begin prepping policy refactor manually

This commit is contained in:
Patricio Whittingslow
2026-08-24 15:59:17 -03:00
parent 263b1ecf11
commit 936790a5d0
8 changed files with 415 additions and 375 deletions
+259
View File
@@ -0,0 +1,259 @@
package rto
import (
"time"
"github.com/soypat/lneto/tcp"
)
// RFC 6298 retransmission-timeout (RTO) parameters. The algorithm keeps a
// single retransmission timer per connection (RFC 6298 §5): the timer is
// (re)started whenever new data is acknowledged while data remains in flight,
// stopped when all data is acknowledged, and on expiry the oldest unacknowledged
// segment is retransmitted and the RTO is doubled (exponential backoff, §5.5).
const (
// rtoInitial is the RTO used before the first RTT measurement (RFC 6298 §2.1).
rtoInitial = time.Second
// rtoMin clamps the lower bound of the RTO. RFC 6298 §2.4 recommends a
// minimum of 1s, but that is punishing on the low-latency links lneto
// targets; like Linux we use a smaller floor so recovery on LAN/embedded
// links is timely.
rtoMin = 200 * time.Millisecond
// rtoMax clamps the upper bound across exponential backoff (RFC 6298 §5.5
// permits a maximum of at least 60s).
rtoMax = 60 * time.Second
// rttGainShift (alpha = 1/8) and rttvarGainShift (beta = 1/4) are the
// smoothing gains of RFC 6298 §2.3, applied as integer shifts.
rttGainShift = 3 // alpha = 1/8
rttvarGainShift = 2 // beta = 1/4
// rttvarK is the RTTVAR multiplier in RTO = SRTT + K*RTTVAR (RFC 6298 §2.3).
rttvarK = 4
// backoffMax caps the exponential-backoff doublings so RTO arithmetic cannot
// overflow and a wedged connection keeps probing at rtoMax.
backoffMax = 12
)
// 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.
//
// 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).
//
// 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]
// advances the highest sequence acknowledged. This is what lets it manage the
// timer (RFC 6298 §5.2/§5.3) without reaching into the tcp state machine, and it
// is also how retransmissions are distinguished for Karn's algorithm — a segment
// whose sequence space is not beyond the shadow snd.NXT is a retransmission and
// is never RTT-sampled.
type Timer struct {
srtt time.Duration // smoothed round-trip time (SRTT).
rttvar time.Duration // round-trip-time variation (RTTVAR).
rto time.Duration // current retransmission timeout.
haveRTT bool // false until the first RTT sample is taken.
// Shadow of the send sequence space, derived from observed segments.
haveSeq bool // false until the first data segment is observed.
sndUNA tcp.Value // highest acknowledged sequence number seen on the wire.
sndNXT tcp.Value // one past the highest sequence number sent.
// RTT sampling state (Karn's algorithm, RFC 6298 §3): at most one segment is
// timed at a time and retransmitted segments are never sampled.
timing bool
timedSeq tcp.Value // ACK at or beyond this value completes the sample.
timedAt int64 // send time (monotonic ns) of the timed segment.
// Retransmission timer state.
running bool
deadline int64 // time (monotonic ns) at which the timer expires.
backoff uint8 // consecutive timeouts, for exponential backoff.
// 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.
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} }
// SmoothedRTT returns the current smoothed round-trip time (SRTT), or zero
// before the first RTT measurement. It is concrete-type introspection and is
// intentionally not part of [tcp.Policy].
func (r *Timer) SmoothedRTT() time.Duration { return r.srtt }
// CurrentRTO returns the timeout currently in effect, clamped to [rtoMin, rtoMax].
func (r *Timer) CurrentRTO() time.Duration {
rto := r.rto
if rto < rtoMin {
rto = rtoMin
} else if rto > rtoMax {
rto = rtoMax
}
return rto
}
// Running reports whether the retransmission timer is currently armed.
func (r *Timer) Running() bool { return r.running }
// Expirations returns how many times the retransmission timer has expired since
// [Timer.Reset]. A policy that shares this timer rather than driving it watches
// this for a change to learn that a timeout happened, since it never sees the
// timer's own directive. It is concrete-type introspection and is intentionally
// not part of [tcp.Policy].
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].
func (r *Timer) NextDeadline() int64 {
if !r.running {
return 0
}
return r.deadline
}
// 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}
}
// 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) {
return
}
ack := incoming.ACK
if r.timing && !ack.LessThan(r.timedSeq) {
// ACK covers the timed segment: take the RTT sample (§4). A valid
// measurement collapses the backoff (§5.7).
r.updateRTT(time.Duration(now - r.timedAt))
r.timing = false
r.backoff = 0
}
if r.sndUNA.LessThan(ack) && !r.sndNXT.LessThan(ack) {
// ACK advances snd.UNA and does not exceed what we have sent.
r.sndUNA = ack
}
if r.sndUNA == r.sndNXT {
r.running = false // §5.3: all outstanding data acknowledged.
