kevy-alloc 6.4.0

Per-shard, mmap-backed, header-free allocator: graded size classes, span ownership, thread-local caching, and exact six-term memory accounting.
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
//! Behavioural tests for the heap.
//!
//! These are the assertions `bench/allocgate.sh` names: M3 (the
//! accounting identity), M4 (reclaim actually returns pages), M6 (an
//! exhausted class refuses rather than hands back a wild pointer).
//! Every test skips cleanly where mapping is unavailable, because a test
//! that silently passes on a target it never ran on is worse than one
//! that says it did not run.

use crate::class;
use crate::class::SPAN_BYTES;
use crate::heap::Heap;
use crate::os;

/// Skip the body on targets without anonymous mapping.
macro_rules! require_mapping {
    () => {
        if !os::available() {
            eprintln!("skipped: no anonymous mapping on this target");
            return;
        }
    };
}

#[test]
fn a_round_trip_leaves_nothing_live() {
    require_mapping!();
    let mut heap = Heap::new(0);
    let p = heap.alloc(400, 8).expect("a fresh heap can serve 400 bytes");
    let before = heap.snapshot();
    assert_eq!(before.live, 400);
    assert!(before.balanced(), "{before:?}");
    // SAFETY: `p` came from this heap with this size.
    unsafe { heap.dealloc(p, 400, 8) };
    let after = heap.snapshot();
    assert_eq!(after.live, 0);
    assert_eq!(after.rounding, 0);
    assert!(after.balanced(), "{after:?}");
}

#[test]
fn the_identity_holds_across_a_churn() {
    require_mapping!();
    let mut heap = Heap::new(0);
    let sizes = [16usize, 33, 128, 400, 999, 4096, 8192];
    let mut live: Vec<(core::ptr::NonNull<u8>, usize)> = Vec::new();
    let mut x = 0x2545_F491_4F6C_DD1Du64;
    for step in 0..4000 {
        x ^= x << 13;
        x ^= x >> 7;
        x ^= x << 17;
        let take = !(x as usize).is_multiple_of(3) || live.is_empty();
        if take {
            let size = sizes[(x as usize >> 8) % sizes.len()];
            if let Some(p) = heap.alloc(size, 8) {
                live.push((p, size));
            }
        } else {
            let ix = (x as usize >> 16) % live.len();
            let (p, size) = live.swap_remove(ix);
            // SAFETY: allocated by this heap with this size, and dropped
            // from the live set here.
            unsafe { heap.dealloc(p, size, 8) };
        }
        if step % 500 == 0 {
            let st = heap.snapshot();
            assert!(st.balanced(), "step {step}: {st:?}");
        }
    }
    let expected: u64 = live.iter().map(|(_, s)| *s as u64).sum();
    let st = heap.snapshot();
    assert_eq!(st.live, expected, "live bytes drifted from the truth");
    assert!(st.balanced(), "{st:?}");
    for (p, size) in live {
        // SAFETY: still-live allocations from this heap.
        unsafe { heap.dealloc(p, size, 8) };
    }
}

#[test]
fn slots_are_distinct_and_usable() {
    require_mapping!();
    let mut heap = Heap::new(0);
    let mut given = Vec::new();
    for step in 0..500usize {
        // Markers repeat every 256 slots, which is fine: an overlap
        // between two slots 256 apart would still show up, and any
        // overlap between neighbours certainly would.
        let i = (step % 251) as u8;
        let p = heap.alloc(48, 8).expect("48-byte allocations should succeed");
        // Write a marker through the whole slot: overlapping slots would
        // corrupt a neighbour and the read-back below would catch it.
        // SAFETY: the slot is at least 48 bytes and ours alone.
        unsafe { core::ptr::write_bytes(p.as_ptr(), i, 48) };
        given.push((p, i));
    }
    for (p, marker) in &given {
        // SAFETY: still live, written above.
        let seen = unsafe { p.as_ptr().read() };
        assert_eq!(seen, *marker, "a slot was handed out twice");
    }
    for (p, _) in given {
        // SAFETY: allocated here with this size.
        unsafe { heap.dealloc(p, 48, 8) };
    }
}

