aboutsummaryrefslogtreecommitdiff
path: root/libbuild2/scheduler.cxx
blob: 03842ecc1125ff28f86b5b4cfbaf0c3e14ab5abc (plain)
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
// file      : libbuild2/scheduler.cxx -*- C++ -*-
// license   : MIT; see accompanying LICENSE file

#include <libbuild2/scheduler.hxx>

#if defined(__linux__) || defined(__FreeBSD__) || defined(__NetBSD__) || defined(__APPLE__)
#  include <pthread.h>
#  ifdef __FreeBSD__
#    include <pthread_np.h> // pthread_attr_get_np() (in <pthread.h> on NetBSD)
#  endif
#endif

#ifndef _WIN32
#  include <thread> // this_thread::sleep_for()
#else
#  include <libbutl/win32-utility.hxx>

#  include <chrono>
#endif

#include <cerrno>

#include <libbuild2/diagnostics.hxx>

using namespace std;

namespace build2
{
  // TLS cache of thread's task queue.
  //
  // Note that scheduler::task_queue struct is private.
  //
  static
#ifdef __cpp_thread_local
  thread_local
#else
  __thread
#endif
  void* scheduler_queue = nullptr;

  scheduler::task_queue* scheduler::
  queue () noexcept
  {
    return static_cast<scheduler::task_queue*> (scheduler_queue);
  }

  void scheduler::
  queue (scheduler::task_queue* q) noexcept
  {
    scheduler_queue = q;
  }

  size_t scheduler::
  wait (size_t start_count, const atomic_count& task_count, work_queue wq)
  {
    // Note that task_count is a synchronization point.
    //
    size_t tc;

    if ((tc = task_count.load (memory_order_acquire)) <= start_count)
      return tc;

    assert (max_active_ != 1); // Serial execution, nobody to wait for.

    // See if we can run some of our own tasks.
    //
    if (wq != work_none)
    {
      // If we are waiting on someone else's task count then there migh still
      // be no queue (set by async()).
      //
      if (task_queue* tq = queue ())
      {
        for (lock ql (tq->mutex); !tq->shutdown && !empty_back (*tq); )
        {
          pop_back (*tq, ql);

          if (wq == work_one)
          {
            if ((tc = task_count.load (memory_order_acquire)) <= start_count)
              return tc;
          }
        }

        // Note that empty task queue doesn't automatically mean the task
        // count has been decremented (some might still be executing
        // asynchronously).
        //
        if ((tc = task_count.load (memory_order_acquire)) <= start_count)
          return tc;
      }
    }

    return suspend (start_count, task_count);
  }

  void scheduler::
  deactivate (bool external)
  {
    if (max_active_ == 1) // Serial execution.
      return;

    lock l (mutex_);

    active_--;
    waiting_++;
    if (external)
      external_++;
    progress_.fetch_add (1, memory_order_relaxed);

    if (waiting_ > stat_max_waiters_)
      stat_max_waiters_ = waiting_;

    // A spare active thread has become available. If there are ready masters
    // or eager helpers, wake someone up.
    //
    if (ready_ != 0)
    {
      ready_condv_.notify_one ();
    }
    else if (queued_task_count_.load (std::memory_order_consume) != 0 &&
             activate_helper (l))
      ;
    else if (active_ == 0 && external_ == 0)
    {
      // Note that we tried to handle this directly in this thread but that
      // wouldn't work for the phase lock case where we call deactivate and
      // then go wait on a condition variable: we would be doing deadlock
      // detection while holding the lock that prevents other threads from
      // making progress! So it has to be a separate monitoring thread.
      //
      dead_condv_.notify_one ();
    }
  }

  void scheduler::
  activate (bool external, bool collision)
  {
    if (max_active_ == 1) // Serial execution.
      return;

    lock l (mutex_);

    if (collision)
      stat_wait_collisions_++;

