fix(udps): stop packets being dated from an earlier time base

Reported as samples sporadically carrying a previous packet's timestamp: holes on one side of the stream and collisions on the other, in both the Go and the MARTe2 receiver. That it appeared in both is what located it -- the shared cause is upstream of either client. Four independent defects, all of which end in a packet's values being placed at a time that is not theirs. Reassembly slot exhaustion (the "Reassembly slots full; evicting oldest" flood). Chunk size was learnt only from fragment 0, so an out-of-order burst destroyed a packet whose bytes had all arrived and left the slot occupied until the 2 s GC. Slots were keyed on the counter alone, but DATA and CONFIG number independently, so equal counters merged the two streams. The 32-byte received-mask covered 256 of the 512 fragments the client accepts, so a duplicate above 255 was counted as new and the packet was delivered with a hole of stale bytes in it. And one datagram was read per Execute(), which cannot drain a fast producer. Fixed with a pendingTail deferral, (counter, type) keying, a 64-byte mask, a 256-datagram drain, counter-age slot reclamation, and a 1 Hz aggregated warning in place of the per-eviction flood. UDPStreamer dropping whole Accumulate batches. EventSem::ResetWait is Reset-then-Wait, so a Post() landing while the sender thread was inside ServiceClients()/SendData() was destroyed by the next Reset. The batch was then skipped with dataReady false, readyFill was never cleared, and the following flush overwrote it: an entire run of RT cycles never reached the wire. The record of pending work now lives in the buffers rather than in the semaphore edge, which also removes up to UDPS_DATA_WAIT_MS of latency; genuine backpressure overwrites are counted and reported. Against the unfixed code the new test sees 2999/3000 batches never consumed. Period inflation after loss. Accumulated scalars carry no SamplingRate, so the receiver derives dt from the sender-clock gap -- but dividing it by the previous packet's sample count is only right while nothing is lost. One loss doubles the reported period, which spreads a batch a full batch past its own end and into the range the next packet claims. That is the hole and the collision, exactly. Inferring the cycle count from the estimate's own period is not a way out: it has a stable fixed point wherever gap/dt is an integer, so a real rate change locks it at the old one for good (AccumDtGTest.FollowsSustainedRateChange). The packet counter removes the ambiguity, so all three receivers now order on it: a DATA packet that does not advance the counter is dropped rather than delivered, because its values are older than data already handed over. Ordering is on the signed difference so it survives the uint32 wrap, and the sequence resets on reconnect, where the producer's counter restarts independently of ours. The loss count that falls out of the same delta feeds the period estimate as cycles = prevN * (1 + lost), which reduces exactly to gap/prevN when nothing is lost and therefore still tracks a genuine rate change. UDPSClient::AcceptDataCounter (C++), udpsprotocol.SequenceGate (Go), decode_data (C). The C client's existing gap counter was wrap-unsafe and let a stale packet rewind last_counter, which made every subsequent gap wrong; it uses the same code now. Docs/Protocol.md gains an Ordering DATA section stating the requirement for any receiver, including ones outside this repository. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-09-02 01:14:40 +02:00
parent 13fac79400
commit deabd257e5
23 changed files with 1975 additions and 163 deletions
@@ -0,0 +1,146 @@
+package udpsprotocol
+
+// Accumulate mode ships one full snapshot of EVERY signal per RT cycle —
+// arrays included.  See UDPStreamer.cpp pass 5 ("ALL signals (scalars and
+// arrays alike) are tagged accumulated = true") and SerializeAccumulated,
+// which writes, for each signal in CONFIG order, numSamples consecutive
+// snapshots of that signal's full element set.
+//
+// The tests below build a payload byte-for-byte the way the C++ producer
+// does, so a decoding regression shows up here rather than as a mangled
+// waveform three components downstream.
+
+import (
+	"encoding/binary"
+	"math"
+	"testing"
+	"time"
+)
+
+// buildAccumulatePayload lays out an Accumulate DATA payload exactly as
+// UDPStreamer::SerializeAccumulated does:
+//
+//	[8 HRT][4 numSamples] then, per signal, numSamples × NumElements float64.
