zudp 0.4.2

Zero-overhead UDP protocol with optional reliability, fragmentation, and relay support
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
use std::{
    collections::HashMap,
    net::SocketAddr,
    sync::{
        Arc,
        atomic::{AtomicBool, AtomicU64, Ordering},
    },
    time::Duration,
};

use bytes::{Bytes, BytesMut};
use dashmap::DashMap;
use parking_lot::{Mutex, RwLock};
use tokio::{sync::mpsc, task::JoinSet, time};

use crate::{
    Config, PeerAddr,
    event::{DisconnectReason, PeerEvent},
    frag::FragAssembler,
    frame::Frame,
    peer::{Inbound, PeerState, RecvState},
    socket::{self, RawSocket},
};

/// Most `SessionUnknown` budgets tracked at once (relayed peers' keys are chosen by the
/// relay, so the map must not grow with them).
#[cfg(feature = "security")]
pub(crate) const MAX_HINT_KEYS: usize = 4_096;

/// Maximum concurrent NACK retransmit tasks across the engine.
///
/// Each task handles one NACK batch (one peer, one burst of missing seqs).  Capping
/// prevents memory/CPU runaway when a peer floods NACKs faster than retransmits drain.
const MAX_RETRANSMIT_TASKS: usize = 32;

/// Tail-loss probes sent per idle period before giving up (backoff doubles each time).
const TAIL_PROBES: u32 = 3;
/// Probe timeout bounds: `2 × srtt`, clamped; [`TAIL_PROBE_DEFAULT`] before the first RTT sample.
const TAIL_PROBE_MIN: Duration = Duration::from_millis(10);
const TAIL_PROBE_MAX: Duration = Duration::from_secs(1);
const TAIL_PROBE_DEFAULT: Duration = Duration::from_millis(100);

/// Capacity of the bounded inbound channel between the engine task and the socket handle.
///
/// At 1 400 B/frame and 8 192 slots this is ~11 MB peak backpressure before the engine
/// starts dropping frames.  The engine uses `try_send` so the recv loop is never stalled.
const INBOUND_CAP: usize = 8_192;

/// Consecutive hard receive errors before an ephemeral-port socket is replaced.
const REBIND_AFTER_ERRORS: u32 = 10;
/// Handshake frames accepted per second per source IP, separate from the data budget.
const HANDSHAKE_PPS_PER_IP: f64 = 20.0;

/// Back-off unit between retries after a hard receive error.
const RECV_ERROR_BACKOFF: Duration = Duration::from_millis(20);

/// Smallest per-stream receive window, whatever `sent_buffer_frames` is set to.
const MIN_RECV_WINDOW: usize = 1_024;

/// Receive side of one (peer, stream).
pub(crate) struct StreamRx {
    /// [`PeerState::generation`] this state belongs to; a mismatch means the peer was reset.
    generation: u64,
    recv: RecvState,
    frag: FragAssembler,
}

pub(crate) type RecvStates = HashMap<(PeerAddr, u16), StreamRx>;

/// State shared between the engine task and the [`crate::ZudpSocket`] / [`crate::ZudpConn`] handles.
pub(crate) struct EngineInner {
    /// Current socket; replaced atomically on rebind.  Use [`get_socket`] to obtain a cheap clone.
    pub socket: RwLock<RawSocket>,
    pub peers: DashMap<PeerAddr, Arc<PeerState>>,
    /// Reverse index: remote engine instance id → peer, for detecting address migrations.
    pub sessions: DashMap<u64, Arc<PeerState>>,
    /// Random id of this engine, sent in every Ping and Pong.
    pub instance_id: u64,
    /// Lifecycle notifications; see [`PeerEvent`].
    pub events: tokio::sync::broadcast::Sender<PeerEvent>,
    pub config: Arc<Config>,
    /// Noise keypair for the `security` feature; `None` when security is disabled.
    #[cfg(feature = "security")]
    pub keypair: Option<crate::security::Keypair>,
    /// Set by the engine task before it exits; callers read this to distinguish a
    /// clean shutdown from an unexpected crash (rebind failure, etc.).
    pub shutdown_reason: Mutex<Option<String>>,
    pub dropped_rate_limited: AtomicU64,
    pub dropped_peer_cap: AtomicU64,
    /// Plaintext data frames dropped because a keypair is configured.
    pub dropped_unauthenticated: AtomicU64,
    /// Set when a reliable stream could not deliver because the inbound channel was full;
    /// the engine then waits for capacity and drains every stream.
    pub delivery_stalled: AtomicBool,
    /// Handshakes in progress, keyed by remote address.
    #[cfg(feature = "security")]
    pub pending: Mutex<crate::handshake::PendingMap>,
    /// Our session index → peer, so Secure frames find their session from any address.
    #[cfg(feature = "security")]
    pub by_index: DashMap<u64, Arc<PeerState>>,
    /// Budget for `SessionUnknown` replies per source IP.
    #[cfg(feature = "security")]
    pub hint_limiter: Mutex<crate::rate::RateLimiter<PeerAddr>>,
    /// Remote static key → direct session, so a relay node finds the client a key names.
    #[cfg(feature = "security")]
    pub by_key: DashMap<[u8; 32], Arc<PeerState>>,
    /// Punched direct address → the relayed peer whose traffic now arrives from it.
    #[cfg(feature = "security")]
    pub punched: DashMap<SocketAddr, Arc<PeerState>>,
    /// Relayed frames dropped because the relay's sender claim did not match the session.
    pub dropped_forged: AtomicU64,
    /// Relayed frames dropped because they came through a relay we never accepted.
    pub dropped_relay_unaccepted: AtomicU64,
    /// Relays allowed to deliver new inbound relayed sessions to us.
    #[cfg(feature = "security")]
    pub inbound_relays: RwLock<std::collections::HashSet<SocketAddr>>,
    /// Static keys of relayed peers the application approved for hole punching.
    #[cfg(feature = "security")]
    pub punch_allowed: RwLock<std::collections::HashSet<[u8; 32]>>,
    /// Budget for relayed handshake initiations per relay IP.
    #[cfg(feature = "security")]
    pub relayed_hs_limiter: Mutex<crate::rate::RateLimiter>,
    /// Relay node state; `Some` when built with [`crate::Zudp::relay_node`].
    #[cfg(feature = "relay")]
    pub relay: Option<crate::relay::node::RelayNode>,
}

impl EngineInner {
    /// Cheap handle to the current socket (clones the inner `Arc<UdpSocket>`).
    pub fn get_socket(&self) -> RawSocket {
        self.socket.read().clone()
    }

    /// `true` when a Noise keypair is configured.
    #[cfg_attr(not(feature = "security"), allow(clippy::unused_self))]
    pub fn secured(&self) -> bool {
        #[cfg(feature = "security")]
        {
            self.keypair.is_some()
        }
        #[cfg(not(feature = "security"))]
        {
            false
        }
    }

    pub fn get_or_create_peer(self: &Arc<Self>, addr: PeerAddr) -> Arc<PeerState> {
        self.peer_for_incoming(addr).0
    }

    /// Like [`get_or_create_peer`](Self::get_or_create_peer); also says whether it was created.
    pub fn peer_for_incoming(self: &Arc<Self>, addr: PeerAddr) -> (Arc<PeerState>, bool) {
        if let Some(p) = self.peers.get(&addr) {
            return (p.value().clone(), false);
        }
        self.make_room(addr);
        let mut created = false;
        let peer = self
            .peers
            .entry(addr)
            .or_insert_with(|| {
                created = true;
                Arc::new(PeerState::new(addr))
            })
            .value()
            .clone();
        (peer, created)
    }