#[test]
fn a_freed_slot_comes_back() {
    require_mapping!();
    let mut heap = Heap::new(0);
    let first = heap.alloc(64, 8).unwrap();
    // SAFETY: ours, this size.
    unsafe { heap.dealloc(first, 64, 8) };
    let again = heap.alloc(64, 8).unwrap();
    assert_eq!(first, again, "the free list should hand the slot back");
    // SAFETY: ours, this size.
    unsafe { heap.dealloc(again, 64, 8) };
}

#[test]
fn large_requests_bypass_the_classes() {
    require_mapping!();
    let mut heap = Heap::new(0);
    let size = class::MAX_SMALL + 1;
    // Direct mappings are counted for the process, not the heap: a large
    // block has no segment and so no owner to route a foreign free to.
    let before = crate::large_stats();
    let p = heap.alloc(size, 8).expect("the direct-mapping path serves this");
    let during = crate::large_stats();
    assert_eq!(during.large_count, before.large_count + 1);
    assert_eq!(during.live, before.live + size as u64);
    assert!(during.balanced(), "{during:?}");
    // The heap's own snapshot covers the small path only, and balances
    // on its own — as does the large figure, so their sum does too.
    assert!(heap.snapshot().balanced());
    // SAFETY: ours, this size.
    unsafe { heap.dealloc(p, size, 8) };
    let after = crate::large_stats();
    assert_eq!(after.large_count, before.large_count);
    assert_eq!(after.live, before.live);
}

#[test]
fn m6_an_exhausted_class_refuses_instead_of_handing_back_a_wild_pointer() {
    require_mapping!();
    // The default cap is a runaway guard set past any real workload, so
    // a tighter one is used here to reach the refusal at all — see
    // `PER_CLASS_CAP` for why the inherited value was wrong.
    const CAP: u32 = 3;
    let mut heap = Heap::with_class_cap(0, CAP);
    // The class is chosen large so the number of allocations stays
    // small: 8192-byte slots give 8 per span.
    let c = class::index_of(class::MAX_SMALL, 8).unwrap();
    let per_span = class::slots_per_span(c);
    let capacity = per_span * CAP as usize;
    let mut given = Vec::new();
    for _ in 0..capacity {
        match heap.alloc(class::MAX_SMALL, 8) {
            Some(p) => given.push(p),
            None => break,
        }
    }
    assert_eq!(given.len(), capacity, "the cap should not bite before it is reached");
    assert!(
        heap.alloc(class::MAX_SMALL, 8).is_none(),
        "past the cap the answer must be None — torajs c2970b6d shipped a null instead"
    );
    let st = heap.snapshot();
    assert!(st.balanced(), "{st:?}");
    for p in given {
        // SAFETY: ours, this size.
        unsafe { heap.dealloc(p, class::MAX_SMALL, 8) };
    }
}

#[test]
fn m4_a_refused_discard_returns_no_pages_from_a_live_span() {
    require_mapping!();
    let mut heap = Heap::new(0);
    let size = 64;
    let per_span = class::slots_per_span(class::index_of(size, 8).unwrap());
    // Two spans' worth, then free only the second half of each: every
    // span keeps live slots, so the sweep takes the page-granular path
    // rather than retiring whole spans. That is the path the previous
    // test never reaches — it frees everything, so every span is empty
    // and goes to the retire branch instead.
    let count = per_span * 2;
    let mut given = Vec::with_capacity(count);
    for _ in 0..count {
        given.push(heap.alloc(size, 8).expect("filling spans"));
    }
    for (i, p) in given.into_iter().enumerate() {
        if i % 2 == 1 {
            // SAFETY: ours, this size.
            unsafe { heap.dealloc(p, size, 8) };
        }
    }
    let before = heap.snapshot();
    assert!(before.live > 0, "every span must still hold live slots");

    heap.reclaim_with(false);
    let after = heap.snapshot();
    assert!(after.balanced(), "{after:?}");
    assert_eq!(after.returned, 0, "a refused discard reported pages as returned");
    assert_eq!(after.live, before.live, "reclaim must not touch live slots");
}

#[test]
fn m4_a_machine_that_cannot_discard_reports_held_not_returned() {
    require_mapping!();
    let mut heap = Heap::new(0);
    let size = 64;
    let count = class::slots_per_span(class::index_of(size, 8).unwrap()) * 12;
    let mut given = Vec::with_capacity(count);
    for _ in 0..count {
        given.push(heap.alloc(size, 8).expect("filling spans"));
    }
    let full = heap.snapshot();
    for p in given {
        // SAFETY: ours, this size.
        unsafe { heap.dealloc(p, size, 8) };
    }
    let idle = heap.snapshot();

