mono/metadata/threadpool-worker-default.c

   1 /*
   2  * threadpool-worker.c: native threadpool worker
   3  *
   4  * Author:
   5  *      Ludovic Henry (ludovic.henry@xamarin.com)
   6  *
   7  * Licensed under the MIT license. See LICENSE file in the project root for full license information.
   8  */
   9
  10 #include <stdlib.h>
  11 #define _USE_MATH_DEFINES // needed by MSVC to define math constants
  12 #include <math.h>
  13 #include <config.h>
  14 #include <glib.h>
  15
  16 #include <mono/metadata/class-internals.h>
  17 #include <mono/metadata/exception.h>
  18 #include <mono/metadata/gc-internals.h>
  19 #include <mono/metadata/object.h>
  20 #include <mono/metadata/object-internals.h>
  21 #include <mono/metadata/threadpool.h>
  22 #include <mono/metadata/threadpool-worker.h>
  23 #include <mono/metadata/threadpool-io.h>
  24 #include <mono/metadata/w32event.h>
  25 #include <mono/utils/atomic.h>
  26 #include <mono/utils/mono-compiler.h>
  27 #include <mono/utils/mono-complex.h>
  28 #include <mono/utils/mono-lazy-init.h>
  29 #include <mono/utils/mono-logger.h>
  30 #include <mono/utils/mono-logger-internals.h>
  31 #include <mono/utils/mono-proclib.h>
  32 #include <mono/utils/mono-threads.h>
  33 #include <mono/utils/mono-time.h>
  34 #include <mono/utils/mono-rand.h>
  35 #include <mono/utils/refcount.h>
  36
  37 #define CPU_USAGE_LOW 80
  38 #define CPU_USAGE_HIGH 95
  39
  40 #define MONITOR_INTERVAL 500 // ms
  41 #define MONITOR_MINIMAL_LIFETIME 60 * 1000 // ms
  42
  43 #define WORKER_CREATION_MAX_PER_SEC 10
  44
  45 /* The exponent to apply to the gain. 1.0 means to use linear gain,
  46  * higher values will enhance large moves and damp small ones.
  47  * default: 2.0 */
  48 #define HILL_CLIMBING_GAIN_EXPONENT 2.0
  49
  50 /* The 'cost' of a thread. 0 means drive for increased throughput regardless
  51  * of thread count, higher values bias more against higher thread counts.
  52  * default: 0.15 */
  53 #define HILL_CLIMBING_BIAS 0.15
  54
  55 #define HILL_CLIMBING_WAVE_PERIOD 4
  56 #define HILL_CLIMBING_MAX_WAVE_MAGNITUDE 20
  57 #define HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER 1.0
  58 #define HILL_CLIMBING_WAVE_HISTORY_SIZE 8
  59 #define HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO 3.0
  60 #define HILL_CLIMBING_MAX_CHANGE_PER_SECOND 4
  61 #define HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE 20
  62 #define HILL_CLIMBING_SAMPLE_INTERVAL_LOW 10
  63 #define HILL_CLIMBING_SAMPLE_INTERVAL_HIGH 200
  64 #define HILL_CLIMBING_ERROR_SMOOTHING_FACTOR 0.01
  65 #define HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT 0.15
  66
  67 typedef enum {
  68         TRANSITION_WARMUP,
  69         TRANSITION_INITIALIZING,
  70         TRANSITION_RANDOM_MOVE,
  71         TRANSITION_CLIMBING_MOVE,
  72         TRANSITION_CHANGE_POINT,
  73         TRANSITION_STABILIZING,
  74         TRANSITION_STARVATION,
  75         TRANSITION_THREAD_TIMED_OUT,
  76         TRANSITION_UNDEFINED,
  77 } ThreadPoolHeuristicStateTransition;
  78
  79 typedef struct {
  80         gint32 wave_period;
  81         gint32 samples_to_measure;
  82         gdouble target_throughput_ratio;
  83         gdouble target_signal_to_noise_ratio;
  84         gdouble max_change_per_second;
  85         gdouble max_change_per_sample;
  86         gint32 max_thread_wave_magnitude;
  87         gint32 sample_interval_low;
  88         gdouble thread_magnitude_multiplier;
  89         gint32 sample_interval_high;
  90         gdouble throughput_error_smoothing_factor;
  91         gdouble gain_exponent;
  92         gdouble max_sample_error;
  93
  94         gdouble current_control_setting;
  95         gint64 total_samples;
  96         gint16 last_thread_count;
  97         gdouble elapsed_since_last_change;
  98         gdouble completions_since_last_change;
  99
 100         gdouble average_throughput_noise;
 101
 102         gdouble *samples;
 103         gdouble *thread_counts;
 104
 105         guint32 current_sample_interval;
 106         gpointer random_interval_generator;
 107
 108         gint32 accumulated_completion_count;
 109         gdouble accumulated_sample_duration;
 110 } ThreadPoolHillClimbing;
 111
 112 typedef struct {
 113         MonoThreadPoolWorkerCallback callback;
 114         gpointer data;
 115 } ThreadPoolWorkItem;
 116
 117 typedef union {
 118         struct {
 119                 gint16 max_working; /* determined by heuristic */
 120                 gint16 starting; /* starting, but not yet in worker_thread */
 121                 gint16 working; /* executing worker_thread */
 122                 gint16 parked; /* parked */
 123         } _;
 124         gint64 as_gint64;
 125 } ThreadPoolWorkerCounter;
 126
 127 typedef MonoInternalThread ThreadPoolWorkerThread;
 128
 129 struct MonoThreadPoolWorker {
 130         MonoRefCount ref;
 131
 132         ThreadPoolWorkerCounter counters;
 133
 134         GPtrArray *threads; // ThreadPoolWorkerThread* []
 135         MonoCoopMutex threads_lock; /* protect access to working_threads and parked_threads */
 136         gint32 parked_threads_count;
 137         MonoCoopCond parked_threads_cond;
 138         MonoCoopCond threads_exit_cond;
 139
 140         ThreadPoolWorkItem *work_items; // ThreadPoolWorkItem []
 141         gint32 work_items_count;
 142         gint32 work_items_size;
 143         MonoCoopMutex work_items_lock;
 144
 145         guint32 worker_creation_current_second;
 146         guint32 worker_creation_current_count;
 147         MonoCoopMutex worker_creation_lock;
 148
 149         gint32 heuristic_completions;
 150         gint64 heuristic_sample_start;
 151         gint64 heuristic_last_dequeue; // ms
 152         gint64 heuristic_last_adjustment; // ms
 153         gint64 heuristic_adjustment_interval; // ms
 154         ThreadPoolHillClimbing heuristic_hill_climbing;
 155         MonoCoopMutex heuristic_lock;
 156
 157         gint32 limit_worker_min;
 158         gint32 limit_worker_max;
 159
 160         MonoCpuUsageState *cpu_usage_state;
 161         gint32 cpu_usage;
 162
 163         /* suspended by the debugger */
 164         gboolean suspended;
 165
 166         gint32 monitor_status;
 167 };
 168
 169 enum {
 170         MONITOR_STATUS_REQUESTED,
 171         MONITOR_STATUS_WAITING_FOR_REQUEST,
 172         MONITOR_STATUS_NOT_RUNNING,
 173 };
 174
 175 #define COUNTER_CHECK(counter) \
 176         do { \
 177                 g_assert (counter._.max_working > 0); \
 178                 g_assert (counter._.starting >= 0); \
 179                 g_assert (counter._.working >= 0); \
 180         } while (0)
 181