    /// Start a new session state for `addr`, dropping everything the old one carried.
    ///
    /// Sequence counters, send buffers, receive state, fragments, congestion state, keys and
    /// the session-id index all belong to one remote incarnation; reusing them across a
    /// restart makes the new incarnation's frames look stale.
    pub fn replace_peer(&self, addr: PeerAddr) -> Arc<PeerState> {
        if !self.peers.contains_key(&addr) {
            self.make_room(addr);
        }
        let fresh = Arc::new(match self.peers.get(&addr) {
            Some(old) => PeerState::with_addr(old.addr_arc()),
            None => PeerState::new(addr),
        });
        if let Some(old) = self.peers.insert(addr, fresh.clone()) {
            self.unindex(&old);
        }
        fresh
    }

    /// Keep the peer table under `max_peers` before `new` is added: evict the least
    /// recently seen peer when it is full.
    fn make_room(&self, new: PeerAddr) {
        if self.evict_lru(new, self.config.max_peers, |_| true) {
            self.dropped_peer_cap.fetch_add(1, Ordering::Relaxed);
        }
    }

    /// When at least `cap` peers match `pred`, evict the least recently seen of them to make
    /// room for `new`.  Returns `true` when a peer was evicted.
    pub fn evict_lru(&self, new: PeerAddr, cap: usize, pred: impl Fn(&PeerAddr) -> bool) -> bool {
        if self.peers.iter().filter(|r| pred(r.key())).count() < cap.max(1) {
            return false;
        }
        let Some(evict) = self
            .peers
            .iter()
            .filter(|r| pred(r.key()) && *r.key() != new)
            .min_by_key(|r| *r.value().last_seen.lock())
            .map(|r| *r.key())
        else {
            return false;
        };
        tracing::warn!(
            target: "zudp::engine",
            evicted = %evict,
            %new,
            "peer cap reached — evicting least-recently-seen peer"
        );
        self.forget_peer(evict);
        self.emit(PeerEvent::Disconnected {
            addr: evict,
            reason: DisconnectReason::Evicted,
        });
        true
    }

    /// Remove `addr` and every index entry pointing at it.  Returns the removed state.
    pub fn forget_peer(&self, addr: PeerAddr) -> Option<Arc<PeerState>> {
        let (_, old) = self.peers.remove(&addr)?;
        self.unindex(&old);
        Some(old)
    }

    pub fn unindex(&self, old: &Arc<PeerState>) {
        #[cfg(feature = "security")]
        if let Some(index) = old.local_index() {
            self.by_index.remove_if(&index, |_, p| Arc::ptr_eq(p, old));
        }
        if let Some(id) = old.their_session_id() {
            self.sessions.remove_if(&id, |_, p| Arc::ptr_eq(p, old));
        }
        #[cfg(feature = "security")]
        {
            if let Some(key) = old.remote_static_key()
                && self
                    .by_key
                    .remove_if(&key, |_, p| Arc::ptr_eq(p, old))
                    .is_some()
            {
                #[cfg(feature = "relay")]
                if let Some(node) = &self.relay {
                    node.forget_client(&key);
                }
            }
            if let Some(direct) = old.direct_path() {
                self.punched.remove_if(&direct, |_, p| Arc::ptr_eq(p, old));
            }
        }
    }

    /// Publish a lifecycle event; nobody listening is fine.
    pub fn emit(&self, event: PeerEvent) {
        let _ = self.events.send(event);
    }

    /// Drop `addr` on request.  Returns `false` when there was no such peer.
    pub fn remove_peer(&self, addr: PeerAddr) -> bool {
        #[cfg(feature = "security")]
        self.pending.lock().remove(&addr);
        if self.forget_peer(addr).is_none() {
            return false;
        }
        self.emit(PeerEvent::Disconnected {
            addr,
            reason: DisconnectReason::Removed,
        });
        true
    }

    /// `true` while `peer` is still the live state for its address.
    pub fn is_current(&self, peer: &Arc<PeerState>) -> bool {
        self.peers
            .get(&peer.addr())
            .is_some_and(|p| Arc::ptr_eq(p.value(), peer))
    }

    /// The datagram that carries `wire` to `peer` along its current path, and where to
    /// send it.  `None` when the path is gone (a relayed peer whose relay session ended).
    pub fn outbound(&self, peer: &PeerState, wire: Bytes) -> Option<(Bytes, SocketAddr)> {
        #[cfg(feature = "security")]
        if let Some(direct) = peer.direct_path() {
            return Some((wire, direct));
        }
        self.outbound_to(peer.addr(), wire)
    }

    /// Like [`outbound`](Self::outbound) for a peer that may have no state yet (handshakes).
    pub fn outbound_to(&self, to: PeerAddr, wire: Bytes) -> Option<(Bytes, SocketAddr)> {
        match to {
            PeerAddr::Direct(addr) => Some((wire, addr)),
            #[cfg(feature = "security")]
            PeerAddr::Relayed { relay, key } => crate::relay::via(self, relay, key, wire),
            #[cfg(not(feature = "security"))]
            PeerAddr::Relayed { .. } => None,
        }
    }

    /// Send `wire` to `peer` along its current path (a punched direct path included),
    /// logging a failure.
    pub async fn send_to_peer(&self, peer: &PeerState, wire: Bytes, what: &'static str) {
        let routed = self.outbound(peer, wire);
        self.send_routed(peer.addr(), routed, what).await;
    }

    /// Like [`send_to_peer`](Self::send_to_peer) for a peer that may have no state yet.
    pub async fn send_to_addr(&self, to: PeerAddr, wire: Bytes, what: &'static str) {
        let peer = self.peers.get(&to).map(|r| r.value().clone());
        match peer {
            Some(peer) => self.send_to_peer(&peer, wire, what).await,
            None => self.send_handshake(to, wire, what).await,
        }
    }

    /// Send a handshake message to `to`: always along the handle's own route (through the
    /// relay for a relayed peer), never along a punched path, which belongs to the session
    /// being replaced.
    pub async fn send_handshake(&self, to: PeerAddr, wire: Bytes, what: &'static str) {
        let routed = self.outbound_to(to, wire);
        self.send_routed(to, routed, what).await;
    }

    async fn send_routed(
        &self,
        to: PeerAddr,
        routed: Option<(Bytes, SocketAddr)>,
        what: &'static str,
    ) {
        let Some((datagram, dest)) = routed else {
            tracing::debug!(target: "zudp::engine", peer = %to, what, "no path to peer");
            return;
        };
        if let Err(e) = self.get_socket().send_to(&datagram, dest).await {
            tracing::warn!(target: "zudp::engine", peer = %to, what, "send failed: {e}");
        }
    }
}

/// Spawn the engine background task and return the inbound receiver.
///
/// `shutdown_rx` resolves when the caller drops the paired `oneshot::Sender`; the engine
/// then drains in-flight retransmits and exits cleanly.
///
/// Each item on the returned channel is `(payload_bytes, sender_addr, stream_id)`.
pub(crate) fn spawn(
    inner: Arc<EngineInner>,
    shutdown_rx: tokio::sync::oneshot::Receiver<()>,
) -> mpsc::Receiver<(Bytes, PeerAddr, u16)> {
    let (inbound_tx, inbound_rx) = mpsc::channel(INBOUND_CAP);
    tokio::spawn(run(inner, inbound_tx, shutdown_rx));
    inbound_rx
}

async fn run(
    inner: Arc<EngineInner>,
    inbound_tx: mpsc::Sender<(Bytes, PeerAddr, u16)>,
    mut shutdown_rx: tokio::sync::oneshot::Receiver<()>,
) {
    // Keyed by (peer_addr, stream_id) so each stream has independent reorder + reassembly.
    let mut recv_states: RecvStates = HashMap::new();
    let keepalive_interval = inner.config.keepalive_interval;
    let sent_prune_age = inner.config.sent_prune_age;

    let mut rate_limiter = crate::rate::RateLimiter::new(
        inner.config.max_pps_per_ip,
        (inner.config.max_pps_per_ip / 5.0).max(1.0),
    );