    // If we have spare active threads, then become active. Otherwise we enter
    // the ready queue.
    //
    if (external)
      external_--;
    waiting_--;
    ready_++;
    progress_.fetch_add (1, memory_order_relaxed);

    while (!shutdown_ && active_ >= max_active_)
      ready_condv_.wait (l);

    ready_--;
    active_++;
    progress_.fetch_add (1, memory_order_relaxed);

    if (shutdown_)
      throw_generic_error (ECANCELED);
  }

  void scheduler::
  sleep (const duration& d)
  {
    deactivate (true /* external */);
    active_sleep (d);
    activate (true /* external */);
  }

  void scheduler::
  active_sleep (const duration& d)
  {
    // MinGW GCC 4.9 doesn't implement this_thread so use Win32 Sleep().
    //
#ifndef _WIN32
    this_thread::sleep_for (d);
#else
    using namespace chrono;
    Sleep (static_cast<DWORD> (duration_cast<milliseconds> (d).count ()));
#endif
  }

  size_t scheduler::
  allocate (size_t n)
  {
    if (max_active_ == 1) // Serial execution.
      return 0;

    lock l (mutex_);

    if (active_ < max_active_)
    {
      size_t d (max_active_ - active_);
      if (n == 0 || d < n)
        n = d;
      active_ -= n;
      return n;
    }
    else
      return 0;
  }

  void scheduler::
  deallocate (size_t n)
  {
    if (max_active_ == 1) // Serial execution.
      return;

    lock l (mutex_);
    active_ += n;
  }

  size_t scheduler::
  suspend (size_t start_count, const atomic_count& task_count)
  {
    wait_slot& s (
      wait_queue_[
        hash<const atomic_count*> () (&task_count) % wait_queue_size_]);

    // This thread is no longer active.
    //
    deactivate (false /* external */);

    // Note that the task count is checked while holding the lock. We also
    // have to notify while holding the lock (see resume()). The aim here
    // is not to end up with a notification that happens between the check
    // and the wait.
    //
    size_t tc (0);
    bool collision;
    {
      lock l (s.mutex);

      // We have a collision if there is already a waiter for a different
      // task count.
      //
      collision = (s.waiters++ != 0 && s.task_count != &task_count);

      // This is nuanced: we want to always have the task count of the last
      // thread to join the queue. Otherwise, if threads are leaving and
      // joining the queue simultaneously, we may end up with a task count of
      // a thread group that is no longer waiting.
      //
      s.task_count = &task_count;

      // We could probably relax the atomic access since we use a mutex for
      // synchronization though this has a different tradeoff (calling wait
      // because we don't see the count).
      //
      while (!(s.shutdown ||
               (tc = task_count.load (memory_order_acquire)) <= start_count))
        s.condv.wait (l);

      s.waiters--;
    }

    // This thread is no longer waiting.
    //
    activate (false /* external */, collision);

    return tc;
  }

  void scheduler::
  resume (const atomic_count& tc)
  {
    if (max_active_ == 1) // Serial execution, nobody to wakeup.
      return;

    wait_slot& s (
      wait_queue_[hash<const atomic_count*> () (&tc) % wait_queue_size_]);

    // See suspend() for why we must hold the lock.
    //
    lock l (s.mutex);

    if (s.waiters != 0)
      s.condv.notify_all ();
  }

  scheduler::
  ~scheduler ()
  {
    try { shutdown (); } catch (system_error&) {}
  }

  auto scheduler::
  wait_idle () -> lock
  {
    lock l (mutex_);

    assert (waiting_ == 0);
    assert (ready_ == 0);

    while (active_ != init_active_ || starting_ != 0)
    {
      l.unlock ();
      this_thread::yield ();
      l.lock ();
    }

    return l;
  }

  size_t scheduler::
  shard_size (size_t mul, size_t div) const
  {
    size_t n (max_threads_ == 1 ? 0 : max_threads_ * mul / div / 4);

    // Return true if the specified number is prime.
    //
    auto prime = [] (size_t n)
    {
      // Check whether any number from 2 to the square root of n evenly
      // divides n, and return false if that's the case.
      //
      // While iterating i till sqrt(n) would be more efficient let's do
      // without floating arithmetic, since it doesn't make much difference
      // for the numbers we evaluate. Note that checking for i <= n / 2 is
      // just as efficient for small numbers but degrades much faster for
      // bigger numbers.
      //
      for (size_t i (2); i * i <= n; ++i)
      {
        if (n % i == 0)
          return false;
      }

      return n > 1;
    };