 182 #define COUNTER_ATOMIC(worker,var,block) \
 183         do { \
 184                 ThreadPoolWorkerCounter __old; \
 185                 do { \
 186                         g_assert (worker); \
 187                         __old = COUNTER_READ (worker); \
 188                         (var) = __old; \
 189                         { block; } \
 190                         COUNTER_CHECK (var); \
 191                 } while (InterlockedCompareExchange64 (&worker->counters.as_gint64, (var).as_gint64, __old.as_gint64) != __old.as_gint64); \
 192         } while (0)
 193
 194 static inline ThreadPoolWorkerCounter
 195 COUNTER_READ (MonoThreadPoolWorker *worker)
 196 {
 197         ThreadPoolWorkerCounter counter;
 198         counter.as_gint64 = InterlockedRead64 (&worker->counters.as_gint64);
 199         return counter;
 200 }
 201
 202 static gpointer
 203 rand_create (void)
 204 {
 205         mono_rand_open ();
 206         return mono_rand_init (NULL, 0);
 207 }
 208
 209 static guint32
 210 rand_next (gpointer *handle, guint32 min, guint32 max)
 211 {
 212         MonoError error;
 213         guint32 val;
 214         mono_rand_try_get_uint32 (handle, &val, min, max, &error);
 215         // FIXME handle error
 216         mono_error_assert_ok (&error);
 217         return val;
 218 }
 219
 220 static void
 221 destroy (gpointer data)
 222 {
 223         MonoThreadPoolWorker *worker;
 224
 225         worker = (MonoThreadPoolWorker*) data;
 226         g_assert (worker);
 227
 228         // FIXME destroy everything
 229
 230         g_free (worker);
 231 }
 232
 233 void
 234 mono_threadpool_worker_init (MonoThreadPoolWorker **worker)
 235 {
 236         MonoThreadPoolWorker *wk;
 237         ThreadPoolHillClimbing *hc;
 238         const char *threads_per_cpu_env;
 239         gint threads_per_cpu;
 240         gint threads_count;
 241
 242         g_assert (worker);
 243
 244         wk = *worker = g_new0 (MonoThreadPoolWorker, 1);
 245
 246         mono_refcount_init (wk, destroy);
 247
 248         wk->threads = g_ptr_array_new ();
 249         mono_coop_mutex_init (&wk->threads_lock);
 250         wk->parked_threads_count = 0;
 251         mono_coop_cond_init (&wk->parked_threads_cond);
 252         mono_coop_cond_init (&wk->threads_exit_cond);
 253
 254         /* wk->work_items_size is inited to 0 */
 255         mono_coop_mutex_init (&wk->work_items_lock);
 256
 257         wk->worker_creation_current_second = -1;
 258         mono_coop_mutex_init (&wk->worker_creation_lock);
 259
 260         wk->heuristic_adjustment_interval = 10;
 261         mono_coop_mutex_init (&wk->heuristic_lock);
 262
 263         mono_rand_open ();
 264
 265         hc = &wk->heuristic_hill_climbing;
 266
 267         hc->wave_period = HILL_CLIMBING_WAVE_PERIOD;
 268         hc->max_thread_wave_magnitude = HILL_CLIMBING_MAX_WAVE_MAGNITUDE;
 269         hc->thread_magnitude_multiplier = (gdouble) HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER;
 270         hc->samples_to_measure = hc->wave_period * HILL_CLIMBING_WAVE_HISTORY_SIZE;
 271         hc->target_throughput_ratio = (gdouble) HILL_CLIMBING_BIAS;
 272         hc->target_signal_to_noise_ratio = (gdouble) HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO;
 273         hc->max_change_per_second = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SECOND;
 274         hc->max_change_per_sample = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE;
 275         hc->sample_interval_low = HILL_CLIMBING_SAMPLE_INTERVAL_LOW;
 276         hc->sample_interval_high = HILL_CLIMBING_SAMPLE_INTERVAL_HIGH;
 277         hc->throughput_error_smoothing_factor = (gdouble) HILL_CLIMBING_ERROR_SMOOTHING_FACTOR;
 278         hc->gain_exponent = (gdouble) HILL_CLIMBING_GAIN_EXPONENT;
 279         hc->max_sample_error = (gdouble) HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT;
 280         hc->current_control_setting = 0;
 281         hc->total_samples = 0;
 282         hc->last_thread_count = 0;
 283         hc->average_throughput_noise = 0;
 284         hc->elapsed_since_last_change = 0;
 285         hc->accumulated_completion_count = 0;
 286         hc->accumulated_sample_duration = 0;
 287         hc->samples = g_new0 (gdouble, hc->samples_to_measure);
 288         hc->thread_counts = g_new0 (gdouble, hc->samples_to_measure);
 289         hc->random_interval_generator = rand_create ();
 290         hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
 291
 292         if (!(threads_per_cpu_env = g_getenv ("MONO_THREADS_PER_CPU")))
 293                 threads_per_cpu = 1;
 294         else
 295                 threads_per_cpu = CLAMP (atoi (threads_per_cpu_env), 1, 50);
 296
 297         threads_count = mono_cpu_count () * threads_per_cpu;
 298
 299         wk->limit_worker_min = threads_count;
 300
 301 #if defined (PLATFORM_ANDROID) || defined (HOST_IOS)
 302         wk->limit_worker_max = CLAMP (threads_count * 100, MIN (threads_count, 200), MAX (threads_count, 200));
 303 #else
 304         wk->limit_worker_max = threads_count * 100;
 305 #endif
 306
 307         wk->counters._.max_working = wk->limit_worker_min;
 308
 309         wk->cpu_usage_state = g_new0 (MonoCpuUsageState, 1);
 310
 311         wk->suspended = FALSE;
 312
 313         wk->monitor_status = MONITOR_STATUS_NOT_RUNNING;
 314 }
 315
 316 void
 317 mono_threadpool_worker_cleanup (MonoThreadPoolWorker *worker)
 318 {
 319         MonoInternalThread *current;
 320
 321         /* we make the assumption along the code that we are
 322          * cleaning up only if the runtime is shutting down */
 323         g_assert (mono_runtime_is_shutting_down ());
 324
 325         current = mono_thread_internal_current ();
 326
 327         while (worker->monitor_status != MONITOR_STATUS_NOT_RUNNING)
 328                 mono_thread_info_sleep (1, NULL);
 329
 330         mono_coop_mutex_lock (&worker->threads_lock);
 331
 332         /* unpark all worker->parked_threads */
 333         mono_coop_cond_broadcast (&worker->parked_threads_cond);
 334
 335         for (;;) {
 336                 ThreadPoolWorkerCounter counter;
 337
 338                 counter = COUNTER_READ (worker);
 339                 if (counter._.starting + counter._.working + counter._.parked == 0)
 340                         break;
 341
 342                 if (counter._.starting + counter._.working + counter._.parked == 1) {
 343                         if (worker->threads->len == 1 && g_ptr_array_index (worker->threads, 0) == current) {
 344                                 /* We are waiting on ourselves */
 345                                 break;
 346                         }
 347                 }