    // Bounded pool for NACK retransmit tasks.  Each task may sleep for CC pacing, so
    // they are spawned off the recv loop to avoid stalling it — but the pool is capped
    // to prevent unbounded task accumulation under heavy loss or adversarial NACKs.
    // Handshakes get their own budget so a data flood from an IP cannot starve that IP's
    // reconnect, while a handshake flood still cannot exhaust the engine.
    let mut handshake_limiter =
        crate::rate::RateLimiter::new(HANDSHAKE_PPS_PER_IP, HANDSHAKE_PPS_PER_IP);

    let mut retransmit_tasks: JoinSet<()> = JoinSet::new();

    // Local recv socket; replaced in-place when the socket is rebound.
    let mut recv_sock = inner.get_socket();
    let mut recv_errors: u32 = 0;

    let mut ticker = time::interval(keepalive_interval / 4);

    loop {
        // Reap completed retransmit tasks without blocking.
        while retransmit_tasks.try_join_next().is_some() {}

        tokio::select! {
            biased;

            _ = &mut shutdown_rx => {
                *inner.shutdown_reason.lock() = Some("shutdown requested".into());
                tracing::info!(target: "zudp::engine", "graceful shutdown — draining retransmits");
                break;
            }

            result = recv_sock.recv_from() => {
                match result {
                    Ok((data, from)) => {
                        recv_errors = 0;
                        let limiter = if is_handshake(&data) {
                            &mut handshake_limiter
                        } else {
                            &mut rate_limiter
                        };
                        // Addresses that completed a handshake are not throttled; the budget is
                        // for traffic nobody has authenticated yet.
                        let trusted = inner
                            .peers
                            .get(&PeerAddr::Direct(from))
                            .is_some_and(|p| p.has_channel())
                            || is_punched(&inner, from);
                        // Handshakes are budgeted per network (an IPv6 /64), so rotating
                        // addresses inside one allocation buys nothing.
                        let key = if is_handshake(&data) {
                            crate::rate::network_of(from.ip())
                        } else {
                            from.ip()
                        };
                        if inner.config.max_pps_per_ip > 0.0 && !trusted && !limiter.allow(key) {
                            tracing::trace!(target: "zudp::engine", %from, "rate limited");
                            inner.dropped_rate_limited.fetch_add(1, Ordering::Relaxed);
                            continue;
                        }
                        handle_incoming(
                            data, from, &inner, &inbound_tx,
                            &mut recv_states,
                            &mut retransmit_tasks,
                        ).await;
                    }
                    Err(e) if socket::is_transient_recv_error(&e) => {
                        tracing::debug!(target: "zudp::engine", "transient recv error ignored: {e}");
                    }
                    Err(e) => {
                        recv_errors += 1;
                        tracing::warn!(target: "zudp::engine", errors = recv_errors, "recv error: {e}");
                        if recv_errors < REBIND_AFTER_ERRORS || inner.config.port != 0 {
                            // A requested port is never given up: keep retrying the socket.
                            time::sleep(RECV_ERROR_BACKOFF * recv_errors.min(50)).await;
                            continue;
                        }
                        recv_errors = 0;
                        match rebind_socket(&inner).await {
                            Ok(new_sock) => {
                                recv_sock = new_sock;
                                // Ping all known peers so they can migrate our address.
                                ping_all_peers(&inner, &recv_sock).await;
                            }
                            Err(bind_err) => {
                                let reason = format!("socket rebind failed: {bind_err}");
                                tracing::error!(target: "zudp::engine", "{reason}");
                                *inner.shutdown_reason.lock() = Some(reason);
                                break;
                            }
                        }
                    }
                }
            }

            permit = inbound_tx.reserve(), if inner.delivery_stalled.load(Ordering::Relaxed) => {
                drop(permit);
                inner.delivery_stalled.store(false, Ordering::Relaxed);
                drain_all_streams(&inner, &mut recv_states, &inbound_tx);
            }

            _ = ticker.tick() => {
                run_background(
                    &inner, &mut recv_states, &mut rate_limiter, keepalive_interval, sent_prune_age,
                );
            }
        }
    }

    // Drain in-flight retransmits before the socket closes.
    retransmit_tasks.shutdown().await;
    tracing::info!(target: "zudp::engine", "engine task exited");
}

/// Replace a failing ephemeral-port socket with a fresh one on a new ephemeral port.
///
/// On success, replaces `inner.socket` so future `send_to` calls use the new interface.
async fn rebind_socket(inner: &Arc<EngineInner>) -> Result<RawSocket, crate::Error> {
    let old_local = inner.get_socket().local_addr()?;
    let mut new_sock = RawSocket::bind(SocketAddr::new(old_local.ip(), 0)).await?;
    new_sock.adopt_blocklist(&inner.get_socket());
    let new_addr = new_sock.local_addr()?;
    *inner.socket.write() = new_sock.clone();
    tracing::info!(target: "zudp::engine", old = %old_local, new = %new_addr, "socket rebound");
    Ok(new_sock)
}

/// Send a Ping carrying our session ID to every known peer so they update our address.
async fn ping_all_peers(inner: &Arc<EngineInner>, sock: &RawSocket) {
    let now_ms = now_micros();
    let peers: Vec<Arc<PeerState>> = inner.peers.iter().map(|r| r.value().clone()).collect();
    for peer in peers {
        let Some(ping) = secure_frame(
            inner,
            &peer,
            Frame::Ping {
                echo: now_ms,
                session_id: inner.instance_id,
            },
        ) else {
            continue;
        };
        let Some((datagram, dest)) = inner.outbound(&peer, ping) else {
            continue;
        };
        if let Err(e) = sock.send_to(&datagram, dest).await {
            tracing::warn!(target: "zudp::engine", peer = %peer.addr(), "post-rebind ping failed: {e}");
        }
    }
}

#[allow(clippy::too_many_arguments)]
async fn handle_incoming(
    data: BytesMut,
    from: SocketAddr,
    inner: &Arc<EngineInner>,
    inbound_tx: &mpsc::Sender<(Bytes, PeerAddr, u16)>,
    recv_states: &mut RecvStates,
    retransmit_tasks: &mut JoinSet<()>,
) {
    let frame = match Frame::decode(data) {
        Ok(f) => f,
        Err(e) => {
            tracing::warn!(target: "zudp::engine", peer = %from, "bad frame: {e}");
            return;
        }
    };

    let known_peer = inner
        .peers
        .get(&PeerAddr::Direct(from))
        .map(|r| r.value().clone());
    // On a secured socket only authenticated frames count as signs of life.
    if let Some(peer) = &known_peer
        && !inner.secured()
    {
        peer.mark_seen();
    }

