implement tcp.Policy and refactor rto to use it

tcp/conn.go

@@ -76,16 +76,10 @@ type ConnConfig struct {
 	// Logger sets the [Conn] logger.
 	// Lower level logging available at [Handler.SetLoggers] via [Conn.InternalHandler].
 	Logger *slog.Logger
-	// LossRecovery is the optional packet-loss recovery algorithm (RTO,
-	// congestion control, ...) for the connection. If set, Nanotime must also be
-	// set (else Configure returns an error). Leaving it nil disables loss
-	// recovery. See [LossRecovery].
-	LossRecovery Policy
-	// Nanotime is the monotonic time source in nanoseconds (the func() int64
-	// convention used across lneto) that drives LossRecovery. It is required when
-	// LossRecovery is set and unused otherwise. The tcp package reads it only to
-	// stamp the loss-recovery hooks; it holds no clock itself.
-	Nanotime func() int64
+	// Policy is the optional transmit-steering algorithm (RTO, congestion
+	// control, ...) for the connection. nil disables it. A Policy needing time
+	// carries its own clock. See [Policy].
+	Policy Policy
 }
 
 // Configure should be called on any newly created connection before usage. See [ConnConfig].
@@ -93,10 +87,6 @@ func (conn *Conn) Configure(config ConnConfig) (err error) {
 	if config.RWBackoff == nil {
 		return lneto.ErrMissingHALConfig
 	}
-	if config.LossRecovery != nil && config.Nanotime == nil {
-		// The tcp package holds no clock: a loss-recovery algorithm cannot run without it.
-		return lneto.ErrInvalidConfig
-	}
 	conn.mu.Lock()
 	defer conn.mu.Unlock()
 	err = conn.h.SetBuffers(config.TxBuf, config.RxBuf, config.TxPacketQueueSize)
@@ -105,7 +95,7 @@ func (conn *Conn) Configure(config ConnConfig) (err error) {
 	}
 	conn._backoff = config.RWBackoff
 	conn.logger.log = config.Logger
-	conn.h.SetPolicy(config.LossRecovery)
+	conn.h.SetPolicy(config.Policy)
 	return nil
 }
 

tcp/control.go

@@ -95,6 +95,12 @@ func (tcb *ControlBlock) RecvWindow() Size { return tcb.rcv.WND }
 // ISS returns the initial sequence number of the connection that was defined on a call to Open by user.
 func (tcb *ControlBlock) ISS() Value { return tcb.snd.ISS }
 
+// SendUNA returns snd.UNA, the oldest sequence number not yet acked by the remote.
+func (tcb *ControlBlock) SendUNA() Value { return tcb.snd.UNA }
+
+// SendNext returns snd.NXT, one past the highest sequence number sent.
+func (tcb *ControlBlock) SendNext() Value { return tcb.snd.NXT }
+
 // MaxInFlightData returns the maximum size of a segment that can be sent by taking into account
 // the send window size and the unacked data. Returns 0 before StateSynRcvd.
 func (tcb *ControlBlock) MaxInFlightData() Size {
@@ -257,8 +263,20 @@ func (tcb *ControlBlock) HasPendingRetransmit() bool {
 	return tcb._state.TxDataOpen() && tcb.dupack >= retransmitAfterDupacks && tcb.nRetransmit <= tcb.dupack-retransmitAfterDupacks
 }
 
+// RetransmitFrom rewinds snd.NXT back to newNxt so the next PendingSegment and
+// Send calls retransmit unacknowledged data from that sequence number onwards.
+// It must be paired with ringTx.RetransmitFrom to rewind the transmit buffer to
+// the same point. Implements RFC 9293 §3.10.8 (RETRANSMISSION TIMEOUT).
+//
+// It reports false and changes nothing when newNxt falls outside the
+// unacknowledged range [snd.UNA, snd.NXT] or the connection cannot send data, so
+// a misbehaving [Policy] cannot corrupt the send sequence space.
 func (tcb *ControlBlock) RetransmitFrom(newNxt Value) bool {
-	panic("not yet implemented")
+	if !tcb._state.TxDataOpen() {
+		return false
+	} else if newNxt.LessThan(tcb.snd.UNA) || tcb.snd.NXT.LessThan(newNxt) {
+		return false
+	}
 	tcb.snd.NXT = newNxt
 	tcb.dupack = 0
 	tcb.nRetransmit = 0

tcp/handler.go

@@ -31,9 +31,8 @@ type Handler struct {
 	optcodec OptionCodec
 	// reasm tracks out-of-order segments staged in bufRx's free region. Always
 	// enabled once buffers are set (see [Handler.SetBuffers]).
-	reasm    reassembly
-	policy   Policy
-	nanotime func() int64
+	reasm  reassembly
+	policy Policy
 
 	closing    bool
 	shutdownRx bool
@@ -74,11 +73,17 @@ func (h *Handler) SetBuffers(txbuf, rxbuf []byte, packets int) error {
 	return h.bufTx.ResetOrReuse(txbuf, packets, 0)
 }
 
+// SetPolicy installs the transmit-steering algorithm. nil disables it.
+// It should be set before the connection is opened. See [Policy].
 func (h *Handler) SetPolicy(policy Policy) {
 	h.policy = policy
 }
 func (h *Handler) policyEnabled() bool { return h.policy != nil }
 
+// ControlBlock returns the state machine underlying the Handler, mainly so a
+// [Policy] can read the sequence spaces. Not for modification.
+func (h *Handler) ControlBlock() *ControlBlock { return &h.scb }
+
 // LocalPort returns the local port of the connection. Returns 0 if the connection is closed and uninitialized.
 func (h *Handler) LocalPort() uint16 {
 	return h.localPort
@@ -144,7 +149,6 @@ func (h *Handler) reset(localPort, remotePort uint16, iss Value) {
 		// Persist configuration across reopen:
 		validator: h.validator,
 		policy:    h.policy,
-		nanotime:  h.nanotime,
 		logger:    h.logger,
 		// persist memory across repoen:
 		bufTx: h.bufTx,
@@ -221,6 +225,9 @@ func (h *Handler) Recv(incomingPacket []byte) error {
 	if prevState != h.scb.State() {
 		h.info("tcp.Handler:rx-statechange", slog.Uint64("port", uint64(h.localPort)), slog.String("old", prevState.String()), slog.String("new", h.scb.State().String()), slog.String("rxflags", segIncoming.Flags.String()))
 	}
+	if h.policyEnabled() {
+		h.policy.PostRx(h, prevState, tfrm)
+	}
 	if segIncoming.DATALEN != 0 && h.shutdownRx && (h.scb.State() == StateFinWait1 || h.scb.State() == StateFinWait2) {
 		// soypat/lneto#50: the application is done in both directions — read side
 		// shut down (CloseRead) and our FIN sent (Close) — so inbound data has no
@@ -356,18 +363,28 @@ func (h *Handler) Send(b []byte) (int, error) {
 	if h.IsTxOver() {
 		return 0, net.ErrClosed
 	}
+	tfrm, err := NewFrame(b)
+	if err != nil {
+		return 0, err
+	}
+	offset := uint8(5)
+	var holdNew bool
 	if h.policyEnabled() {
-		tfrm, err := NewFrame(b)
-		if err != nil {
-			return 0, err
+		// Hand the Policy a defined frame: zeroed header at the minimum offset.
+		// It may append options and raise the offset, which is read back below.
+		tfrm.ClearHeader()
+		tfrm.SetOffsetAndFlags(offset, 0)
+		rtxFrom, doRtx, hold := h.policy.PreTx(h, tfrm)
+		holdNew = hold
+		if doRtx && h.scb.RetransmitFrom(rtxFrom) {
+			// Retransmission directed by the Policy: rewind the transmit buffer
+			// to match the send sequence so unacknowledged data is resent. Done
+			// before the early short-circuit below so an expired RTO
+			// retransmits even with no new data queued.
+			h.bufTx.RetransmitFrom(rtxFrom)
 		}
-		rtxFrom, doRtx, _ := h.policy.PreTx(h, tfrm)
-		if doRtx {
-			// Go-back-N retransmission directed by loss recovery: rewind the
-			// send sequence and transmit buffer so unacknowledged data is resent
-			// from snd.UNA. Done before the early short-circuit below so an
-			// expired RTO retransmits even with no new data queued.
-			h.scb.RetransmitFrom(rtxFrom)
+		if o, _ := tfrm.OffsetAndFlags(); o > offset && int(o)*4 < len(b) {
+			offset = o
 		}
 	}
 	awaitingSyn := h.AwaitingSynSend()
@@ -383,29 +400,27 @@ func (h *Handler) Send(b []byte) (int, error) {
 		// Early nop short circuit.
 		return 0, nil
 	}
-	tfrm, err := NewFrame(b)
-	if err != nil {
-		return 0, err
-	}
 	if buffered == 0 && h.closing && (h.scb.State() != StateCloseWait || !h.scb.HasPending()) {
 		// If Close called and no more data to be sent, terminate connection.
 		// In CLOSE-WAIT: wait until the pending ACK is sent first, since scb.Close()
 		// overwrites pending with [FIN|ACK] (unlike ESTABLISHED which merges via bitmask).
 		h.closing = false
-		err = h.scb.Close()
+		err := h.scb.Close()
 		if err != nil {
 			h.logerr("tcp.Handler.Close", slog.String("err", errstr(err)), slog.String("state", h.State().String()))
 			h.Abort()
 			return 0, io.EOF
 		}
 	}
-	offset := uint8(5)
-	mss := uint16(len(b) - sizeHeaderTCP)
+	// optHead is where the Handler's own options begin: after the fixed header
+	// and after any options the Policy already wrote, so neither clobbers the other.
+	optHead := int(offset) * 4
+	mss := uint16(len(b) - optHead)
 	var segment Segment
 	if awaitingSyn || requeueControl && h.scb.State() == StateSynSent {
 		// Handling init syn segment.
 		segment = ClientSynSegment(h.bufTx.iss, Size(h.bufRx.Size()))
-		h.optcodec.PutOption16(b[sizeHeaderTCP:], OptMaxSegmentSize, mss)
+		h.optcodec.PutOption16(b[optHead:], OptMaxSegmentSize, mss)
 		offset++
 		if requeueControl {
 			h.info("tcp.Handler:requeue-syn", slog.Uint64("port", uint64(h.localPort)), slog.Uint64("rport", uint64(h.remotePort)))
@@ -417,7 +432,7 @@ func (h *Handler) Send(b []byte) (int, error) {
 			WND:   Size(h.bufRx.Free()),
 			Flags: synack,
 		}
-		h.optcodec.PutOption16(b[sizeHeaderTCP:], OptMaxSegmentSize, mss)
+		h.optcodec.PutOption16(b[optHead:], OptMaxSegmentSize, mss)
 		offset++
 		h.info("tcp.Handler:requeue-synack", slog.Uint64("port", uint64(h.localPort)), slog.Uint64("rport", uint64(h.remotePort)))
 	} else if requeueControl {
@@ -425,17 +440,22 @@ func (h *Handler) Send(b []byte) (int, error) {
 		return 0, nil
 	} else {
 		var ok bool
-		maxPayload := len(b) - sizeHeaderTCP
+		maxPayload := len(b) - optHead
+		if holdNew && !h.nextSegmentIsRetransmit() {
+			// Policy is holding new data back (congestion window exhausted).
+			// A retransmission it directed in this same call still proceeds.
+			maxPayload = 0
+		}
 		segment, ok = h.scb.PendingSegment(maxPayload)
 		segment.WND = h.recvWindow()
 		if !ok {
 			// No pending control segment or data to send. Yield.
 			return 0, nil
 		} else if segment.Flags == synack {
-			h.optcodec.PutOption16(b[sizeHeaderTCP:], OptMaxSegmentSize, mss)
+			h.optcodec.PutOption16(b[optHead:], OptMaxSegmentSize, mss)
 			offset++
 		} else if segment.DATALEN > 0 {
-			n, err := h.bufTx.MakePacket(b[sizeHeaderTCP:sizeHeaderTCP+segment.DATALEN], segment.SEQ)
+			n, err := h.bufTx.MakePacket(b[optHead:optHead+int(segment.DATALEN)], segment.SEQ)
 			if err != nil {
 				return 0, err
 			}
@@ -452,15 +472,19 @@ func (h *Handler) Send(b []byte) (int, error) {
 	} else if prevState != h.scb.State() && h.logenabled(slog.LevelInfo) {
 		h.info("tcp.Handler:tx-statechange", slog.Uint64("port", uint64(h.localPort)), slog.String("oldState", prevState.String()), slog.String("newState", h.scb.State().String()), slog.String("txflags", segment.Flags.String()))
 	}
-	if h.policyEnabled() {
-		h.policy.PostTx(h, tfrm)
-	}
 	h.requeueControl = false
 	tfrm.SetSourcePort(h.localPort)
 	tfrm.SetDestinationPort(h.remotePort)
 	tfrm.SetSegment(segment, offset)
 	tfrm.SetUrgentPtr(0)
 	datalen := int(offset)*4 + int(segment.DATALEN)
+	if h.policyEnabled() {
+		// Frame trimmed to what is actually emitted so the Policy's Payload()
+		// is the segment data and nothing more.
+		if sent, err := NewFrame(b[:datalen]); err == nil {
+			h.policy.PostTx(h, sent)
+		}
+	}
 	closedSuccess := prevState == StateTimeWait && segment.Flags.HasAny(FlagACK)
 	if closedSuccess {
 		h.reset(0, 0, 0)
@@ -472,6 +496,14 @@ func (h *Handler) Send(b []byte) (int, error) {
 	return datalen, nil
 }
 