 348
 349                 mono_coop_cond_wait (&worker->threads_exit_cond, &worker->threads_lock);
 350         }
 351
 352         mono_coop_mutex_unlock (&worker->threads_lock);
 353
 354         mono_refcount_dec (worker);
 355 }
 356
 357 static void
 358 work_item_lock (MonoThreadPoolWorker *worker)
 359 {
 360         mono_coop_mutex_lock (&worker->work_items_lock);
 361 }
 362
 363 static void
 364 work_item_unlock (MonoThreadPoolWorker *worker)
 365 {
 366         mono_coop_mutex_unlock (&worker->work_items_lock);
 367 }
 368
 369 static void
 370 work_item_push (MonoThreadPoolWorker *worker, MonoThreadPoolWorkerCallback callback, gpointer data)
 371 {
 372         ThreadPoolWorkItem work_item;
 373
 374         g_assert (worker);
 375         g_assert (callback);
 376
 377         work_item.callback = callback;
 378         work_item.data = data;
 379
 380         work_item_lock (worker);
 381
 382         g_assert (worker->work_items_count <= worker->work_items_size);
 383
 384         if (G_UNLIKELY (worker->work_items_count == worker->work_items_size)) {
 385                 worker->work_items_size += 64;
 386                 worker->work_items = g_renew (ThreadPoolWorkItem, worker->work_items, worker->work_items_size);
 387         }
 388
 389         g_assert (worker->work_items);
 390
 391         worker->work_items [worker->work_items_count ++] = work_item;
 392
 393         // printf ("[push] worker->work_items = %p, worker->work_items_count = %d, worker->work_items_size = %d\n",
 394         //      worker->work_items, worker->work_items_count, worker->work_items_size);
 395
 396         work_item_unlock (worker);
 397 }
 398
 399 static gboolean
 400 work_item_try_pop (MonoThreadPoolWorker *worker, ThreadPoolWorkItem *work_item)
 401 {
 402         g_assert (worker);
 403         g_assert (work_item);
 404
 405         work_item_lock (worker);
 406
 407         // printf ("[pop]  worker->work_items = %p, worker->work_items_count = %d, worker->work_items_size = %d\n",
 408         //      worker->work_items, worker->work_items_count, worker->work_items_size);
 409
 410         if (worker->work_items_count == 0) {
 411                 work_item_unlock (worker);
 412                 return FALSE;
 413         }
 414
 415         *work_item = worker->work_items [-- worker->work_items_count];
 416
 417         if (G_UNLIKELY (worker->work_items_count >= 64 * 3 && worker->work_items_count < worker->work_items_size / 2)) {
 418                 worker->work_items_size -= 64;
 419                 worker->work_items = g_renew (ThreadPoolWorkItem, worker->work_items, worker->work_items_size);
 420         }
 421
 422         work_item_unlock (worker);
 423
 424         return TRUE;
 425 }
 426
 427 static gint32
 428 work_item_count (MonoThreadPoolWorker *worker)
 429 {
 430         gint32 count;
 431
 432         work_item_lock (worker);
 433         count = worker->work_items_count;
 434         work_item_unlock (worker);
 435
 436         return count;
 437 }
 438
 439 static void worker_request (MonoThreadPoolWorker *worker);
 440
 441 void
 442 mono_threadpool_worker_enqueue (MonoThreadPoolWorker *worker, MonoThreadPoolWorkerCallback callback, gpointer data)
 443 {
 444         work_item_push (worker, callback, data);
 445
 446         worker_request (worker);
 447 }
 448
 449 static void
 450 worker_wait_interrupt (gpointer data)
 451 {
 452         MonoThreadPoolWorker *worker;
 453
 454         worker = (MonoThreadPoolWorker*) data;
 455         g_assert (worker);
 456
 457         mono_coop_mutex_lock (&worker->threads_lock);
 458         mono_coop_cond_signal (&worker->parked_threads_cond);
 459         mono_coop_mutex_unlock (&worker->threads_lock);
 460
 461         mono_refcount_dec (worker);
 462 }
 463
 464 /* return TRUE if timeout, FALSE otherwise (worker unpark or interrupt) */
 465 static gboolean
 466 worker_park (MonoThreadPoolWorker *worker)
 467 {
 468         gboolean timeout = FALSE;
 469
 470         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker parking", mono_native_thread_id_get ());
 471
 472         mono_coop_mutex_lock (&worker->threads_lock);
 473
 474         if (!mono_runtime_is_shutting_down ()) {
 475                 static gpointer rand_handle = NULL;
 476                 MonoInternalThread *thread;
 477                 gboolean interrupted = FALSE;
 478                 ThreadPoolWorkerCounter counter;
 479
 480                 if (!rand_handle)
 481                         rand_handle = rand_create ();
 482                 g_assert (rand_handle);
 483
 484                 thread = mono_thread_internal_current ();
 485                 g_assert (thread);
 486
 487                 COUNTER_ATOMIC (worker, counter, {
 488                         counter._.working --;
 489                         counter._.parked ++;
 490                 });
 491
 492                 worker->parked_threads_count += 1;
 493
 494                 mono_thread_info_install_interrupt (worker_wait_interrupt, mono_refcount_inc (worker), &interrupted);
 495                 if (interrupted) {
 496                         mono_refcount_dec (worker);
 497                         goto done;
 498                 }
 499
 500                 if (mono_coop_cond_timedwait (&worker->parked_threads_cond, &worker->threads_lock, rand_next (&rand_handle, 5 * 1000, 60 * 1000)) != 0)
 501                         timeout = TRUE;
 502
 503                 mono_thread_info_uninstall_interrupt (&interrupted);
 504                 if (!interrupted)
 505                         mono_refcount_dec (worker);
 506
 507 done:
 508                 worker->parked_threads_count -= 1;
 509
 510                 COUNTER_ATOMIC (worker, counter, {
 511                         counter._.working ++;
 512                         counter._.parked --;
 513                 });
 514         }
 515
 516         mono_coop_mutex_unlock (&worker->threads_lock);
 517
 518         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker unparking, timeout? %s", mono_native_thread_id_get (), timeout ? "yes" : "no");
 519
 520         return timeout;
 521 }
 522
 523 static gboolean
 524 worker_try_unpark (MonoThreadPoolWorker *worker)
 525 {
 526         gboolean res = FALSE;
 527
 528         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker", mono_native_thread_id_get ());
 529
 530         mono_coop_mutex_lock (&worker->threads_lock);
 531         if (worker->parked_threads_count > 0) {
 532                 mono_coop_cond_signal (&worker->parked_threads_cond);
 533                 res = TRUE;
 534         }
 535         mono_coop_mutex_unlock (&worker->threads_lock);
 536