    // Return a prime number that is not less than the specified number.
    //
    auto next_prime = [&prime] (size_t n)
    {
      // Note that there is always a prime number in [n, 2 * n).
      //
      for (;; ++n)
      {
        if (prime (n))
          return n;
      }
    };

    // Experience shows that we want something close to 2x for small numbers,
    // then reduce to 1.5x in-between, and 1x for large ones.
    //
    // Note that Intel Xeons are all over the map when it comes to cores (6,
    // 8, 10, 12, 14, 16, 18, 20, 22).
    //
    // HW threads x arch-bits (see max_threads below).
    //
    return n ==  0 ? 1                      : // serial
           n ==  1 ? 3                      : // odd prime number
           n <= 16 ? next_prime (n * 2)     : // {2, 4} x 4, 2 x 8
           n <= 80 ? next_prime (n * 3 / 2) : // {4, 6, 8, 10} x 8
                     next_prime (n)         ; // {12, 14, 16, ...} x 8, ...
  }

  void scheduler::
  startup (size_t max_active,
           size_t init_active,
           size_t max_threads,
           size_t queue_depth,
           optional<size_t> max_stack)
  {
    // Lock the mutex to make sure our changes are visible in (other) active
    // threads.
    //
    lock l (mutex_);

    max_stack_ = max_stack;

    // Use 8x max_active on 32-bit and 32x max_active on 64-bit. Unless we
    // were asked to run serially.
    //
    if (max_threads == 0)
      max_threads = (max_active == 1    ? 1 :
                     sizeof (void*) < 8 ? 8 : 32) * max_active;

    assert (shutdown_ &&
            init_active != 0 &&
            init_active <= max_active &&
            max_active <= max_threads);

    active_ = init_active_ = init_active;
    max_active_ = orig_max_active_ = max_active;
    max_threads_ = max_threads;

    // This value should be proportional to the amount of hardware concurrency
    // we have (no use queing things up if helpers cannot keep up). Note that
    // the queue entry is quite sizable.
    //
    // The relationship is as follows: we want to have a deeper queue if the
    // tasks take long (e.g., compilation) and shorter if they are quick (e.g,
    // test execution). If the tasks are quick then the synchronization
    // overhead required for queuing/dequeuing things starts to dominate.
    //
    task_queue_depth_ = queue_depth != 0
      ? queue_depth
      : max_active * 8;

    queued_task_count_.store (0, memory_order_relaxed);

    if ((wait_queue_size_ = max_threads == 1 ? 0 : shard_size ()) != 0)
      wait_queue_.reset (new wait_slot[wait_queue_size_]);

    // Reset counters.
    //
    stat_max_waiters_     = 0;
    stat_wait_collisions_ = 0;

    progress_.store (0, memory_order_relaxed);

    for (size_t i (0); i != wait_queue_size_; ++i)
      wait_queue_[i].shutdown = false;

    shutdown_ = false;

    if (max_active_ != 1)
      dead_thread_ = thread (deadlock_monitor, this);
  }

  size_t scheduler::
  tune (size_t max_active)
  {
    // Note that if we tune a parallel scheduler to run serially, we will
    // still have the deadlock monitoring thread running.

    // With multiple initial active threads we will need to make changes to
    // max_active_ visible to other threads and which we currently say can be
    // accessed between startup and shutdown without a lock.
    //
    assert (init_active_ == 1);

    if (max_active == 0)
      max_active = orig_max_active_;

    if (max_active != max_active_)
    {
      assert (max_active >= init_active_ &&
              max_active <= orig_max_active_);

      // The scheduler must not be active though some threads might still be
      // comming off from finishing a task. So we busy-wait for them.
      //
      lock l (wait_idle ());

      swap (max_active_, max_active);
    }

    return max_active == orig_max_active_ ? 0 : max_active;
  }

  auto scheduler::
  shutdown () -> stat
  {
    // Our overall approach to shutdown is not to try and stop everything as
    // quickly as possible but rather to avoid performing any tasks. This
    // avoids having code littered with if(shutdown) on every other line.

    stat r;
    lock l (mutex_);

    if (!shutdown_)
    {
      // Collect statistics.
      //
      r.thread_helpers = helpers_;