    // The branch this machine may not be able to take. Every Apple
    // Silicon Mac takes it — `sysconf` answers 16384 while the span-page
    // arithmetic is written at 4096 — and there page-granular reclaim
    // hands back nothing. The accounting has to say so: nothing
    // `returned`, and the emptied spans `hysteresis`, which is what
    // "retained rather than released" means.
    heap.reclaim_with(false);
    let after = heap.snapshot();
    assert!(after.balanced(), "{after:?}");
    assert_eq!(after.returned, 0, "a refused discard was reported as a return");
    assert!(
        after.hysteresis > idle.hysteresis,
        "the emptied spans went somewhere other than held: {} -> {}",
        idle.hysteresis,
        after.hysteresis
    );
    assert_eq!(
        after.predicted_resident(),
        full.predicted_resident(),
        "residency was predicted to fall on a sweep that returned nothing"
    );
}

#[test]
fn m4_emptied_spans_have_their_pages_returned() {
    require_mapping!();
    let mut heap = Heap::new(0);
    let size = 64;
    let per_span = class::slots_per_span(class::index_of(size, 8).unwrap());
    // Fill several spans' worth so there is something to reclaim beyond
    // the hysteresis the policy deliberately keeps.
    let count = per_span * 12;
    let mut given = Vec::with_capacity(count);
    for _ in 0..count {
        given.push(heap.alloc(size, 8).expect("filling spans"));
    }
    let full = heap.snapshot();
    assert!(full.balanced(), "{full:?}");
    for p in given {
        // SAFETY: ours, this size.
        unsafe { heap.dealloc(p, size, 8) };
    }
    let idle = heap.snapshot();
    assert_eq!(idle.live, 0);
    heap.reclaim();
    let after = heap.snapshot();
    assert!(after.balanced(), "{after:?}");
    // This used to read `after.hysteresis > idle.hysteresis`, with a
    // message that said "reclaim returned nothing" — the assertion said
    // hysteresis and meant returned, which is how the two terms stayed
    // swapped for the whole v5 arc. Worse, it was true on both sides of
    // the branch below, so it passed on a machine where reclaim hands
    // back nothing at all and reported that as the feature working.
    //
    // Which branch holds is a property of the machine, not of the
    // allocator, so the test asserts the one that applies and names it.
    if crate::os::page_size_matches() {
        assert!(
            after.returned > idle.returned,
            "reclaim returned nothing: returned {} -> {}",
            idle.returned,
            after.returned
        );
        // And the policy really does hold some back rather than
        // releasing everything: `EMPTY_SPAN_HYSTERESIS` spans stay
        // assigned to their class, resident, per sweep. A version that
        // released the lot would satisfy the assertion above and be an
        // mmap storm.
        assert!(
            after.hysteresis > 0,
            "the whole pool was released, so nothing absorbs the next burst"
        );
        assert!(
            after.predicted_resident() < full.predicted_resident(),
            "predicted residency did not fall: {} -> {}",
            full.predicted_resident(),
            after.predicted_resident()
        );
    } else {
        // A 16 KiB-page machine (every Apple Silicon Mac) is one: the
        // span-page arithmetic is written at `os::PAGE` = 4096, so the
        // ranges are not page-aligned, `madvise` reclaims nothing, and
        // `discard` is refused rather than lying about it. Page-granular
        // reclaim is INERT here — the emptied spans stay resident, and
        // saying so is the point of the branch.
        assert_eq!(after.returned, 0, "a refused discard reported a return");
        assert!(
            after.hysteresis > idle.hysteresis,
            "the emptied spans went somewhere other than held: {} -> {}",
            idle.hysteresis,
            after.hysteresis
        );
        // Nothing went back, so nothing should predict that it did.
        // This assertion sat outside the branch and passed here, which
        // is the third thing the swapped terms bought: `hysteresis` was
        // subtracted from the prediction, so held memory read as
        // released, so M4 — the gate for the property the whole
        // experiment rests on — was green on a machine where the
        // property does not hold at all.
        assert_eq!(
            after.predicted_resident(),
            full.predicted_resident(),
            "residency was predicted to fall on a machine that returned nothing"
        );
    }
}