 537         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker, success? %s", mono_native_thread_id_get (), res ? "yes" : "no");
 538
 539         return res;
 540 }
 541
 542 static gsize WINAPI
 543 worker_thread (gpointer data)
 544 {
 545         MonoThreadPoolWorker *worker;
 546         MonoError error;
 547         MonoInternalThread *thread;
 548         ThreadPoolWorkerCounter counter;
 549
 550         mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker starting", mono_native_thread_id_get ());
 551
 552         worker = (MonoThreadPoolWorker*) data;
 553         g_assert (worker);
 554
 555         COUNTER_ATOMIC (worker, counter, {
 556                 counter._.starting --;
 557                 counter._.working ++;
 558         });
 559
 560         thread = mono_thread_internal_current ();
 561         g_assert (thread);
 562
 563         mono_coop_mutex_lock (&worker->threads_lock);
 564         g_ptr_array_add (worker->threads, thread);
 565         mono_coop_mutex_unlock (&worker->threads_lock);
 566
 567         mono_thread_set_name_internal (thread, mono_string_new (mono_get_root_domain (), "Threadpool worker"), FALSE, &error);
 568         mono_error_assert_ok (&error);
 569
 570         while (!mono_runtime_is_shutting_down ()) {
 571                 ThreadPoolWorkItem work_item;
 572
 573                 if (mono_thread_interruption_checkpoint ())
 574                         continue;
 575
 576                 if (!work_item_try_pop (worker, &work_item)) {
 577                         gboolean timeout;
 578
 579                         timeout = worker_park (worker);
 580                         if (timeout)
 581                                 break;
 582
 583                         continue;
 584                 }
 585
 586                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker executing %p (%p)",
 587                         mono_native_thread_id_get (), work_item.callback, work_item.data);
 588
 589                 work_item.callback (work_item.data);
 590         }
 591
 592         mono_coop_mutex_lock (&worker->threads_lock);
 593
 594         COUNTER_ATOMIC (worker, counter, {
 595                 counter._.working --;
 596         });
 597
 598         g_ptr_array_remove (worker->threads, thread);
 599
 600         mono_coop_cond_signal (&worker->threads_exit_cond);
 601
 602         mono_coop_mutex_unlock (&worker->threads_lock);
 603
 604         mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker finishing", mono_native_thread_id_get ());
 605
 606         mono_refcount_dec (worker);
 607
 608         return 0;
 609 }
 610
 611 static gboolean
 612 worker_try_create (MonoThreadPoolWorker *worker)
 613 {
 614         MonoError error;
 615         MonoInternalThread *thread;
 616         gint64 current_ticks;
 617         gint32 now;
 618         ThreadPoolWorkerCounter counter;
 619
 620         if (mono_runtime_is_shutting_down ())
 621                 return FALSE;
 622
 623         mono_coop_mutex_lock (&worker->worker_creation_lock);
 624
 625         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker", mono_native_thread_id_get ());
 626
 627         current_ticks = mono_100ns_ticks ();
 628         if (0 == current_ticks) {
 629                 g_warning ("failed to get 100ns ticks");
 630         } else {
 631                 now = current_ticks / (10 * 1000 * 1000);
 632                 if (worker->worker_creation_current_second != now) {
 633                         worker->worker_creation_current_second = now;
 634                         worker->worker_creation_current_count = 0;
 635                 } else {
 636                         g_assert (worker->worker_creation_current_count <= WORKER_CREATION_MAX_PER_SEC);
 637                         if (worker->worker_creation_current_count == WORKER_CREATION_MAX_PER_SEC) {
 638                                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: maximum number of worker created per second reached, current count = %d",
 639                                         mono_native_thread_id_get (), worker->worker_creation_current_count);
 640                                 mono_coop_mutex_unlock (&worker->worker_creation_lock);
 641                                 return FALSE;
 642                         }
 643                 }
 644         }
 645
 646         COUNTER_ATOMIC (worker, counter, {
 647                 if (counter._.working >= counter._.max_working) {
 648                         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: maximum number of working threads reached",
 649                                 mono_native_thread_id_get ());
 650                         mono_coop_mutex_unlock (&worker->worker_creation_lock);
 651                         return FALSE;
 652                 }
 653                 counter._.starting ++;
 654         });
 655
 656         thread = mono_thread_create_internal (mono_get_root_domain (), worker_thread, mono_refcount_inc (worker), TRUE, 0, &error);
 657         if (!thread) {
 658                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: could not create thread due to %s", mono_native_thread_id_get (), mono_error_get_message (&error));
 659                 mono_error_cleanup (&error);
 660
 661                 COUNTER_ATOMIC (worker, counter, {
 662                         counter._.starting --;
 663                 });
 664
 665                 mono_coop_mutex_unlock (&worker->worker_creation_lock);
 666
 667                 mono_refcount_dec (worker);
 668
 669                 return FALSE;
 670         }
 671
 672         worker->worker_creation_current_count += 1;
 673
 674         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, created %p, now = %d count = %d",
 675                 mono_native_thread_id_get (), (gpointer) thread->tid, now, worker->worker_creation_current_count);
 676
 677         mono_coop_mutex_unlock (&worker->worker_creation_lock);
 678         return TRUE;
 679 }
 680
 681 static void monitor_ensure_running (MonoThreadPoolWorker *worker);
 682
 683 static void
 684 worker_request (MonoThreadPoolWorker *worker)
 685 {
 686         g_assert (worker);
 687
 688         if (worker->suspended)
 689                 return;
 690
 691         monitor_ensure_running (worker);
 692
 693         if (worker_try_unpark (worker)) {
 694                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, unparked", mono_native_thread_id_get ());
 695                 return;
 696         }
 697
 698         if (worker_try_create (worker)) {
 699                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, created", mono_native_thread_id_get ());