    #[cfg(feature = "security")]
    {
        match frame {
            Frame::Handshake {
                step,
                version,
                payload,
            } => {
                handle_handshake(step, version, &payload, PeerAddr::Direct(from), inner).await;
            }
            Frame::Secure {
                index,
                nonce,
                ciphertext,
            } => {
                let Some((peer, inner_frame)) =
                    open_secure(inner, from, index, nonce, &ciphertext, recv_states).await
                else {
                    return;
                };
                let Some((peer, inner_frame)) = crate::relay::open(inner, peer, inner_frame).await
                else {
                    return;
                };
                // Dispatch under the session's (possibly just migrated) address so a late
                // datagram from the old address does not split its state.
                dispatch_frame(
                    inner_frame,
                    peer.addr(),
                    inner,
                    inbound_tx,
                    recv_states,
                    retransmit_tasks,
                )
                .await;
            }
            Frame::SessionUnknown { index } => {
                on_session_unknown(inner, PeerAddr::Direct(from), index);
            }
            frame => {
                if inner.secured() && !plaintext_allowed(&frame) {
                    tracing::trace!(target: "zudp::engine", peer = %from, "plaintext frame on a secured socket dropped");
                    inner
                        .dropped_unauthenticated
                        .fetch_add(1, Ordering::Relaxed);
                    return;
                }
                dispatch_frame(
                    frame,
                    PeerAddr::Direct(from),
                    inner,
                    inbound_tx,
                    recv_states,
                    retransmit_tasks,
                )
                .await;
            }
        }
        announce_plain_peer(inner, PeerAddr::Direct(from));
    }

    #[cfg(not(feature = "security"))]
    dispatch_frame(
        frame,
        PeerAddr::Direct(from),
        inner,
        inbound_tx,
        recv_states,
        retransmit_tasks,
    )
    .await;

    #[cfg(not(feature = "security"))]
    announce_plain_peer(inner, PeerAddr::Direct(from));
}

// All parameters are genuinely distinct state that the frame dispatch needs to mutate or read;
// grouping them into a context struct would add complexity without architectural benefit.
#[allow(clippy::too_many_arguments, clippy::too_many_lines)]
async fn dispatch_frame(
    frame: Frame,
    from: PeerAddr,
    inner: &Arc<EngineInner>,
    inbound_tx: &mpsc::Sender<(Bytes, PeerAddr, u16)>,
    recv_states: &mut RecvStates,
    retransmit_tasks: &mut JoinSet<()>,
) {
    match frame {
        Frame::Datagram(payload) => {
            let peer = data_peer(inner, from).await;
            peer.rx_bytes
                .fetch_add(payload.len() as u64, std::sync::atomic::Ordering::Relaxed);
            if inbound_tx.try_send((payload, from, 0)).is_err() {
                tracing::warn!(target: "zudp::engine", %from, "inbound channel full — dropping datagram");
            }
        }

        Frame::Stream {
            seq,
            stream_id,
            payload,
        } => {
            let peer = data_peer(inner, from).await;
            ingest_reliable(
                inner,
                &peer,
                from,
                stream_id,
                seq,
                Inbound::Data(payload),
                recv_states,
                inbound_tx,
            )
            .await;
        }

        Frame::Nack { stream_id, seqs } => {
            if let Some(peer) = inner.peers.get(&from).map(|r| r.value().clone()) {
                let to_retransmit = peer.frames_for_retransmit(stream_id, &seqs);
                if !to_retransmit.is_empty() {
                    if retransmit_tasks.len() >= MAX_RETRANSMIT_TASKS {
                        tracing::warn!(
                            target: "zudp::engine",
                            peer = %from, stream_id,
                            cap = MAX_RETRANSMIT_TASKS,
                            "retransmit pool full — dropping NACK batch"
                        );
                    } else {
                        // Spawn so the engine recv loop isn't stalled by CC pacing sleeps.
                        let sock = inner.get_socket();
                        let wires: Vec<(u64, Bytes, SocketAddr)> = to_retransmit
                            .into_iter()
                            .filter_map(|(seq, plain)| {
                                let (wire, dest) =
                                    inner.outbound(&peer, seal(inner, &peer, &plain)?)?;
                                Some((seq, wire, dest))
                            })
                            .collect();
                        retransmit_tasks.spawn(async move {
                            for (seq, wire, dest) in wires {
                                // Extract delay before awaiting — MutexGuard is not Send.
                                let delay = peer.cc.lock().consume(wire.len());
                                if let Some(d) = delay {
                                    tokio::time::sleep(d).await;
                                }
                                tracing::debug!(
                                    target: "zudp::engine",
                                    peer = %from, seq, stream_id,
                                    "retransmit on NACK"
                                );
                                if let Err(e) = sock.send_to(&wire, dest).await {
                                    tracing::warn!(
                                        target: "zudp::engine",
                                        peer = %from,
                                        "retransmit failed: {e}"
                                    );
                                }
                            }
                        });
                    }
                }
            }
        }

        Frame::Ping { echo, session_id } => {
            let Some(peer) = observe_session(from, session_id, inner, recv_states, true) else {
                return;
            };
            let Some(pong) = secure_frame(
                inner,
                &peer,
                Frame::Pong {
                    echo,
                    session_id: inner.instance_id,
                },
            ) else {
                return;
            };
            peer.mark_sent();
            inner.send_to_peer(&peer, pong, "pong").await;
        }

        Frame::Pong { echo, session_id } => {
            if let Some(peer) = observe_session(from, session_id, inner, recv_states, false) {
                // echo carries a µs timestamp; RTT is the round-trip in microseconds.
                let now_us = now_micros();
                if echo <= now_us {
                    let rtt_us = now_us - echo;
                    let pacing_rate = {
                        let mut cc = peer.cc.lock();
                        cc.on_rtt_sample(rtt_us);
                        // Bounded to [10_000, 100_000_000] by the controller.
                        #[allow(clippy::cast_possible_truncation, clippy::cast_sign_loss)]
                        let rate = cc.pacing_rate() as u64;
                        rate
                    };
                    tracing::debug!(
                        target: "zudp::engine",
                        peer = %from,
                        rtt_us,
                        pacing_rate,
                        "RTT sample"
                    );
                }
            }
        }

        Frame::Fragment {
            msg_id,
            frag_idx,
            frag_total,
            seq,
            stream_id,
            payload,
        } => {
            let peer = data_peer(inner, from).await;
            let inbound = Inbound::Fragment {
                msg_id,
                frag_idx,
                frag_total,
                data: payload,
            };
            ingest_reliable(
                inner,
                &peer,
                from,
                stream_id,
                seq,
                inbound,
                recv_states,
                inbound_tx,
            )
            .await;
        }

        Frame::MtuProbe { probe_id, .. } => {
            // Reflect probe_id in a tiny MtuAck — always gets through regardless of path MTU.
            // Plaintext when there is no session: the ack reveals nothing and probes precede it.
            let ack = inner
                .peers
                .get(&from)
                .and_then(|peer| secure_frame(inner, &peer, Frame::MtuAck { probe_id }))
                .unwrap_or_else(|| Frame::MtuAck { probe_id }.encode());
            inner.send_to_addr(from, ack, "mtu ack").await;
        }

        Frame::MtuAck { probe_id } => {
            if let Some(peer) = inner.peers.get(&from).map(|r| r.value().clone())
                && let Some(tx) = peer.probe_acks.lock().remove(&probe_id)
            {
                let _ = tx.send(());
            }
        }

        Frame::Probe { .. } | Frame::Beacon { .. } => {}

        // Relay frames only mean something inside a session with a direct peer, where
        // `relay::open` consumed them; plaintext ones are noise.
        // Path validation is handled where a relayed peer's direct frames are opened.
        #[cfg(feature = "security")]
        Frame::Via { .. }
        | Frame::ViaFrom { .. }
        | Frame::Unreachable { .. }
        | Frame::IntroRequest { .. }
        | Frame::Intro { .. }
        | Frame::PathChallenge { .. }
        | Frame::PathResponse { .. } => {}

