quic: send and receive datagrams

Add the ability for Conns to send and receive datagrams.
No socket handling yet; this only functions in tests for now.

Extend testConn to permit tests to send packets to Conns and
observe the packets Conns send.

There's a circular dependency here: We can't test Handshake and 1-RTT
packets until we have the handshake implemented, but we can't
implement the handshake without the ability to send and receive
Handshake and 1-RTT packets. This CL adds the ability to send and
receive those packets; tests for those paths will follow with the
handshake implementation.

For golang/go#58547

Change-Id: I4e7f88f5f039baf7e01f68a53639022866786af9
Reviewed-on: https://go-review.googlesource.com/c/net/+/509017
Run-TryBot: Damien Neil <dneil@google.com>
Reviewed-by: Jonathan Amsterdam <jba@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>

internal/quic/conn.go

@@ -9,42 +9,87 @@ package quic
 import (
 	"errors"
 	"fmt"
+	"net/netip"
 	"time"
 )
 
 // A Conn is a QUIC connection.
 //
 // Multiple goroutines may invoke methods on a Conn simultaneously.
 type Conn struct {
+	side      connSide
+	listener  connListener
+	testHooks connTestHooks
+	peerAddr  netip.AddrPort
+
 	msgc   chan any
 	donec  chan struct{} // closed when conn loop exits
 	exited bool          // set to make the conn loop exit immediately
 
-	testHooks connTestHooks
+	w           packetWriter
+	acks        [numberSpaceCount]ackState // indexed by number space
+	connIDState connIDState
+	tlsState    tlsState
+	loss        lossState
 
 	// idleTimeout is the time at which the connection will be closed due to inactivity.
 	// https://www.rfc-editor.org/rfc/rfc9000#section-10.1
 	maxIdleTimeout time.Duration
 	idleTimeout    time.Time
+
+	peerAckDelayExponent int8 // -1 when unknown
+
+	// Tests only: Send a PING in a specific number space.
+	testSendPingSpace numberSpace
+	testSendPing      sentVal
+}
+
+// The connListener is the Conn's Listener.
+// Defined as an interface so we can swap it out in tests.
+type connListener interface {
+	sendDatagram(p []byte, addr netip.AddrPort) error
 }
 
 // connTestHooks override conn behavior in tests.
 type connTestHooks interface {
 	nextMessage(msgc chan any, nextTimeout time.Time) (now time.Time, message any)
 }
 
-func newConn(now time.Time, hooks connTestHooks) (*Conn, error) {
+func newConn(now time.Time, side connSide, initialConnID []byte, peerAddr netip.AddrPort, l connListener, hooks connTestHooks) (*Conn, error) {
 	c := &Conn{
-		donec:          make(chan struct{}),
-		testHooks:      hooks,
-		maxIdleTimeout: defaultMaxIdleTimeout,
-		idleTimeout:    now.Add(defaultMaxIdleTimeout),
+		side:                 side,
+		listener:             l,
+		peerAddr:             peerAddr,
+		msgc:                 make(chan any, 1),
+		donec:                make(chan struct{}),
+		testHooks:            hooks,
+		maxIdleTimeout:       defaultMaxIdleTimeout,
+		idleTimeout:          now.Add(defaultMaxIdleTimeout),
+		peerAckDelayExponent: -1,
 	}
 
 	// A one-element buffer allows us to wake a Conn's event loop as a
 	// non-blocking operation.
 	c.msgc = make(chan any, 1)
 