+// nextSegmentIsRetransmit reports whether the next data segment would resend
+// already-transmitted bytes rather than open new sequence space. Used to let a
+// retransmission through while a [Policy] holds new data back.
+func (h *Handler) nextSegmentIsRetransmit() bool {
+	endSeq, hasSent := h.bufTx.sentEndSeq()
+	return hasSent && h.scb.snd.NXT.LessThan(endSeq)
+}
+
 // Write implements [io.Writer] by copying b to a internal buffer to be sent over the network on the next
 // [Handler.Send] call that can send data to remote peer. Use [Handler.Free] to know the maximum length the argument slice can be before erroring.
 func (h *Handler) Write(b []byte) (int, error) {

tcp/loss_test.go

@@ -1,262 +0,0 @@
-package tcp
-
-import (
-	"math/rand"
-	"testing"
-
-	"github.com/soypat/lneto/ethernet"
-)
-
-// recordingLoss is a test LossRecovery that records every hook invocation and
-// lets the test steer the directives returned to the Handler. It is the
-// interface counterpart driven by the Handler under test.
-type recordingLoss struct {
-	resets   int
-	preRx    []hookCall
-	preTx    []int64
-	postTx   []hookCall
-	deadline int64 // value NextDeadline reports back.
-
-	// Directives handed back to the Handler.
-	keep bool        // PreRx result. Default true (see newRecordingLoss).
-	tx   TxDirective // PreTx result.
-}
-
-type hookCall struct {
-	seg Segment
-	now int64
-}
-
-func newRecordingLoss() *recordingLoss { return &recordingLoss{keep: true} }
-
-var _ LossRecovery = (*recordingLoss)(nil)
-
-func (l *recordingLoss) Reset()              { l.resets++ }
-func (l *recordingLoss) NextDeadline() int64 { return l.deadline }
-
-func (l *recordingLoss) PreRx(incoming Segment, now int64) RxDirective {
-	l.preRx = append(l.preRx, hookCall{seg: incoming, now: now})
-	return RxDirective{Keep: l.keep}
-}
-
-func (l *recordingLoss) PreTx(now int64) TxDirective {
-	l.preTx = append(l.preTx, now)
-	return l.tx
-}
-
-func (l *recordingLoss) PostTx(outgoing Segment, now int64) {
-	l.postTx = append(l.postTx, hookCall{seg: outgoing, now: now})
-}
-
-// TestLossRecovery_DisabledByDefault verifies the Handler runs normally with no
-// loss recovery installed: NextDeadline reports no deadline and the transmit/
-// receive paths never touch a nil LossRecovery.
-func TestLossRecovery_DisabledByDefault(t *testing.T) {
-	const mtu = ethernet.MaxMTU
-	rng := rand.New(rand.NewSource(1))
-	client, server := newHandler(t, mtu, 3), newHandler(t, mtu, 3)
-	setupClientServer(t, rng, client, server)
-
-	if d := client.NextDeadline(); d != 0 {
-		t.Fatalf("NextDeadline with no loss recovery = %d, want 0", d)
-	}
-	var buf [mtu]byte
-	establish(t, client, server, buf[:]) // must not panic on nil loss recovery.
-}
-
-// TestLossRecovery_HooksInvoked verifies the Handler drives the full hook
-// contract across a handshake: Reset on open, PreTx+PostTx on every transmit,
-// PreRx on every receive, each stamped with the configured monotonic clock.
-func TestLossRecovery_HooksInvoked(t *testing.T) {
-	const mtu = ethernet.MaxMTU
-	rng := rand.New(rand.NewSource(2))
-	client, server := newHandler(t, mtu, 3), newHandler(t, mtu, 3)
-
-	loss := newRecordingLoss()
-	const clockNow = 1_000_000
-	client.SetLossRecovery(loss, func() int64 { return clockNow })
-
-	setupClientServer(t, rng, client, server) // OpenActive → reset → Reset().
-	if loss.resets == 0 {
-		t.Fatal("Reset not called on open")
-	}
-
-	var buf [mtu]byte
-	establish(t, client, server, buf[:])
-
-	// Client emitted SYN and the final ACK: both paths must have hit PreTx/PostTx.
-	if len(loss.preTx) == 0 {
-		t.Fatal("PreTx never called on transmit")
-	}
-	if len(loss.postTx) == 0 {
-		t.Fatal("PostTx never called on transmit")
-	}
-	if len(loss.preTx) != len(loss.postTx) {
-		t.Fatalf("PreTx calls=%d, PostTx calls=%d, want equal", len(loss.preTx), len(loss.postTx))
-	}
-	// Client received the SYN-ACK: PreRx must have seen it.
-	if len(loss.preRx) == 0 {
-		t.Fatal("PreRx never called on receive")
-	}
-
-	// The Handler holds no clock: every hook must be stamped from the supplied
-	// nanotime source.
-	for i, c := range loss.postTx {
-		if c.now != clockNow {
-			t.Fatalf("PostTx[%d].now = %d, want clock %d", i, c.now, clockNow)
-		}
-	}
-	for i, now := range loss.preTx {
-		if now != clockNow {
-			t.Fatalf("PreTx[%d].now = %d, want clock %d", i, now, clockNow)
-		}
-	}
-	for i, c := range loss.preRx {
-		if c.now != clockNow {
-			t.Fatalf("PreRx[%d].now = %d, want clock %d", i, c.now, clockNow)
-		}
-	}
-
-	// PostTx receives the segment actually emitted: the first is the SYN.
-	if !loss.postTx[0].seg.Flags.HasAny(FlagSYN) {
-		t.Fatalf("first PostTx segment flags=%s, want SYN", loss.postTx[0].seg.Flags)
-	}
-}
-
-// TestLossRecovery_NextDeadlineDelegates verifies NextDeadline is forwarded to
-// the installed LossRecovery unchanged.
-func TestLossRecovery_NextDeadlineDelegates(t *testing.T) {
-	const mtu = ethernet.MaxMTU
-	rng := rand.New(rand.NewSource(3))
-	client, server := newHandler(t, mtu, 3), newHandler(t, mtu, 3)
-
-	loss := newRecordingLoss()
-	loss.deadline = 4242
-	client.SetLossRecovery(loss, func() int64 { return 1 })
-	setupClientServer(t, rng, client, server)
-
-	if d := client.NextDeadline(); d != 4242 {
-		t.Fatalf("NextDeadline = %d, want delegated 4242", d)
-	}
-}
-
-// TestLossRecovery_PreRxDropsSegment verifies a PreRx directive of Keep=false
-// drops the segment before the state machine sees it: the payload is not
-// buffered and connection state is untouched.
-func TestLossRecovery_PreRxDropsSegment(t *testing.T) {
-	const mtu = ethernet.MaxMTU
-	rng := rand.New(rand.NewSource(4))
-	client, server := newHandler(t, mtu, 3), newHandler(t, mtu, 3)
-
-	loss := newRecordingLoss()
-	server.SetLossRecovery(loss, func() int64 { return 1 })
-	setupClientServer(t, rng, client, server)
-	var buf [mtu]byte
-	establish(t, client, server, buf[:]) // keep=true so handshake completes.
-
-	// Now start dropping everything the server receives.
-	loss.keep = false
-	preRxBefore := len(loss.preRx)
-
-	data := []byte("dropme")
-	if _, err := client.Write(data); err != nil {
-		t.Fatal("client write:", err)
-	}
-	clear(buf[:])
-	n, err := client.Send(buf[:])
-	if err != nil {
-		t.Fatal("client send:", err)
-	}
-
-	if err := server.Recv(buf[:n]); err != nil {
-		t.Fatalf("dropped segment must return nil, got %v", err)
-	}
-	if len(loss.preRx) != preRxBefore+1 {
-		t.Fatalf("PreRx calls=%d, want %d (segment must reach PreRx)", len(loss.preRx), preRxBefore+1)
-	}
-	if server.BufferedInput() != 0 {
-		t.Fatalf("dropped segment must not be buffered, got %d bytes", server.BufferedInput())
-	}
-	if server.State() != StateEstablished {
-		t.Fatalf("dropped segment must not change state, got %s", server.State())
-	}
-}
-
-// TestLossRecovery_PreTxRetransmitAll verifies a PreTx directive of
-// RetransmitAll drives go-back-N: the Handler rewinds and re-emits already-sent,
-// unacknowledged data from snd.UNA on the next transmit.
-func TestLossRecovery_PreTxRetransmitAll(t *testing.T) {
-	const mtu = ethernet.MaxMTU
-	rng := rand.New(rand.NewSource(5))
-	client, server := newHandler(t, mtu, 3), newHandler(t, mtu, 3)
-
-	loss := newRecordingLoss()
-	client.SetLossRecovery(loss, func() int64 { return 1 })
-	setupClientServer(t, rng, client, server)
-	var buf [mtu]byte
-	establish(t, client, server, buf[:])
-
-	// Emit one data segment; server never ACKs, so it stays unacknowledged.
-	data := []byte("payload")
-	if _, err := client.Write(data); err != nil {
-		t.Fatal("client write:", err)
-	}
-	clear(buf[:])
-	n, err := client.Send(buf[:])
-	if err != nil {
-		t.Fatal("client send data:", err)
-	}
-	if n <= sizeHeaderTCP {
-		t.Fatal("expected data segment")
-	}
-	firstSeg := mustSegment(t, buf[:n], n-sizeHeaderTCP)
-
-	// Direct go-back-N on the next transmit.
-	loss.tx = TxDirective{RetransmitAll: true}
-	clear(buf[:])
-	n, err = client.Send(buf[:])
-	if err != nil {
-		t.Fatal("client send retransmit:", err)
-	}
-	if n <= sizeHeaderTCP {
-		t.Fatal("expected retransmitted data segment")
-	}
-	rtSeg := mustSegment(t, buf[:n], n-sizeHeaderTCP)
-
-	if rtSeg.SEQ != firstSeg.SEQ {
-		t.Fatalf("retransmit SEQ=%d, want original UNA SEQ=%d (go-back-N)", rtSeg.SEQ, firstSeg.SEQ)
-	}
-	if rtSeg.DATALEN != firstSeg.DATALEN {
-		t.Fatalf("retransmit DATALEN=%d, want %d", rtSeg.DATALEN, firstSeg.DATALEN)
-	}
-}
-
-// TestLossRecovery_ResetOnReopen verifies Reset fires on every (re)open and on
-// Abort, so a single LossRecovery value can be reused across connection reuse.
-func TestLossRecovery_ResetOnReopen(t *testing.T) {
-	const mtu = ethernet.MaxMTU
-	client := newHandler(t, mtu, 3)
-	loss := newRecordingLoss()
-	client.SetLossRecovery(loss, func() int64 { return 1 })
-
-	if err := client.OpenActive(1234, 5678, 0); err != nil {
-		t.Fatal("open 1:", err)
-	}
-	afterOpen := loss.resets
-	if afterOpen == 0 {
-		t.Fatal("Reset not called on first open")
-	}
-
-	client.Abort()
-	if loss.resets <= afterOpen {
-		t.Fatalf("Reset not called on Abort: resets=%d, want >%d", loss.resets, afterOpen)
-	}
-	afterAbort := loss.resets
-
-	if err := client.OpenActive(1234, 5678, 0); err != nil {
-		t.Fatal("open 2:", err)
-	}
-	if loss.resets <= afterAbort {
-		t.Fatalf("Reset not called on reopen: resets=%d, want >%d", loss.resets, afterAbort)
-	}
-}