+//
+// slots[i][k] holds signal i's element set for cycle k.
+func buildAccumulatePayload(hrt uint64, slots [][][]float64) []byte {
+	numSamples := 0
+	if len(slots) > 0 {
+		numSamples = len(slots[0])
+	}
+	out := make([]byte, 12)
+	binary.LittleEndian.PutUint64(out[0:8], hrt)
+	binary.LittleEndian.PutUint32(out[8:12], uint32(numSamples))
+	for _, sig := range slots {
+		for _, elems := range sig {
+			for _, v := range elems {
+				var b [8]byte
+				binary.LittleEndian.PutUint64(b[:], math.Float64bits(v))
+				out = append(out, b[:]...)
+			}
+		}
+	}
+	return out
+}
+
+// TestParseDataAccumulateGivesEachSlotItsOwnArray pins the array case: with an
+// accumulated batch, slot k's array signal must decode to the values the
+// producer captured on cycle k, not to some other cycle's.  Handing every slot
+// slot 0's array would stamp one cycle's data with every slot's timestamp —
+// the same samples drawn repeatedly at advancing times, with the cycles they
+// displaced missing entirely.
+func TestParseDataAccumulateGivesEachSlotItsOwnArray(t *testing.T) {
+	sigs := []SignalInfo{
+		{Name: "Time", TypeCode: 9, NumRows: 1, NumCols: 1, QuantType: QuantNone},
+		{Name: "Wave", TypeCode: 9, NumRows: 4, NumCols: 1, QuantType: QuantNone},
+	}
+	// Three RT cycles.  "Wave" carries a different ramp each cycle so a
+	// mix-up is unambiguous.
+	timeSlots := [][]float64{{10}, {20}, {30}}
+	waveSlots := [][]float64{
+		{1, 2, 3, 4},
+		{5, 6, 7, 8},
+		{9, 10, 11, 12},
+	}
+	payload := buildAccumulatePayload(777, [][][]float64{timeSlots, waveSlots})
+
+	samples, err := ParseData(payload, sigs, PublishModeAccumulate, time.Now())
+	if err != nil {
+		t.Fatalf("ParseData: %v", err)
+	}
+	if len(samples) != 3 {
+		t.Fatalf("expected 3 slots, got %d", len(samples))
+	}
+	for k, s := range samples {
+		got := s.Values["Wave"]
+		want := waveSlots[k]
+		if len(got) != len(want) {
+			t.Fatalf("slot %d: Wave has %d elements, want %d", k, len(got), len(want))
+		}
+		for i := range want {
+			if got[i] != want[i] {
+				t.Fatalf("slot %d: Wave = %v, want %v (slot %d's data has been "+
+					"served for this slot's timestamp)", k, got, want,
+					indexOfSlot(waveSlots, got))
+			}
+		}
+		if tv := s.Values["Time"]; len(tv) != 1 || tv[0] != timeSlots[k][0] {
+			t.Fatalf("slot %d: Time = %v, want %v", k, tv, timeSlots[k])
+		}
+	}
+}
+
+// TestParseDataAccumulateConsumesTheWholeArrayBlock catches the same defect
+// from the other side: a signal following an array must be read at the right
+// offset.  Under-reading the array block slides every later signal backwards
+// into the array's tail, which decodes as plausible-looking but wrong values
+// rather than as an error.