#[test]
fn reclaimed_spans_are_reusable_and_start_clean() {
    require_mapping!();
    let mut heap = Heap::new(0);
    let size = 64;
    let per_span = class::slots_per_span(class::index_of(size, 8).unwrap());
    let mut given: Vec<_> =
        (0..per_span * 12).map(|_| heap.alloc(size, 8).expect("filling spans")).collect();
    // Poison every byte so a reclaimed-and-reused span that failed to
    // reset its cursors would hand back recognisable garbage.
    for p in &given {
        // SAFETY: live slots of at least `size` bytes.
        unsafe { core::ptr::write_bytes(p.as_ptr(), 0xAB, size) };
    }
    for p in given.drain(..) {
        // SAFETY: ours, this size.
        unsafe { heap.dealloc(p, size, 8) };
    }
    heap.reclaim();
    let mut again = Vec::new();
    for _ in 0..per_span * 4 {
        let p = heap.alloc(size, 8).expect("reclaimed spans must be reusable");
        // SAFETY: freshly handed out, ours.
        unsafe { core::ptr::write_bytes(p.as_ptr(), 0x11, size) };
        again.push(p);
    }
    let st = heap.snapshot();
    assert!(st.balanced(), "{st:?}");
    for p in again {
        // SAFETY: ours, this size.
        unsafe { heap.dealloc(p, size, 8) };
    }
}

#[test]
fn spans_are_page_multiples_so_discard_is_legal() {
    // madvise refuses ranges that are not page-aligned, and a silent
    // refusal would make M4 look like a policy choice rather than a bug.
    assert_eq!(SPAN_BYTES % os::PAGE, 0);
}

/// Resident pages, from `/proc/self/statm` field 2 (in pages).
#[cfg(target_os = "linux")]
fn rss_bytes() -> u64 {
    let s = std::fs::read_to_string("/proc/self/statm").expect("procfs");
    let pages: u64 = s.split_whitespace().nth(1).unwrap().parse().unwrap();
    pages * os::PAGE as u64
}

/// M4 in its real form: the kernel's own resident count must fall.
///
/// The model-level test above checks our prediction; this checks the
/// thing the prediction is about. Linux only, and deliberately so —
/// `MADV_DONTNEED` drops pages outright, while macOS's `MADV_FREE` only
/// marks them reclaimable, so a passing assertion there would mean
/// nothing. glibc's brk arena cannot pass this at any page count, which
/// is the whole reason this crate exists.
#[cfg(target_os = "linux")]
#[test]
fn m4_the_kernel_agrees_that_pages_came_back() {
    require_mapping!();
    let mut heap = Heap::new(0);
    let size = 64;
    let per_span = class::slots_per_span(class::index_of(size, 8).unwrap());
    // Enough spans that the returned bytes clear ordinary process noise.
    let count = per_span * 200;
    let mut given = Vec::with_capacity(count);
    for _ in 0..count {
        let p = heap.alloc(size, 8).expect("filling spans");
        // Touch it: untouched pages are not resident, and a test that
        // never made them resident could not observe them leaving.
        // SAFETY: a live slot of at least `size` bytes.
        unsafe { core::ptr::write_bytes(p.as_ptr(), 0x5A, size) };
        given.push(p);
    }
    let peak = rss_bytes();
    for p in given {
        // SAFETY: ours, this size and alignment.
        unsafe { heap.dealloc(p, size, 8) };
    }
    heap.reclaim();
    let after = rss_bytes();
    let touched = (count * size) as u64;
    assert!(
        after + touched / 2 < peak,
        "RSS barely moved: {peak} -> {after} after freeing {touched} bytes across {} spans",
        count / per_span
    );
}

#[test]
fn spans_with_room_are_reused_before_new_ones_are_claimed() {
    require_mapping!();
    let mut heap = Heap::new(0);
    let size = 64;
    let c = class::index_of(size, 8).unwrap();
    let per_span = class::slots_per_span(c);

    // Fill one span exactly, then free every slot in it. The freed slots
    // land on *that* span's list — not on the current span's, because
    // the next allocation will have moved on.
    let first: Vec<_> =
        (0..per_span).map(|_| heap.alloc(size, 8).expect("filling the first span")).collect();
    // One more allocation forces a second span to be claimed.
    let straggler = heap.alloc(size, 8).expect("second span");
    for p in first {
        // SAFETY: ours, this size and alignment.
        unsafe { heap.dealloc(p, size, 8) };
    }
    let before = heap.snapshot();