        // Handshake and Secure are handled before dispatch_frame is called.
        #[cfg(feature = "security")]
        Frame::Handshake { .. } | Frame::Secure { .. } | Frame::SessionUnknown { .. } => {}
    }
}

/// Match the engine instance id in a Ping or Pong from `from` to a peer.
///
/// Handles first contact, a familiar address, a restarted remote on the same address (its
/// state is reset), and a known remote seen from a new address (migration).  Returns
/// `None` only when `create` is false and nothing matched.
fn observe_session(
    from: PeerAddr,
    session_id: u64,
    inner: &Arc<EngineInner>,
    recv_states: &mut RecvStates,
    create: bool,
) -> Option<Arc<PeerState>> {
    if from.is_relayed() {
        // A relayed session is bounded by its handshake and never moves by instance id; the
        // same remote may well hold a direct session too, which owns its id in `sessions`.
        let peer = inner.peers.get(&from).map(|r| r.value().clone())?;
        peer.set_their_session_id(session_id);
        return Some(peer);
    }
    if let Some(peer) = inner.peers.get(&from).map(|r| r.value().clone()) {
        match peer.their_session_id() {
            Some(known) if known == session_id => return Some(peer),
            None => {
                // A state made by a data frame before the remote's id was known: if that id
                // belongs to a peer at another address, it is that peer arriving from here.
                if let Some(moved) = inner
                    .sessions
                    .get(&session_id)
                    .map(|r| r.value().clone())
                    .filter(|p| !Arc::ptr_eq(p, &peer) && p.addr() != from)
                {
                    migrate_peer(inner, &moved, from, recv_states);
                    return Some(moved);
                }
                peer.set_their_session_id(session_id);
                inner.sessions.insert(session_id, peer.clone());
                return Some(peer);
            }
            Some(_) if inner.secured() => {
                // With Noise, sessions are bounded by handshakes, not by this id.
                peer.set_their_session_id(session_id);
                inner.sessions.insert(session_id, peer.clone());
                return Some(peer);
            }
            Some(previous) => {
                tracing::info!(
                    target: "zudp::engine",
                    peer = %from,
                    previous,
                    current = session_id,
                    "remote restarted — resetting its session"
                );
                let fresh = inner.replace_peer(from);
                fresh.set_their_session_id(session_id);
                fresh.announce();
                inner.emit(PeerEvent::Reset {
                    addr: from,
                    key: None,
                });
                inner.sessions.insert(session_id, fresh.clone());
                return Some(fresh);
            }
        }
    }

    if let Some(peer) = inner.sessions.get(&session_id).map(|r| r.value().clone()) {
        migrate_peer(inner, &peer, from, recv_states);
        return Some(peer);
    }

    if !create {
        return None;
    }
    let peer = inner.get_or_create_peer(from);
    peer.set_their_session_id(session_id);
    inner.sessions.insert(session_id, peer.clone());
    if let PeerAddr::Direct(addr) = from {
        spawn_mtu_probe(inner, addr);
    }
    Some(peer)
}

/// Move `peer` to `to`: re-key the peer table and its receive state, and re-probe the MTU.
///
/// Whatever state sat at `to` (a placeholder made by frames that arrived before we knew who
/// was sending) is discarded; anything it buffered is requested again on the moved state.
fn migrate_peer(
    inner: &Arc<EngineInner>,
    peer: &Arc<PeerState>,
    to: PeerAddr,
    recv_states: &mut RecvStates,
) {
    let old_addr = peer.addr();
    if old_addr == to {
        return;
    }
    tracing::info!(target: "zudp::engine", old = %old_addr, new = %to, "peer address migrated");
    peer.migrate_to(to);
    inner
        .peers
        .remove_if(&old_addr, |_, p| Arc::ptr_eq(p, peer));
    if let Some(placeholder) = inner.peers.insert(to, peer.clone())
        && !Arc::ptr_eq(&placeholder, peer)
    {
        inner.unindex(&placeholder);
    }
    recv_states.retain(|(addr, _), _| *addr != to);
    let moved: Vec<u16> = recv_states
        .keys()
        .filter(|(a, _)| *a == old_addr)
        .map(|(_, s)| *s)
        .collect();
    for sid in moved {
        if let Some(state) = recv_states.remove(&(old_addr, sid)) {
            recv_states.insert((to, sid), state);
        }
    }
    peer.set_effective_mtu(0);
    if let PeerAddr::Direct(addr) = to {
        spawn_mtu_probe(inner, addr);
    }
    inner.emit(PeerEvent::Migrated { from: old_addr, to });
}

/// Fire-and-forget PLPMTUD probe toward `peer_addr`.
///
/// Spawned on first contact and on every migration so the server always knows the usable path
/// MTU — not just the client side (which probes inside `connect()`).
pub(crate) fn spawn_mtu_probe(inner: &Arc<EngineInner>, peer_addr: SocketAddr) {
    let weak = Arc::downgrade(inner);
    let initial_mtu = inner.config.mtu;
    let max_mtu = inner.config.max_mtu;
    let _mtu_probe = tokio::spawn(async move {
        let discovered = crate::mtu::probe(&weak, peer_addr, initial_mtu, max_mtu).await;
        let Some(inner) = weak.upgrade() else { return };
        let Some(discovered) = discovered else {
            tracing::debug!(target: "zudp::mtu", peer = %peer_addr, "no MTU probe answered — keeping configured MTU");
            return;
        };
        if let Some(p) = inner.peers.get(&PeerAddr::Direct(peer_addr)) {
            p.set_effective_mtu(discovered);
        }
        tracing::info!(
            target: "zudp::mtu",
            peer = %peer_addr,
            mtu = discovered,
            "path MTU discovered"
        );
    });
}

/// Encode `frame` and seal it for `peer`; see [`seal`].
fn secure_frame(inner: &EngineInner, peer: &PeerState, frame: Frame) -> Option<Bytes> {
    seal(inner, peer, &frame.encode())
}

/// Wrap an encoded frame for the wire.
///
/// With a Noise session it is encrypted under a fresh nonce (so retransmits never reuse
/// one).  Without a keypair it goes out as-is.  With a keypair but no session yet it returns
/// `None`: nothing but handshake traffic may leave a secured socket in plaintext.
pub(crate) fn seal(inner: &EngineInner, peer: &PeerState, plain: &Bytes) -> Option<Bytes> {
    #[cfg(feature = "security")]
    {
        let guard = peer.channel.read();
        if let Some(channel) = guard.as_ref() {
            return match channel.encrypt(plain) {
                Ok((nonce, ct)) => Some(
                    Frame::Secure {
                        index: channel.remote_index(),
                        nonce,
                        ciphertext: Bytes::from(ct),
                    }
                    .encode(),
                ),
                Err(e) => {
                    tracing::warn!(target: "zudp::engine", peer = %peer.addr(), "encrypt failed: {e}");
                    None
                }
            };
        }
    }
    let _ = peer;
    (!inner.secured()).then(|| plain.clone())
}

