aboutsummaryrefslogtreecommitdiff
path: root/build2/scheduler.cxx
blob: fdf41216ad5f16a2de1b067f205da678d504727c (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
// file      : build2/scheduler.cxx -*- C++ -*-
// copyright : Copyright (c) 2014-2017 Code Synthesis Ltd
// license   : MIT; see accompanying LICENSE file

#include <build2/scheduler.hxx>

#include <cerrno>

using namespace std;

namespace build2
{
  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 = task_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 ()
  {
    if (max_active_ == 1) // Serial execution.
      return;

    lock l (mutex_);

    active_--;
    waiting_++;

    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);
  }

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

    lock l (mutex_);
    waiting_--;

    if (collision)
      stat_wait_collisions_++;

    // If we have spare active threads, then become active. Otherwise it
    // enters the ready queue.
    //
    ready_++;

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

    ready_--;

    if (shutdown_)
      throw_generic_error (ECANCELED);

    active_++;
  }

  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 ();

    // 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 (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&) {}
  }

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

    // 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).
    //
    return              // HW threads x arch-bits (see max_threads below)
      n ==   0 ?    1 : // serial
      //
      // 2x
      //
      n ==   1 ?    3 :
      n ==   2 ?    5 :
      n ==   4 ?   11 :
      n ==   6 ?   13 :
      n ==   8 ?   17 : // 2 x 4
      n ==  16 ?   31 : // 4 x 4, 2 x 8
      //
      // 1.5x
      //
      n ==  32 ?   47 : // 4 x 8
      n ==  48 ?   53 : // 6 x 8
      n ==  64 ?   67 : // 8 x 8
      n ==  80 ?   89 : // 10 x 8
      //
      // 1x
      //
      n ==  96 ?  101 : // 12 x 8
      n == 112 ?  127 : // 14 x 8
      n == 128 ?  131 : // 16 x 8
      n == 144 ?  139 : // 18 x 8
      n == 160 ?  157 : // 20 x 8
      n == 176 ?  173 : // 22 x 8
      n == 192 ?  191 : // 24 x 8
      n == 224 ?  223 : // 28 x 8
      n == 256 ?  251 : // 32 x 8
      n == 288 ?  271 : // 36 x 8
      n == 320 ?  313 : // 40 x 8
      n == 352 ?  331 : // 44 x 8
      n == 384 ?  367 : // 48 x 8
      n == 512 ?  499 : // 64 x 8
      n - 1;            // Assume it is even.
  }

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

    // Use 16x 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 * (max_active == 1 ? 1 : sizeof (void*) * 4);

    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 * 4;

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

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

    shutdown_ = false;
  }

  void scheduler::
  tune (size_t max_active)
  {
    lock l (mutex_);

    if (max_active == 0)
      max_active = orig_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.
    //
    while (active_ != init_active_)
    {
      l.unlock ();
      this_thread::yield ();
      l.lock ();
    }

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

    max_active_ = 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 ();
      }

      // 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;
  }

  void scheduler::
  activate_helper (lock& l)
  {
    if (!shutdown_)
    {
      if (idle_ != 0)
        idle_condv_.notify_one ();
      else if (init_active_ + helpers_ < max_threads_ ||
               //
               // 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., work the
               // queues). This, for example, can happen if a thread waits for
               // a task that is in its queue but is below the mark.
               //
               (active_ == 0 &&
                queued_task_count_.load (memory_order_consume) != 0))
      {
        create_helper (l);
      }
    }
  }

  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_};

    thread t (helper, this);
    g.l = nullptr; // Disarm.

    t.detach ();
  }

  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 ();

          for (size_t i (0); i != n; ++i)
          {
            task_queue& tq (*it++);

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

          l.lock ();
        }

        s.active_--;

        // While executing the tasks a thread might have become ready.
        //
        if (s.ready_ != 0)
          s.ready_condv_.notify_one ();
      }

      // Become idle and wait for a notification (note that task_ is false
      // here).
      //
      s.idle_++;
      s.idle_condv_.wait (l);
      s.idle_--;
    }

    s.helpers_--;
  }

#ifdef __cpp_thread_local
    thread_local
#else
    __thread
#endif
  scheduler::task_queue* scheduler::task_queue_ = 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_;
    }

    task_queue_ = tq;
    return *tq;
  }
}