subc-daemon 0.49.3

Embeddable subc daemon: bootstrap, module supervision, and opaque-byte splice routing.
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
//! Every supervised process this daemon has spawned and not yet reaped, and the
//! daemon's own end-of-life stop for them.
//!
//! Each supervised module runs in its own process group (see
//! `spawn_child_in_slot`). That keeps a service manager's process-group kill
//! from reaching modules before they see EOF on their control connection, but
//! it also means nothing outside this daemon will end a child that outlives it.
//! A `protocol: "none"` child (the NATS server) has no control connection and
//! never sees EOF at all; left alone it would survive the daemon as an orphan
//! and the next daemon would start a second one that fights it for its port.
//! So the daemon ends its own children on the announced-shutdown path, and this
//! roster is how it finds them: the child handles themselves are owned by the
//! per-module supervisor tasks, which keep reaping them while shutdown runs.

use std::{
    collections::HashMap,
    path::PathBuf,
    sync::{
        atomic::{AtomicBool, AtomicU64, Ordering},
        Arc, Mutex, MutexGuard, OnceLock,
    },
    time::Duration,
};

use subc_control::ModuleProtocol;
use tracing::warn;

use crate::live_children::{ExecutableIdentity, LiveChild};

/// One live supervised process.
///
/// Signalled only by the Unix shutdown stop; Windows child lifetime is a
/// job-object concern, so there the entry only feeds the live-children record.
#[cfg_attr(not(unix), allow(dead_code))]
#[derive(Debug, Clone)]
pub(crate) struct RosterEntry {
    pub(crate) module_id: String,
    pub(crate) pid: u32,
    pub(crate) protocol: ModuleProtocol,
    /// Kernel start time where the platform exposes one, used to refuse a
    /// signal to a different process that has reused a reaped child's pid.
    pub(crate) start_time: Option<u64>,
    /// What the live-children record says about this process, for the next
    /// daemon's orphan sweep if this one dies without stopping it.
    recorded: RecordedIdentity,
    /// The module's resolved drain budget, shared with its supervisor so a
    /// configuration rescan that changes it is seen at shutdown.
    drain_budget: Arc<Mutex<Duration>>,
}

/// The identity facts the live-children record keeps for one process beyond
/// its module id, pid and protocol. See `live_children::LiveChild`.
#[derive(Debug, Clone, Default)]
pub(crate) struct RecordedIdentity {
    pub(crate) start_time: Option<u64>,
    pub(crate) executable: Option<ExecutableIdentity>,
    pub(crate) cgroup_name: Option<String>,
    pub(crate) container: Option<crate::live_children::ContainerIdentity>,
    #[cfg(unix)]
    pub(crate) carried: Option<std::sync::Weak<crate::supervise::containment::CarriedContainment>>,
    /// The owning subtree for shutdown tree kills; not persisted in the record.
    #[cfg(target_os = "linux")]
    pub(crate) cgroup_placement: Option<subc_cgroup::Placement>,
}

/// Set once, when the daemon begins its announced shutdown, and never cleared.
///
/// Shared by the roster (which refuses spawns once it is set), every module's
/// terminal ring (which records any exit after it as `daemon_shutdown`) and the
/// control handler's capability-requirement log (which reports requirement
/// changes after it at INFO with `cause=daemon_shutdown`), so the reap path,
/// the spawn path, the record and the log all read the same flag.
#[derive(Debug, Clone, Default)]
pub(crate) struct DaemonShutdownFlag(Arc<AtomicBool>);

impl DaemonShutdownFlag {
    pub(crate) fn is_set(&self) -> bool {
        self.0.load(Ordering::SeqCst)
    }

    #[cfg(unix)]
    fn set(&self) {
        self.0.store(true, Ordering::SeqCst);
    }

    /// Set the flag without a roster, so a test can stand in for the start of
    /// daemon shutdown on any platform.
    #[cfg(test)]
    pub(crate) fn set_for_test(&self) {
        self.0.store(true, Ordering::SeqCst);
    }
}

#[derive(Debug, Default)]
struct RosterInner {
    #[cfg(target_os = "linux")]
    cgroup_placement: Mutex<Option<subc_cgroup::Placement>>,
    next_key: AtomicU64,
    closed: DaemonShutdownFlag,
    closed_notify: tokio::sync::Notify,
    live: Mutex<HashMap<u64, RosterEntry>>,
    record_write: Mutex<()>,
    /// Where the live-children record is written; unset means no record, the
    /// default for an in-process daemon (see `BootstrapConfig`).
    record_path: OnceLock<PathBuf>,
    recovery_block: OnceLock<String>,
    pending_admissions: Mutex<HashMap<String, (u32, String)>>,
    unquarantined_record: std::sync::atomic::AtomicBool,
    retained: Mutex<Vec<(LiveChild, crate::live_children::RetainedReason)>>,
    #[cfg(target_os = "macos")]
    privacy_trampoline: OnceLock<Result<PathBuf, String>>,
    /// Test-only: called before the record is serialized and written, so a
    /// test can block there to stand in for a slow write.
    #[cfg(test)]
    record_write_hook: Mutex<Option<RecordWriteHook>>,
    #[cfg(all(test, unix))]
    admission_timeout: Mutex<Option<Duration>>,
    #[cfg(all(test, unix))]
    admission_hook: Mutex<Option<AdmissionHook>>,
    #[cfg(test)]
    fail_adoption: AtomicBool,
    #[cfg(all(test, unix))]
    placement_pause: Mutex<HashMap<String, subc_os::admission_gate::PlacementPause>>,
}

#[cfg(test)]
type RecordWriteFn = dyn Fn(&HashMap<u64, RosterEntry>) + Send + Sync;

#[cfg(test)]
#[derive(Clone)]
struct RecordWriteHook(Arc<RecordWriteFn>);