/// Recover from tail loss on a reliable stream.
///
/// The receiver only NACKs a gap when a later frame reveals it, so if the last frames of a
/// burst are lost nothing ever asks for them.  Once `stream_id` goes quiet for a probe
/// timeout, re-send its newest frame (TCP's tail loss probe): a duplicate is discarded,
/// a missing tail is delivered, and a hole before it gets requested again as usual.
///
/// Call after recording a reliable send.  At most one task runs per (peer, stream).
pub(crate) fn arm_tail_probe(inner: &Arc<EngineInner>, peer: &Arc<PeerState>, stream_id: u16) {
    if !peer.tail_probe_armed.lock().insert(stream_id) {
        return;
    }
    // Weak: a probe task must not keep the socket bound after the engine stops.
    let weak = Arc::downgrade(inner);
    let peer = peer.clone();
    tokio::spawn(async move {
        let mut probed_seq = 0;
        let mut probes = 0;
        loop {
            let pto = probe_timeout(&peer);
            let newest = peer.newest_sent(stream_id);
            match &newest {
                Some((_, _, sent_at)) if sent_at.elapsed() < pto => {
                    tokio::time::sleep(pto.saturating_sub(sent_at.elapsed())).await;
                }
                Some((seq, frame, _)) if *seq != probed_seq || probes < TAIL_PROBES => {
                    if *seq != probed_seq {
                        probed_seq = *seq;
                        probes = 0;
                    }
                    let Some(inner) = weak.upgrade().filter(|i| i.is_current(&peer)) else {
                        // Engine gone or the peer was reset: these frames belong to nobody now.
                        peer.tail_probe_armed.lock().remove(&stream_id);
                        return;
                    };
                    send_tail_probe(&inner, &peer, frame, stream_id, *seq).await;
                    drop(inner);
                    probes += 1;
                    tokio::time::sleep(pto * 2u32.pow(probes)).await;
                }
                _ => {
                    // Disarm under the lock: a send recorded after this check re-arms.
                    let mut armed = peer.tail_probe_armed.lock();
                    if peer.newest_sent(stream_id).map(|n| n.0) == newest.as_ref().map(|n| n.0) {
                        armed.remove(&stream_id);
                        return;
                    }
                }
            }
        }
    });
}

fn probe_timeout(peer: &PeerState) -> Duration {
    peer.cc.lock().srtt().map_or(TAIL_PROBE_DEFAULT, |srtt| {
        (srtt * 2).clamp(TAIL_PROBE_MIN, TAIL_PROBE_MAX)
    })
}

async fn send_tail_probe(
    inner: &Arc<EngineInner>,
    peer: &PeerState,
    plain_frame: &Bytes,
    stream_id: u16,
    seq: u64,
) {
    let Some(wire) = seal(inner, peer, plain_frame) else {
        return;
    };
    peer.cc.lock().consume(wire.len());
    tracing::debug!(target: "zudp::engine", peer = %peer.addr(), seq, stream_id, "tail loss probe");
    inner.send_to_peer(peer, wire, "tail loss probe").await;
}

async fn send_nack_if_needed(
    nack_seqs: Vec<u64>,
    stream_id: u16,
    to: PeerAddr,
    inner: &Arc<EngineInner>,
) {
    if nack_seqs.is_empty() {
        return;
    }
    tracing::debug!(target: "zudp::engine", peer = %to, count = nack_seqs.len(), stream_id, "sending NACK");
    let nack = Frame::Nack {
        stream_id,
        seqs: nack_seqs,
    };
    let wire = match inner.peers.get(&to).map(|r| r.value().clone()) {
        Some(peer) => secure_frame(inner, &peer, nack),
        None => (!inner.secured()).then(|| nack.encode()),
    };
    let Some(wire) = wire else { return };
    inner.send_to_addr(to, wire, "nack").await;
}

#[allow(clippy::too_many_arguments)]
fn run_background(
    inner: &Arc<EngineInner>,
    recv_states: &mut RecvStates,
    rate_limiter: &mut crate::rate::RateLimiter,
    keepalive_interval: Duration,
    sent_prune_age: Duration,
) {
    let peer_timeout = inner.config.effective_peer_timeout();
    let peers: Vec<Arc<PeerState>> = inner.peers.iter().map(|r| r.value().clone()).collect();

    for peer in peers {
        if peer.since_seen() >= peer_timeout {
            let addr = peer.addr();
            if inner
                .peers
                .remove_if(&addr, |_, p| Arc::ptr_eq(p, &peer))
                .is_some()
            {
                inner.unindex(&peer);
                tracing::info!(target: "zudp::engine", peer = %addr, "peer timed out");
                inner.emit(PeerEvent::Disconnected {
                    addr,
                    reason: DisconnectReason::TimedOut,
                });
            }
            continue;
        }
        #[cfg(feature = "security")]
        crate::relay::tick(inner, &peer);
        // Ping when we have been quiet, and also when they have: a peer that only receives
        // must still hear from us, and we must still hear from it, within the timeout.
        if peer.since_sent() >= keepalive_interval || peer.since_seen() >= keepalive_interval {
            // Spawn each keepalive independently so sequential `send_to` calls
            // don't stall the recv loop when many peers need pinging at once.
            let sock = inner.get_socket();
            let addr = peer.addr();
            let Some((ping, dest)) = ping_for(inner, &peer).and_then(|p| inner.outbound(&peer, p))
            else {
                peer.prune_sent(sent_prune_age);
                continue;
            };
            peer.mark_sent();
            tokio::spawn(async move {
                if let Err(e) = sock.send_to(&ping, dest).await {
                    tracing::warn!(target: "zudp::engine", peer = %addr, "keepalive failed: {e}");
                }
                tracing::trace!(target: "zudp::engine", peer = %addr, "sent keepalive ping");
            });
        }
        peer.prune_sent(sent_prune_age);
    }

    recv_states.retain(|(addr, _), rx| {
        inner
            .peers
            .get(addr)
            .is_some_and(|p| p.generation == rx.generation)
    });
    for ((addr, stream_id), rx) in recv_states.iter_mut() {
        rx.frag.prune(sent_prune_age);
        if rx.recv.prune_gap(sent_prune_age) {
            tracing::warn!(
                target: "zudp::engine",
                peer = %addr,
                stream_id,
                "gap timeout — missing seq discarded, stream resyncing"
            );
        }
    }

    rate_limiter.tick_prune();

    #[cfg(feature = "relay")]
    if let Some(node) = &inner.relay {
        node.prune();
    }
    #[cfg(feature = "security")]
    crate::handshake::expire(inner);
    #[cfg(feature = "security")]
    {
        inner.hint_limiter.lock().tick_prune();
        inner.relayed_hs_limiter.lock().tick_prune();
    }
}

fn now_micros() -> u64 {
    std::time::SystemTime::now()
        .duration_since(std::time::UNIX_EPOCH)
        .map_or(0u64, |d| u64::try_from(d.as_micros()).unwrap_or(u64::MAX))
}

/// Feed one handshake message to the driver, answer it, and install a finished session.
#[cfg(feature = "security")]
pub(crate) async fn handle_handshake(
    step: u8,
    version: u8,
    payload: &Bytes,
    from: PeerAddr,
    inner: &Arc<EngineInner>,
) {
    let (reply, done) = crate::handshake::on_message(inner, from, step, version, payload);
    if let Some((established, waiter)) = done {
        let peer = install_session(inner, from, established);
        if let Some(waiter) = waiter {
            let _ = waiter.send(crate::handshake::Outcome::Ready {
                generation: peer.generation,
            });
        }
    }
    if let Some(reply) = reply {
        inner.send_handshake(from, reply, "handshake").await;
    }
}

