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
//! 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>,
/// 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 {
next_key: AtomicU64,
closed: DaemonShutdownFlag,
live: Mutex<HashMap<u64, RosterEntry>>,
/// Where the live-children record is written; unset means no record, the
/// default for an in-process daemon (see `BootstrapConfig`).
record_path: OnceLock<PathBuf>,
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(test)]
type RecordWriteFn = dyn Fn(&HashMap<u64, RosterEntry>) + Send + Sync;
#[cfg(test)]
#[derive(Clone)]
struct RecordWriteHook(Arc<RecordWriteFn>);
#[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 taking the `live` lock as well as the write: the lock
/// is held for the whole write, so a caller waiting for it waits as long.
///
/// 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 the record from `live`. Called with the `live` lock held, so
/// concurrent admits and releases write their snapshots in the order they
/// changed the roster and the last write on disk is the current roster.
fn write_record(&self, live: &HashMap<u64, RosterEntry>) -> std::io::Result<()> {
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(),
})
.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))
.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) {
// Take `live` before `retained`, matching record writes made when a
// newly spawned child joins the roster. This prevents pruning from
// racing a write that must preserve a previous instance.
let live = lock(&self.live);
let changed = {
let mut retained = lock(&self.retained);
let before = retained.len();
retained.retain(|(entry, _)| {
#[cfg(unix)]
{
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
};
if changed {
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>>,
}
impl Default for ChildRoster {
fn default() -> Self {
Self {
inner: Arc::default(),
drain_budget: Arc::new(Mutex::new(crate::supervise::DEFAULT_DRAIN_TIMEOUT)),
}
}
}
/// 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 Drop for RosterGuard {
fn drop(&mut self) {
self.inner.without_stalling_runtime(|| {
let mut live = lock(&self.inner.live);
if live.remove(&self.key).is_some() {
let _ = self.inner.write_record(&live);
}
});
}
}
#[cfg(target_os = "macos")]
impl RosterGuard {
/// Publish an executable only after the privacy trampoline has exec'd and
/// the kernel image agrees with the configured module. Before then shutdown
/// owns the pid, but orphan cleanup has no image it could mistakenly trust.
pub(crate) fn confirm_executable(
&self,
observation: subc_os::Observation,
) -> std::io::Result<()> {
self.inner.without_stalling_runtime(|| {
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);
self.inner.write_record(&live)?;
}
Ok(())
})
}
}
impl ChildRoster {
#[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,
}
}
/// 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()
}
/// 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) {
let (retained, reasons) =
crate::live_children::read_record_with_reasons(&path).unwrap_or_default();
let first_record = self.inner.record_path.set(path).is_ok();
if first_record {
*lock(&self.inner.retained) = retained
.into_iter()
.map(|entry| {
let reason = reasons
.get(&entry.pid)
.copied()
.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 previous_instance_pid(&self, module_id: &str) -> Option<u32> {
self.previous_instance(module_id).map(|(pid, _)| pid)
}
pub(crate) fn previous_instance(
&self,
module_id: &str,
) -> Option<(u32, crate::live_children::RetainedReason)> {
self.inner.recheck_retained();
lock(&self.inner.retained)
.iter()
.find(|(entry, _)| entry.module_id == module_id)
.map(|(entry, reason)| (entry.pid, *reason))
}
pub(crate) fn admit(
&self,
module_id: String,
pid: u32,
protocol: ModuleProtocol,
start_time: Option<u64>,
recorded: RecordedIdentity,
) -> std::io::Result<RosterGuard> {
let key = self.inner.next_key.fetch_add(1, Ordering::Relaxed);
self.inner.without_stalling_runtime(|| {
let mut live = lock(&self.inner.live);
live.insert(
key,
RosterEntry {
module_id,
pid,
protocol,
start_time,
recorded,
drain_budget: Arc::clone(&self.drain_budget),
},
);
if let Err(error) = self.inner.write_record(&live) {
live.remove(&key);
return Err(error);
}
Ok(())
})?;
Ok(RosterGuard {
inner: Arc::clone(&self.inner),
key,
})
}
#[cfg(unix)]
fn live(&self) -> Vec<(u64, RosterEntry)> {
self.inner.without_stalling_runtime(|| {
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();
}
}
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;
#[cfg(unix)]
mod unix_shutdown {
use std::{collections::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.
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 = HashSet::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);
}
}
}
loop {
let live = roster.live();
if live.is_empty() {
return;
}
let now = Instant::now();
for (key, entry) in &live {
if killed.contains(key) {
continue;
}
let deadline = started + budget(entry);
let kill_at = match entry.protocol {
ModuleProtocol::None => deadline,
ModuleProtocol::Subc => deadline + TERM_TO_KILL,
};
if escalated || now >= kill_at {
warn!(
module_id = %entry.module_id,
pid = entry.pid,
escalated,
"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 now >= deadline && termed.insert(*key) {
warn!(
module_id = %entry.module_id,
pid = entry.pid,
"supervised module still running at its shutdown deadline after EOF; sending SIGTERM"
);
signal(entry, Signal::TERM);
}
}
// 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)
{
return;
}
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) => {}
}
}
}
fn budget(entry: &RosterEntry) -> Duration {
(*lock(&entry.drain_budget)).min(CHILD_SHUTDOWN_CAP)
}
fn signal(entry: &RosterEntry, signal: Signal) {
// 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"
);
}
}
}