 700                 return;
 701         }
 702
 703         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, failed", mono_native_thread_id_get ());
 704 }
 705
 706 static gboolean
 707 monitor_should_keep_running (MonoThreadPoolWorker *worker)
 708 {
 709         static gint64 last_should_keep_running = -1;
 710
 711         g_assert (worker->monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || worker->monitor_status == MONITOR_STATUS_REQUESTED);
 712
 713         if (InterlockedExchange (&worker->monitor_status, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST) {
 714                 gboolean should_keep_running = TRUE, force_should_keep_running = FALSE;
 715
 716                 if (mono_runtime_is_shutting_down ()) {
 717                         should_keep_running = FALSE;
 718                 } else {
 719                         if (work_item_count (worker) == 0)
 720                                 should_keep_running = FALSE;
 721
 722                         if (!should_keep_running) {
 723                                 if (last_should_keep_running == -1 || mono_100ns_ticks () - last_should_keep_running < MONITOR_MINIMAL_LIFETIME * 1000 * 10) {
 724                                         should_keep_running = force_should_keep_running = TRUE;
 725                                 }
 726                         }
 727                 }
 728
 729                 if (should_keep_running) {
 730                         if (last_should_keep_running == -1 || !force_should_keep_running)
 731                                 last_should_keep_running = mono_100ns_ticks ();
 732                 } else {
 733                         last_should_keep_running = -1;
 734                         if (InterlockedCompareExchange (&worker->monitor_status, MONITOR_STATUS_NOT_RUNNING, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST)
 735                                 return FALSE;
 736                 }
 737         }
 738
 739         g_assert (worker->monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || worker->monitor_status == MONITOR_STATUS_REQUESTED);
 740
 741         return TRUE;
 742 }
 743
 744 static gboolean
 745 monitor_sufficient_delay_since_last_dequeue (MonoThreadPoolWorker *worker)
 746 {
 747         gint64 threshold;
 748
 749         g_assert (worker);
 750
 751         if (worker->cpu_usage < CPU_USAGE_LOW) {
 752                 threshold = MONITOR_INTERVAL;
 753         } else {
 754                 ThreadPoolWorkerCounter counter;
 755                 counter = COUNTER_READ (worker);
 756                 threshold = counter._.max_working * MONITOR_INTERVAL * 2;
 757         }
 758
 759         return mono_msec_ticks () >= worker->heuristic_last_dequeue + threshold;
 760 }
 761
 762 static void hill_climbing_force_change (MonoThreadPoolWorker *worker, gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition);
 763
 764 static gsize WINAPI
 765 monitor_thread (gpointer data)
 766 {
 767         MonoThreadPoolWorker *worker;
 768         MonoInternalThread *internal;
 769         guint i;
 770
 771         worker = (MonoThreadPoolWorker*) data;
 772         g_assert (worker);
 773
 774         internal = mono_thread_internal_current ();
 775         g_assert (internal);
 776
 777         mono_cpu_usage (worker->cpu_usage_state);
 778
 779         // printf ("monitor_thread: start\n");
 780
 781         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, started", mono_native_thread_id_get ());
 782
 783         do {
 784                 ThreadPoolWorkerCounter counter;
 785                 gboolean limit_worker_max_reached;
 786                 gint32 interval_left = MONITOR_INTERVAL;
 787                 gint32 awake = 0; /* number of spurious awakes we tolerate before doing a round of rebalancing */
 788
 789                 g_assert (worker->monitor_status != MONITOR_STATUS_NOT_RUNNING);
 790
 791                 // counter = COUNTER_READ (worker);
 792                 // printf ("monitor_thread: starting = %d working = %d parked = %d max_working = %d\n",
 793                 //      counter._.starting, counter._.working, counter._.parked, counter._.max_working);
 794
 795                 do {
 796                         gint64 ts;
 797                         gboolean alerted = FALSE;
 798
 799                         if (mono_runtime_is_shutting_down ())
 800                                 break;
 801
 802                         ts = mono_msec_ticks ();
 803                         if (mono_thread_info_sleep (interval_left, &alerted) == 0)
 804                                 break;
 805                         interval_left -= mono_msec_ticks () - ts;
 806
 807                         g_assert (!(internal->state & ThreadState_StopRequested));
 808                         mono_thread_interruption_checkpoint ();
 809                 } while (interval_left > 0 && ++awake < 10);
 810
 811                 if (mono_runtime_is_shutting_down ())
 812                         continue;
 813
 814                 if (worker->suspended)
 815                         continue;
 816
 817                 if (work_item_count (worker) == 0)
 818                         continue;
 819
 820                 worker->cpu_usage = mono_cpu_usage (worker->cpu_usage_state);
 821
 822                 if (!monitor_sufficient_delay_since_last_dequeue (worker))
 823                         continue;
 824
 825                 limit_worker_max_reached = FALSE;
 826
 827                 COUNTER_ATOMIC (worker, counter, {
 828                         if (counter._.max_working >= worker->limit_worker_max) {
 829                                 limit_worker_max_reached = TRUE;
 830                                 break;
 831                         }
 832                         counter._.max_working ++;
 833                 });
 834
 835                 if (limit_worker_max_reached)
 836                         continue;
 837
 838                 hill_climbing_force_change (worker, counter._.max_working, TRANSITION_STARVATION);
 839
 840                 for (i = 0; i < 5; ++i) {
 841                         if (mono_runtime_is_shutting_down ())
 842                                 break;
 843
 844                         if (worker_try_unpark (worker)) {
 845                                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, unparked", mono_native_thread_id_get ());
 846                                 break;
 847                         }
 848
 849                         if (worker_try_create (worker)) {
 850                                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, created", mono_native_thread_id_get ());