/// Make `established` the live session for `addr`.
#[cfg(feature = "security")]
fn install_session(
    inner: &Arc<EngineInner>,
    addr: PeerAddr,
    established: crate::handshake::Established,
) -> Arc<PeerState> {
    let had_session = inner.peers.get(&addr).is_some_and(|p| p.has_channel());
    let key = established.remote_static;
    if !inner.peers.contains_key(&addr) {
        limit_sessions(inner, addr);
    }
    // A new session starts on the relay; a punched path belonged to the old one.
    let was_punched = match addr {
        PeerAddr::Relayed { relay, .. } => inner
            .peers
            .get(&addr)
            .and_then(|p| p.direct_path())
            .map(|_| relay),
        PeerAddr::Direct(_) => None,
    };
    let peer = inner.replace_peer(addr);
    if let Some(relay) = was_punched {
        inner.emit(PeerEvent::PathChanged {
            addr,
            path: crate::Path::Relayed(relay),
        });
    }
    let local_index = established.channel.local_index();
    peer.set_channel(
        established.channel,
        key,
        established.finish_wire,
        established.initiator,
    );
    inner.by_index.insert(local_index, peer.clone());
    match (addr, key) {
        (PeerAddr::Direct(_), Some(key)) => {
            inner.by_key.insert(key, peer.clone());
        }
        (PeerAddr::Relayed { .. }, _) => crate::relay::auto_punch(inner, &peer),
        (PeerAddr::Direct(_), None) => {}
    }
    peer.announce();
    inner.emit(if had_session {
        PeerEvent::Reset { addr, key }
    } else {
        PeerEvent::Connected { addr, key }
    });
    tracing::info!(target: "zudp::engine", peer = %addr, initiator = established.initiator, "Noise XX handshake complete");
    peer
}

/// Bound what one source can make us hold before a new session at `addr` is installed:
/// sessions through one relay may take at most half of `max_peers`, and a relay node keeps
/// at most `max_sessions_per_ip` sessions per source network and `max_sessions_per_prefix`
/// per IPv6 /48.  Over a bound, that source's least recently seen session makes room.  A
/// relay node whose table is full makes room in its heaviest prefix, never at the expense
/// of a lighter one, so a flood from one allocation only ever displaces itself.
#[cfg(feature = "security")]
fn limit_sessions(inner: &EngineInner, addr: PeerAddr) {
    match addr {
        PeerAddr::Relayed { relay, .. } => {
            let cap = inner.config.max_peers / 2;
            inner.evict_lru(
                addr,
                cap,
                |p| matches!(p, PeerAddr::Relayed { relay: r, .. } if *r == relay),
            );
        }
        PeerAddr::Direct(direct) => {
            #[cfg(not(feature = "relay"))]
            let _ = direct;
            #[cfg(feature = "relay")]
            if let Some(node) = &inner.relay {
                use crate::rate::{network_of, prefix_of};
                let (per_network, per_prefix) = node.session_caps();
                let network = network_of(direct.ip());
                let prefix = prefix_of(direct.ip());
                let in_network =
                    |p: &PeerAddr| p.direct().is_some_and(|a| network_of(a.ip()) == network);
                let in_prefix =
                    |p: &PeerAddr| p.direct().is_some_and(|a| prefix_of(a.ip()) == prefix);
                let mut evicted = inner.evict_lru(addr, per_network, in_network);
                evicted |= inner.evict_lru(addr, per_prefix, in_prefix);
                if inner.peers.len() >= inner.config.max_peers {
                    let heaviest = crate::rate::heaviest_prefixes(
                        inner
                            .peers
                            .iter()
                            .filter_map(|r| r.key().direct())
                            .map(|a| prefix_of(a.ip())),
                        prefix,
                    );
                    let in_heaviest = |p: &PeerAddr| {
                        p.direct()
                            .is_some_and(|a| heaviest.contains(&prefix_of(a.ip())))
                    };
                    evicted |= inner.evict_lru(addr, 1, in_heaviest);
                }
                if evicted {
                    node.session_evicted();
                }
            }
        }
    }
}

/// A keepalive Ping for `peer`, encrypted when a session exists.
pub(crate) fn ping_for(inner: &EngineInner, peer: &PeerState) -> Option<Bytes> {
    secure_frame(
        inner,
        peer,
        Frame::Ping {
            echo: now_micros(),
            session_id: inner.instance_id,
        },
    )
}

/// `true` when the datagram is a Noise handshake frame (type byte is last on the wire).
fn is_handshake(data: &[u8]) -> bool {
    #[cfg(feature = "security")]
    {
        data.last() == Some(&crate::frame::TYPE_HANDSHAKE)
    }
    #[cfg(not(feature = "security"))]
    {
        let _ = data;
        false
    }
}

/// Retry delivery on every stream after the application freed inbound capacity.
fn drain_all_streams(
    inner: &Arc<EngineInner>,
    recv_states: &mut RecvStates,
    inbound_tx: &mpsc::Sender<(Bytes, PeerAddr, u16)>,
) {
    for (&(addr, stream_id), rx) in recv_states.iter_mut() {
        let Some(peer) = inner
            .peers
            .get(&addr)
            .map(|r| r.value().clone())
            .filter(|p| p.generation == rx.generation)
        else {
            continue;
        };
        deliver_ready(
            &mut rx.recv,
            &mut rx.frag,
            &peer,
            addr,
            stream_id,
            inbound_tx,
            &inner.delivery_stalled,
        );
    }
}
/// Buffer one reliable frame, NACK what is missing before it, and deliver whatever is now
/// in order.
#[allow(clippy::too_many_arguments)]
async fn ingest_reliable(
    inner: &Arc<EngineInner>,
    peer: &PeerState,
    from: PeerAddr,
    stream_id: u16,
    seq: u64,
    inbound: Inbound,
    recv_states: &mut RecvStates,
    inbound_tx: &mpsc::Sender<(Bytes, PeerAddr, u16)>,
) {
    let window = inner.config.sent_buffer_frames.max(MIN_RECV_WINDOW);
    let fresh = || StreamRx {
        generation: peer.generation,
        recv: RecvState::new(window),
        frag: FragAssembler::new(),
    };
    let rx = recv_states.entry((from, stream_id)).or_insert_with(fresh);
    if rx.generation != peer.generation {
        *rx = fresh();
    }
    // Until a plain-mode peer's session is known, its frames may belong to a previous
    // incarnation of the remote, so only accept the very next one: nothing stale gets
    // buffered and nothing gets NACKed from a sender that is about to reset.
    let allow_reorder = inner.secured() || peer.their_session_id().is_some();
    let nack_seqs = rx.recv.ingest(seq, inbound, allow_reorder);
    deliver_ready(
        &mut rx.recv,
        &mut rx.frag,
        peer,
        from,
        stream_id,
        inbound_tx,
        &inner.delivery_stalled,
    );
    send_nack_if_needed(nack_seqs, stream_id, from, inner).await;
}

/// Hand in-order payloads to the application while the inbound channel has room.
///
/// A payload is only taken out of the receive buffer once a channel slot is reserved for
/// it, so a slow application stalls the stream instead of silently losing reliable data.
fn deliver_ready(
    recv: &mut RecvState,
    frag: &mut FragAssembler,
    peer: &PeerState,
    from: PeerAddr,
    stream_id: u16,
    inbound_tx: &mpsc::Sender<(Bytes, PeerAddr, u16)>,
    stalled: &AtomicBool,
) {
    while recv.has_ready() {
        let Ok(permit) = inbound_tx.try_reserve() else {
            tracing::trace!(target: "zudp::engine", %from, stream_id, "inbound channel full — holding stream");
            stalled.store(true, Ordering::Relaxed);
            return;
        };
        let complete = match recv.pop_ready() {
            Some(Inbound::Data(bytes)) => Some(bytes),
            Some(Inbound::Fragment {
                msg_id,
                frag_idx,
                frag_total,
                data,
            }) => frag.insert(msg_id, frag_idx, frag_total, data),
            None => None,
        };
        if let Some(bytes) = complete {
            peer.rx_bytes
                .fetch_add(bytes.len() as u64, Ordering::Relaxed);
            permit.send((bytes, from, stream_id));
        }
    }
}

