./...   # quick run
go test -race ./...

None of these need cgo. The tests against other uTP implementations are in their own modules under internal/integration; the one against BitTorrent's C implementation needs cgo.

This project began as a fork of zen-eth/utp-go. See NOTICE.

MIT. See LICENSE and NOTICE.