tcp/policy.go

@@ -1,15 +1,24 @@
 package tcp
 
+// Policy observes segment traffic and steers transmit behaviour: RTO,
+// congestion control and the like (discussion #157). The tcp package holds no
+// clock, so a Policy needing time carries its own (issue #140).
+// Introspection stays off the interface; put it on the concrete type.
 type Policy interface {
+	// Reset returns the Policy to its pre-connection state. Called on every
+	// (re)open and Abort. Must preserve configuration such as a clock.
 	Reset()
 	// PreTx is called before writing to a frame.
 	// The outgoing frame options can be set by the Policy and will be respected if Frame offset >5.
+	// rtxFrom is ignored unless within [snd.UNA, snd.NXT]. Nothing is committed
+	// until PostTx: a transmit attempt may emit no segment at all.
 	PreTx(h *Handler, outgoingOpts Frame) (rtxFrom Value, retransmit, holdNew bool)
 	// PreRx is called by [Handler] on every incoming segment.
 	// PreRx can choose to drop segment if it returns keep=false.
 	PreRx(h *Handler, incoming Frame) (keep bool)
 	// PostRx is called by [Handler] after accepting an incoming segment.
+	// TODO: congestion control will also want the pre-Recv snd.UNA here.
 	PostRx(h *Handler, prevState State, accepted Frame)
-	// PostTx called on leaving the transmit path.
+	// PostTx called on leaving the transmit path with the fully written frame.
 	PostTx(h *Handler, outgoing Frame)
 }