 851                                 break;
 852                         }
 853                 }
 854         } while (monitor_should_keep_running (worker));
 855
 856         // printf ("monitor_thread: stop\n");
 857
 858         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, finished", mono_native_thread_id_get ());
 859
 860         return 0;
 861 }
 862
 863 static void
 864 monitor_ensure_running (MonoThreadPoolWorker *worker)
 865 {
 866         MonoError error;
 867         for (;;) {
 868                 switch (worker->monitor_status) {
 869                 case MONITOR_STATUS_REQUESTED:
 870                         // printf ("monitor_thread: requested\n");
 871                         return;
 872                 case MONITOR_STATUS_WAITING_FOR_REQUEST:
 873                         // printf ("monitor_thread: waiting for request\n");
 874                         InterlockedCompareExchange (&worker->monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_WAITING_FOR_REQUEST);
 875                         break;
 876                 case MONITOR_STATUS_NOT_RUNNING:
 877                         // printf ("monitor_thread: not running\n");
 878                         if (mono_runtime_is_shutting_down ())
 879                                 return;
 880                         if (InterlockedCompareExchange (&worker->monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_NOT_RUNNING) == MONITOR_STATUS_NOT_RUNNING) {
 881                                 // printf ("monitor_thread: creating\n");
 882                                 if (!mono_thread_create_internal (mono_get_root_domain (), monitor_thread, worker, TRUE, SMALL_STACK, &error)) {
 883                                         // printf ("monitor_thread: creating failed\n");
 884                                         worker->monitor_status = MONITOR_STATUS_NOT_RUNNING;
 885                                         mono_error_cleanup (&error);
 886                                 }
 887                                 return;
 888                         }
 889                         break;
 890                 default: g_assert_not_reached ();
 891                 }
 892         }
 893 }
 894
 895 static void
 896 hill_climbing_change_thread_count (MonoThreadPoolWorker *worker, gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
 897 {
 898         ThreadPoolHillClimbing *hc;
 899
 900         g_assert (worker);
 901
 902         hc = &worker->heuristic_hill_climbing;
 903
 904         mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] hill climbing, change max number of threads %d", mono_native_thread_id_get (), new_thread_count);
 905
 906         hc->last_thread_count = new_thread_count;
 907         hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
 908         hc->elapsed_since_last_change = 0;
 909         hc->completions_since_last_change = 0;
 910 }
 911
 912 static void
 913 hill_climbing_force_change (MonoThreadPoolWorker *worker, gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
 914 {
 915         ThreadPoolHillClimbing *hc;
 916
 917         g_assert (worker);
 918
 919         hc = &worker->heuristic_hill_climbing;
 920
 921         if (new_thread_count != hc->last_thread_count) {
 922                 hc->current_control_setting += new_thread_count - hc->last_thread_count;
 923                 hill_climbing_change_thread_count (worker, new_thread_count, transition);
 924         }
 925 }
 926
 927 static double_complex
 928 hill_climbing_get_wave_component (MonoThreadPoolWorker *worker, gdouble *samples, guint sample_count, gdouble period)
 929 {
 930         ThreadPoolHillClimbing *hc;
 931         gdouble w, cosine, sine, coeff, q0, q1, q2;
 932         guint i;
 933
 934         g_assert (worker);
 935         g_assert (sample_count >= period);
 936         g_assert (period >= 2);
 937
 938         hc = &worker->heuristic_hill_climbing;
 939
 940         w = 2.0 * M_PI / period;
 941         cosine = cos (w);
 942         sine = sin (w);
 943         coeff = 2.0 * cosine;
 944         q0 = q1 = q2 = 0;
 945
 946         for (i = 0; i < sample_count; ++i) {
 947                 q0 = coeff * q1 - q2 + samples [(hc->total_samples - sample_count + i) % hc->samples_to_measure];
 948                 q2 = q1;
 949                 q1 = q0;
 950         }
 951
 952         return mono_double_complex_scalar_div (mono_double_complex_make (q1 - q2 * cosine, (q2 * sine)), ((gdouble)sample_count));
 953 }
 954
 955 static gint16
 956 hill_climbing_update (MonoThreadPoolWorker *worker, gint16 current_thread_count, guint32 sample_duration, gint32 completions, gint64 *adjustment_interval)
 957 {
 958         ThreadPoolHillClimbing *hc;
 959         ThreadPoolHeuristicStateTransition transition;
 960         gdouble throughput;
 961         gdouble throughput_error_estimate;
 962         gdouble confidence;
 963         gdouble move;
 964         gdouble gain;
 965         gint sample_index;
 966         gint sample_count;
 967         gint new_thread_wave_magnitude;
 968         gint new_thread_count;
 969         double_complex thread_wave_component;
 970         double_complex throughput_wave_component;
 971         double_complex ratio;
 972
 973         g_assert (worker);
 974         g_assert (adjustment_interval);
 975
 976         hc = &worker->heuristic_hill_climbing;
 977
 978         /* If someone changed the thread count without telling us, update our records accordingly. */
 979         if (current_thread_count != hc->last_thread_count)
 980                 hill_climbing_force_change (worker, current_thread_count, TRANSITION_INITIALIZING);
 981
 982         /* Update the cumulative stats for this thread count */
 983         hc->elapsed_since_last_change += sample_duration;
 984         hc->completions_since_last_change += completions;
 985
 986         /* Add in any data we've already collected about this sample */
 987         sample_duration += hc->accumulated_sample_duration;
 988         completions += hc->accumulated_completion_count;
 989
 990         /* We need to make sure we're collecting reasonably accurate data. Since we're just counting the end
 991          * of each work item, we are goinng to be missing some data about what really happened during the
 992          * sample interval. The count produced by each thread includes an initial work item that may have
 993          * started well before the start of the interval, and each thread may have been running some new
 994          * work item for some time before the end of the interval, which did not yet get counted. So
 995          * our count is going to be off by +/- threadCount workitems.