    // Now allocate a whole span's worth again. Every one of these can be
    // served from the emptied first span.
    let again: Vec<_> =
        (0..per_span).map(|_| heap.alloc(size, 8).expect("reusing the emptied span")).collect();
    let after = heap.snapshot();

    assert_eq!(
        after.spans_assigned, before.spans_assigned,
        "a third span was claimed while the first sat empty — reusable \
         spans must be adopted before new ones are taken"
    );
    assert!(after.balanced(), "{after:?}");

    for p in again {
        // SAFETY: ours, this size and alignment.
        unsafe { heap.dealloc(p, size, 8) };
    }
    // SAFETY: ours, this size and alignment.
    unsafe { heap.dealloc(straggler, size, 8) };
}

#[test]
fn a_foreign_free_leaves_the_owner_balanced_before_and_after_draining() {
    require_mapping!();
    // Two heaps in one thread stand in for two shards: the second has a
    // different identity, so a free through it takes the foreign path
    // exactly as a cross-thread free would — without the timing.
    //
    // This is the regression test for the defect the standard library
    // found the first time it ran on this allocator: the freeing side
    // was subtracting from *its own* counters bytes that had been
    // counted on the owner's, so its totals went negative. Single-thread
    // tests could not reach it, and the byte identity is what catches
    // it, so the assertion is on the identity rather than on a symptom.
    let mut owner = Heap::new(1);
    let mut other = Heap::new(2);

    let size = 400;
    let held: Vec<_> =
        (0..1_000).map(|_| owner.alloc(size, 8).expect("owner serves these")).collect();
    let full = owner.snapshot();
    assert_eq!(full.live, 1_000 * size as u64);
    assert!(full.balanced(), "{full:?}");

    // Hand every one of them back through the wrong heap. Foreign frees
    // land in the freeing heap's outbound ring and cross to the owner's
    // segment only in batches — that amortisation is the whole point
    // (M1: the per-op version cost cross-shard KV 18–39 %).
    for p in held {
        // SAFETY: allocated with this size and alignment; `dealloc`
        // ships a foreign slot home in batches.
        unsafe { other.dealloc(p, size, 8) };
    }

    // The other heap allocated nothing, so nothing may have moved on it.
    let intruder = other.snapshot();
    assert_eq!(intruder.live, 0, "the freeing heap counted bytes it never handed out");
    assert!(intruder.balanced(), "{intruder:?}");

    // Before the freeing side flushes, a partial batch may still sit in
    // its ring. Those slots are still counted by the owner as live —
    // the documented staleness window — and the identity holds exactly
    // through it on both sides.
    let slot = class::size_of(class::index_of(size, 8).unwrap()) as u64;
    let mid = owner.snapshot();
    assert!(mid.balanced(), "{mid:?}");
    let shipped = mid.cache / slot;
    let pending = mid.live / size as u64;
    assert_eq!(
        shipped + pending,
        1_000,
        "every slot is either parked (shipped) or still covered as live (pending): {mid:?}"
    );
    assert_eq!(
        mid.rounding,
        pending * (slot - size as u64),
        "pending rounding rides with pending live"
    );

    // The freeing side's tick flushes the ring; now everything is
    // parked on the owner.
    other.reclaim();
    let parked = owner.snapshot();
    assert_eq!(parked.cache, 1_000 * slot, "slot bytes should be parked");
    assert_eq!(parked.live, 0, "the owner should no longer count them as live");
    assert!(parked.balanced(), "{parked:?}");

    owner.drain_foreign();
    let settled = owner.snapshot();
    assert_eq!(settled.cache, 0, "the list should be empty after draining");
    assert_eq!(settled.live, 0);
    assert_eq!(settled.rounding, 0);
    assert!(settled.balanced(), "{settled:?}");
}