+func TestParseDataAccumulateConsumesTheWholeArrayBlock(t *testing.T) {
+	sigs := []SignalInfo{
+		{Name: "Wave", TypeCode: 9, NumRows: 4, NumCols: 1, QuantType: QuantNone},
+		{Name: "Tail", TypeCode: 9, NumRows: 1, NumCols: 1, QuantType: QuantNone},
+	}
+	waveSlots := [][]float64{
+		{1, 2, 3, 4},
+		{5, 6, 7, 8},
+		{9, 10, 11, 12},
+	}
+	tailSlots := [][]float64{{100}, {200}, {300}}
+	payload := buildAccumulatePayload(0, [][][]float64{waveSlots, tailSlots})
+
+	samples, err := ParseData(payload, sigs, PublishModeAccumulate, time.Now())
+	if err != nil {
+		t.Fatalf("ParseData: %v", err)
+	}
+	if len(samples) != 3 {
+		t.Fatalf("expected 3 slots, got %d", len(samples))
+	}
+	for k, s := range samples {
+		tv := s.Values["Tail"]
+		if len(tv) != 1 || tv[0] != tailSlots[k][0] {
+			t.Fatalf("slot %d: Tail = %v, want %v — the array block before it "+
+				"was not fully consumed", k, tv, tailSlots[k])
+		}
+	}
+}
+
+// indexOfSlot reports which slot's data a decoded array actually matches, so a
+// failure message can name the culprit instead of just showing numbers.
+func indexOfSlot(slots [][]float64, got []float64) int {
+	for k, want := range slots {
+		if len(want) != len(got) {
+			continue
+		}
+		same := true
+		for i := range want {
+			if want[i] != got[i] {
+				same = false
+				break
+			}
+		}
+		if same {
+			return k
+		}
+	}
+	return -1
+}
@@ -290,28 +290,37 @@ type DataSample struct {
 	HRTTimestamp uint64
 	WallTime     time.Time            // wall-clock time at UDP arrival; used as x-axis
 	Values       map[string][]float64 // key = signal name, value = []float64 with NumElements entries
+	// Lost is the number of DATA packets missing between the previous sample
+	// and this one, taken from the producer's packet counter (see
+	// SequenceGate). Consumers that derive a per-element period from the
+	// inter-packet gap need it: the gap widens with every lost packet, and
+	// dividing it by this packet's element count alone reports a period too
+	// long by exactly that factor — which walks the packet's elements past
+	// their own end and into the range the next packet claims.
+	Lost uint32
 }

 // parseElems reads n elements for sig from payload at offset, advancing offset.
 // Returns the slice of float64 values and the new offset.
 func parseElems(payload []byte, offset, n int, sig SignalInfo) ([]float64, int, error) {
+	sz := rawTypeSize(sig.TypeCode)
+	if sig.QuantType != QuantNone {
+		sz = quantSize(sig.QuantType)
+	}
+	// Bounds-check before allocating.  In Accumulate mode n is numSamples ×
+	// NumElements, so a malformed packet could otherwise ask for an allocation
+	// far larger than its own payload could ever justify.
+	if n < 0 || n > (len(payload)-offset)/sz {
+		return nil, offset, fmt.Errorf("data payload truncated for signal %q", sig.Name)
+	}
 	elems := make([]float64, n)
+	needed := n * sz
 	if sig.QuantType == QuantNone {
-		sz := rawTypeSize(sig.TypeCode)
-		needed := n * sz
-		if offset+needed > len(payload) {
-			return nil, offset, fmt.Errorf("data payload truncated for signal %q", sig.Name)
-		}
 		for i := 0; i < n; i++ {
 			elems[i] = readRawElement(payload, offset+i*sz, sig.TypeCode)
 		}
 		offset += needed
 	} else {
-		sz := quantSize(sig.QuantType)
-		needed := n * sz
-		if offset+needed > len(payload) {
-			return nil, offset, fmt.Errorf("data payload truncated (quant) for signal %q", sig.Name)
-		}
 		for i := 0; i < n; i++ {
 			var raw uint16
 			if sz == 1 {
@@ -331,7 +340,13 @@ func parseElems(payload []byte, offset, n int, sig SignalInfo) ([]float64, int,
 //
 // For PublishModeAccumulate the payload format is:
 //
-//	[8 HRT][4 numSamples][for each signal: accumulated scalars → numSamples elems; arrays → NumElements elems]
+//	[8 HRT][4 numSamples][for each signal: numSamples × NumElements elems]
+//
+// Every signal is accumulated, arrays included: the producer captures one full
+// snapshot of the whole signal set per RT cycle and lays the cycles out
+// contiguously per signal (UDPStreamer::SerializeAccumulated).  Reading only
+// NumElements for an array would hand every slot the first cycle's data and
+// slide all later signals into that array's tail.