} else {
// §5.3: new (but not all) data acknowledged — restart the timer.
r.running = true
r.deadline = now + int64(r.CurrentRTO())
}
}
// 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
if !r.running || now < r.deadline || r.sndUNA == r.sndNXT {
return tcp.TxDirective{}
}
r.expirations++
r.timing = false // §5.4: do not sample a retransmitted segment.
if r.backoff < backoffMax {
r.backoff++
r.rto = min(r.CurrentRTO()*2, rtoMax) // §5.5: RTO = RTO * 2.
}
r.running = true
r.deadline = now + int64(r.CurrentRTO())
return tcp.TxDirective{Retransmit: true, RetransmitFrom: intent.UNA}
}
// PostTx records an emitted segment: it advances the shadow send sequence,
// begins timing newly transmitted data (RFC 6298 §3) and arms the timer (§5.1).
// 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) {
if outgoing.DATALEN == 0 {
return // only data segments are timed / arm the RTO.
}
segStart := outgoing.SEQ
segEnd := segStart + tcp.Value(outgoing.LEN())
if !r.haveSeq {
r.haveSeq = true
r.sndUNA = segStart
r.sndNXT = segStart
}
if !r.sndNXT.LessThan(segEnd) {
// tcp.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
return
}
r.sndNXT = segEnd
if !r.timing {
r.timing = true
r.timedSeq = segEnd
r.timedAt = now
}
if !r.running {
r.running = true
r.deadline = now + int64(r.CurrentRTO())
}
}
// ObserveRTT folds a round-trip measurement taken by other means into the
// estimator, for a policy that composes this timer and can measure the round trip
// more accurately than acknowledgement timing allows. The RFC 7323 timestamp echo
// is the case this exists for.
//
// Unlike the timer's own sampling this does not apply Karn's algorithm, because a
// sample derived from an echoed timestamp is unambiguous even when the segment
// carrying it was a retransmission (RFC 7323 §4.1). Non-positive samples are
// ignored.
func (r *Timer) ObserveRTT(rtt time.Duration) { r.updateRTT(rtt) }
// updateRTT folds a round-trip measurement into SRTT/RTTVAR/RTO using the
// integer-shift form of RFC 6298 §2.2/§2.3.
func (r *Timer) updateRTT(sample time.Duration) {
if sample <= 0 {
return
}
if !r.haveRTT {
// First measurement (RFC 6298 §2.2).
r.srtt = sample
r.rttvar = sample / 2
r.haveRTT = true
} else {
// Subsequent measurements (RFC 6298 §2.3):
// RTTVAR = (1-beta)*RTTVAR + beta*|SRTT-R|
// SRTT = (1-alpha)*SRTT + alpha*R
diff := r.srtt - sample
if diff < 0 {
diff = -diff
}
r.rttvar += (diff - r.rttvar) >> rttvarGainShift
r.srtt += (sample - r.srtt) >> rttGainShift
}
r.rto = r.srtt + rttvarK*r.rttvar
}
+286
View File
@@ -0,0 +1,286 @@
package rto
import (
"testing"
"time"
"github.com/soypat/lneto/tcp"
)
const rtoMs = int64(time.Millisecond)
// dataSeg builds a data segment of datalen octets starting at seq.
func dataSeg(seq uint32, datalen int) tcp.Segment {
return tcp.Segment{SEQ: tcp.Value(seq), DATALEN: tcp.Size(datalen), Flags: tcp.FlagPSH | tcp.FlagACK}
}
// ackSeg builds a bare ACK acknowledging up to ack.
func ackSeg(ack uint32) tcp.Segment {
return tcp.Segment{ACK: tcp.Value(ack), Flags: tcp.FlagACK}
}
func newRTO() *Timer {
var r Timer
r.Reset()
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}
}
// 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_Reset(t *testing.T) {
var r Timer
r.Reset()
if r.rto != rtoInitial {
t.Errorf("initial rto=%v, want %v", r.rto, rtoInitial)
}
if r.CurrentRTO() != rtoInitial {
t.Errorf("CurrentRTO=%v, want %v", r.CurrentRTO(), rtoInitial)
}
if r.haveRTT {
t.Error("haveRTT should be false before first sample")
}
if r.Running() || r.NextDeadline() != 0 {
t.Error("timer must be disarmed after Reset")
}
}
// TestRTO_ArmOnSendSampleOnAck sends data, verifies the timer arms, then acks it
// and verifies an RTT sample is taken and the timer stops once all data is acked.
func TestRTO_ArmOnSendSampleOnAck(t *testing.T) {
r := newRTO()
const iss = uint32(1000)
r.PostTx(dataSeg(iss, 100), 0)
if !r.Running() {
t.Fatal("timer must arm after sending data")
}
if r.NextDeadline() != int64(rtoInitial) {
t.Errorf("deadline=%d, want %d", r.NextDeadline(), int64(rtoInitial))
}
// ACK arrives one RTT (40ms) later covering all sent data.
