windows-file-watcher 0.2.4

Memory-safe, queue-mediated Windows path-change watching over ReadDirectoryChangesW, with a coarse fallback.
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
// Copyright (c) 2026 Mike Grier
//! Unit tests for the servicing path.
//!
//! What is under test is the machinery's four promises: every request is serviced
//! exactly once, in submission order, never concurrently with another, and
//! teardown converges. The request type is deliberately trivial -- these are
//! properties of the path, not of what travels it.

use std::sync::Arc;
use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
use std::time::{Duration, Instant};

use super::Servicer;

/// Upper bound for waiting on work the pool really should run.
const SERVICE_TIMEOUT: Duration = Duration::from_secs(30);

/// How long a deliberately-parked handler will wait to be released before giving
/// up and returning.
///
/// The bound is not politeness. Teardown waits for a running handler, so a
/// handler that parks unconditionally turns any assertion failure *before* the
/// release into a deadlock rather than a test failure -- the panic unwinds into
/// `Drop`, which waits for a handler nothing will ever free. That wedged a whole
/// nextest run rather than reporting one failing assertion.
const PARK_LIMIT: Duration = Duration::from_secs(10);

/// Park until `gate` is set, or until [`PARK_LIMIT`] elapses.
fn park_until(gate: &std::sync::atomic::AtomicBool) {
    let deadline = Instant::now() + PARK_LIMIT;
    while !gate.load(Ordering::SeqCst) && Instant::now() < deadline {
        std::thread::sleep(Duration::from_millis(1));
    }
}

/// Records what the handler saw, so a test can assert on order and count without
/// running assertions on the servicing path itself.
#[derive(Default)]
struct Log {
    seen: std::sync::Mutex<Vec<u64>>,
    /// How many handler calls are in flight. Peaks above one would mean the path
    /// is not serialising.
    inside: AtomicUsize,
    peak: AtomicUsize,
}

impl Log {
    fn record(&self, item: u64) {
        let now = self.inside.fetch_add(1, Ordering::SeqCst) + 1;
        self.peak.fetch_max(now, Ordering::SeqCst);
        self.seen.lock().expect("record").push(item);
        self.inside.fetch_sub(1, Ordering::SeqCst);
    }

    fn items(&self) -> Vec<u64> {
        self.seen.lock().expect("read").clone()
    }

    fn peak(&self) -> usize {
        self.peak.load(Ordering::SeqCst)
    }
}

/// A servicer that appends every request to a shared log.
fn logging() -> (Servicer<u64>, Arc<Log>) {
    let log = Arc::new(Log::default());
    let handler = Arc::clone(&log);
    let servicer =
        Servicer::new(move |item: u64| handler.record(item)).expect("create the servicing path");
    (servicer, log)
}

/// Block until `predicate` holds, failing rather than hanging if it never does.
fn wait_until<F>(what: &str, predicate: F)
where
    F: Fn() -> bool,
{
    let deadline = Instant::now() + SERVICE_TIMEOUT;
    while !predicate() {
        assert!(Instant::now() < deadline, "timed out waiting for {what}");
        std::thread::sleep(Duration::from_millis(1));
    }
}

#[test]
fn a_submitted_request_is_serviced() {
    let (servicer, log) = logging();
    servicer.submit(7).expect("accepted");
    wait_until("the request", || log.items() == [7]);
}

#[test]
fn requests_are_serviced_in_submission_order() {
    let (servicer, log) = logging();
    for item in 0..64 {
        servicer.submit(item).expect("accepted");
    }
    wait_until("every request", || log.items().len() == 64);

    let expected: Vec<u64> = (0..64).collect();
    assert_eq!(log.items(), expected, "the queue must be FIFO");
}

#[test]
fn every_request_is_serviced_exactly_once() {
    let (servicer, log) = logging();
    for item in 0..500 {
        servicer.submit(item).expect("accepted");
    }
    wait_until("every request", || log.items().len() == 500);

    let mut seen = log.items();
    seen.sort_unstable();
    seen.dedup();
    assert_eq!(seen.len(), 500, "no request may be serviced twice or lost");
}

#[test]
fn handlers_never_run_concurrently() {
    let (servicer, log) = logging();
    for item in 0..500 {
        servicer.submit(item).expect("accepted");
    }
    wait_until("every request", || log.items().len() == 500);

    // The guarantee D-2 actually needs. Coalescing alone would not give it: a
    // drain that has emptied the queue but is still running its last handler
    // leaves the queue observably empty, so an empty/non-empty doorbell would
    // start a second drain alongside the first.
    assert_eq!(
        log.peak(),
        1,
        "resident-state mutations must be serialised, saw {} concurrent handlers",
        log.peak()
    );
}