+	if c.side == clientSide {
+		if err := c.connIDState.initClient(newRandomConnID); err != nil {
+			return nil, err
+		}
+		initialConnID = c.connIDState.dstConnID()
+	} else {
+		if err := c.connIDState.initServer(newRandomConnID, initialConnID); err != nil {
+			return nil, err
+		}
+	}
+
+	// The smallest allowed maximum QUIC datagram size is 1200 bytes.
+	// TODO: PMTU discovery.
+	const maxDatagramSize = 1200
+	c.loss.init(c.side, maxDatagramSize, now)
+
+	c.tlsState.init(c.side, initialConnID)
+
 	go c.loop(now)
 	return c, nil
 }
@@ -76,7 +121,14 @@ func (c *Conn) loop(now time.Time) {
 	}
 
 	for !c.exited {
-		nextTimeout := c.idleTimeout
+		sendTimeout := c.maybeSend(now) // try sending
+
+		// Note that we only need to consider the ack timer for the App Data space,
+		// since the Initial and Handshake spaces always ack immediately.
+		nextTimeout := sendTimeout
+		nextTimeout = firstTime(nextTimeout, c.idleTimeout)
+		nextTimeout = firstTime(nextTimeout, c.loss.timer)
+		nextTimeout = firstTime(nextTimeout, c.acks[appDataSpace].nextAck)
 
 		var m any
 		if hooks != nil {
@@ -100,6 +152,9 @@ func (c *Conn) loop(now time.Time) {
 			now = time.Now()
 		}
 		switch m := m.(type) {
+		case *datagram:
+			c.handleDatagram(now, m)
+			m.recycle()
 		case timerEvent:
 			// A connection timer has expired.
 			if !now.Before(c.idleTimeout) {
@@ -109,6 +164,7 @@ func (c *Conn) loop(now time.Time) {
 				c.exited = true
 				return
 			}
+			c.loss.advance(now, c.handleAckOrLoss)
 		case func(time.Time, *Conn):
 			// Send a func to msgc to run it on the main Conn goroutine
 			m(now, c)
@@ -146,10 +202,30 @@ func (c *Conn) runOnLoop(f func(now time.Time, c *Conn)) error {
 	return nil
 }
 
+// abort terminates a connection with an error.
+func (c *Conn) abort(now time.Time, err error) {
+	// TODO: Send CONNECTION_CLOSE frames.
+	c.exit()
+}
+
 // exit fully terminates a connection immediately.
 func (c *Conn) exit() {
 	c.runOnLoop(func(now time.Time, c *Conn) {
 		c.exited = true
 	})
 	<-c.donec
 }
+
+// firstTime returns the earliest non-zero time, or zero if both times are zero.
+func firstTime(a, b time.Time) time.Time {
+	switch {
+	case a.IsZero():
+		return b
+	case b.IsZero():
+		return a
+	case a.Before(b):
+		return a
+	default:
+		return b
+	}
+}

internal/quic/conn_loss.go

@@ -0,0 +1,46 @@
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//go:build go1.21
+
+package quic
+
+import "fmt"
+
+// handleAckOrLoss deals with the final fate of a packet we sent:
+// Either the peer acknowledges it, or we declare it lost.
+//
+// In order to handle packet loss, we must retain any information sent to the peer
+// until the peer has acknowledged it.
+//
+// When information is acknowledged, we can discard it.
+//
+// When information is lost, we mark it for retransmission.
+// See RFC 9000, Section 13.3 for a complete list of information which is retransmitted on loss.
+// https://www.rfc-editor.org/rfc/rfc9000#section-13.3
+func (c *Conn) handleAckOrLoss(space numberSpace, sent *sentPacket, fate packetFate) {
+	// The list of frames in a sent packet is marshaled into a buffer in the sentPacket
+	// by the packetWriter. Unmarshal that buffer here. This code must be kept in sync with
+	// packetWriter.append*.
+	//
+	// A sent packet meets its fate (acked or lost) only once, so it's okay to consume
+	// the sentPacket's buffer here.
+	for !sent.done() {
+		switch f := sent.next(); f {
+		default:
+			panic(fmt.Sprintf("BUG: unhandled lost frame type %x", f))
+		case frameTypeAck:
+			// Unlike most information, loss of an ACK frame does not trigger
+			// retransmission. ACKs are sent in response to ack-eliciting packets,
+			// and always contain the latest information available.
+			//
+			// Acknowledgement of an ACK frame may allow us to discard information
+			// about older packets.
+			largest := packetNumber(sent.nextInt())
+			if fate == packetAcked {
+				c.acks[space].handleAck(largest)
+			}
+		}
+	}
+}