#[cfg(all(test, unix))]
type AdmissionHookFn = dyn Fn(&str, &'static str, Option<u32>) -> bool + Send + Sync;
#[cfg(all(test, unix))]
#[derive(Clone)]
struct AdmissionHook(Arc<AdmissionHookFn>);
#[cfg(all(test, unix))]
impl std::fmt::Debug for AdmissionHook {
    fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        formatter.write_str("AdmissionHook")
    }
}

#[cfg(test)]
impl std::fmt::Debug for RecordWriteHook {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.write_str("RecordWriteHook")
    }
}

impl RosterInner {
    /// Run roster work that may wait on the record write without holding up
    /// the rest of the async runtime.
    ///
    /// Rewriting the live-children record syncs the new file and then its
    /// directory to disk, and on a busy disk each sync can take seconds. The
    /// roster is called synchronously from async tasks: a module's supervisor
    /// admits each child it spawns and releases each child it reaps. Blocking
    /// a tokio worker thread for that long stalls more than the calling task.
    /// A task woken by I/O or a timer normally runs on the worker that was
    /// polling the runtime's I/O and timer driver, and while that worker is
    /// blocked no other worker takes the driver over, so every timer and
    /// socket in the daemon waits. That includes the authentication handshake
    /// of the module just spawned, and the Rust SDK gives up on that handshake
    /// after 2 seconds, so a restart on a busy machine crashed its own
    /// replacement.
    ///
    /// The wait covers the record-write gate as well as the write. Readers of
    /// the live roster do not take that gate, so shutdown can still poll them.
    ///
    /// On a multi-threaded runtime `block_in_place` hands this worker's queued
    /// tasks to another thread for the duration. A current-thread runtime
    /// (some in-process test daemons) cannot do that, so the work runs inline
    /// there. Without a record path nothing slow happens under the lock, and
    /// the work also runs inline.
    fn without_stalling_runtime<T>(&self, work: impl FnOnce() -> T) -> T {
        let multi_threaded = tokio::runtime::Handle::try_current().is_ok_and(|handle| {
            handle.runtime_flavor() == tokio::runtime::RuntimeFlavor::MultiThread
        });
        if multi_threaded && self.record_path.get().is_some() {
            tokio::task::block_in_place(work)
        } else {
            work()
        }
    }

    /// Rewrite a snapshot taken after acquiring the record-write gate. Taking
    /// each snapshot inside that gate prevents an older snapshot from replacing
    /// a newer record, without locking the live roster across disk syncs.
    fn write_record(&self, live: &HashMap<u64, RosterEntry>) -> std::io::Result<()> {
        if self.unquarantined_record.load(Ordering::Acquire) {
            return Err(std::io::Error::other(
                "unusable recovery record could not be quarantined; refusing to overwrite it",
            ));
        }
        let Some(path) = self.record_path.get() else {
            return Ok(());
        };
        #[cfg(test)]
        {
            let hook = lock(&self.record_write_hook).clone();
            if let Some(hook) = hook {
                (hook.0)(live);
            }
        }
        let mut entries: Vec<(&u64, &RosterEntry)> = live.iter().collect();
        entries.sort_by_key(|(key, _)| **key);
        let children: Vec<LiveChild> = entries
            .into_iter()
            .map(|(_, entry)| LiveChild {
                module_id: entry.module_id.clone(),
                pid: entry.pid,
                protocol: entry.protocol,
                start_time: entry.recorded.start_time,
                executable: entry.recorded.executable,
                cgroup_name: entry.recorded.cgroup_name.clone(),
                container: entry.recorded.container.clone(),
            })
            .collect();
        let children: Vec<_> = children
            .into_iter()
            .chain(lock(&self.retained).iter().map(|(entry, _)| entry.clone()))
            .collect();
        let reasons = lock(&self.retained)
            .iter()
            .map(|(entry, reason)| (entry.pid, reason.clone()))
            .collect();
        if let Err(error) =
            crate::live_children::write_record_with_reasons(path, &children, reasons)
        {
            warn!(
                path = %path.display(),
                %error,
                "could not rewrite the live-children record; a crash now could leave orphans the next boot cannot find"
            );
            return Err(error);
        }
        Ok(())
    }

    fn recheck_retained(&self) {
        self.without_stalling_runtime(|| self.recheck_retained_now());
    }

    fn recheck_retained_now(&self) {
        // Serialize pruning with admits and releases, but leave the live roster
        // readable while probing retained processes and syncing the record.
        let _write = lock(&self.record_write);
        let changed = {
            let mut retained = lock(&self.retained);
            let before = retained.len();
            #[cfg(unix)]
            let mut updated = false;
            #[cfg(not(unix))]
            let updated = false;
            retained.retain_mut(|(entry, reason)| {
                #[cfg(not(unix))]
                let _ = reason;
                #[cfg(unix)]
                {
                    if entry.container.is_some() {
                        return match crate::live_children::recheck_container(entry,
                            #[cfg(target_os = "linux")]
                            lock(&self.cgroup_placement).as_ref(),
                        ) {
                            Some(next) => {
                                if reason.blocks_fresh_copy() && !next.blocks_fresh_copy() {
                                    tracing::info!(module_id = %entry.module_id, container = ?entry.container,
                                        "container survivors are gone; fresh starts are no longer blocked");
                                }
                                updated |= *reason != next;
                                *reason = next;
                                true
                            }
                            None => {
                                if reason.blocks_fresh_copy() {
                                    tracing::info!(module_id = %entry.module_id, container = ?entry.container,
                                        "container survivors are gone; fresh starts are no longer blocked");
                                }
                                false
                            }
                        };
                    }
                    let Some(pid) = i32::try_from(entry.pid)
                        .ok()
                        .and_then(rustix::process::Pid::from_raw)
                    else {
                        return true;
                    };
                    // Only signal-zero ESRCH permits dropping an unidentifiable
                    // previous instance. Permission and other errors keep it blocked.
                    rustix::process::test_kill_process(pid) != Err(rustix::io::Errno::SRCH)
                }
                #[cfg(windows)]
                {
                    match subc_os::Process::open(entry.pid) {
                        Ok(None) => false,
                        Ok(Some(process)) => {
                            !matches!(process.wait_for_exit(Duration::ZERO), Ok(true))
                        }
                        Err(_) => true,
                    }
                }
                #[cfg(not(any(unix, windows)))]
                {
                    let _ = entry;
                    true
                }
            });
            retained.len() != before || updated
        };
        if changed {
            let live = lock(&self.live).clone();
            let _ = self.write_record(&live);
        }
    }
}