#[test]
fn many_producers_are_all_serviced_and_never_overlap() {
    let (servicer, log) = logging();
    let servicer = Arc::new(servicer);

    let producers: Vec<_> = (0..8)
        .map(|producer| {
            let servicer = Arc::clone(&servicer);
            std::thread::spawn(move || {
                for index in 0..100 {
                    servicer.submit(producer * 100 + index).expect("accepted");
                }
            })
        })
        .collect();
    for producer in producers {
        producer.join().expect("producer");
    }

    wait_until("every request", || log.items().len() == 800);
    assert_eq!(log.peak(), 1, "concurrent producers, one servicer");

    let mut seen = log.items();
    seen.sort_unstable();
    seen.dedup();
    assert_eq!(seen.len(), 800);
}

#[test]
fn the_doorbell_coalesces_a_burst_of_submissions() {
    // The concrete failure this exists to prevent: subscribing to 500 paths must
    // not queue 500 drains, 499 of which find the queue already emptied.
    //
    // Parking the handler is what makes the count exact rather than a bound. A
    // burst against an idle servicer would ring an unpredictable number of times
    // -- legitimately, since a drain that outpaces the producer finds nothing to
    // coalesce -- so asserting a small number there would be asserting the
    // machine's speed. With a handler held mid-flight the backlog is guaranteed,
    // and the flag is set at *submission* time rather than when the drain runs, so
    // the answer is one regardless of how the pool schedules anything.
    let released = Arc::new(std::sync::atomic::AtomicBool::new(false));
    let gate = Arc::clone(&released);
    let count = Arc::new(AtomicUsize::new(0));
    let counter = Arc::clone(&count);

    let servicer = Servicer::new(move |_: u64| {
        park_until(&gate);
        counter.fetch_add(1, Ordering::SeqCst);
    })
    .expect("create the servicing path");

    for item in 0..500 {
        servicer.submit(item).expect("accepted");
    }
    assert_eq!(
        servicer.rings(),
        1,
        "500 submissions behind one running handler must queue exactly one drain"
    );

    released.store(true, Ordering::SeqCst);
    wait_until("every request", || count.load(Ordering::SeqCst) == 500);

    // Servicing the backlog rings nothing further: the drain loops until empty
    // rather than needing a ring per request.
    assert_eq!(servicer.rings(), 1);
}

#[test]
fn an_idle_servicer_has_not_rung() {
    let (servicer, log) = logging();
    assert_eq!(servicer.rings(), 0);
    assert_eq!(servicer.pending(), 0);
    assert!(log.items().is_empty());
}

#[test]
fn a_request_submitted_from_a_handler_is_serviced() {
    // Re-entrant submission is what M3.5's cancel-during-subscribe path will do,
    // and it must not deadlock: `submit` holds the queue lock only for the
    // enqueue, and the drain holds it only to pop.
    let log = Arc::new(Log::default());
    let handler = Arc::clone(&log);
    let inner: Arc<std::sync::OnceLock<Servicer<u64>>> = Arc::new(std::sync::OnceLock::new());
    // Weak, not a second `Arc`: the strong form would be a cycle -- the servicer
    // owns the work object, which owns this closure -- so the servicer would never
    // drop and its teardown would never be exercised.
    let self_ref = Arc::downgrade(&inner);

    let servicer = Servicer::new(move |item: u64| {
        handler.record(item);
        if item == 1 {
            let inner = self_ref
                .upgrade()
                .expect("the servicer outlives its own handler");
            let servicer = inner
                .get()
                .expect("the servicer is set before any submission");
            servicer.submit(2).expect("accepted");
        }
    })
    .expect("create the servicing path");
    inner
        .set(servicer)
        .unwrap_or_else(|_| unreachable!("set exactly once"));

    inner.get().expect("set").submit(1).expect("accepted");
    wait_until("both requests", || log.items() == [1, 2]);
}

#[test]
fn a_slow_handler_does_not_block_producers() {
    let released = Arc::new(std::sync::atomic::AtomicBool::new(false));
    let gate = Arc::clone(&released);
    let count = Arc::new(AtomicUsize::new(0));
    let counter = Arc::clone(&count);

    let servicer = Servicer::new(move |_: u64| {
        // The first request parks, so every later submission happens while a
        // handler is genuinely mid-flight.
        park_until(&gate);
        counter.fetch_add(1, Ordering::SeqCst);
    })
    .expect("create the servicing path");

    servicer.submit(0).expect("accepted");
    // Wait until that request has actually been taken off the queue, so the count
    // asserted below is exact rather than a race against the drain starting.
    wait_until("the handler to start", || servicer.pending() == 0);