tcp/policy_test.go

@@ -0,0 +1,434 @@
+package tcp
+
+import (
+	"math/rand"
+	"testing"
+
+	"github.com/soypat/lneto/ethernet"
+)
+
+// recordingPolicy records every hook invocation and lets the test steer what is
+// returned to the Handler. It is the [Policy] counterpart driven by the Handler
+// under test.
+type recordingPolicy struct {
+	resets int
+	preRx  []Segment
+	preTx  int
+	postRx []Segment
+	postTx []txRecord
+
+	// Values handed back to the Handler.
+	keep       bool // PreRx result. Default true (see newRecordingPolicy).
+	rtxFrom    Value
+	retransmit bool
+	holdNew    bool
+	// writeOpts, when non-empty, is appended as TCP options by PreTx.
+	writeOpts []byte
+}
+
+// txRecord is what PostTx observed on the emitted frame.
+type txRecord struct {
+	seg    Segment
+	offset uint8
+	sport  uint16
+	dport  uint16
+}
+
+func newRecordingPolicy() *recordingPolicy { return &recordingPolicy{keep: true} }
+
+var _ Policy = (*recordingPolicy)(nil)
+
+func (p *recordingPolicy) Reset() { p.resets++ }
+
+func (p *recordingPolicy) PreRx(h *Handler, incoming Frame) bool {
+	p.preRx = append(p.preRx, incoming.Segment(len(incoming.Payload())))
+	return p.keep
+}
+
+func (p *recordingPolicy) PostRx(h *Handler, prevState State, accepted Frame) {
+	p.postRx = append(p.postRx, accepted.Segment(len(accepted.Payload())))
+}
+
+func (p *recordingPolicy) PreTx(h *Handler, outgoingOpts Frame) (Value, bool, bool) {
+	p.preTx++
+	if len(p.writeOpts) > 0 {
+		// Raise the offset first: Options() is sized from it.
+		words := uint8(5 + (len(p.writeOpts)+3)/4)
+		outgoingOpts.SetOffsetAndFlags(words, 0)
+		copy(outgoingOpts.Options(), p.writeOpts)
+	}
+	return p.rtxFrom, p.retransmit, p.holdNew
+}
+
+func (p *recordingPolicy) PostTx(h *Handler, outgoing Frame) {
+	offset, _ := outgoing.OffsetAndFlags()
+	p.postTx = append(p.postTx, txRecord{
+		seg:    outgoing.Segment(len(outgoing.Payload())),
+		offset: offset,
+		sport:  outgoing.SourcePort(),
+		dport:  outgoing.DestinationPort(),
+	})
+}
+
+// TestPolicy_DisabledByDefault verifies the Handler runs normally with no Policy
+// installed: the transmit and receive paths never touch a nil Policy.
+func TestPolicy_DisabledByDefault(t *testing.T) {
+	const mtu = ethernet.MaxMTU
+	rng := rand.New(rand.NewSource(1))
+	client, server := newHandler(t, mtu, 3), newHandler(t, mtu, 3)
+	setupClientServer(t, rng, client, server)
+
+	var buf [mtu]byte
+	establish(t, client, server, buf[:]) // must not panic on nil Policy.
+}
+
+// TestPolicy_HooksInvoked verifies the Handler drives the full hook contract
+// across a handshake: Reset on open, PreTx+PostTx on transmit, PreRx+PostRx on
+// receive.
+func TestPolicy_HooksInvoked(t *testing.T) {
+	const mtu = ethernet.MaxMTU
+	rng := rand.New(rand.NewSource(2))
+	client, server := newHandler(t, mtu, 3), newHandler(t, mtu, 3)
+
+	pol := newRecordingPolicy()
+	client.SetPolicy(pol)
+
+	setupClientServer(t, rng, client, server) // OpenActive → reset → Reset().
+	if pol.resets == 0 {
+		t.Fatal("Reset not called on open")
+	}
+
+	var buf [mtu]byte
+	establish(t, client, server, buf[:])
+
+	if pol.preTx == 0 {
+		t.Fatal("PreTx never called on transmit")
+	}
+	if len(pol.postTx) == 0 {
+		t.Fatal("PostTx never called on transmit")
+	}
+	if pol.preTx < len(pol.postTx) {
+		t.Fatalf("PreTx calls=%d < PostTx calls=%d: PostTx must never fire without PreTx", pol.preTx, len(pol.postTx))
+	}
+	// Client received the SYN-ACK and accepted it.
+	if len(pol.preRx) == 0 {
+		t.Fatal("PreRx never called on receive")
+	}
+	if len(pol.postRx) == 0 {
+		t.Fatal("PostRx never called on accepted receive")
+	}
+	// PostTx receives the segment actually emitted: the first is the SYN.
+	if !pol.postTx[0].seg.Flags.HasAny(FlagSYN) {
+		t.Fatalf("first PostTx segment flags=%s, want SYN", pol.postTx[0].seg.Flags)
+	}
+}
+
+// TestPolicy_PostTxSeesWrittenFrame verifies PostTx observes the fully populated
+// frame — ports, sequence numbers and payload length as emitted — and not the
+// frame as it stood before the segment was written into it.
+func TestPolicy_PostTxSeesWrittenFrame(t *testing.T) {
+	const mtu = ethernet.MaxMTU
+	rng := rand.New(rand.NewSource(6))
+	client, server := newHandler(t, mtu, 3), newHandler(t, mtu, 3)
+
+	pol := newRecordingPolicy()
+	client.SetPolicy(pol)
+	setupClientServer(t, rng, client, server)
+	var buf [mtu]byte
+	establish(t, client, server, buf[:])
+
+	data := []byte("payload")
+	if _, err := client.Write(data); err != nil {
+		t.Fatal("client write:", err)
+	}
+	clear(buf[:])
+	n, err := client.Send(buf[:])
+	if err != nil {
+		t.Fatal("client send:", err)
+	}
+	last := pol.postTx[len(pol.postTx)-1]
+	wantSeg := mustSegment(t, buf[:n], n-int(last.offset)*4)
+	if last.seg != wantSeg {
+		t.Fatalf("PostTx segment=%+v, want emitted %+v", last.seg, wantSeg)
+	}
+	if int(last.seg.DATALEN) != len(data) {
+		t.Fatalf("PostTx DATALEN=%d, want %d", last.seg.DATALEN, len(data))
+	}
+	if last.sport != client.LocalPort() || last.dport != client.RemotePort() {
+		t.Fatalf("PostTx ports=%d→%d, want %d→%d", last.sport, last.dport, client.LocalPort(), client.RemotePort())
+	}
+}
+
+// TestPolicy_NoPostTxWithoutSegment verifies a transmit attempt that emits
+// nothing still runs PreTx but never PostTx, so a Policy cannot mistake a
+// no-op Send for a segment on the wire.
+func TestPolicy_NoPostTxWithoutSegment(t *testing.T) {
+	const mtu = ethernet.MaxMTU
+	rng := rand.New(rand.NewSource(7))
+	client, server := newHandler(t, mtu, 3), newHandler(t, mtu, 3)
+
+	pol := newRecordingPolicy()
+	client.SetPolicy(pol)
+	setupClientServer(t, rng, client, server)
+	var buf [mtu]byte
+	establish(t, client, server, buf[:])
+
+	preTxBefore, postTxBefore := pol.preTx, len(pol.postTx)
+	n, err := client.Send(buf[:]) // Nothing queued: no segment.
+	if err != nil {
+		t.Fatal("client send:", err)
+	}
+	if n != 0 {
+		t.Fatalf("expected no segment, got %d bytes", n)
+	}
+	if pol.preTx != preTxBefore+1 {
+		t.Fatalf("PreTx calls=%d, want %d: PreTx must run on every attempt", pol.preTx, preTxBefore+1)
+	}
+	if len(pol.postTx) != postTxBefore {
+		t.Fatalf("PostTx calls=%d, want %d: no segment was emitted", len(pol.postTx), postTxBefore)
+	}
+}
+
+// TestPolicy_PreTxOptions verifies options written by PreTx survive to the wire:
+// the data offset accounts for them and the payload starts after them.
+func TestPolicy_PreTxOptions(t *testing.T) {
+	const mtu = ethernet.MaxMTU
+	rng := rand.New(rand.NewSource(8))
+	client, server := newHandler(t, mtu, 3), newHandler(t, mtu, 3)
+
+	pol := newRecordingPolicy()
+	client.SetPolicy(pol)
+	setupClientServer(t, rng, client, server)
+	var buf [mtu]byte
+	establish(t, client, server, buf[:])
+
+	// One 4-byte option word: NOP,NOP,NOP,EOL.
+	opts := []byte{1, 1, 1, 0}
+	pol.writeOpts = opts
+
+	data := []byte("payload")
+	if _, err := client.Write(data); err != nil {
+		t.Fatal("client write:", err)
+	}
+	clear(buf[:])
+	n, err := client.Send(buf[:])
+	if err != nil {
+		t.Fatal("client send:", err)
+	}
+	frm, err := NewFrame(buf[:n])
+	if err != nil {
+		t.Fatal("frame:", err)
+	}
+	offset, _ := frm.OffsetAndFlags()
+	if offset != 6 {
+		t.Fatalf("data offset=%d, want 6 (header + one option word)", offset)
+	}
+	if got := frm.Options(); string(got) != string(opts) {
+		t.Fatalf("options=%v, want %v", got, opts)
+	}
+	if got := frm.Payload(); string(got) != string(data) {
+		t.Fatalf("payload=%q, want %q: options must not overlap data", got, data)
+	}
+	if n != int(offset)*4+len(data) {
+		t.Fatalf("frame length=%d, want %d", n, int(offset)*4+len(data))
+	}
+}
+
+// TestPolicy_PreRxDropsSegment verifies keep=false drops the segment before the
+// state machine sees it: the payload is not buffered, connection state is
+// untouched and PostRx never fires.
+func TestPolicy_PreRxDropsSegment(t *testing.T) {
+	const mtu = ethernet.MaxMTU
+	rng := rand.New(rand.NewSource(4))
+	client, server := newHandler(t, mtu, 3), newHandler(t, mtu, 3)
+
+	pol := newRecordingPolicy()
+	server.SetPolicy(pol)
+	setupClientServer(t, rng, client, server)
+	var buf [mtu]byte
+	establish(t, client, server, buf[:]) // keep=true so handshake completes.
+
+	// Now start dropping everything the server receives.
+	pol.keep = false
+	preRxBefore, postRxBefore := len(pol.preRx), len(pol.postRx)
+
+	data := []byte("dropme")
+	if _, err := client.Write(data); err != nil {
+		t.Fatal("client write:", err)
+	}
+	clear(buf[:])
+	n, err := client.Send(buf[:])
+	if err != nil {
+		t.Fatal("client send:", err)
+	}
+
+	if err := server.Recv(buf[:n]); err != nil {
+		t.Fatalf("dropped segment must return nil, got %v", err)
+	}
+	if len(pol.preRx) != preRxBefore+1 {
+		t.Fatalf("PreRx calls=%d, want %d (segment must reach PreRx)", len(pol.preRx), preRxBefore+1)
+	}
+	if len(pol.postRx) != postRxBefore {
+		t.Fatalf("PostRx calls=%d, want %d: a dropped segment was never accepted", len(pol.postRx), postRxBefore)
+	}
+	if server.BufferedInput() != 0 {
+		t.Fatalf("dropped segment must not be buffered, got %d bytes", server.BufferedInput())
+	}
+	if server.State() != StateEstablished {
+		t.Fatalf("dropped segment must not change state, got %s", server.State())
+	}
+}
+
+// TestPolicy_PreTxRetransmit verifies a PreTx retransmit directive drives
+// go-back-N: the Handler rewinds the send sequence and the transmit buffer
+// together and re-emits already-sent, unacknowledged data from snd.UNA.
+func TestPolicy_PreTxRetransmit(t *testing.T) {
+	const mtu = ethernet.MaxMTU
+	rng := rand.New(rand.NewSource(5))
+	client, server := newHandler(t, mtu, 3), newHandler(t, mtu, 3)
+
+	pol := newRecordingPolicy()
+	client.SetPolicy(pol)
+	setupClientServer(t, rng, client, server)
+	var buf [mtu]byte
+	establish(t, client, server, buf[:])
+
+	// Emit one data segment; server never ACKs, so it stays unacknowledged.
+	data := []byte("payload")
+	if _, err := client.Write(data); err != nil {
+		t.Fatal("client write:", err)
+	}
+	clear(buf[:])
+	n, err := client.Send(buf[:])
+	if err != nil {
+		t.Fatal("client send data:", err)
+	}
+	if n <= sizeHeaderTCP {
+		t.Fatal("expected data segment")
+	}
+	firstSeg := mustSegment(t, buf[:n], n-sizeHeaderTCP)
+	firstData := append([]byte(nil), buf[sizeHeaderTCP:n]...)
+
+	// Direct go-back-N on the next transmit.
+	pol.rtxFrom, pol.retransmit = client.ControlBlock().SendUNA(), true
+	clear(buf[:])
+	n, err = client.Send(buf[:])
+	if err != nil {
+		t.Fatal("client send retransmit:", err)
+	}
+	if n <= sizeHeaderTCP {
+		t.Fatal("expected retransmitted data segment")
+	}
+	rtSeg := mustSegment(t, buf[:n], n-sizeHeaderTCP)
+
+	if rtSeg.SEQ != firstSeg.SEQ {
+		t.Fatalf("retransmit SEQ=%d, want original UNA SEQ=%d (go-back-N)", rtSeg.SEQ, firstSeg.SEQ)
+	}
+	if rtSeg.DATALEN != firstSeg.DATALEN {
+		t.Fatalf("retransmit DATALEN=%d, want %d", rtSeg.DATALEN, firstSeg.DATALEN)
+	}
+	if got := buf[sizeHeaderTCP:n]; string(got) != string(firstData) {
+		t.Fatalf("retransmit payload=%q, want %q", got, firstData)
+	}
+}
+
+// TestPolicy_PreTxRetransmitOutOfRange verifies an out-of-range rtxFrom is
+// refused, leaving the send sequence and transmit buffer untouched.
+func TestPolicy_PreTxRetransmitOutOfRange(t *testing.T) {
+	const mtu = ethernet.MaxMTU
+	rng := rand.New(rand.NewSource(9))
+	client, server := newHandler(t, mtu, 3), newHandler(t, mtu, 3)
+
+	pol := newRecordingPolicy()
+	client.SetPolicy(pol)
+	setupClientServer(t, rng, client, server)
+	var buf [mtu]byte
+	establish(t, client, server, buf[:])
+
+	if _, err := client.Write([]byte("payload")); err != nil {
+		t.Fatal("client write:", err)
+	}
+	clear(buf[:])
+	if _, err := client.Send(buf[:]); err != nil {
+		t.Fatal("client send data:", err)
+	}
+	nxtBefore := client.ControlBlock().SendNext()
+
+	// Well beyond snd.NXT: must be refused.
+	pol.rtxFrom, pol.retransmit = nxtBefore+1000, true
+	clear(buf[:])
+	if _, err := client.Send(buf[:]); err != nil {
+		t.Fatal("client send:", err)
+	}
+	if got := client.ControlBlock().SendNext(); got != nxtBefore {
+		t.Fatalf("snd.NXT=%d, want unchanged %d: out-of-range rtxFrom must be refused", got, nxtBefore)
+	}
+}
+
+// TestPolicy_HoldNew verifies holdNew suppresses new data while leaving control
+// segments free to go out.
+func TestPolicy_HoldNew(t *testing.T) {
+	const mtu = ethernet.MaxMTU
+	rng := rand.New(rand.NewSource(10))
+	client, server := newHandler(t, mtu, 3), newHandler(t, mtu, 3)
+
+	pol := newRecordingPolicy()
+	client.SetPolicy(pol)
+	setupClientServer(t, rng, client, server)
+	var buf [mtu]byte
+	establish(t, client, server, buf[:])
+
+	pol.holdNew = true
+	if _, err := client.Write([]byte("payload")); err != nil {
+		t.Fatal("client write:", err)
+	}
+	clear(buf[:])
+	n, err := client.Send(buf[:])
+	if err != nil {
+		t.Fatal("client send:", err)
+	}
+	if n > sizeHeaderTCP {
+		t.Fatalf("holdNew must suppress new data, got %d payload bytes", n-sizeHeaderTCP)
+	}
+
+	// Releasing the hold lets the same data out.
+	pol.holdNew = false
+	clear(buf[:])
+	n, err = client.Send(buf[:])
+	if err != nil {
+		t.Fatal("client send after hold:", err)
+	}
+	if n <= sizeHeaderTCP {
+		t.Fatal("data must flow once holdNew is cleared")
+	}
+}
+
+// TestPolicy_ResetOnReopen verifies Reset fires on every (re)open and on Abort,
+// so a single Policy value can be reused across connection reuse.
+func TestPolicy_ResetOnReopen(t *testing.T) {
+	const mtu = ethernet.MaxMTU
+	client := newHandler(t, mtu, 3)
+	pol := newRecordingPolicy()
+	client.SetPolicy(pol)
+
+	if err := client.OpenActive(1234, 5678, 0); err != nil {
+		t.Fatal("open 1:", err)
+	}
+	afterOpen := pol.resets
+	if afterOpen == 0 {
+		t.Fatal("Reset not called on first open")
+	}
+
+	client.Abort()
+	if pol.resets <= afterOpen {
+		t.Fatalf("Reset not called on Abort: resets=%d, want >%d", pol.resets, afterOpen)
+	}
+	afterAbort := pol.resets
+
+	if err := client.OpenActive(1234, 5678, 0); err != nil {
+		t.Fatal("open 2:", err)
+	}
+	if pol.resets <= afterAbort {
+		t.Fatalf("Reset not called on reopen: resets=%d, want >%d", pol.resets, afterAbort)
+	}
+}