 996          *
 997          * The exception is that the thread that reported to us last time definitely wasn't running any work
 998          * at that time, and the thread that's reporting now definitely isn't running a work item now. So
 999          * we really only need to consider threadCount-1 threads.
1000          *
1001          * Thus the percent error in our count is +/- (threadCount-1)/numCompletions.
1002          *
1003          * We cannot rely on the frequency-domain analysis we'll be doing later to filter out this error, because
1004          * of the way it accumulates over time. If this sample is off by, say, 33% in the negative direction,
1005          * then the next one likely will be too. The one after that will include the sum of the completions
1006          * we missed in the previous samples, and so will be 33% positive. So every three samples we'll have
1007          * two "low" samples and one "high" sample. This will appear as periodic variation right in the frequency
1008          * range we're targeting, which will not be filtered by the frequency-domain translation. */
1009         if (hc->total_samples > 0 && ((current_thread_count - 1.0) / completions) >= hc->max_sample_error) {
1010                 /* Not accurate enough yet. Let's accumulate the data so
1011                  * far, and tell the MonoThreadPoolWorker to collect a little more. */
1012                 hc->accumulated_sample_duration = sample_duration;
1013                 hc->accumulated_completion_count = completions;
1014                 *adjustment_interval = 10;
1015                 return current_thread_count;
1016         }
1017
1018         /* We've got enouugh data for our sample; reset our accumulators for next time. */
1019         hc->accumulated_sample_duration = 0;
1020         hc->accumulated_completion_count = 0;
1021
1022         /* Add the current thread count and throughput sample to our history. */
1023         throughput = ((gdouble) completions) / sample_duration;
1024
1025         sample_index = hc->total_samples % hc->samples_to_measure;
1026         hc->samples [sample_index] = throughput;
1027         hc->thread_counts [sample_index] = current_thread_count;
1028         hc->total_samples ++;
1029
1030         /* Set up defaults for our metrics. */
1031         thread_wave_component = mono_double_complex_make(0, 0);
1032         throughput_wave_component = mono_double_complex_make(0, 0);
1033         throughput_error_estimate = 0;
1034         ratio = mono_double_complex_make(0, 0);
1035         confidence = 0;
1036
1037         transition = TRANSITION_WARMUP;
1038
1039         /* How many samples will we use? It must be at least the three wave periods we're looking for, and it must also
1040          * be a whole multiple of the primary wave's period; otherwise the frequency we're looking for will fall between
1041          * two frequency bands in the Fourier analysis, and we won't be able to measure it accurately. */
1042         sample_count = ((gint) MIN (hc->total_samples - 1, hc->samples_to_measure) / hc->wave_period) * hc->wave_period;
1043
1044         if (sample_count > hc->wave_period) {
1045                 guint i;
1046                 gdouble average_throughput;
1047                 gdouble average_thread_count;
1048                 gdouble sample_sum = 0;
1049                 gdouble thread_sum = 0;
1050
1051                 /* Average the throughput and thread count samples, so we can scale the wave magnitudes later. */
1052                 for (i = 0; i < sample_count; ++i) {
1053                         guint j = (hc->total_samples - sample_count + i) % hc->samples_to_measure;
1054                         sample_sum += hc->samples [j];
1055                         thread_sum += hc->thread_counts [j];
1056                 }
1057
1058                 average_throughput = sample_sum / sample_count;
1059                 average_thread_count = thread_sum / sample_count;
1060
1061                 if (average_throughput > 0 && average_thread_count > 0) {
1062                         gdouble noise_for_confidence, adjacent_period_1, adjacent_period_2;
1063
1064                         /* Calculate the periods of the adjacent frequency bands we'll be using to
1065                          * measure noise levels. We want the two adjacent Fourier frequency bands. */
1066                         adjacent_period_1 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) + 1);
1067                         adjacent_period_2 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) - 1);
1068
1069                         /* Get the the three different frequency components of the throughput (scaled by average
1070                          * throughput). Our "error" estimate (the amount of noise that might be present in the
1071                          * frequency band we're really interested in) is the average of the adjacent bands. */
1072                         throughput_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (worker, hc->samples, sample_count, hc->wave_period), average_throughput);
1073                         throughput_error_estimate = cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (worker, hc->samples, sample_count, adjacent_period_1), average_throughput));
1074
1075                         if (adjacent_period_2 <= sample_count) {
1076                                 throughput_error_estimate = MAX (throughput_error_estimate, cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (
1077                                         worker, hc->samples, sample_count, adjacent_period_2), average_throughput)));
1078                         }
1079
1080                         /* Do the same for the thread counts, so we have something to compare to. We don't
1081                          * measure thread count noise, because there is none; these are exact measurements. */
1082                         thread_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (worker, hc->thread_counts, sample_count, hc->wave_period), average_thread_count);
1083
1084                         /* Update our moving average of the throughput noise. We'll use this
1085                          * later as feedback to determine the new size of the thread wave. */
1086                         if (hc->average_throughput_noise == 0) {
1087                                 hc->average_throughput_noise = throughput_error_estimate;
1088                         } else {
1089                                 hc->average_throughput_noise = (hc->throughput_error_smoothing_factor * throughput_error_estimate)
1090                                         + ((1.0 + hc->throughput_error_smoothing_factor) * hc->average_throughput_noise);
1091                         }
1092
1093                         if (cabs (thread_wave_component) > 0) {
1094                                 /* Adjust the throughput wave so it's centered around the target wave,
1095                                  * and then calculate the adjusted throughput/thread ratio. */
1096                                 ratio = mono_double_complex_div (mono_double_complex_sub (throughput_wave_component, mono_double_complex_scalar_mul(thread_wave_component, hc->target_throughput_ratio)), thread_wave_component);
1097                                 transition = TRANSITION_CLIMBING_MOVE;
1098                         } else {
1099                                 ratio = mono_double_complex_make (0, 0);
1100                                 transition = TRANSITION_STABILIZING;
1101                         }
1102
1103                         noise_for_confidence = MAX (hc->average_throughput_noise, throughput_error_estimate);
1104                         if (noise_for_confidence > 0) {
1105                                 confidence = cabs (thread_wave_component) / noise_for_confidence / hc->target_signal_to_noise_ratio;
1106                         } else {
1107                                 /* there is no noise! */
1108                                 confidence = 1.0;
1109                         }
1110                 }
1111         }
1112
1113         /* We use just the real part of the complex ratio we just calculated. If the throughput signal
1114          * is exactly in phase with the thread signal, this will be the same as taking the magnitude of
1115          * the complex move and moving that far up. If they're 180 degrees out of phase, we'll move
1116          * backward (because this indicates that our changes are having the opposite of the intended effect).