/// Shared by every clone of one `Supervisor` and every module task it starts.
/// A module task's copy also carries that module's drain budget, which every
/// process it spawns is admitted with.
#[derive(Debug, Clone)]
pub(crate) struct ChildRoster {
    inner: Arc<RosterInner>,
    drain_budget: Arc<Mutex<Duration>>,
    pub(crate) containment: crate::supervise::containment::Slot,
    pub(crate) counters: crate::observability::DaemonCounters,
}

impl Default for ChildRoster {
    fn default() -> Self {
        Self {
            inner: Arc::default(),
            drain_budget: Arc::new(Mutex::new(crate::supervise::DEFAULT_DRAIN_TIMEOUT)),
            containment: Arc::default(),
            counters: crate::observability::DaemonCounters::default(),
        }
    }
}

/// A cached diagnostic view of containment; it does not own the resource.
#[derive(Debug, Clone)]
pub(crate) struct RetainedView(std::sync::Weak<RosterInner>);

impl RetainedView {
    pub(crate) fn has_carried_container(&self, module_id: &str) -> bool {
        let Some(inner) = self.0.upgrade() else {
            return false;
        };
        let live = lock(&inner.live);
        live.values()
            .any(|entry| entry.module_id == module_id && entry.recorded.container.is_some())
    }

    pub(crate) fn recovery_block(&self) -> Option<String> {
        self.0.upgrade()?.recovery_block.get().cloned()
    }

    pub(crate) fn waiting_detail(&self, module_id: &str) -> Option<(u32, String)> {
        let inner = self.0.upgrade()?;
        if let Some((pid, reason)) = lock(&inner.pending_admissions).get(module_id) {
            return Some((*pid, reason.clone()));
        }
        let retained = lock(&inner.retained);
        retained
            .iter()
            .find(|(entry, reason)| {
                entry.module_id == module_id
                    && entry.container.is_some()
                    && reason.blocks_fresh_copy()
            })
            .map(|(entry, reason)| (entry.pid, reason.waiting_message(entry.pid)))
    }
}

/// Holds a child's roster entry. Dropped when the child is reaped, or when its
/// handle is dropped (which kills it), so the roster never outlives the pid.
#[derive(Debug)]
pub(crate) struct RosterGuard {
    inner: Arc<RosterInner>,
    key: u64,
}

impl RosterGuard {
    /// Remove a cancelled, proven-never-executed attempt only after the new
    /// record is durable. An error requires quarantining the recovery store.
    #[cfg(unix)]
    pub(crate) fn release_durably(self) -> std::io::Result<()> {
        self.inner.without_stalling_runtime(|| {
            let _write = lock(&self.inner.record_write);
            let snapshot = {
                let mut live = lock(&self.inner.live);
                live.remove(&self.key);
                live.clone()
            };
            self.inner.write_record(&snapshot)
        })
    }
    #[cfg(unix)]
    pub(crate) fn retain(self, reason: crate::live_children::RetainedReason) {
        self.inner.without_stalling_runtime(|| {
            let _write = lock(&self.inner.record_write);
            let Some(entry) = lock(&self.inner.live).remove(&self.key) else {
                return;
            };
            let recorded = LiveChild {
                module_id: entry.module_id,
                pid: entry.pid,
                protocol: entry.protocol,
                start_time: entry.recorded.start_time,
                executable: entry.recorded.executable,
                cgroup_name: entry.recorded.cgroup_name,
                container: entry.recorded.container,
            };
            if reason.blocks_fresh_copy() {
                tracing::error!(module_id = %recorded.module_id, container = ?recorded.container,
                    reason = %reason.waiting_message(recorded.pid),
                    "container survivors block a fresh instance; retaining the container record");
            }
            lock(&self.inner.retained).push((recorded, reason));
            let live = lock(&self.inner.live).clone();
            let _ = self.inner.write_record(&live);
        });
    }
}

#[derive(Debug)]
pub(crate) struct AdmissionError {
    source: std::io::Error,
    guard: RosterGuard,
}

impl AdmissionError {
    /// Transfer ownership to cleanup that may have to retain an unresolved
    /// container. Dropping the error itself would discard that recovery guard.
    pub(crate) fn into_parts(self) -> (std::io::Error, RosterGuard) {
        (self.source, self.guard)
    }

    /// Keep shutdown ownership until the caller has killed and reaped the
    /// rejected child. A failed record write is not proof that the child exited.
    pub(crate) fn with_cleanup<T>(self, cleanup: impl FnOnce(std::io::Error) -> T) -> T {
        let result = cleanup(self.source);
        drop(self.guard);
        result
    }
}

impl Drop for RosterGuard {
    fn drop(&mut self) {
        self.inner.without_stalling_runtime(|| {
            let _write = lock(&self.inner.record_write);
            let snapshot = {
                let mut live = lock(&self.inner.live);
                live.remove(&self.key).map(|_| live.clone())
            };
            if let Some(snapshot) = snapshot {
                let _ = self.inner.write_record(&snapshot);
            }
        });
    }
}