if !r.PreRx(rxAt(ackSeg(iss+100), 40*rtoMs)).Keep {
t.Error("PreRx must keep the segment")
}
r.PostRx(acceptedAt(ackSeg(iss+100), 40*rtoMs))
if r.Running() {
t.Error("timer must stop once all data is acknowledged")
}
if r.SmoothedRTT() != 40*time.Millisecond {
t.Errorf("srtt=%v, want 40ms", r.SmoothedRTT())
}
}
// TestRTO_RetransmitOnTimeout verifies PreTx directs a go-back-N retransmit once
// the deadline passes with data outstanding, and backs the RTO off.
func TestRTO_RetransmitOnTimeout(t *testing.T) {
r := newRTO()
const iss = uint32(1000)
r.PostTx(dataSeg(iss, 100), 0)
if r.PreTx(txAt(int64(rtoInitial) - 1)).Retransmit {
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 {
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 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))
if r.timing {
t.Error("retransmitted segment must not be RTT-sampled (Karn)")
}
}
// TestRTO_KarnNoSampleOnRetransmittedAck verifies that after a retransmission the
// ACK does not produce an RTT sample (Karn's algorithm).
func TestRTO_KarnNoSampleOnRetransmittedAck(t *testing.T) {
r := newRTO()
const iss = uint32(1000)
r.PostTx(dataSeg(iss, 100), 0)
// Timeout and retransmit.
r.PreTx(txAt(int64(rtoInitial)))
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))
if r.haveRTT {
t.Error("no RTT sample should exist after a retransmission (Karn)")
}
}
// TestRTO_TimerRestartsWhilePartiallyAcked verifies the timer restarts (not
// stops) when an ACK advances UNA but data remains in flight (RFC 6298 §5.3).
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.PostRx(acceptedAt(ackSeg(iss+100), 40*rtoMs)) // acks first 100 only.
if !r.Running() {
t.Fatal("timer must remain armed while data is still in flight")
}
if r.NextDeadline() != 40*rtoMs+int64(r.CurrentRTO()) {
t.Errorf("deadline=%d, want %d", r.NextDeadline(), 40*rtoMs+int64(r.CurrentRTO()))
}
}
// TestRTO_NoArmWithoutData verifies control-only segments neither arm the timer
// 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.
if r.Running() || r.timing {
t.Error("pure control segment must not arm the timer or start a sample")
}
}
// TestRTO_BackoffCollapsesOnValidSample verifies a valid RTT measurement
// collapses the exponential backoff counter (RFC 6298 §5.7).
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).
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))
if r.backoff != 0 {
t.Errorf("backoff=%d, want 0 after a valid RTT sample", r.backoff)
}
}
// TestRTO_Clamped verifies CurrentRTO is clamped to [rtoMin, rtoMax].
func TestRTO_Clamped(t *testing.T) {
var r Timer
r.Reset()
r.rto = time.Nanosecond
if got := r.CurrentRTO(); got != rtoMin {
t.Errorf("CurrentRTO=%v, want floor %v", got, rtoMin)
}
r.rto = time.Hour
if got := r.CurrentRTO(); got != rtoMax {
t.Errorf("CurrentRTO=%v, want ceiling %v", got, rtoMax)
}
}
// 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.updateRTT(100 * time.Millisecond)
if r.srtt != 100*time.Millisecond {
t.Errorf("srtt=%v, want 100ms", r.srtt)
}
if r.rttvar != 50*time.Millisecond {
t.Errorf("rttvar=%v, want 50ms", r.rttvar)
}
// RTO = SRTT + K*RTTVAR = 100 + 4*50 = 300ms.
if r.rto != 300*time.Millisecond {
t.Errorf("rto=%v, want 300ms", r.rto)
}
}
// 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")
}
if !lr.PreRx(rxAt(ackSeg(1100), 10*rtoMs)).Keep {
t.Error("PreRx must keep")
}
lr.PostRx(acceptedAt(ackSeg(1100), 10*rtoMs))
if lr.NextDeadline() != 0 {
t.Error("expected disarmed timer after full ack")
}
}
// 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) {
r := newRTO()
const iss = uint32(1000)
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 {
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)
}
if r.SmoothedRTT() != 0 {
t.Errorf("took an RTT sample of %v from a refused segment", r.SmoothedRTT())
}
if !r.Running() {
t.Error("timer disarmed by a refused acknowledgement")
}
}
// TestRTO_RetransmitsZeroWindowProbe verifies the timer takes over the periodic
// probing of a closed send window. A zero-window probe is a single octet the peer
// cannot accept, so it goes unacknowledged; the timer must keep resending it, with
// exponential backoff, which is the persist-timer behaviour of RFC 9293 §3.8.6.1.
// The tcp package relies on this and refuses to probe without a policy installed.
func TestRTO_RetransmitsZeroWindowProbe(t *testing.T) {
r := newRTO()
const iss = uint32(5000)
probe := dataSeg(iss, 1) // The one-octet probe.
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 {
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 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)
now += int64(prevRTO)
}
}