 //
 // The function returns one DataSample per accumulated snapshot so the hub can
 // process each slot independently with its own timestamp.
@@ -357,28 +372,18 @@ func ParseData(payload []byte, sigs []SignalInfo, publishMode uint8, arrivalTime
 		}

 		// Parse per-signal data blocks (all slots for a signal are contiguous).
-		accumVals := make(map[string][]float64, len(sigs)) // scalars: numSamples values
-		fixedVals := make(map[string][]float64, len(sigs)) // arrays: NumElements values
+		accumVals := make(map[string][]float64, len(sigs)) // numSamples × NumElements
+		accumElems := make(map[string]int, len(sigs))

 		for _, sig := range sigs {
 			n := sig.NumElements()
-			if n == 1 {
-				// Accumulated scalar: read numSamples back-to-back elements.
-				elems, newOff, err := parseElems(payload, offset, numSamples, sig)
-				if err != nil {
-					return nil, err
-				}
-				offset = newOff
-				accumVals[sig.Name] = elems
-			} else {
-				// Fixed array (non-accumulated): one set of NumElements values.
-				elems, newOff, err := parseElems(payload, offset, n, sig)
-				if err != nil {
-					return nil, err
-				}
-				offset = newOff
-				fixedVals[sig.Name] = elems
+			elems, newOff, err := parseElems(payload, offset, numSamples*n, sig)
+			if err != nil {
+				return nil, err
 			}
+			offset = newOff
+			accumVals[sig.Name] = elems
+			accumElems[sig.Name] = n
 		}

 		// Build one DataSample per slot.
@@ -386,10 +391,10 @@ func ParseData(payload []byte, sigs []SignalInfo, publishMode uint8, arrivalTime
 		for k := 0; k < numSamples; k++ {
 			vals := make(map[string][]float64, len(sigs))
 			for sigName, av := range accumVals {
-				vals[sigName] = []float64{av[k]}
-			}
-			for sigName, fv := range fixedVals {
-				vals[sigName] = fv // shared read-only reference; hub does not modify
+				n := accumElems[sigName]
+				// Sub-slice of the decoded block; the hub treats values as
+				// read-only, so no copy is needed.
+				vals[sigName] = av[k*n : (k+1)*n : (k+1)*n]
 			}
 			samples[k] = DataSample{HRTTimestamp: hrt, WallTime: arrivalTime, Values: vals}
 		}
@@ -0,0 +1,49 @@
+package udpsprotocol
+
+// SequenceGate orders DATA packets by the producer's packet counter.
+//
+// Reassembly completes in arrival order, not counter order, so a packet that
+// was delayed or duplicated on the wire is handed up after a newer one has
+// already been consumed. Its samples then carry an older time base than the
+// data already in the ring: they land on top of samples that are already
+// there, and the span they should have filled stays empty. That is a hole on
+// one side and a collision on the other, from a packet that is entirely
+// well-formed — the counter is the only thing that distinguishes it.
+//
+// A SequenceGate is not safe for concurrent use; each receive loop owns one.
+type SequenceGate struct {
+	last  uint32
+	valid bool
+	// Stale counts packets rejected for not advancing the counter (reordered
+	// or duplicated), for diagnostics.
+	Stale uint64
+}
+
+// Reset forgets the sequence. Call it on (re)connect: the producer's counter
+// restarts independently of ours, so a counter carried over from the previous
+// connection would reject the whole new stream.
+func (g *SequenceGate) Reset() {
+	g.last = 0
+	g.valid = false
+}
+
+// Accept reports whether a DATA packet with this counter should be delivered,
+// and how many packets went missing immediately before it.