#[test]
fn v2_pages_return_while_the_span_still_lives() {
    require_mapping!();
    // The measurement that forced v2: a span of the 416-byte class holds
    // 157 slots, and under the whole-span rule all 157 had to die before
    // one page came back. Here the span keeps survivors — only the
    // slots overlapping its LAST page stay live — and the pages below
    // them must come back anyway.
    let mut heap = Heap::new(0);
    let size = 400;
    let slot = class::size_of(class::index_of(size, 8).unwrap());
    let per_span = class::slots_per_span(class::index_of(size, 8).unwrap());
    let given: Vec<_> =
        (0..per_span).map(|_| heap.alloc(size, 8).expect("filling one span")).collect();
    // Keep everything whose slot touches the last page; free the rest.
    let last_page_start = (crate::pagemap::PAGES_PER_SPAN - 1) * os::PAGE;
    let mut survivors = Vec::new();
    for (i, p) in given.into_iter().enumerate() {
        if (i + 1) * slot > last_page_start {
            survivors.push(p);
        } else {
            // SAFETY: ours, this size and alignment.
            unsafe { heap.dealloc(p, size, 8) };
        }
    }
    assert!(!survivors.is_empty(), "the last page must hold live slots");
    let before = heap.snapshot();
    assert_eq!(before.returned, 0, "nothing returned before the sweep");
    heap.reclaim();
    let after = heap.snapshot();
    assert!(after.balanced(), "{after:?}");
    // `returned` is the accounting, not the kernel. On a system whose
    // page size is not `os::PAGE` the reclaim path refuses (it would
    // otherwise report pages it cannot return), so the behaviour under
    // test is not available here — and this assertion, which reads the
    // accounting, would have passed on such a machine right up until
    // the refusal landed. Say which case ran rather than pass either
    // way.
    if os::page_size_matches() {
        assert!(
            after.returned > 0,
            "a span with survivors returned nothing — the v1 failure, back: {after:?}"
        );
    } else {
        assert_eq!(
            after.returned, 0,
            "the reclaim path must not account pages it refused to return: {after:?}"
        );
        eprintln!(
            "NOT EXERCISED: page return needs a {}-byte page; this system disagrees",
            os::PAGE
        );
    }
    if !os::page_size_matches() {
        // The rest of this test is about how much came back, and nothing
        // did. Returning here is the honest end — the alternative is a
        // second reading of the same accounting under a different name.
        return;
    }
    // Almost all of the span's free bytes should be returned: only the
    // pages pinned by survivors (and slot-straddling edges) stay.
    assert!(
        after.returned > (crate::pagemap::PAGES_PER_SPAN as u64 / 2) * os::PAGE as u64,
        "returned only {} bytes of a {}-byte span",
        after.returned,
        class::SPAN_BYTES
    );
    assert!(
        after.predicted_resident() < before.predicted_resident(),
        "prediction did not fall: {} -> {}",
        before.predicted_resident(),
        after.predicted_resident()
    );
    for p in survivors {
        // SAFETY: still live, ours.
        unsafe { heap.dealloc(p, size, 8) };
    }
}

#[test]
fn v2_densification_migrates_free_space_into_whole_pages() {
    require_mapping!();
    // Lowest-first allocation is a claim about CHURN, not about one
    // round: any single kill-pattern can leave a survivor pinning every
    // page (the first version of this test did exactly that, freeing
    // alternate slots — the untouched alternates pinned everything).
    // What densification promises is that as churn continues, deaths
    // eventually visit the high slots and their replacements all land
    // low, so free space migrates upward into whole pages. A LIFO free
    // list refills wherever death last struck and promises nothing.
    let mut heap = Heap::new(0);
    let size = 400;
    let c = class::index_of(size, 8).unwrap();
    let per_span = class::slots_per_span(c);
    let mut live: Vec<_> = (0..per_span * 4).map(|_| heap.alloc(size, 8).expect("fill")).collect();
    // Rounds of "half die, a quarter arrive": net shrinkage under a
    // death order that decorrelates from slot position as reallocated
    // (low) slots mix into the vec.
    for _ in 0..8 {
        let mut i = 0usize;
        let mut freed = 0usize;
        live.retain(|p| {
            i += 1;
            if i.is_multiple_of(2) {
                // SAFETY: ours, this size and alignment.
                unsafe { heap.dealloc(*p, size, 8) };
                freed += 1;
                false
            } else {
                true
            }
        });
        for _ in 0..freed / 4 {
            live.push(heap.alloc(size, 8).expect("refill"));
        }
    }
    heap.reclaim();
    let st = heap.snapshot();
    assert!(st.balanced(), "{st:?}");
    // Same split as `v2_pages_return_while_the_span_still_lives`: this
    // reads the accounting, and the accounting is only about the kernel
    // on a system with the page size the geometry was built for.
    if os::page_size_matches() {
        assert!(
            st.returned > 0,
            "an interleaved churn produced no returnable page — densification is not happening: {st:?}"
        );
    } else {
        assert_eq!(st.returned, 0, "accounted a page the reclaim path refused: {st:?}");
        eprintln!("NOT EXERCISED: densification's page return needs a {}-byte page", os::PAGE);
    }
    for p in live {
        // SAFETY: ours.
        unsafe { heap.dealloc(p, size, 8) };
    }
}