    // If `submit` waited on servicing, this would not return until the handler
    // was released -- so completing it at all is the assertion.
    let started = Instant::now();
    for item in 1..200 {
        servicer.submit(item).expect("accepted");
    }
    assert!(
        started.elapsed() < Duration::from_secs(5),
        "submitting blocked behind a running handler"
    );
    assert_eq!(servicer.pending(), 199);

    released.store(true, Ordering::SeqCst);
    wait_until("every request", || count.load(Ordering::SeqCst) == 200);
}

#[test]
fn shutdown_refuses_further_requests_and_hands_them_back() {
    let (servicer, _log) = logging();
    servicer.shut_down();

    assert!(!servicer.is_open());
    let rejected = servicer
        .submit(99)
        .expect_err("a shut-down path must refuse");
    assert_eq!(
        rejected.0, 99,
        "the request comes back so it can be reported"
    );
}

#[test]
fn shutdown_is_idempotent_and_drop_after_it_is_safe() {
    let (servicer, _log) = logging();
    servicer.submit(1).expect("accepted");
    servicer.shut_down();
    servicer.shut_down();
    servicer.shut_down();
    assert!(!servicer.is_open());
    drop(servicer);
}

#[test]
fn shutdown_discards_pending_requests_rather_than_servicing_them() {
    let released = Arc::new(std::sync::atomic::AtomicBool::new(false));
    let gate = Arc::clone(&released);
    let count = Arc::new(AtomicUsize::new(0));
    let counter = Arc::clone(&count);

    let servicer = Servicer::new(move |_: u64| {
        park_until(&gate);
        counter.fetch_add(1, Ordering::SeqCst);
    })
    .expect("create the servicing path");

    servicer.submit(0).expect("accepted");
    wait_until("the handler to start", || servicer.pending() == 0);
    for item in 1..50 {
        servicer.submit(item).expect("accepted");
    }
    assert_eq!(servicer.pending(), 49);

    // Release before shutting down, or shutdown would wait on a handler that
    // never returns -- which is a property of this test's handler, not of the
    // path.
    released.store(true, Ordering::SeqCst);
    servicer.shut_down();

    assert_eq!(servicer.pending(), 0);
    assert!(
        servicer.discarded() > 0,
        "pending requests must be discarded, not serviced, at teardown"
    );
    assert!(
        count.load(Ordering::SeqCst) < 50,
        "teardown serviced the whole backlog instead of discarding it"
    );
}

#[test]
fn shutdown_waits_for_a_running_handler() {
    let finished = Arc::new(std::sync::atomic::AtomicBool::new(false));
    let flag = Arc::clone(&finished);

    let servicer = Servicer::new(move |_: u64| {
        std::thread::sleep(Duration::from_millis(50));
        flag.store(true, Ordering::SeqCst);
    })
    .expect("create the servicing path");

    servicer.submit(0).expect("accepted");
    // Wait until the handler is genuinely running, so the shutdown below has
    // something in flight to wait for rather than racing the submission.
    wait_until("the handler to start", || servicer.pending() == 0);
    servicer.shut_down();

    assert!(
        finished.load(Ordering::SeqCst),
        "shutdown returned while a handler was still running"
    );
}

#[test]
fn dropping_an_idle_servicer_is_prompt() {
    let (servicer, _log) = logging();
    let started = Instant::now();
    drop(servicer);
    assert!(
        started.elapsed() < Duration::from_secs(5),
        "dropping an idle servicing path must not wait for anything"
    );
}

#[test]
fn dropping_waits_for_a_running_handler() {
    let finished = Arc::new(std::sync::atomic::AtomicBool::new(false));
    let flag = Arc::clone(&finished);

    let servicer = Servicer::new(move |_: u64| {
        std::thread::sleep(Duration::from_millis(50));
        flag.store(true, Ordering::SeqCst);
    })
    .expect("create the servicing path");

    servicer.submit(0).expect("accepted");
    wait_until("the handler to start", || servicer.pending() == 0);
    drop(servicer);

    // The context the handler borrows lives in the work object, so returning from
    // `drop` before the handler finished would be a use-after-free rather than
    // merely untidy.
    assert!(finished.load(Ordering::SeqCst));
}