      // Signal shutdown.
      //
      shutdown_ = true;

      for (size_t i (0); i != wait_queue_size_; ++i)
      {
        wait_slot& ws (wait_queue_[i]);
        lock l (ws.mutex);
        ws.shutdown = true;
      }

      for (task_queue& tq: task_queues_)
      {
        lock ql (tq.mutex);
        r.task_queue_full += tq.stat_full;
        tq.shutdown = true;
      }

      // Wait for all the helpers to terminate waking up any thread that
      // sleeps.
      //
      while (helpers_ != 0)
      {
        bool i (idle_ != 0);
        bool r (ready_ != 0);
        bool w (waiting_ != 0);

        l.unlock ();

        if (i)
          idle_condv_.notify_all ();

        if (r)
          ready_condv_.notify_all ();

        if (w)
          for (size_t i (0); i != wait_queue_size_; ++i)
            wait_queue_[i].condv.notify_all ();

        this_thread::yield ();
        l.lock ();
      }

      assert (external_ == 0);

      // Wait for the deadlock monitor (the only remaining thread).
      //
      if (orig_max_active_ != 1) // See tune() for why not max_active_.
      {
        l.unlock ();
        dead_condv_.notify_one ();
        dead_thread_.join ();
      }

      // Free the memory.
      //
      wait_queue_.reset ();
      task_queues_.clear ();

      r.thread_max_active     = orig_max_active_;
      r.thread_max_total      = max_threads_;
      r.thread_max_waiting    = stat_max_waiters_;

      r.task_queue_depth      = task_queue_depth_;
      r.task_queue_remain     = queued_task_count_.load (memory_order_consume);

      r.wait_queue_slots      = wait_queue_size_;
      r.wait_queue_collisions = stat_wait_collisions_;
    }

    return r;
  }

  scheduler::monitor_guard scheduler::
  monitor (atomic_count& c, size_t t, function<size_t (size_t)> f)
  {
    assert (monitor_count_ == nullptr && t != 0);

    // While the scheduler must not be active, some threads might still be
    // comming off from finishing a task and trying to report progress. So we
    // busy-wait for them (also in ~monitor_guard()).
    //
    lock l (wait_idle ());

    monitor_count_ = &c;
    monitor_tshold_.store (t, memory_order_relaxed);
    monitor_init_ = c.load (memory_order_relaxed);
    monitor_func_ = move (f);

    return monitor_guard (this);
  }

  bool scheduler::
  activate_helper (lock& l)
  {
    if (shutdown_)
      return false;

    if (idle_ != 0)
    {
      idle_condv_.notify_one ();
    }
    //
    // Ignore the max_threads value if we have queued tasks but no active
    // threads. This means everyone is waiting for something to happen but
    // nobody is doing anything (e.g., working the queues). This, for
    // example, can happen if a thread waits for a task that is in its queue
    // but is below the mark.
    //
    else if (init_active_ + helpers_ < max_threads_ ||
             (active_ == 0 &&
              queued_task_count_.load (memory_order_consume) != 0))
    {
      create_helper (l);
    }
    else
      return false;

    return true;
  }

  void scheduler::
  create_helper (lock& l)
  {
    helpers_++;
    starting_++;
    l.unlock ();

    // Restore the counters if the thread creation fails.
    //
    struct guard
    {
      lock* l;
      size_t& h;
      size_t& s;

      ~guard () {if (l != nullptr) {l->lock (); h--; s--;}}

    } g {&l, helpers_, starting_};