/// The kernel's own verdict on v2, Linux only (macOS MADV_FREE gives no
/// prompt guarantee — same reasoning as the whole-span M4 test).
#[cfg(target_os = "linux")]
#[test]
fn v2_the_kernel_reclaims_pages_from_spans_with_survivors() {
    require_mapping!();
    let mut heap = Heap::new(0);
    let size = 400;
    let c = class::index_of(size, 8).unwrap();
    let slot = class::size_of(c);
    let per_span = class::slots_per_span(c);
    let spans = 200;
    let mut given = Vec::with_capacity(per_span * spans);
    for _ in 0..per_span * spans {
        let p = heap.alloc(size, 8).expect("fill");
        // Touch, so the pages are resident and their leaving is visible.
        // SAFETY: live slot of at least `size` bytes.
        unsafe { core::ptr::write_bytes(p.as_ptr(), 0x5A, size) };
        given.push(p);
    }
    let peak = rss_bytes();
    // Free all but each span's last-page slots — every span keeps
    // survivors, so the v1 whole-span rule would return NOTHING here.
    let last_page_start = (crate::pagemap::PAGES_PER_SPAN - 1) * os::PAGE;
    let mut survivors = Vec::new();
    for (n, p) in given.into_iter().enumerate() {
        let in_span = n % per_span;
        if (in_span + 1) * slot > last_page_start {
            survivors.push(p);
        } else {
            // SAFETY: ours, this size and alignment.
            unsafe { heap.dealloc(p, size, 8) };
        }
    }
    heap.reclaim();
    let after = rss_bytes();
    let st = heap.snapshot();
    assert!(st.balanced(), "{st:?}");
    assert!(
        after + st.returned / 2 < peak,
        "kernel RSS barely moved with survivors pinning every span: {peak} -> {after} (returned={})",
        st.returned
    );
    for p in survivors {
        // SAFETY: ours.
        unsafe { heap.dealloc(p, size, 8) };
    }
}

#[test]
fn spliced_chains_from_two_freeing_heaps_arrive_complete() {
    require_mapping!();
    // Two foreign heaps free into the same owner concurrently-ish: their
    // batches splice onto one segment list. Nothing may be lost, and
    // after the owner drains, every slot must be reusable again.
    let mut owner = Heap::new(1);
    let mut b = Heap::new(2);
    let mut c = Heap::new(3);
    let size = 400;
    let n = 600; // several batches' worth from each side
    let held: Vec<_> =
        (0..n * 2).map(|_| owner.alloc(size, 8).expect("owner serves these")).collect();
    for (i, p) in held.into_iter().enumerate() {
        // SAFETY: ours, this size and alignment; alternating freers.
        unsafe {
            if i.is_multiple_of(2) {
                b.dealloc(p, size, 8);
            } else {
                c.dealloc(p, size, 8);
            }
        }
    }
    b.reclaim();
    c.reclaim();
    let parked = owner.snapshot();
    let slot = class::size_of(class::index_of(size, 8).unwrap()) as u64;
    assert_eq!(parked.cache, (n * 2) as u64 * slot, "a spliced batch went missing: {parked:?}");
    assert!(parked.balanced(), "{parked:?}");

    owner.drain_foreign();
    let settled = owner.snapshot();
    assert_eq!(settled.live, 0);
    assert_eq!(settled.cache, 0);
    assert!(settled.balanced(), "{settled:?}");
    // And the memory is genuinely reusable: refill without new segments.
    let before_spans = settled.spans_assigned;
    let again: Vec<_> =
        (0..n * 2).map(|_| owner.alloc(size, 8).expect("drained slots serve again")).collect();
    assert_eq!(owner.snapshot().spans_assigned, before_spans, "drained slots were not reused");
    for p in again {
        // SAFETY: ours.
        unsafe { owner.dealloc(p, size, 8) };
    }
}