#[cfg(unix)]
impl RosterGuard {
    /// Publish the committed executable after exec. A carried child is recorded
    /// before its exec (it waits in the admission barrier), so that early record
    /// must not persist the daemon's pre-exec image, nor the macOS privacy
    /// trampoline before it has replaced itself with the configured module.
    pub(crate) fn confirm_executable(
        &self,
        observation: subc_os::Observation,
    ) -> std::io::Result<()> {
        self.inner.without_stalling_runtime(|| {
            let _write = lock(&self.inner.record_write);
            let snapshot = {
                let mut live = lock(&self.inner.live);
                if let Some(entry) = live.get_mut(&self.key) {
                    entry.recorded.start_time = Some(observation.start_time);
                    entry.recorded.executable =
                        observation.executable.map(ExecutableIdentity::from);
                    Some(live.clone())
                } else {
                    None
                }
            };
            if let Some(snapshot) = snapshot {
                self.inner.write_record(&snapshot)?;
            }
            Ok(())
        })
    }
}

impl ChildRoster {
    #[cfg(target_os = "linux")]
    pub(crate) fn set_cgroup_placement(&self, placement: Option<subc_cgroup::Placement>) {
        *lock(&self.inner.cgroup_placement) = placement;
    }

    pub(crate) fn retained_view(&self) -> RetainedView {
        RetainedView(Arc::downgrade(&self.inner))
    }
    #[cfg(target_os = "macos")]
    pub(crate) fn set_privacy_trampoline(&self, result: Result<PathBuf, String>) {
        let _ = self.inner.privacy_trampoline.set(result);
    }

    #[cfg(target_os = "macos")]
    pub(crate) fn privacy_trampoline(&self) -> Result<PathBuf, String> {
        self.inner
            .privacy_trampoline
            .get()
            .cloned()
            .unwrap_or_else(|| Err("no privacy trampoline configured".to_string()))
    }
    /// The same roster, admitting children under one module's drain budget.
    pub(crate) fn for_module(&self, drain_budget: Arc<Mutex<Duration>>) -> Self {
        Self {
            inner: Arc::clone(&self.inner),
            drain_budget,
            containment: Arc::default(),
            counters: self.counters.clone(),
        }
    }

    /// True once daemon shutdown has begun. A spawn after this point would
    /// create a child the shutdown stop may already have finished looking for,
    /// so spawning refuses instead (the supervisor would otherwise restart each
    /// module as it exits on EOF).
    pub(crate) fn is_closed(&self) -> bool {
        self.inner.closed.is_set()
    }

    pub(crate) async fn closed(&self) {
        loop {
            let changed = self.inner.closed_notify.notified();
            tokio::pin!(changed);
            changed.as_mut().enable();
            if self.is_closed() {
                return;
            }
            changed.await;
        }
    }

    /// The flag [`Self::close`] sets, for the terminal rings to read.
    pub(crate) fn shutdown_flag(&self) -> DaemonShutdownFlag {
        self.inner.closed.clone()
    }

    /// Keep the live-children record at `path` from now on. Set once, before
    /// anything is admitted; a second call is ignored.
    pub(crate) fn record_to(&self, path: PathBuf) {
        if self.inner.record_path.set(path.clone()).is_err() {
            return;
        }
        let (retained, reasons) = match crate::live_children::read_record_with_reasons(&path) {
            Ok(record) => record,
            Err(error) => {
                if let Err(quarantine_error) =
                    crate::live_children::quarantine_record(&path, &error)
                {
                    self.inner
                        .unquarantined_record
                        .store(true, Ordering::Release);
                    tracing::error!(path = %path.display(), %error, %quarantine_error,
                        "cannot quarantine recovery identity; refusing to overwrite the evidence");
                }
                let _ = self.inner.recovery_block.set(format!(
                    "unknown previous container identity: unusable recovery record {}: {error}; offline clearance is required",
                    path.display()
                ));
                Default::default()
            }
        };
        if let Some(reason) = crate::live_children::quarantine_block(&path) {
            tracing::error!(%reason, "carried modules blocked by quarantined recovery identity");
            let _ = self.inner.recovery_block.set(reason);
        }
        {
            *lock(&self.inner.retained) = retained
                .into_iter()
                .map(|entry| {
                    let reason = reasons
                        .get(&entry.pid)
                        .cloned()
                        .unwrap_or(crate::live_children::RetainedReason::Hidden);
                    (entry, reason)
                })
                .collect();
            if let Ok(runtime) = tokio::runtime::Handle::try_current() {
                let weak = Arc::downgrade(&self.inner);
                runtime.spawn(async move {
                    loop {
                        tokio::time::sleep(Duration::from_secs(5)).await;
                        let Some(inner) = weak.upgrade() else {
                            return;
                        };
                        inner.recheck_retained();
                    }
                });
            }
        }
    }

    pub(crate) fn recovery_block(&self) -> Option<&str> {
        self.inner.recovery_block.get().map(String::as_str)
    }

    pub(crate) fn previous_instance_pid(&self, module_id: &str) -> Option<u32> {
        self.previous_instance(module_id)
            .map(|(pid, _)| pid)
            .filter(|pid| *pid > 0)
    }

    pub(crate) fn previous_instance(
        &self,
        module_id: &str,
    ) -> Option<(u32, crate::live_children::RetainedReason)> {
        if let Some((pid, reason)) = lock(&self.inner.pending_admissions).get(module_id) {
            return Some((
                *pid,
                crate::live_children::RetainedReason::ContainerUnverifiable(reason.clone()),
            ));
        }
        // A different module's stalled write cannot change an absent retained
        // identity into this module's previous instance. Avoid the write gate
        // until there is actually retained state for this module to recheck.
        if !lock(&self.inner.retained)
            .iter()
            .any(|(entry, _)| entry.module_id == module_id)
        {
            return None;
        }
        self.inner.recheck_retained();
        lock(&self.inner.retained)
            .iter()
            .find(|(entry, reason)| entry.module_id == module_id && reason.blocks_fresh_copy())
            .map(|(entry, reason)| (entry.pid, reason.clone()))
    }

    pub(crate) fn admission_pending(&self, module_id: &str) -> bool {
        lock(&self.inner.pending_admissions).contains_key(module_id)
    }