tcp/rto/timer.go

@@ -3,6 +3,7 @@ package rto
 import (
 	"time"
 
+	"github.com/soypat/lneto"
 	"github.com/soypat/lneto/tcp"
 )
 
@@ -35,15 +36,14 @@ const (
 )
 
 // Timer implements the RFC 6298 round-trip-time estimator and the single
-// retransmission timer as a [tcp.Policy]. Construct it with new(Timer) and hand
-// it to [tcp.ConnConfig.Policy]; the connection calls [Timer.Reset] on open, so
-// the zero value is ready to use.
+// retransmission timer as a [tcp.Policy]. Construct it with [NewTimer] and hand
+// it to [tcp.ConnConfig.Policy].
 //
 // Timer is a pure, reactive state machine: it observes the segments a connection
-// sends and receives (via the tcp.Policy hooks) and the monotonic time handed
-// in at each hook, and from those alone derives RTT estimates and retransmission
-// decisions. It holds no clock and allocates nothing, which keeps it
-// deterministic for unit testing (see issue #140).
+// sends and receives (via the tcp.Policy hooks) and from those alone derives RTT
+// estimates and retransmission decisions. The tcp package holds no clock, so the
+// Timer carries its own; injecting it keeps the estimator deterministic for unit
+// testing (see issue #140).
 //
 // Timer tracks its own shadow of the send sequence space purely from the segments
 // it observes: [Timer.PostTx] advances the highest sequence sent and [Timer.PreRx]
@@ -53,6 +53,9 @@ const (
 // whose sequence space is not beyond the shadow snd.NXT is a retransmission and
 // is never RTT-sampled.
 type Timer struct {
+	// nanotime is the monotonic time source in nanoseconds. Preserved by Reset.
+	nanotime func() int64
+
 	srtt    time.Duration // smoothed round-trip time (SRTT).
 	rttvar  time.Duration // round-trip-time variation (RTTVAR).
 	rto     time.Duration // current retransmission timeout.
@@ -76,17 +79,28 @@ type Timer struct {
 
 	// expirations counts timeouts since Reset. It exists so a policy sharing this
 	// timer can notice a timeout it did not itself drive: a congestion controller
-	// must collapse its window on one, and when the timer is a peer in a
-	// [tcp.Composite] the controller never sees the timer's directive.
+	// must collapse its window on one, and a policy that composes the timer as a
+	// peer never sees the timer's own directive.
 	expirations uint32
 }
 
 var _ tcp.Policy = (*Timer)(nil)
 
-// Reset returns the estimator to its pre-connection state with the initial RTO.
-// It implements [tcp.Policy] and is called when the connection opens or aborts
-// so the estimator can be reused across connection reuse.
-func (r *Timer) Reset() { *r = Timer{rto: rtoInitial} }
+// Configure prepares the Timer for use with nanotime, the monotonic time source
+// in nanoseconds (the func() int64 convention used across lneto). It must be
+// called before the connection is opened.
+func (r *Timer) Configure(nanotime func() int64) error {
+	if nanotime == nil {
+		return lneto.ErrMissingHALConfig // The estimator cannot run without a clock.
+	}
+	*r = Timer{rto: rtoInitial, nanotime: nanotime}
+	return nil
+}
+
+// Reset returns the estimator to its pre-connection state with the initial RTO,
+// preserving the configured clock. It implements [tcp.Policy] and is called when
+// the connection opens or aborts so the estimator survives connection reuse.
+func (r *Timer) Reset() { *r = Timer{rto: rtoInitial, nanotime: r.nanotime} }
 
 // SmoothedRTT returns the current smoothed round-trip time (SRTT), or zero
 // before the first RTT measurement. It is concrete-type introspection and is
@@ -115,7 +129,9 @@ func (r *Timer) Running() bool { return r.running }
 func (r *Timer) Expirations() uint32 { return r.expirations }
 
 // NextDeadline returns the monotonic-nanosecond instant at which the timer
-// expires, or 0 when it is not armed. It implements [tcp.Policy].
+// expires, or 0 when it is not armed. It is concrete-type introspection, not
+// part of [tcp.Policy]: an event loop that wants to schedule against the RTO
+// holds the Timer it configured and reads this.
 func (r *Timer) NextDeadline() int64 {
 	if !r.running {
 		return 0
@@ -126,19 +142,22 @@ func (r *Timer) NextDeadline() int64 {
 // PreRx keeps every segment: the estimator never drops traffic and records
 // nothing before the connection has decided whether the segment counts. It
 // implements [tcp.Policy].
-func (r *Timer) PreRx(rx tcp.RxMeta) tcp.RxDirective {
-	return tcp.RxDirective{Keep: true}
+func (r *Timer) PreRx(h *tcp.Handler, incoming tcp.Frame) bool {
+	return true
 }
 
 // PostRx samples the RTT and manages the retransmission timer from a segment the
 // connection accepted (RFC 6298 §5.2/§5.3). It implements [tcp.Policy].
 //
-// A refused segment is ignored. Acting on one would let an acknowledgement the
-// state machine rejected, for data never sent, collapse the backoff and take a
-// bogus RTT sample.
-func (r *Timer) PostRx(event tcp.RxEvent) {
-	incoming, now := event.Segment, event.Now
-	if !event.Accepted || !r.haveSeq || !incoming.Flags.HasAny(tcp.FlagACK) {
+// Only accepted segments reach here. Acting on a refused one would let an
+// acknowledgement the state machine rejected, for data never sent, collapse the
+// backoff and take a bogus RTT sample.
+func (r *Timer) PostRx(h *tcp.Handler, prevState tcp.State, accepted tcp.Frame) {
+	r.postRx(accepted.Segment(len(accepted.Payload())), r.nanotime())
+}
+
+func (r *Timer) postRx(incoming tcp.Segment, now int64) {
+	if !r.haveSeq || !incoming.Flags.HasAny(tcp.FlagACK) {
 		return
 	}
 	ack := incoming.ACK
@@ -162,19 +181,18 @@ func (r *Timer) PostRx(event tcp.RxEvent) {
 	}
 }
 
-// WriteOptions adds no TCP options: retransmission timing needs none of its
-// own. It implements [tcp.Policy].
-func (r *Timer) WriteOptions(plan tcp.TxPlan, opts []byte) uint8 { return 0 }
-
 // PreTx reports whether the retransmission timer has expired and, if so, applies
 // the RFC 6298 §5.4–§5.6 timeout response — discard the outstanding RTT sample
-// (Karn), back the RTO off exponentially and restart the timer — returning a
-// directive that asks the connection to retransmit from snd.UNA (go-back-N). It
-// implements [tcp.Policy].
-func (r *Timer) PreTx(intent tcp.TxIntent) tcp.TxDirective {
-	now := intent.Now
+// (Karn), back the RTO off exponentially and restart the timer — and asks the
+// connection to retransmit from snd.UNA (go-back-N). It writes no TCP options:
+// retransmission timing needs none of its own. It implements [tcp.Policy].
+func (r *Timer) PreTx(h *tcp.Handler, outgoingOpts tcp.Frame) (rtxFrom tcp.Value, retransmit, holdNew bool) {
+	return r.preTx(r.nanotime(), h.ControlBlock().SendUNA())
+}
+
+func (r *Timer) preTx(now int64, una tcp.Value) (rtxFrom tcp.Value, retransmit, holdNew bool) {
 	if !r.running || now < r.deadline || r.sndUNA == r.sndNXT {
-		return tcp.TxDirective{}
+		return 0, false, false
 	}
 	r.expirations++
 	r.timing = false // §5.4: do not sample a retransmitted segment.
@@ -184,7 +202,7 @@ func (r *Timer) PreTx(intent tcp.TxIntent) tcp.TxDirective {
 	}
 	r.running = true
 	r.deadline = now + int64(r.CurrentRTO())
-	return tcp.TxDirective{Retransmit: true, RetransmitFrom: intent.UNA}
+	return una, true, false
 }
 