/// The peer a data frame from `from` belongs to.
///
/// A peer first seen through a data frame is pinged at once so its engine instance id, and
/// with it restart and migration detection, is known within one round trip.
async fn data_peer(inner: &Arc<EngineInner>, from: PeerAddr) -> Arc<PeerState> {
    let (peer, created) = inner.peer_for_incoming(from);
    if created && let Some(ping) = ping_for(inner, &peer) {
        inner.send_to_peer(&peer, ping, "first-contact ping").await;
    }
    peer
}

/// Frames a secured socket accepts outside a Noise session: handshakes, discovery and path
/// MTU probing.  Everything else must arrive inside a `Secure` frame.
#[cfg(feature = "security")]
fn plaintext_allowed(frame: &Frame) -> bool {
    matches!(
        frame,
        Frame::Handshake { .. }
            | Frame::Secure { .. }
            | Frame::SessionUnknown { .. }
            | Frame::Probe { .. }
            | Frame::Beacon { .. }
            | Frame::MtuProbe { .. }
            | Frame::MtuAck { .. }
    )
}

/// Without security a session starts with the first frame heard from an address.
fn announce_plain_peer(inner: &EngineInner, from: PeerAddr) {
    if inner.secured() {
        return;
    }
    if let Some(peer) = inner.peers.get(&from).map(|r| r.value().clone())
        && peer.announce()
    {
        inner.emit(PeerEvent::Connected {
            addr: from,
            key: None,
        });
    }
}

/// Authenticate a Secure frame and find its session by index.
///
/// A frame that decrypts and carries a newer nonce than any before moves the session to
/// `from` (roaming).  A frame for an index we do not know earns the sender a rate-limited
/// `SessionUnknown` so it can re-handshake instead of talking into the void.
#[cfg(feature = "security")]
async fn open_secure(
    inner: &Arc<EngineInner>,
    from: SocketAddr,
    index: u64,
    nonce: u64,
    ciphertext: &Bytes,
    recv_states: &mut RecvStates,
) -> Option<(Arc<PeerState>, Frame)> {
    let Some(peer) = inner.by_index.get(&index).map(|r| r.value().clone()) else {
        if !crate::handshake::is_pending_index(inner, index) {
            send_session_unknown(inner, PeerAddr::Direct(from), index).await;
        }
        return None;
    };
    // A relayed session only moves onto a direct path when punching is on.
    let relayed = peer.addr().is_relayed();
    if relayed && !inner.config.punching {
        return None;
    }
    let plain = decrypt(&peer, nonce, ciphertext)?;
    let newest = peer.advance_nonce(nonce);
    peer.mark_seen();
    peer.mark_confirmed();
    if relayed {
        let (peer, frame) = inner_frame(peer, &plain)?;
        return crate::relay::on_direct(inner, peer, from, newest, frame).await;
    }
    if peer.addr() != from && newest {
        migrate_peer(inner, &peer, PeerAddr::Direct(from), recv_states);
    }
    inner_frame(peer, &plain)
}

/// Decrypt a Secure frame in `peer`'s session; `None` when it does not authenticate or
/// is a replay.
#[cfg(feature = "security")]
pub(crate) fn decrypt(peer: &PeerState, nonce: u64, ciphertext: &[u8]) -> Option<Vec<u8>> {
    let guard = peer.channel.read();
    match guard.as_ref()?.decrypt(nonce, ciphertext) {
        Ok(plain) => Some(plain),
        Err(e) => {
            tracing::debug!(target: "zudp::engine", peer = %peer.addr(), "decrypt failed: {e}");
            None
        }
    }
}

/// Decode the plaintext of a Secure frame.  Session-layer frames never travel encrypted,
/// so one found inside is dropped.
#[cfg(feature = "security")]
pub(crate) fn inner_frame(peer: Arc<PeerState>, plain: &[u8]) -> Option<(Arc<PeerState>, Frame)> {
    match Frame::decode(BytesMut::from(plain)) {
        Ok(Frame::Handshake { .. } | Frame::Secure { .. } | Frame::SessionUnknown { .. }) => None,
        Ok(frame) => Some((peer, frame)),
        Err(e) => {
            tracing::warn!(target: "zudp::engine", peer = %peer.addr(), "inner frame decode failed: {e}");
            None
        }
    }
}

/// `true` when `addr` is the punched direct path of a relayed session.
fn is_punched(inner: &EngineInner, addr: SocketAddr) -> bool {
    #[cfg(feature = "security")]
    {
        inner.punched.contains_key(&addr)
    }
    #[cfg(not(feature = "security"))]
    {
        let _ = (inner, addr);
        false
    }
}

#[cfg(feature = "security")]
pub(crate) async fn send_session_unknown(inner: &Arc<EngineInner>, to: PeerAddr, index: u64) {
    // Per source network (an IPv6 /64) for direct peers.  Relayed peers get their own budget each, so peers
    // behind one relay do not starve each other, under an aggregate per relay, because the
    // relay chooses the keys.  Port 0 / 1 keep the direct and the aggregate keys apart.
    {
        let mut limiter = inner.hint_limiter.lock();
        let allowed = match to {
            PeerAddr::Direct(addr) => limiter.allow(PeerAddr::Direct(SocketAddr::new(
                crate::rate::network_of(addr.ip()),
                0,
            ))),
            PeerAddr::Relayed { relay, .. } => {
                limiter.allow(PeerAddr::Direct(SocketAddr::new(relay.ip(), 1))) && limiter.allow(to)
            }
        };
        if !allowed {
            return;
        }
    }
    tracing::debug!(target: "zudp::engine", peer = %to, index, "secure frame for an unknown session");
    let wire = Frame::SessionUnknown { index }.encode();
    inner.send_to_addr(to, wire, "session-unknown").await;
}

/// The remote lost the session we initiated (it restarted): handshake again in the
/// background.  The hint is unauthenticated, so it only ever starts a handshake — the old
/// session stays until a new one replaces it.
#[cfg(feature = "security")]
pub(crate) fn on_session_unknown(inner: &Arc<EngineInner>, from: PeerAddr, index: u64) {
    // The session at `from` itself, or a relayed peer we punched through that answers from
    // the same address — both can exist at once; the index tells which one was lost.
    let direct = inner.peers.get(&from).map(|r| r.value().clone());
    let punched = from
        .direct()
        .and_then(|addr| inner.punched.get(&addr).map(|r| r.value().clone()));
    let Some(peer) = [direct, punched]
        .into_iter()
        .flatten()
        .find(|p| p.remote_index() == Some(index))
    else {
        return;
    };
    let from = peer.addr();
    if !peer.is_initiator()
        || peer.remote_index() != Some(index)
        || !peer.begin_rehandshake(inner.config.handshake_timeout)
    {
        return;
    }
    tracing::info!(target: "zudp::engine", peer = %from, "remote lost our session — re-handshaking");
    let weak = Arc::downgrade(inner);
    tokio::spawn(async move {
        if let Err(e) = crate::handshake::initiate(&weak, from).await {
            tracing::warn!(target: "zudp::engine", peer = %from, "re-handshake failed: {e}");
        }
    });
}