    #[cfg(all(test, unix))]
    pub(crate) fn set_admission_timeout_for_test(&self, timeout: Duration) {
        *lock(&self.inner.admission_timeout) = Some(timeout);
    }

    #[cfg(all(test, unix))]
    pub(crate) fn admission_timeout_for_test(&self) -> Option<Duration> {
        *lock(&self.inner.admission_timeout)
    }

    #[cfg(all(test, unix))]
    pub(crate) fn set_admission_hook_for_test(
        &self,
        hook: impl Fn(&str, &'static str, Option<u32>) -> bool + Send + Sync + 'static,
    ) {
        *lock(&self.inner.admission_hook) = Some(AdmissionHook(Arc::new(hook)));
    }

    #[cfg(all(test, unix))]
    pub(crate) fn admission_event_for_test(
        &self,
        module: &str,
        stage: &'static str,
        pid: Option<u32>,
    ) -> bool {
        let hook = lock(&self.inner.admission_hook).clone();
        hook.is_none_or(|hook| (hook.0)(module, stage, pid))
    }

    #[cfg(all(test, unix))]
    pub(crate) fn fail_adoption_for_test(&self) {
        self.inner.fail_adoption.store(true, Ordering::Release);
    }

    #[cfg(all(test, unix))]
    pub(crate) fn set_placement_pause_for_test(
        &self,
        module: &str,
        pause: subc_os::admission_gate::PlacementPause,
    ) {
        lock(&self.inner.placement_pause).insert(module.into(), pause);
    }

    #[cfg(all(test, unix))]
    pub(crate) fn take_placement_pause_for_test(
        &self,
        module: &str,
    ) -> Option<subc_os::admission_gate::PlacementPause> {
        lock(&self.inner.placement_pause).remove(module)
    }

    #[cfg(test)]
    pub(crate) fn take_adoption_failure_for_test(&self) -> bool {
        self.inner.fail_adoption.swap(false, Ordering::AcqRel)
    }

    #[cfg(unix)]
    pub(crate) fn flush_admissions(&self) -> std::io::Result<()> {
        self.inner.without_stalling_runtime(|| {
            let _write = lock(&self.inner.record_write);
            let snapshot = lock(&self.inner.live).clone();
            self.inner.write_record(&snapshot)
        })
    }

    #[cfg(unix)]
    pub(crate) fn begin_pending_admission(&self, module_id: &str, pid: u32) {
        lock(&self.inner.pending_admissions).insert(module_id.into(), (pid,
            "carried admission cleanup pending; the child and durable record must be resolved before another start".into()));
    }

    #[cfg(unix)]
    pub(crate) fn finish_pending_admission(&self, module_id: &str) {
        lock(&self.inner.pending_admissions).remove(module_id);
    }

    #[cfg(unix)]
    pub(crate) fn quarantine_failed_admission(&self, module_id: &str, source: &std::io::Error) {
        let reason = format!(
            "failed late admission write for {module_id}: {source}; offline clearance is required"
        );
        let _ = self.inner.recovery_block.set(reason.clone());
        tracing::error!(module_id, %source, "failed late admission write; quarantining the recovery store");
        self.inner.without_stalling_runtime(|| {
            let _write = lock(&self.inner.record_write);
            if let Some(path) = self.inner.record_path.get() {
                if let Err(error) = crate::live_children::quarantine_failed_admission(path, &reason, source) {
                    self.inner.unquarantined_record.store(true, Ordering::Release);
                    tracing::error!(%error, %reason, "cannot persist admission failure quarantine; preserving recovery evidence");
                }
            }
        });
    }

    pub(crate) fn admit(
        &self,
        module_id: String,
        pid: u32,
        protocol: ModuleProtocol,
        start_time: Option<u64>,
        recorded: RecordedIdentity,
    ) -> Result<RosterGuard, AdmissionError> {
        let key = self.inner.next_key.fetch_add(1, Ordering::Relaxed);
        // The process is already spawned. Publish ownership before waiting for
        // another writer, so shutdown can stop it even while that writer syncs.
        lock(&self.inner.live).insert(
            key,
            RosterEntry {
                module_id,
                pid,
                protocol,
                start_time,
                recorded,
                drain_budget: Arc::clone(&self.drain_budget),
            },
        );
        let guard = RosterGuard {
            inner: Arc::clone(&self.inner),
            key,
        };
        let recorded = self.inner.without_stalling_runtime(|| {
            let _write = lock(&self.inner.record_write);
            let snapshot = lock(&self.inner.live).clone();
            self.inner.write_record(&snapshot)
        });
        match recorded {
            Ok(()) => Ok(guard),
            Err(source) => Err(AdmissionError { source, guard }),
        }
    }

    #[cfg(all(test, unix))]
    pub(crate) fn live_count_for_test(&self) -> usize {
        self.live().len()
    }

    #[cfg(unix)]
    fn live(&self) -> Vec<(u64, RosterEntry)> {
        lock(&self.inner.live)
            .iter()
            .map(|(key, entry)| (*key, entry.clone()))
            .collect()
    }

    /// Run `hook` at every record write from now on (when a record path is
    /// set), before the roster is serialized and written, with the roster
    /// being written. Only the Unix restart test uses it.
    #[cfg(all(test, unix))]
    pub(crate) fn set_record_write_hook_for_test(
        &self,
        hook: impl Fn(&HashMap<u64, RosterEntry>) + Send + Sync + 'static,
    ) {
        *lock(&self.inner.record_write_hook) = Some(RecordWriteHook(Arc::new(hook)));
    }

    /// Mark daemon shutdown as begun. Idempotent.
    ///
    /// Called at the very start of the announced shutdown, before the notice
    /// and before any connection is closed: from then on no module is
    /// respawned, and every exit is recorded as `daemon_shutdown`.
    #[cfg(unix)]
    pub(crate) fn close(&self) {
        self.inner.closed.set();
        self.inner.closed_notify.notify_waiters();
    }
}

fn lock<T>(mutex: &Mutex<T>) -> MutexGuard<'_, T> {
    mutex
        .lock()
        .unwrap_or_else(|poisoned| poisoned.into_inner())
}