 // PostTx records an emitted segment: it advances the shadow send sequence,
@@ -192,7 +210,11 @@ func (r *Timer) PreTx(intent tcp.TxIntent) tcp.TxDirective {
 // Segments that do not extend the send sequence are retransmissions and are
 // never RTT-sampled (Karn's algorithm). Control-only segments (no data) are
 // ignored. It implements [tcp.Policy].
-func (r *Timer) PostTx(outgoing tcp.Segment, now int64) {
+func (r *Timer) PostTx(h *tcp.Handler, outgoing tcp.Frame) {
+	r.postTx(outgoing.Segment(len(outgoing.Payload())), r.nanotime())
+}
+
+func (r *Timer) postTx(outgoing tcp.Segment, now int64) {
 	if outgoing.DATALEN == 0 {
 		return // only data segments are timed / arm the RTO.
 	}
@@ -204,7 +226,7 @@ func (r *Timer) PostTx(outgoing tcp.Segment, now int64) {
 		r.sndNXT = segStart
 	}
 	if !r.sndNXT.LessThan(segEnd) {
-		// tcp.Segment does not extend the send sequence: it is a retransmission.
+		// Segment does not extend the send sequence: it is a retransmission.
 		// Discard any outstanding RTT sample per Karn's algorithm. The timer was
 		// already (re)armed by PreTx on the timeout that triggered this resend.
 		r.timing = false

tcp/rto/timer_test.go

@@ -21,25 +21,42 @@ func ackSeg(ack uint32) tcp.Segment {
 
 func newRTO() *Timer {
 	var r Timer
-	r.Reset()
+	if err := r.Configure(func() int64 { return 0 }); err != nil {
+		panic(err)
+	}
 	return &r
 }
 
-// rxAt builds the minimal tcp.RxMeta for driving PreRx directly.
-func rxAt(seg tcp.Segment, now int64) tcp.RxMeta { return tcp.RxMeta{Segment: seg, Now: now} }
-
-// acceptedAt builds the event for a segment the connection accepted, which is what
-// drives the estimator. Timing state is only allowed to move for those.
-func acceptedAt(seg tcp.Segment, now int64) tcp.RxEvent {
-	return tcp.RxEvent{Segment: seg, Now: now, Accepted: true}
+// frameOf renders a segment as the wire frame the [tcp.Policy] hooks receive.
+func frameOf(t *testing.T, s tcp.Segment) tcp.Frame {
+	t.Helper()
+	frm, err := tcp.NewFrame(make([]byte, 20+int(s.DATALEN)))
+	if err != nil {
+		t.Fatal(err)
+	}
+	frm.SetSegment(s, 5)
+	return frm
 }
 
-// txAt builds the minimal tcp.TxIntent for driving Timer.PreTx directly: the timer
-// tracks the send sequence itself via PostTx and only reads the clock.
-func txAt(now int64) tcp.TxIntent { return tcp.TxIntent{Now: now} }
+func TestRTO_Configure(t *testing.T) {
+	var r Timer
+	if err := r.Configure(nil); err == nil {
+		t.Error("Configure must reject a nil clock")
+	}
+	if err := r.Configure(func() int64 { return 0 }); err != nil {
+		t.Fatal(err)
+	}
+	if r.nanotime == nil {
+		t.Fatal("clock not stored")
+	}
+	r.Reset()
+	if r.nanotime == nil {
+		t.Error("Reset must preserve the configured clock")
+	}
+}
 
 func TestRTO_Reset(t *testing.T) {
-	var r Timer
+	r := newRTO()
 	r.Reset()
 	if r.rto != rtoInitial {
 		t.Errorf("initial rto=%v, want %v", r.rto, rtoInitial)
@@ -61,7 +78,7 @@ func TestRTO_ArmOnSendSampleOnAck(t *testing.T) {
 	r := newRTO()
 	const iss = uint32(1000)
 
-	r.PostTx(dataSeg(iss, 100), 0)
+	r.postTx(dataSeg(iss, 100), 0)
 	if !r.Running() {
 		t.Fatal("timer must arm after sending data")
 	}
@@ -70,10 +87,10 @@ func TestRTO_ArmOnSendSampleOnAck(t *testing.T) {
 	}
 
 	// ACK arrives one RTT (40ms) later covering all sent data.
-	if !r.PreRx(rxAt(ackSeg(iss+100), 40*rtoMs)).Keep {
+	if !r.PreRx(nil, frameOf(t, ackSeg(iss+100))) {
 		t.Error("PreRx must keep the segment")
 	}
-	r.PostRx(acceptedAt(ackSeg(iss+100), 40*rtoMs))
+	r.postRx(ackSeg(iss+100), 40*rtoMs)
 	if r.Running() {
 		t.Error("timer must stop once all data is acknowledged")
 	}
@@ -87,23 +104,26 @@ func TestRTO_ArmOnSendSampleOnAck(t *testing.T) {
 func TestRTO_RetransmitOnTimeout(t *testing.T) {
 	r := newRTO()
 	const iss = uint32(1000)
-	r.PostTx(dataSeg(iss, 100), 0)
+	r.postTx(dataSeg(iss, 100), 0)
 
-	if r.PreTx(txAt(int64(rtoInitial) - 1)).Retransmit {
+	if _, rtx, _ := r.preTx(int64(rtoInitial)-1, tcp.Value(iss)); rtx {
 		t.Fatal("must not retransmit before the deadline")
 	}
-	dir := r.PreTx(tcp.TxIntent{Now: int64(rtoInitial), UNA: tcp.Value(iss), NXT: tcp.Value(iss + 100)})
-	if !dir.Retransmit {
+	from, rtx, hold := r.preTx(int64(rtoInitial), tcp.Value(iss))
+	if !rtx {
 		t.Fatal("RTO must fire at the deadline with data outstanding")
 	}
-	if dir.RetransmitFrom != tcp.Value(iss) {
-		t.Errorf("retransmit from %d, want snd.UNA=%d", dir.RetransmitFrom, iss)
+	if hold {
+		t.Error("the estimator never holds new data back")
+	}
+	if from != tcp.Value(iss) {
+		t.Errorf("retransmit from %d, want snd.UNA=%d", from, iss)
 	}
 	if r.CurrentRTO() != 2*rtoInitial {
 		t.Errorf("rto=%v after one backoff, want %v", r.CurrentRTO(), 2*rtoInitial)
 	}
-	// The connection resends from snd.UNA; PostTx sees a retransmission.
-	r.PostTx(dataSeg(iss, 100), int64(rtoInitial))
+	// The connection resends from snd.UNA; postTx sees a retransmission.
+	r.postTx(dataSeg(iss, 100), int64(rtoInitial))
 	if r.timing {
 		t.Error("retransmitted segment must not be RTT-sampled (Karn)")
 	}
@@ -114,12 +134,12 @@ func TestRTO_RetransmitOnTimeout(t *testing.T) {
 func TestRTO_KarnNoSampleOnRetransmittedAck(t *testing.T) {
 	r := newRTO()
 	const iss = uint32(1000)
-	r.PostTx(dataSeg(iss, 100), 0)
+	r.postTx(dataSeg(iss, 100), 0)
 	// Timeout and retransmit.
-	r.PreTx(txAt(int64(rtoInitial)))
-	r.PostTx(dataSeg(iss, 100), int64(rtoInitial))
+	r.preTx(int64(rtoInitial), tcp.Value(iss))
+	r.postTx(dataSeg(iss, 100), int64(rtoInitial))
 	// ACK now arrives; no sample should be taken since timing was discarded.
-	r.PostRx(acceptedAt(ackSeg(iss+100), int64(rtoInitial)+10*rtoMs))
+	r.postRx(ackSeg(iss+100), int64(rtoInitial)+10*rtoMs)
 	if r.haveRTT {
 		t.Error("no RTT sample should exist after a retransmission (Karn)")
 	}
@@ -130,10 +150,10 @@ func TestRTO_KarnNoSampleOnRetransmittedAck(t *testing.T) {
 func TestRTO_TimerRestartsWhilePartiallyAcked(t *testing.T) {
 	r := newRTO()
 	const iss = uint32(1000)
-	r.PostTx(dataSeg(iss, 100), 0)
-	r.PostTx(dataSeg(iss+100, 100), 0) // 200 octets outstanding, iss..iss+200.
+	r.postTx(dataSeg(iss, 100), 0)
+	r.postTx(dataSeg(iss+100, 100), 0) // 200 octets outstanding, iss..iss+200.
 
-	r.PostRx(acceptedAt(ackSeg(iss+100), 40*rtoMs)) // acks first 100 only.
+	r.postRx(ackSeg(iss+100), 40*rtoMs) // acks first 100 only.
 	if !r.Running() {
 		t.Fatal("timer must remain armed while data is still in flight")
 	}
@@ -146,7 +166,7 @@ func TestRTO_TimerRestartsWhilePartiallyAcked(t *testing.T) {
 // nor start an RTT sample.
 func TestRTO_NoArmWithoutData(t *testing.T) {
 	r := newRTO()
-	r.PostTx(tcp.Segment{SEQ: 1000, Flags: tcp.FlagACK}, 0) // pure ACK, DATALEN==0.
+	r.postTx(tcp.Segment{SEQ: 1000, Flags: tcp.FlagACK}, 0) // pure ACK, DATALEN==0.
 	if r.Running() || r.timing {
 		t.Error("pure control segment must not arm the timer or start a sample")
 	}
@@ -157,15 +177,15 @@ func TestRTO_NoArmWithoutData(t *testing.T) {
 func TestRTO_BackoffCollapsesOnValidSample(t *testing.T) {
 	r := newRTO()
 	const iss = uint32(1000)
-	r.PostTx(dataSeg(iss, 100), 0)
-	r.PreTx(txAt(int64(rtoInitial)))               // one timeout: backoff=1.
-	r.PostTx(dataSeg(iss, 100), int64(rtoInitial)) // retransmit (no sample).
+	r.postTx(dataSeg(iss, 100), 0)
+	r.preTx(int64(rtoInitial), tcp.Value(iss))     // one timeout: backoff=1.
+	r.postTx(dataSeg(iss, 100), int64(rtoInitial)) // retransmit (no sample).
 	if r.backoff != 1 {
 		t.Fatalf("backoff=%d, want 1 after a timeout", r.backoff)
 	}
 	// New data sent and freshly sampled, then acked.
-	r.PostTx(dataSeg(iss+100, 100), int64(rtoInitial)+rtoMs)
-	r.PostRx(acceptedAt(ackSeg(iss+200), int64(rtoInitial)+30*rtoMs))
+	r.postTx(dataSeg(iss+100, 100), int64(rtoInitial)+rtoMs)
+	r.postRx(ackSeg(iss+200), int64(rtoInitial)+30*rtoMs)
 	if r.backoff != 0 {
 		t.Errorf("backoff=%d, want 0 after a valid RTT sample", r.backoff)
 	}
@@ -173,8 +193,7 @@ func TestRTO_BackoffCollapsesOnValidSample(t *testing.T) {
 