+//
+// The counter is a wrapping uint32, so ordering is done on the signed
+// difference: a plain comparison would call the first packet after the wrap
+// stale and reject everything from then on.
+func (g *SequenceGate) Accept(counter uint32) (ok bool, lost uint32) {
+	if !g.valid {
+		g.valid = true
+		g.last = counter
+		return true, 0
+	}
+	delta := int32(counter - g.last)
+	if delta <= 0 {
+		g.Stale++
+		return false, 0
+	}
+	g.last = counter
+	return true, uint32(delta) - 1
+}
@@ -0,0 +1,82 @@
+package udpsprotocol
+
+import "testing"
+
+// A packet older than one already delivered carries an older time base. Its
+// samples land on top of data that is already in the ring and leave the span
+// they should have filled empty, so it must not get through.
+func TestSequenceGateRejectsStaleAndDuplicate(t *testing.T) {
+	var g SequenceGate
+
+	if ok, lost := g.Accept(10); !ok || lost != 0 {
+		t.Fatalf("first packet: got (%v, %d), want (true, 0)", ok, lost)
+	}
+	if ok, _ := g.Accept(11); !ok {
+		t.Fatal("counter 11 advances past 10 and must be accepted")
+	}
+	if ok, _ := g.Accept(9); ok {
+		t.Error("counter 9 is older than the delivered 11 and must be dropped")
+	}
+	if ok, _ := g.Accept(11); ok {
+		t.Error("a repeat of the delivered counter must be dropped")
+	}
+	if g.Stale != 2 {
+		t.Errorf("Stale = %d, want 2", g.Stale)
+	}
+	// The rejections must not have moved the sequence on.
+	if ok, lost := g.Accept(12); !ok || lost != 0 {
+		t.Errorf("after rejections: got (%v, %d), want (true, 0)", ok, lost)
+	}
+}
+
+// The loss count is what lets a consumer tell a widened gap from a slowed
+// producer, so it must exclude the packet being delivered and must not persist
+// into the next one.
+func TestSequenceGateReportsLoss(t *testing.T) {
+	var g SequenceGate
+
+	g.Accept(100)
+	if _, lost := g.Accept(104); lost != 3 {
+		t.Errorf("101..103 missing: lost = %d, want 3", lost)
+	}
+	if _, lost := g.Accept(105); lost != 0 {
+		t.Errorf("consecutive packet: lost = %d, want 0", lost)
+	}
+	if g.Stale != 0 {
+		t.Errorf("Stale = %d, want 0", g.Stale)
+	}
+}
+
+// The counter is a wrapping uint32. Ordering it by plain comparison would call
+// every packet after the wrap older than 0xFFFFFFFF and kill the stream.
+func TestSequenceGateSurvivesWraparound(t *testing.T) {
+	var g SequenceGate
+
+	for _, c := range []uint32{0xFFFFFFFD, 0xFFFFFFFE, 0xFFFFFFFF, 0, 1, 2} {
+		ok, lost := g.Accept(c)
+		if !ok {
+			t.Fatalf("counter %#x rejected across the wrap", c)
+		}
+		if lost != 0 {
+			t.Errorf("counter %#x: lost = %d, want 0", c, lost)
+		}
+	}
+	// Loss must still be measured correctly across the wrap.
+	var h SequenceGate
+	h.Accept(0xFFFFFFFE)
+	if _, lost := h.Accept(1); lost != 2 {
+		t.Errorf("0xFFFFFFFF and 0 missing: lost = %d, want 2", lost)
+	}
+}
+
+// A reconnect restarts the producer's counter independently of ours; a carried
+// over counter would reject the entire new stream.
+func TestSequenceGateResetAcceptsLowerCounter(t *testing.T) {
+	var g SequenceGate
+
+	g.Accept(5000)
+	g.Reset()
+	if ok, lost := g.Accept(3); !ok || lost != 0 {
+		t.Errorf("after Reset: got (%v, %d), want (true, 0)", ok, lost)
+	}
+}