#[cfg(unix)]
pub(crate) use unix_shutdown::{end_children_for_daemon_shutdown, TERM_TO_KILL};

#[cfg(unix)]
mod unix_shutdown {
    use std::{
        collections::{HashMap, HashSet},
        future::Future,
        time::Duration,
    };

    use rustix::process::{kill_process, Pid, Signal};
    use tokio::time::{sleep, Instant};
    use tracing::{debug, info, warn};

    use super::{lock, ChildRoster, RosterEntry};
    use subc_control::ModuleProtocol;

    /// The longest any one child is waited on before escalation, whatever its
    /// drain budget says.
    ///
    /// Modules are outside the daemon's process group, so if the service
    /// manager SIGKILLs the daemon before this stop finishes, whatever the
    /// daemon has not ended survives it. A subc module has already had its EOF
    /// and finishes its own teardown regardless, but a `protocol: "none"` child
    /// would be orphaned. So the whole shutdown must fit inside the service
    /// manager's stop timeout: the 0.5 s notice and 2 s drain before this, this
    /// cap, [`TERM_TO_KILL`] and [`CHILD_REAP_BOUND`] after it come to 28.25 s,
    /// under the 35 s `ExitTimeOut` / `TimeoutStopSec` that `ck setup` writes
    /// (see `desired_definition`). 25 s covers BROCA's teardown (10 s run grace,
    /// then a seal, inside a 20 s budget; 12 s measured) with room to spare.
    const CHILD_SHUTDOWN_CAP: Duration = Duration::from_secs(25);
    /// For a subc module still running at its deadline: the time between the
    /// SIGTERM sent then and the SIGKILL. Long enough for a handler to write a
    /// last line and exit.
    pub(crate) const TERM_TO_KILL: Duration = Duration::from_millis(500);
    /// After a SIGKILL, how long to wait for the supervisor tasks to reap.
    /// SIGKILL cannot be ignored, so this only covers scheduling; it bounds the
    /// wait even if a reap never lands.
    const CHILD_REAP_BOUND: Duration = Duration::from_millis(250);
    const POLL: Duration = Duration::from_millis(10);

    /// End every supervised child before the daemon exits.
    ///
    /// The caller has already closed every connection, so each subc module has
    /// had its EOF: that is its one stop request, and it is left alone to run
    /// its own teardown. A `protocol: "none"` child has no connection, so its
    /// one stop request is SIGTERM, sent here immediately (the same stop the
    /// supervisor sends it on restart).
    ///
    /// Escalation happens per child, at that child's own deadline: its
    /// resolved drain budget (per-module `drain_timeout_ms`, else the daemon
    /// default), capped at [`CHILD_SHUTDOWN_CAP`]. At the deadline a subc
    /// module still running gets SIGTERM and, [`TERM_TO_KILL`] later, SIGKILL;
    /// a `protocol: "none"` child, already asked, gets SIGKILL. All children are
    /// waited on concurrently, so the whole stop takes as long as the longest
    /// single deadline, and returns as soon as every child has been reaped.
    /// A fixed short bound for everyone would SIGKILL a module whose teardown
    /// legitimately takes seconds (BROCA seals in-flight runs) on exactly the
    /// stops where it has work in flight.
    ///
    /// Spawns are refused first, so a module exiting on EOF is not restarted.
    /// `already_escalated` or `escalate` resolving (a second SIGTERM to the
    /// daemon) skips every remaining wait and kills what is left: the
    /// operator has said stop waiting, and leaving children behind would be
    /// the orphan this exists to prevent.
    pub(crate) async fn end_children_for_daemon_shutdown(
        roster: &ChildRoster,
        already_escalated: bool,
        escalate: impl Future<Output = ()>,
    ) {
        roster.close();
        tokio::pin!(escalate);
        let started = Instant::now();
        let mut termed = HashMap::new();
        let mut killed = HashSet::new();
        let mut last_kill: Option<Instant> = None;
        let mut escalated = already_escalated;

        if !escalated {
            for (key, entry) in roster.live() {
                if entry.protocol == ModuleProtocol::None {
                    signal(&entry, Signal::TERM);
                    termed.insert(key, Instant::now());
                }
            }
        }

        loop {
            let live = roster.live();
            if live.is_empty() {
                break;
            }
            let now = Instant::now();
            for (key, entry) in &live {
                if killed.contains(key) {
                    continue;
                }
                let deadline = started + budget(entry);
                let due = shutdown_signal(
                    entry.protocol,
                    now,
                    deadline,
                    termed.get(key).copied(),
                    escalated,
                );
                if due == Some(Signal::KILL) {
                    warn!(
                        module_id = %entry.module_id,
                        pid = entry.pid,
                        escalated,
                        child_stop_elapsed_ms = started.elapsed().as_millis() as u64,
                        term_to_kill_ms = termed.get(key).map(|at| at.elapsed().as_millis() as u64),
                        "supervised child did not exit during daemon shutdown; sending SIGKILL"
                    );
                    #[cfg(target_os = "linux")]
                    if entry.start_time.is_none_or(|expected| {
                        crate::provenance::process_start_time(entry.pid) == Some(expected)
                    }) {
                        if let Some(name) = &entry.recorded.cgroup_name {
                            crate::supervise::kill_module_cgroup(
                                entry.recorded.cgroup_placement.as_ref(),
                                name,
                            );
                        }
                    }
                    signal(entry, Signal::KILL);
                    killed.insert(*key);
                    last_kill = Some(now);
                } else if due == Some(Signal::TERM) {
                    warn!(
                        module_id = %entry.module_id,
                        pid = entry.pid,
                        child_stop_elapsed_ms = started.elapsed().as_millis() as u64,
                        "supervised module still running at its shutdown deadline after EOF; sending SIGTERM"
                    );
                    signal(entry, Signal::TERM);
                    termed.insert(*key, Instant::now());
                }
            }
            // Everything left has been SIGKILLed: wait only for the reaps.
            if live.iter().all(|(key, _)| killed.contains(key))
                && last_kill.is_some_and(|at| now >= at + CHILD_REAP_BOUND)
            {
                break;
            }
            if escalated {
                sleep(POLL).await;
                continue;
            }
            tokio::select! {
                biased;
                _ = escalate.as_mut() => {
                    info!("second SIGTERM: killing remaining supervised children without further grace");
                    escalated = true;
                }
                _ = sleep(POLL) => {}
            }
        }
        info!(
            child_stop_elapsed_ms = started.elapsed().as_millis() as u64,
            "supervised child shutdown wait complete"
        );
    }