 // TestRTO_Clamped verifies CurrentRTO is clamped to [rtoMin, rtoMax].
 func TestRTO_Clamped(t *testing.T) {
-	var r Timer
-	r.Reset()
+	r := newRTO()
 	r.rto = time.Nanosecond
 	if got := r.CurrentRTO(); got != rtoMin {
 		t.Errorf("CurrentRTO=%v, want floor %v", got, rtoMin)
@@ -188,8 +207,7 @@ func TestRTO_Clamped(t *testing.T) {
 // TestRTO_UpdateRTTFirstSample verifies the first-measurement initialization of
 // SRTT/RTTVAR (RFC 6298 §2.2).
 func TestRTO_UpdateRTTFirstSample(t *testing.T) {
-	var r Timer
-	r.Reset()
+	r := newRTO()
 	r.updateRTT(100 * time.Millisecond)
 	if r.srtt != 100*time.Millisecond {
 		t.Errorf("srtt=%v, want 100ms", r.srtt)
@@ -203,53 +221,61 @@ func TestRTO_UpdateRTTFirstSample(t *testing.T) {
 	}
 }
 
-// TestRTO_ImplementsPolicy exercises Timer through the [tcp.Policy]
-// interface: sending data arms a deadline and a full ACK disarms it.
-func TestRTO_ImplementsPolicy(t *testing.T) {
-	var lr tcp.Policy = newRTO()
-	lr.Reset()
-	lr.PostTx(dataSeg(1000, 100), 0)
-	if lr.NextDeadline() == 0 {
-		t.Error("expected an armed deadline after sending data")
+// TestRTO_PolicyHooksDeriveFromFrame exercises Timer through the [tcp.Policy]
+// hooks, verifying it reads the segment out of the frame it is handed: sending
+// data arms a deadline and a full ACK disarms it and yields the RTT sample.
+func TestRTO_PolicyHooksDeriveFromFrame(t *testing.T) {
+	var clock int64
+	var r Timer
+	if err := r.Configure(func() int64 { return clock }); err != nil {
+		t.Fatal(err)
 	}
-	if !lr.PreRx(rxAt(ackSeg(1100), 10*rtoMs)).Keep {
+	var pol tcp.Policy = &r
+	pol.Reset()
+
+	pol.PostTx(nil, frameOf(t, dataSeg(1000, 100)))
+	if r.NextDeadline() == 0 {
+		t.Fatal("expected an armed deadline after sending data")
+	}
+	clock = 10 * rtoMs
+	if !pol.PreRx(nil, frameOf(t, ackSeg(1100))) {
 		t.Error("PreRx must keep")
 	}
-	lr.PostRx(acceptedAt(ackSeg(1100), 10*rtoMs))
-	if lr.NextDeadline() != 0 {
+	pol.PostRx(nil, tcp.StateEstablished, frameOf(t, ackSeg(1100)))
+	if r.NextDeadline() != 0 {
 		t.Error("expected disarmed timer after full ack")
 	}
+	if r.SmoothedRTT() != 10*time.Millisecond {
+		t.Errorf("srtt=%v, want 10ms sampled through the hooks", r.SmoothedRTT())
+	}
 }
 
-// TestRTO_IgnoresRejectedSegment verifies the estimator does not act on a segment
-// the connection refused. PreRx runs before the state machine has judged the
-// segment, so an acknowledgement for data never sent would otherwise collapse the
-// backoff and take a bogus round-trip sample.
-func TestRTO_IgnoresRejectedSegment(t *testing.T) {
+// TestRTO_PreRxNeverDrops verifies the estimator keeps every segment and records
+// nothing at PreRx time. Dropping is not its business, and the connection has not
+// yet judged the segment: an acknowledgement for data never sent would otherwise
+// collapse the backoff and take a bogus round-trip sample. Only accepted segments
+// reach PostRx, which the Handler guarantees.
+func TestRTO_PreRxNeverDrops(t *testing.T) {
 	r := newRTO()
 	const iss = uint32(1000)
-	r.PostTx(dataSeg(iss, 100), 0)
+	r.postTx(dataSeg(iss, 100), 0)
 	armed := r.NextDeadline()
 	if armed == 0 {
 		t.Fatal("timer must be armed after sending data")
 	}
 
-	// An acknowledgement far beyond anything sent, refused by the connection.
-	bogus := ackSeg(iss + 100000)
-	if !r.PreRx(rxAt(bogus, 40*rtoMs)).Keep {
+	// An acknowledgement far beyond anything sent, which the connection refuses.
+	if !r.PreRx(nil, frameOf(t, ackSeg(iss+100000))) {
 		t.Error("PreRx must keep: dropping is not the estimator's business")
 	}
-	r.PostRx(tcp.RxEvent{Segment: bogus, Now: 40 * rtoMs, Accepted: false})
-
 	if r.NextDeadline() != armed {
-		t.Errorf("deadline moved to %d on a refused segment, want it left at %d",
-			r.NextDeadline(), armed)
+		t.Errorf("deadline moved to %d at PreRx, want it left at %d", r.NextDeadline(), armed)
 	}
 	if r.SmoothedRTT() != 0 {
-		t.Errorf("took an RTT sample of %v from a refused segment", r.SmoothedRTT())
+		t.Errorf("took an RTT sample of %v at PreRx", r.SmoothedRTT())
 	}
 	if !r.Running() {
-		t.Error("timer disarmed by a refused acknowledgement")
+		t.Error("timer disarmed at PreRx")
 	}
 }
 
@@ -262,25 +288,24 @@ func TestRTO_RetransmitsZeroWindowProbe(t *testing.T) {
 	r := newRTO()
 	const iss = uint32(5000)
 	probe := dataSeg(iss, 1) // The one-octet probe.
-	r.PostTx(probe, 0)
+	r.postTx(probe, 0)
 
 	now := int64(rtoInitial)
 	prevRTO := r.CurrentRTO()
 	for attempt := 1; attempt <= 4; attempt++ {
-		dir := r.PreTx(tcp.TxIntent{Now: now, UNA: tcp.Value(iss), NXT: tcp.Value(iss + 1)})
-		if !dir.Retransmit {
+		from, rtx, _ := r.preTx(now, tcp.Value(iss))
+		if !rtx {
 			t.Fatalf("attempt %d: timer did not fire; the probe would never be resent", attempt)
 		}
-		if dir.RetransmitFrom != tcp.Value(iss) {
-			t.Errorf("attempt %d: retransmit from %d, want the probe octet at %d",
-				attempt, dir.RetransmitFrom, iss)
+		if from != tcp.Value(iss) {
+			t.Errorf("attempt %d: retransmit from %d, want the probe octet at %d", attempt, from, iss)
 		}
 		if got := r.CurrentRTO(); got <= prevRTO {
 			t.Errorf("attempt %d: rto %v did not back off past %v", attempt, got, prevRTO)
 		}
 		prevRTO = r.CurrentRTO()
 		// The peer still cannot accept the octet, so it stays unacknowledged.
-		r.PostTx(probe, now)
+		r.postTx(probe, now)
 		now += int64(prevRTO)
 	}
 }

tcp/txqueue.go

@@ -227,18 +227,39 @@ func (rtx *ringTx) RetransmitFromUNA() {
 	if oldest == nil {
 		return // Nothing in the retransmission queue.
 	}
-	unaSeq := oldest.seq
-	if rtx.sentend != 0 {
-		// Merge sent region [sentoff, sentend) back into unsent.
-		rtx.unsentoff = rtx.sentoff
-		if rtx.unsentend == 0 {
-			rtx.unsentend = rtx.sentend
-		}
-		rtx.sentoff = 0
-		rtx.sentend = 0
+	rtx.RetransmitFrom(oldest.seq)
+}
+
+// RetransmitFrom rewinds the transmit queue so sent-but-unacked data at and
+// after seq becomes unsent again; the next MakePacket calls re-send it. seq is
+// snapped down to the start of the packet containing it — the retransmission
+// queue tracks whole packets, so sub-packet rewind is not representable. It is
+// a no-op when seq is not covered by any queued packet (nothing to resend).
+//
+// Callers must pair this with [ControlBlock.RetransmitFrom] using the same seq
+// so the send sequence space and the transmit buffer rewind together.
+func (rtx *ringTx) RetransmitFrom(seq Value) {
+	pkt := rtx.slist.packetContaining(seq)
+	if pkt == nil {
+		return // seq not in the retransmission queue.
 	}
-	// Clear packet metadata; sequence tracking restarts from UNA.
-	rtx.slist.Reset(cap(rtx.slist.pkts), unaSeq)
+	rewindOff, rewindSeq := pkt.off, pkt.seq
+	// The write position is unsentend, except when the unsent region is empty
+	// (unsentend==0) in which case data ends where the sent region ends. Capture
+	// it before reopening the unsent region over the rewound packets.
+	writeEnd := rtx.unsentend
+	if writeEnd == 0 {
+		writeEnd = rtx.sentend
+	}
+	if rewindOff == rtx.sentoff {
+		rtx.sentoff = 0 // Whole queue rewound: sent region becomes empty.
+		rtx.sentend = 0
+	} else {
+		rtx.sentend = rewindOff
+	}
+	rtx.unsentoff = rewindOff
+	rtx.unsentend = writeEnd
+	rtx.slist.truncateFrom(rewindSeq)
 }
 
 func (rtx *ringTx) consolidateBufs() {
@@ -331,6 +352,36 @@ func (sl *sentlist) Free() int {
 	return cap(sl.pkts) - len(sl.pkts)
 }
 