    // For some platforms/compilers the default stack size for newly created
    // threads may differ from that of the main thread. Here are the default
    // main/new thread sizes (in KB) for some of them:
    //
    // Linux   :   8192 / 8196
    // FreeBSD : 524288 / 2048
    // MacOS   :   8192 /  512
    // MinGW   :   2048 / 2048
    // VC      :   1024 / 1024
    //
    // Provided the main thread size is less-equal than
    // LIBBUILD2_SANE_STACK_SIZE (which defaults to
    // sizeof(void*)*LIBBUILD2_DEFAULT_STACK_SIZE), we make sure that the new
    // thread stack is the same as for the main thread. Otherwise, we cap it
    // at LIBBUILD2_DEFAULT_STACK_SIZE (default: 8MB). This can also be
    // overridden at runtime with the --max-stack build2 driver option
    // (remember to update its documentation of changing anything here).
    //
    // On Windows the stack size is the same for all threads and is customized
    // at the linking stage (see build2/buildfile). Thus neither *_STACK_SIZE
    // nor --max-stack have any effect here.
    //
    // On Linux, *BSD and MacOS there is no way to change it once and for
    // all newly created threads. Thus we will use pthreads, creating threads
    // with the stack size of the current thread. This way all threads will
    // inherit the main thread's stack size (since the first helper is always
    // created by the main thread).
    //
    // Note also the interaction with our backtrace functionality: in order to
    // get the complete stack trace we let unhandled exceptions escape the
    // thread function expecting the runtime to still call std::terminate. In
    // particular, having a noexcept function anywhere on the exception's path
    // causes the stack trace to be truncated, at least on Linux.
    //
#if defined(__linux__) || defined(__FreeBSD__) || defined(__NetBSD__) || defined(__APPLE__)

#ifndef LIBBUILD2_DEFAULT_STACK_SIZE
#  define LIBBUILD2_DEFAULT_STACK_SIZE 8388608 // 8MB
#endif

#ifndef LIBBUILD2_SANE_STACK_SIZE
#  define LIBBUILD2_SANE_STACK_SIZE (sizeof(void*) * LIBBUILD2_DEFAULT_STACK_SIZE)
#endif

    // Auto-deleter.
    //
    struct attr_deleter
    {
      void
      operator() (pthread_attr_t* a) const
      {
        int r (pthread_attr_destroy (a));

        // We should be able to destroy the valid attributes object, unless
        // something is severely damaged.
        //
        assert (r == 0);
      }
    };

    // Calculate the current thread stack size. Don't forget to update #if
    // conditions above when adding the stack size customization for a new
    // platforms/compilers.
    //
    size_t stack_size;
    {
#ifdef __linux__
      // Note that the attributes must not be initialized.
      //
      pthread_attr_t attr;
      int r (pthread_getattr_np (pthread_self (), &attr));

      if (r != 0)
        throw_system_error (r);

      unique_ptr<pthread_attr_t, attr_deleter> ad (&attr);
      r = pthread_attr_getstacksize (&attr, &stack_size);

      if (r != 0)
        throw_system_error (r);

#elif defined(__FreeBSD__) || defined(__NetBSD__)
      pthread_attr_t attr;
      int r (pthread_attr_init (&attr));

      if (r != 0)
        throw_system_error (r);

      unique_ptr<pthread_attr_t, attr_deleter> ad (&attr);
      r = pthread_attr_get_np (pthread_self (), &attr);

      if (r != 0)
        throw_system_error (r);

      r = pthread_attr_getstacksize (&attr, &stack_size);

      if (r != 0)
        throw_system_error (r);

#else // defined(__APPLE__)
      stack_size = pthread_get_stacksize_np (pthread_self ());
#endif
    }

    // Cap the size if necessary.
    //
    if (max_stack_)
    {
      if (*max_stack_ != 0 && stack_size > *max_stack_)
        stack_size = *max_stack_;
    }
    else if (stack_size > LIBBUILD2_SANE_STACK_SIZE)
      stack_size = LIBBUILD2_DEFAULT_STACK_SIZE;

    pthread_attr_t attr;
    int r (pthread_attr_init (&attr));

    if (r != 0)
      throw_system_error (r);

    unique_ptr<pthread_attr_t, attr_deleter> ad (&attr);

    // Create the thread already detached.
    //
    r = pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED);

    if (r != 0)
      throw_system_error (r);

    r = pthread_attr_setstacksize (&attr, stack_size);

    if (r != 0)
      throw_system_error (r);

    pthread_t t;
    r = pthread_create (&t, &attr, helper, this);

    if (r != 0)
      throw_system_error (r);
#else
    thread t (helper, this);
    t.detach ();
#endif

    g.l = nullptr; // Disarm.
  }

  void* scheduler::
  helper (void* d)
  {
    scheduler& s (*static_cast<scheduler*> (d));