    fn budget(entry: &RosterEntry) -> Duration {
        (*lock(&entry.drain_budget)).min(CHILD_SHUTDOWN_CAP)
    }

    fn shutdown_signal(
        protocol: ModuleProtocol,
        now: Instant,
        deadline: Instant,
        term_sent: Option<Instant>,
        escalated: bool,
    ) -> Option<Signal> {
        // A delayed poll must not skip SIGTERM or consume its grace before the
        // signal was sent. The second SIGTERM deliberately bypasses that grace.
        let kill_due = match protocol {
            ModuleProtocol::None => now >= deadline,
            ModuleProtocol::Subc => term_sent.is_some_and(|at| now >= at + TERM_TO_KILL),
        };
        if escalated || kill_due {
            Some(Signal::KILL)
        } else if now >= deadline && term_sent.is_none() {
            Some(Signal::TERM)
        } else {
            None
        }
    }

    #[test]
    fn release_record_sync_does_not_lock_the_shutdown_roster() {
        let dir = cortexkit_test_support::ScratchDir::new("roster-release-sync");
        let path = dir.join("children.json");
        let roster = ChildRoster::default();
        roster.record_to(path.clone());
        let guard = roster
            .admit(
                "record-only".into(),
                std::process::id(),
                ModuleProtocol::Subc,
                None,
                super::RecordedIdentity::default(),
            )
            .unwrap();
        let entered = std::sync::Arc::new(std::sync::Barrier::new(2));
        let release = std::sync::Arc::new(std::sync::Barrier::new(2));
        roster.set_record_write_hook_for_test({
            let entered = entered.clone();
            let release = release.clone();
            move |live| {
                if live.is_empty() {
                    entered.wait();
                    release.wait();
                }
            }
        });
        let writer = std::thread::spawn(move || drop(guard));
        entered.wait();
        let readable = roster.inner.live.try_lock().is_ok();
        release.wait();
        writer.join().unwrap();
        assert!(
            readable,
            "a record sync must not block shutdown's live-child snapshot"
        );
        assert!(roster.live().is_empty());
        assert!(crate::live_children::read_record_with_reasons(&path)
            .unwrap()
            .0
            .is_empty());
    }

    #[test]
    fn failed_admission_keeps_ownership_until_child_is_killed_and_reaped() {
        struct OwnedChild(std::process::Child);
        impl Drop for OwnedChild {
            fn drop(&mut self) {
                if self.0.try_wait().is_ok_and(|status| status.is_none()) {
                    let _ = self.0.kill();
                    let _ = self.0.wait();
                }
            }
        }
        let dir = cortexkit_test_support::ScratchDir::new("failed-admission-owner");
        let record = dir.join("children.json");
        let roster = ChildRoster::default();
        roster.record_to(record.clone());
        // Startup quarantine can move an unusable path aside. Make the
        // destination a directory afterwards so admission's file write fails,
        // preserving the ordinary child's ownership through kill and reap.
        std::fs::create_dir(&record).unwrap();
        assert!(roster.recovery_block().is_none());
        let mut child = OwnedChild(
            std::process::Command::new("sleep")
                .arg("60")
                .spawn()
                .unwrap(),
        );
        let pid = child.0.id();
        let failure = roster
            .admit(
                "rejected".into(),
                pid,
                ModuleProtocol::Subc,
                None,
                super::RecordedIdentity::default(),
            )
            .unwrap_err();
        failure.with_cleanup(|_source| {
            assert!(
                roster.live().iter().any(|(_, entry)| entry.pid == pid),
                "failed admission released ownership before rejection cleanup"
            );
            child.0.kill().unwrap();
            assert!(!child.0.wait().unwrap().success());
            assert!(
                roster.live().iter().any(|(_, entry)| entry.pid == pid),
                "ownership must survive through the reap"
            );
        });
        assert!(
            child.0.try_wait().unwrap().is_some(),
            "the rejected child must be reaped"
        );
        assert!(
            roster.live().is_empty(),
            "cleanup must release the rejected child's roster entry"
        );
    }

    #[test]
    fn queued_admission_remains_visible_to_shutdown() {
        let roster = ChildRoster::default();
        let write = lock(&roster.inner.record_write);
        let worker_roster = roster.clone();
        let worker = std::thread::spawn(move || {
            worker_roster
                .admit(
                    "queued".into(),
                    std::process::id(),
                    ModuleProtocol::Subc,
                    None,
                    super::RecordedIdentity::default(),
                )
                .unwrap()
        });
        // This is a fixture scheduling guard, not a shutdown-policy deadline.
        let deadline = std::time::Instant::now() + Duration::from_secs(15);
        let visible = loop {
            if !roster.live().is_empty() {
                break true;
            }
            if std::time::Instant::now() >= deadline {
                break false;
            }
            std::thread::sleep(Duration::from_millis(1));
        };
        drop(write);
        let guard = worker.join().unwrap();
        assert!(
            visible,
            "shutdown must see a spawned child queued behind a record write"
        );
        drop(guard);
    }