#[test]
fn many_servicers_tear_down_concurrently_without_wedging() {
    let workers: Vec<_> = (0..16)
        .map(|_| {
            std::thread::spawn(|| {
                let (servicer, log) = logging();
                for item in 0..50 {
                    servicer.submit(item).expect("accepted");
                }
                servicer.shut_down();
                drop(servicer);
                log.peak()
            })
        })
        .collect();

    let started = Instant::now();
    for worker in workers {
        let peak = worker.join().expect("worker");
        assert!(peak <= 1, "each path serialises independently");
    }
    assert!(
        started.elapsed() < Duration::from_secs(30),
        "concurrent teardown wedged"
    );
}

#[test]
fn teardown_from_a_thread_other_than_the_creator_is_safe() {
    let (servicer, log) = logging();
    servicer.submit(1).expect("accepted");
    wait_until("the request", || log.items() == [1]);

    // The shape the monitor uses when it is dropped on a thread that did not
    // create it, which is the ordinary case for a monitor handed to a worker.
    std::thread::spawn(move || drop(servicer))
        .join()
        .expect("teardown thread");
}

/// `shut_down` waits for a producer that passed the gate but has not published.
///
/// **The window is between two locks, so nothing in the queue state can see
/// it.** `submit` releases the lock after deciding to ring and only then calls
/// `ThreadpoolWork::submit`; a producer preempted in between has committed to
/// publishing and is invisible to anyone holding the lock. Without the
/// `publishing` count, `shut_down` could close the queue and run its whole
/// cancellation inside that window, and the publication would land after it
/// returned -- contradicting its own documented guarantee that no drain can
/// start afterwards.
///
/// Asserted by handshake rather than by elapsed time: shutdown runs on its own
/// thread and must not report itself finished while the producer is parked. A
/// correct implementation is blocked on exactly that producer, so no amount of
/// looking can manufacture a failure, and an incorrect one finishes at once.
#[test]
fn shutting_down_waits_for_a_producer_that_has_not_published_yet() {
    let ran = Arc::new(AtomicUsize::new(0));
    let seen = Arc::clone(&ran);
    let servicer = Arc::new(
        Servicer::<u32>::new(move |_| {
            seen.fetch_add(1, Ordering::SeqCst);
        })
        .expect("create the servicer"),
    );

    servicer.pause_gate().arm();
    let producer = {
        let servicer = Arc::clone(&servicer);
        std::thread::spawn(move || servicer.submit(1))
    };
    servicer.pause_gate().await_parked();

    // Shutdown runs on its own thread, so the test can observe that it has NOT
    // returned while the producer is still parked.
    //
    // **No elapsed-time assertion, and the first version of this had one.**
    // Measuring that `shut_down` took at least 40ms both failed to prove what it
    // claimed -- elapsed time does not say what a call was blocked *on* -- and
    // could fail on correct code: if this thread were descheduled past the
    // releaser, shutdown would start after the publication was already unblocked
    // and return at once. The handshake below has no such direction. A correct
    // `shut_down` *cannot* set `finished` while the producer is parked, because
    // it is blocked on exactly that producer, so waiting longer can never
    // manufacture a failure; an incorrect one sets it immediately.
    let finished = Arc::new(AtomicBool::new(false));
    let shutdown = {
        let servicer = Arc::clone(&servicer);
        let finished = Arc::clone(&finished);
        std::thread::spawn(move || {
            servicer.shut_down();
            finished.store(true, Ordering::SeqCst);
        })
    };

    // Give a shutdown that does not wait every chance to finish. This bound is
    // generous because it only costs time on a *passing* run of broken code; a
    // correct implementation is parked here regardless of how long we look.
    let deadline = Instant::now() + Duration::from_millis(250);
    while Instant::now() < deadline {
        assert!(
            !finished.load(Ordering::SeqCst),
            "shut_down returned while a producer was still parked before its \
             publication, so a drain could start after it had returned and \
             promised that none could"
        );
        std::thread::sleep(Duration::from_millis(5));
    }

    // Now let the publication happen; shutdown must then complete.
    servicer.pause_gate().release();
    shutdown.join().expect("shutdown thread");
    assert!(
        finished.load(Ordering::SeqCst),
        "shut_down must return once the parked producer has published"
    );

    assert_eq!(
        servicer.rings(),
        1,
        "the producer decided to ring before the close, so the ring is counted"
    );
    assert!(!servicer.is_open(), "shut_down closed the servicing path");

    producer.join().expect("producer").expect("submit accepted");

    // Nothing may run after shut_down returned: the queue was cleared under the
    // lock, so the late callback finds it empty and the handler never fires.
    assert_eq!(
        ran.load(Ordering::SeqCst),
        0,
        "the request was discarded by the close, so no handler may run"
    );
    assert_eq!(servicer.discarded(), 1, "and it is counted as discarded");
}