    // Note that this thread can be in an in-between state (not active or
    // idle) but only while holding the lock. Which means that if we have the
    // lock then we can account for all of them (this is important during
    // shutdown). Except when the thread is just starting, before acquiring
    // the lock for the first time, which we handle with the starting count.
    //
    lock l (s.mutex_);
    s.starting_--;

    while (!s.shutdown_)
    {
      // If there is a spare active thread, become active and go looking for
      // some work.
      //
      if (s.active_ < s.max_active_)
      {
        s.active_++;

        while (s.queued_task_count_.load (memory_order_consume) != 0)
        {
          // Queues are never removed which means we can get the current range
          // and release the main lock while examining each of them.
          //
          auto it (s.task_queues_.begin ());
          size_t n (s.task_queues_.size ()); // Different to end().
          l.unlock ();

          // Note: we have to be careful not to advance the iterator past the
          // last element (since what's past could be changing).
          //
          for (size_t i (0);; ++it)
          {
            task_queue& tq (*it);

            for (lock ql (tq.mutex); !tq.shutdown && !s.empty_front (tq); )
              s.pop_front (tq, ql);

            if (++i == n)
              break;
          }

          l.lock ();
        }

        s.active_--;

        // While executing the tasks a thread might have become ready
        // (equivalent logic to deactivate()).
        //
        if (s.ready_ != 0)
          s.ready_condv_.notify_one ();
        else if (s.active_ == 0 && s.external_ == 0)
          s.dead_condv_.notify_one ();
      }

      // Become idle and wait for a notification.
      //
      s.idle_++;
      s.idle_condv_.wait (l);
      s.idle_--;
    }

    s.helpers_--;
    return nullptr;
  }

  auto scheduler::
  create_queue () -> task_queue&
  {
    // Note that task_queue_depth is immutable between startup() and
    // shutdown() (but see join()).
    //
    task_queue* tq;
    {
      lock l (mutex_);
      task_queues_.emplace_back (task_queue_depth_);
      tq = &task_queues_.back ();
      tq->shutdown = shutdown_;
    }

    queue (tq);
    return *tq;
  }

  void* scheduler::
  deadlock_monitor (void* d)
  {
    using namespace chrono;

    scheduler& s (*static_cast<scheduler*> (d));

    lock l (s.mutex_);
    while (!s.shutdown_)
    {
      s.dead_condv_.wait (l);

      while (s.active_ == 0 && s.external_ == 0 && !s.shutdown_)
      {
        // We may have a deadlock which can happen because of dependency
        // cycles.
        //
        // Relying on the active_ count alone is not precise enough, however:
        // some threads might be transitioning between active/waiting/ready
        // states. Carefully accounting for this is not trivial, to say the
        // least (especially in the face of spurious wakeups). So we are going
        // to do a "fuzzy" deadlock detection by measuring "progress". The
        // idea is that those transitions should be pretty short-lived and so
        // if we wait for a few thousand context switches, then we should be
        // able to distinguish a real deadlock from the transition case.
        //
        size_t op (s.progress_.load (memory_order_relaxed)), np (op);

        l.unlock ();
        for (size_t i (0), n (10000), m (9990); op == np && i != n; ++i)
        {
          // On the last few iterations sleep a bit instead of yielding (in
          // case yield() is a noop; we use the consume order for the same
          // reason).
          //
          if (i <= m)
            this_thread::yield ();
          else
            active_sleep ((i - m) * 20ms);

          np = s.progress_.load (memory_order_consume);
        }
        l.lock ();

        // Re-check active/external counts for good measure (in case we were
        // spinning too fast).
        //
        if (np == op && s.active_ == 0 && s.external_ == 0 && !s.shutdown_)
        {
          // Shutting things down cleanly is tricky: we could have handled it
          // in the scheduler (e.g., by setting a flag and then waking
          // everyone up, similar to shutdown). But there could also be
          // "external waiters" that have called deactivate() -- we have no
          // way to wake those up. So for now we are going to abort (the nice
          // thing about abort is if this is not a dependency cycle, then we
          // have a core to examine).
          //
          error << "deadlock suspected, aborting" <<
            info << "deadlocks are normally caused by dependency cycles" <<
            info << "re-run with -s to diagnose dependency cycles";

          terminate (false /* trace */);
        }
      }
    }

    return nullptr;
  }
}