    #[test]
    fn late_shutdown_poll_still_gives_sigterm_its_grace() {
        let deadline = Instant::now();
        let late_poll = deadline + Duration::from_secs(1);
        assert_eq!(
            shutdown_signal(ModuleProtocol::Subc, late_poll, deadline, None, false),
            Some(Signal::TERM)
        );
        assert_eq!(
            shutdown_signal(
                ModuleProtocol::Subc,
                late_poll,
                deadline,
                Some(late_poll),
                false
            ),
            None
        );
        assert_eq!(
            shutdown_signal(
                ModuleProtocol::Subc,
                late_poll + TERM_TO_KILL,
                deadline,
                Some(late_poll),
                false
            ),
            Some(Signal::KILL)
        );
        assert_eq!(
            shutdown_signal(
                ModuleProtocol::Subc,
                late_poll,
                deadline,
                Some(late_poll),
                true
            ),
            Some(Signal::KILL)
        );
    }

    fn signal(entry: &RosterEntry, signal: Signal) {
        if let Some(carried) = &entry.recorded.carried {
            if let Some(container) = carried.upgrade() {
                if signal == Signal::KILL {
                    container.hard_kill();
                }
            }
            return;
        }
        // A reaped child's pid can be reused. Where the kernel start time is
        // known, refuse to signal a process that is not the one spawned.
        if let Some(expected) = entry.start_time {
            if crate::provenance::process_start_time(entry.pid) != Some(expected) {
                debug!(
                    module_id = %entry.module_id,
                    pid = entry.pid,
                    "supervised child already gone; not signalling its pid"
                );
                return;
            }
        }
        let Some(pid) = i32::try_from(entry.pid).ok().and_then(Pid::from_raw) else {
            return;
        };
        if let Err(error) = kill_process(pid, signal) {
            debug!(
                module_id = %entry.module_id,
                pid = entry.pid,
                ?signal,
                %error,
                "signal to supervised child failed; it has most likely already exited"
            );
        }
    }
}

#[cfg(all(test, unix))]
mod retained_container_tests {
    use super::*;
    use crate::live_children::{ContainerIdentity, RetainedReason};
    use cortexkit_test_support::ScratchDir;
    use std::os::unix::process::CommandExt;
    use std::process::{Child, Command};

    struct OwnedChild(Child);
    impl OwnedChild {
        fn spawn(group: u32) -> Self {
            Self(
                Command::new("/bin/sleep")
                    .arg("600")
                    .process_group(group as i32)
                    .spawn()
                    .unwrap(),
            )
        }
        fn id(&self) -> u32 {
            self.0.id()
        }
    }
    impl Drop for OwnedChild {
        fn drop(&mut self) {
            let _ = self.0.kill();
            let _ = self.0.wait();
        }
    }

    fn fixture(label: &str) -> (ScratchDir, ChildRoster, OwnedChild, u32) {
        let dir = ScratchDir::new(label);
        let roster = ChildRoster::default();
        roster.record_to(dir.join("children.json"));
        let mut first = OwnedChild::spawn(0);
        let identity = subc_os::process_group::inspect(first.id())
            .unwrap()
            .unwrap();
        let holder = OwnedChild::spawn(first.id());
        let guard = roster
            .admit(
                label.into(),
                first.id(),
                ModuleProtocol::None,
                Some(identity.start_time),
                RecordedIdentity {
                    container: Some(ContainerIdentity {
                        pgid: Some(first.id()),
                        created_start_time: Some(identity.start_time),
                        cgroup_path: None,
                        ..ContainerIdentity::test_metadata()
                    }),
                    ..RecordedIdentity::default()
                },
            )
            .unwrap();
        first.0.kill().unwrap();
        first.0.wait().unwrap();
        guard.retain(RetainedReason::ContainerSurvivors(vec![holder.id()]));
        (dir, roster, holder, first.id())
    }

    #[test]
    fn retained_container_survivor_blocks_until_its_death() {
        let label = "retained-container-live";
        let (dir, roster, mut holder, pgid) = fixture(label);
        let recorded = crate::live_children::read_record(&dir.join("children.json")).unwrap();
        assert_eq!(
            recorded.len(),
            1,
            "the dead leader's container record must survive guard release"
        );
        assert_eq!(recorded[0].container.as_ref().unwrap().pgid, Some(pgid));
        let (pid, reason) = roster
            .previous_instance(label)
            .expect("a live container member blocks a fresh copy");
        assert_eq!(pid, pgid);
        assert_eq!(
            reason,
            RetainedReason::ContainerSurvivors(vec![holder.id()])
        );
        let detail = reason.waiting_message(pid);
        assert!(
            detail.contains("container survivors") && detail.contains(&holder.id().to_string()),
            "{detail}"
        );
        // The recheck queued SIGKILL after verifying this owned child. Confirm
        // its death before checking that the roster no longer blocks a fresh copy.
        holder.0.wait().unwrap();
        assert_eq!(roster.previous_instance(label), None);
        assert!(
            crate::live_children::read_record(&dir.join("children.json"))
                .unwrap()
                .is_empty()
        );
    }

    #[test]
    fn retained_container_zombie_does_not_block_a_fresh_copy() {
        let label = "retained-container-zombie";
        let (_dir, roster, mut holder, _) = fixture(label);
        holder.0.kill().unwrap();
        let deadline = std::time::Instant::now() + Duration::from_secs(2);
        while subc_os::child_exit::observe(holder.id()).unwrap().is_none() {
            assert!(
                std::time::Instant::now() < deadline,
                "owned child must reach zombie state"
            );
            std::thread::sleep(Duration::from_millis(1));
        }
        assert!(subc_os::process_group::inspect(holder.id())
            .unwrap()
            .is_none_or(|facts| facts.exited));
        assert_eq!(
            roster.previous_instance(label),
            None,
            "an exited unreaped helper has no live resources"
        );
    }
}