+// packetContaining returns the queued packet whose sequence range covers seq, or
+// nil when no packet does. It is the floor lookup a retransmission rewind needs:
+// seq lands inside a packet and the whole packet is resent.
+func (sl *sentlist) packetContaining(seq Value) *ringidx {
+	for i := range sl.pkts {
+		pkt := &sl.pkts[i]
+		if pkt.seq.LessThanEq(seq) && seq.LessThan(pkt.endSeq()) {
+			return pkt
+		}
+	}
+	return nil
+}
+
+// truncateFrom drops the packet starting at seq and every packet sent after it,
+// so their data can be re-queued as unsent. seq must be a packet start sequence
+// (see [sentlist.packetContaining]). When no packet survives, the auxiliary
+// sequence counter is rewound to seq so [sentlist.EndSeq] keeps reporting where
+// the next packet begins.
+func (sl *sentlist) truncateFrom(seq Value) {
+	for i := range sl.pkts {
+		if sl.pkts[i].seq == seq {
+			sl.pkts = sl.pkts[:i]
+			if i == 0 {
+				sl.ssn = seq
+			}
+			return
+		}
+	}
+}
+
 func (sl *sentlist) AddPacket(datalen, off, bufsize int, seq Value) *ringidx {
 	free := sl.Free()
 	if free == 0 {

tcp/txqueue_retransmit_test.go

@@ -0,0 +1,215 @@
+package tcp
+
+import (
+	"bytes"
+	"testing"
+)
+
+// newRetransmitQueue builds a queue holding npkt sent packets of pktlen octets
+// each, starting at iss, plus any leftover unsent data. It returns the queue and
+// the full byte stream that was written.
+func newRetransmitQueue(t *testing.T, bufsize, maxPkts, npkt, pktlen, unsent int, iss Value) (*ringTx, []byte) {
+	t.Helper()
+	var rtx ringTx
+	if err := rtx.Reset(make([]byte, bufsize), maxPkts, iss); err != nil {
+		t.Fatal(err)
+	}
+	stream := make([]byte, npkt*pktlen+unsent)
+	for i := range stream {
+		stream[i] = byte(i + 1) // Non-zero so a stale ring shows up as a mismatch.
+	}
+	if n, err := rtx.Write(stream); err != nil || n != len(stream) {
+		t.Fatalf("write n=%d err=%v", n, err)
+	}
+	seq := iss
+	scratch := make([]byte, pktlen)
+	for i := 0; i < npkt; i++ {
+		n, err := rtx.MakePacket(scratch, seq)
+		if err != nil {
+			t.Fatalf("packet %d: %v", i, err)
+		}
+		if n != pktlen {
+			t.Fatalf("packet %d: n=%d, want %d", i, n, pktlen)
+		}
+		seq += Value(n)
+	}
+	testQueueSanity(t, &rtx)
+	return &rtx, stream
+}
+
+// mustRemake asserts the queue re-emits datalen octets at seq matching want.
+func mustRemake(t *testing.T, rtx *ringTx, seq Value, want []byte) {
+	t.Helper()
+	got := make([]byte, len(want))
+	n, err := rtx.MakePacket(got, seq)
+	if err != nil {
+		t.Fatalf("MakePacket at seq %d: %v", seq, err)
+	}
+	if n != len(want) {
+		t.Fatalf("MakePacket at seq %d: n=%d, want %d", seq, n, len(want))
+	}
+	if !bytes.Equal(got, want) {
+		t.Fatalf("MakePacket at seq %d: got %v, want %v", seq, got, want)
+	}
+}
+
+// TestRingTx_RetransmitFromBoundary rewinds to the start of the second of three
+// sent packets: the first stays sent, the rest become unsent and re-emit their
+// original bytes.
+func TestRingTx_RetransmitFromBoundary(t *testing.T) {
+	const iss, pktlen = Value(100), 4
+	rtx, stream := newRetransmitQueue(t, 64, 4, 3, pktlen, 0, iss)
+
+	sentBefore := rtx.BufferedSent()
+	rtx.RetransmitFrom(iss + pktlen) // Start of packet 2.
+	testQueueSanity(t, rtx)
+
+	if got := rtx.BufferedSent(); got != pktlen {
+		t.Fatalf("sent=%d, want %d (only packet 1 remains sent)", got, pktlen)
+	}
+	if got := rtx.BufferedUnsent(); got != sentBefore-pktlen {
+		t.Fatalf("unsent=%d, want %d", got, sentBefore-pktlen)
+	}
+	mustRemake(t, rtx, iss+pktlen, stream[pktlen:2*pktlen])
+	testQueueSanity(t, rtx)
+	mustRemake(t, rtx, iss+2*pktlen, stream[2*pktlen:3*pktlen])
+	testQueueSanity(t, rtx)
+}
+
+// TestRingTx_RetransmitFromMidPacket verifies a sequence inside a packet is
+// snapped down to that packet's start: the queue tracks whole packets.
+func TestRingTx_RetransmitFromMidPacket(t *testing.T) {
+	const iss, pktlen = Value(100), 4
+	rtx, stream := newRetransmitQueue(t, 64, 4, 3, pktlen, 0, iss)
+
+	rtx.RetransmitFrom(iss + pktlen + 2) // Two octets into packet 2.
+	testQueueSanity(t, rtx)
+
+	if got := rtx.BufferedSent(); got != pktlen {
+		t.Fatalf("sent=%d, want %d: rewind must floor to the packet start", got, pktlen)
+	}
+	mustRemake(t, rtx, iss+pktlen, stream[pktlen:2*pktlen])
+}
+
+// TestRingTx_RetransmitFromOldest rewinds the whole queue, which must match
+// RetransmitFromUNA.
+func TestRingTx_RetransmitFromOldest(t *testing.T) {
+	const iss, pktlen, npkt = Value(100), 4, 3
+	rtx, stream := newRetransmitQueue(t, 64, 4, npkt, pktlen, 0, iss)
+	rtx.RetransmitFrom(iss)
+	testQueueSanity(t, rtx)
+
+	viaUNA, _ := newRetransmitQueue(t, 64, 4, npkt, pktlen, 0, iss)
+	viaUNA.RetransmitFromUNA()
+	testQueueSanity(t, viaUNA)
+
+	if rtx.BufferedSent() != 0 {
+		t.Fatalf("sent=%d, want 0 after a full rewind", rtx.BufferedSent())
+	}
+	if rtx.BufferedUnsent() != npkt*pktlen {
+		t.Fatalf("unsent=%d, want %d", rtx.BufferedUnsent(), npkt*pktlen)
+	}
+	if rtx.BufferedSent() != viaUNA.BufferedSent() || rtx.BufferedUnsent() != viaUNA.BufferedUnsent() {
+		t.Fatal("RetransmitFrom(oldest) must match RetransmitFromUNA")
+	}
+	mustRemake(t, rtx, iss, stream[:pktlen])
+}
+
+// TestRingTx_RetransmitFromUnknownSeq verifies a sequence covered by no queued
+// packet leaves the queue untouched.
+func TestRingTx_RetransmitFromUnknownSeq(t *testing.T) {
+	const iss, pktlen, npkt = Value(100), 4, 3
+	rtx, _ := newRetransmitQueue(t, 64, 4, npkt, pktlen, 0, iss)
+	sent, unsent := rtx.BufferedSent(), rtx.BufferedUnsent()
+
+	rtx.RetransmitFrom(iss - 1)           // Before the queue.
+	rtx.RetransmitFrom(iss + npkt*pktlen) // One past the last octet sent.
+	rtx.RetransmitFrom(iss + 1000)        // Far beyond.
+	testQueueSanity(t, rtx)
+
+	if rtx.BufferedSent() != sent || rtx.BufferedUnsent() != unsent {
+		t.Fatalf("queue moved: sent %d→%d, unsent %d→%d", sent, rtx.BufferedSent(), unsent, rtx.BufferedUnsent())
+	}
+}
+
+// TestRingTx_RetransmitWithUnsentTail verifies a rewind reopens the unsent region
+// over the rewound packets without losing the unsent tail behind them.
+func TestRingTx_RetransmitWithUnsentTail(t *testing.T) {
+	const iss, pktlen, npkt, tail = Value(100), 4, 2, 5
+	rtx, stream := newRetransmitQueue(t, 64, 4, npkt, pktlen, tail, iss)
+	if got := rtx.BufferedUnsent(); got != tail {
+		t.Fatalf("unsent tail=%d, want %d", got, tail)
+	}
+
+	rtx.RetransmitFrom(iss + pktlen) // Rewind the second packet only.
+	testQueueSanity(t, rtx)
+
+	if got := rtx.BufferedUnsent(); got != pktlen+tail {
+		t.Fatalf("unsent=%d, want %d (rewound packet plus the tail)", got, pktlen+tail)
+	}
+	// The rewound packet re-emits first, then the tail follows in order.
+	mustRemake(t, rtx, iss+pktlen, stream[pktlen:2*pktlen])
+	testQueueSanity(t, rtx)
+	mustRemake(t, rtx, iss+2*pktlen, stream[2*pktlen:])
+}
+
+// TestRingTx_RetransmitAfterDrainedUnsent pins the write-position recovery: when
+// every octet written has been packetized the unsent region is empty, so the
+// rewind must reconstruct where data ends from the sent region.
+func TestRingTx_RetransmitAfterDrainedUnsent(t *testing.T) {
+	const iss, pktlen, npkt = Value(100), 4, 3
+	rtx, stream := newRetransmitQueue(t, 64, 4, npkt, pktlen, 0, iss)
+	if got := rtx.BufferedUnsent(); got != 0 {
+		t.Fatalf("unsent=%d, want 0: all written data was packetized", got)
+	}
+
+	rtx.RetransmitFrom(iss + pktlen)
+	testQueueSanity(t, rtx)
+
+	if got := rtx.BufferedUnsent(); got != 2*pktlen {
+		t.Fatalf("unsent=%d, want %d: rewind lost the end of the data", got, 2*pktlen)
+	}
+	mustRemake(t, rtx, iss+pktlen, stream[pktlen:2*pktlen])
+	testQueueSanity(t, rtx)
+	mustRemake(t, rtx, iss+2*pktlen, stream[2*pktlen:3*pktlen])
+}
+
+// TestRingTx_RetransmitWrapped exercises a rewind on a queue whose regions wrap
+// the end of the ring buffer.
+func TestRingTx_RetransmitWrapped(t *testing.T) {
+	const bufsize, pktlen = 16, 4
+	const iss = Value(100)
+	var rtx ringTx
+	if err := rtx.Reset(make([]byte, bufsize), 4, iss); err != nil {
+		t.Fatal(err)
+	}
+	// Push the queue most of the way around the ring, acking as we go.
+	seq := iss
+	scratch := make([]byte, pktlen)
+	for round := 0; round < 3; round++ {
+		chunk := make([]byte, pktlen)
+		for i := range chunk {
+			chunk[i] = byte(round*pktlen + i + 1)
+		}
+		if _, err := rtx.Write(chunk); err != nil {
+			t.Fatal(err)
+		}
+		if _, err := rtx.MakePacket(scratch, seq); err != nil {
+			t.Fatal(err)
+		}
+		seq += Value(pktlen)
+		if round < 2 {
+			if err := rtx.RecvACK(seq); err != nil {
+				t.Fatal(err)
+			}
+		}
+		testQueueSanity(t, &rtx)
+	}
+	// Two packets outstanding, straddling the wrap. Rewind the newest.
+	rewindSeq := seq - Value(pktlen)
+	want := append([]byte(nil), scratch...)
+	rtx.RetransmitFrom(rewindSeq)
+	testQueueSanity(t, &rtx)
+	mustRemake(t, &rtx, rewindSeq, want)
+	testQueueSanity(t, &rtx)
+}