/// The collection-write shape the claimed word exists for: short-lived
/// small allocations recycling inside one word, heap-locally. The
/// observable contract: pointers stay lowest-word-stable (position
/// awareness), the accounting identity balances mid-claim, and a
/// retire hands unused bits back so densification sees them.
#[test]
fn claimed_word_recycles_locally_and_retires_honestly() {
    let mut h = Heap::new(1);
    // Churn one size class hard: alloc/free pairs like a hash-node
    // path. Every pointer must come from the same low word while the
    // claim holds (position-aware recycling, not LIFO wander).
    let mut last = None;
    for _ in 0..1000 {
        let p = h.alloc(48, 8).unwrap();
        if let Some(prev) = last {
            assert_eq!(p, prev, "short-lived churn must reuse the same lowest slot");
        }
        last = Some(p);
        // SAFETY: `p` came from `h.alloc` with this same size and alignment and has not
        // been freed yet, which is `dealloc`'s contract.
        unsafe { h.dealloc(p, 48, 8) };
    }
    // Identity balances with a claim in flight (no flush).
    let st = h.snapshot();
    assert_eq!(st.live, 0, "everything was freed");
    assert!(st.balanced(), "{st:?}");
    // After a flush the span sees the truth and reclaim can sweep.
    h.flush_claims();
    let st = h.snapshot();
    assert!(st.balanced(), "{st:?}");
    h.reclaim();
    let st = h.snapshot();
    assert!(st.balanced(), "{st:?}");
}

/// Filling past one word forces claim → refill → claim on the next
/// word; freeing everything then reclaiming must return the pages —
/// the claim must never strand occupancy.
#[test]
fn claims_span_words_and_never_strand_occupancy() {
    let mut h = Heap::new(1);
    let mut ptrs = Vec::new();
    for _ in 0..200 {
        ptrs.push(h.alloc(64, 8).unwrap()); // > 64 slots ⇒ multiple words
    }
    for p in ptrs.drain(..) {
        // SAFETY: `p` came from `h.alloc` with this same size and alignment and has not
        // been freed yet, which is `dealloc`'s contract.
        unsafe { h.dealloc(p, 64, 8) };
    }
    h.flush_claims();
    let st = h.snapshot();
    assert_eq!(st.live, 0);
    assert!(st.balanced(), "{st:?}");
    h.reclaim();
    assert!(h.snapshot().balanced());
}

// ── the reclaim path may not report pages it did not return ──────────

/// `os::PAGE` is a constant and the running system's page size is not.
///
/// This machine reports 16384 and the bench box reports 4096, and the
/// reclaim path computes `madvise` ranges at 4096 granularity. On the
/// larger page those ranges are not page-aligned; macOS answers 0
/// regardless and reclaims nothing, so a reclaim that ran anyway would
/// mark the pages discarded and count them in `returned` for memory the
/// kernel still holds.
///
/// This asserts the check exists and agrees with the system, which is
/// the only part that can be checked from inside the process. What it
/// cannot check is whether `madvise` did anything — that needs RSS, and
/// `bench/allocgate-mem.sh` is where that lives.
#[test]
fn the_page_size_check_agrees_with_the_system() {
    let matches = crate::os::page_size_matches();
    #[cfg(any(target_os = "linux", target_os = "macos"))]
    {
        // An independent reading, not the same call: the constant this
        // crate was built around is 4096, and a system that disagrees
        // must make `page_size_matches` false.
        let out = std::process::Command::new("getconf").arg("PAGE_SIZE").output();
        if let Ok(o) = out
            && let Ok(text) = String::from_utf8(o.stdout)
            && let Ok(sys) = text.trim().parse::<usize>()
        {
            assert_eq!(
                matches,
                sys == crate::os::PAGE,
                "getconf says {sys}, the crate assumes {}, and page_size_matches() says {matches}",
                crate::os::PAGE
            );
            return;
        }
    }
    // No independent reading available: assert only that the answer is
    // stable, since it is cached and a flapping answer would be worse
    // than either value.
    assert_eq!(matches, crate::os::page_size_matches());
}