1117          * If they're 90 degrees out of phase, we won't move at all, because we can't tell wether we're
1118          * having a negative or positive effect on throughput. */
1119         move = creal (ratio);
1120         move = CLAMP (move, -1.0, 1.0);
1121
1122         /* Apply our confidence multiplier. */
1123         move *= CLAMP (confidence, -1.0, 1.0);
1124
1125         /* Now apply non-linear gain, such that values around zero are attenuated, while higher values
1126          * are enhanced. This allows us to move quickly if we're far away from the target, but more slowly
1127         * if we're getting close, giving us rapid ramp-up without wild oscillations around the target. */
1128         gain = hc->max_change_per_second * sample_duration;
1129         move = pow (fabs (move), hc->gain_exponent) * (move >= 0.0 ? 1 : -1) * gain;
1130         move = MIN (move, hc->max_change_per_sample);
1131
1132         /* If the result was positive, and CPU is > 95%, refuse the move. */
1133         if (move > 0.0 && worker->cpu_usage > CPU_USAGE_HIGH)
1134                 move = 0.0;
1135
1136         /* Apply the move to our control setting. */
1137         hc->current_control_setting += move;
1138
1139         /* Calculate the new thread wave magnitude, which is based on the moving average we've been keeping of the
1140          * throughput error.  This average starts at zero, so we'll start with a nice safe little wave at first. */
1141         new_thread_wave_magnitude = (gint)(0.5 + (hc->current_control_setting * hc->average_throughput_noise
1142                 * hc->target_signal_to_noise_ratio * hc->thread_magnitude_multiplier * 2.0));
1143         new_thread_wave_magnitude = CLAMP (new_thread_wave_magnitude, 1, hc->max_thread_wave_magnitude);
1144
1145         /* Make sure our control setting is within the MonoThreadPoolWorker's limits. */
1146         hc->current_control_setting = CLAMP (hc->current_control_setting, worker->limit_worker_min, worker->limit_worker_max - new_thread_wave_magnitude);
1147
1148         /* Calculate the new thread count (control setting + square wave). */
1149         new_thread_count = (gint)(hc->current_control_setting + new_thread_wave_magnitude * ((hc->total_samples / (hc->wave_period / 2)) % 2));
1150
1151         /* Make sure the new thread count doesn't exceed the MonoThreadPoolWorker's limits. */
1152         new_thread_count = CLAMP (new_thread_count, worker->limit_worker_min, worker->limit_worker_max);
1153
1154         if (new_thread_count != current_thread_count)
1155                 hill_climbing_change_thread_count (worker, new_thread_count, transition);
1156
1157         if (creal (ratio) < 0.0 && new_thread_count == worker->limit_worker_min)
1158                 *adjustment_interval = (gint)(0.5 + hc->current_sample_interval * (10.0 * MAX (-1.0 * creal (ratio), 1.0)));
1159         else
1160                 *adjustment_interval = hc->current_sample_interval;
1161
1162         return new_thread_count;
1163 }
1164
1165 static gboolean
1166 heuristic_should_adjust (MonoThreadPoolWorker *worker)
1167 {
1168         if (worker->heuristic_last_dequeue > worker->heuristic_last_adjustment + worker->heuristic_adjustment_interval) {
1169                 ThreadPoolWorkerCounter counter;
1170                 counter = COUNTER_READ (worker);
1171                 if (counter._.working <= counter._.max_working)
1172                         return TRUE;
1173         }
1174
1175         return FALSE;
1176 }
1177
1178 static void
1179 heuristic_adjust (MonoThreadPoolWorker *worker)
1180 {
1181         if (mono_coop_mutex_trylock (&worker->heuristic_lock) == 0) {
1182                 gint32 completions = InterlockedExchange (&worker->heuristic_completions, 0);
1183                 gint64 sample_end = mono_msec_ticks ();
1184                 gint64 sample_duration = sample_end - worker->heuristic_sample_start;
1185
1186                 if (sample_duration >= worker->heuristic_adjustment_interval / 2) {
1187                         ThreadPoolWorkerCounter counter;
1188                         gint16 new_thread_count;
1189
1190                         counter = COUNTER_READ (worker);
1191                         new_thread_count = hill_climbing_update (worker, counter._.max_working, sample_duration, completions, &worker->heuristic_adjustment_interval);
1192
1193                         COUNTER_ATOMIC (worker, counter, {
1194                                 counter._.max_working = new_thread_count;
1195                         });
1196
1197                         if (new_thread_count > counter._.max_working)
1198                                 worker_request (worker);
1199
1200                         worker->heuristic_sample_start = sample_end;
1201                         worker->heuristic_last_adjustment = mono_msec_ticks ();
1202                 }
1203
1204                 mono_coop_mutex_unlock (&worker->heuristic_lock);
1205         }
1206 }
1207
1208 static void
1209 heuristic_notify_work_completed (MonoThreadPoolWorker *worker)
1210 {
1211         g_assert (worker);
1212
1213         InterlockedIncrement (&worker->heuristic_completions);
1214         worker->heuristic_last_dequeue = mono_msec_ticks ();
1215
1216         if (heuristic_should_adjust (worker))
1217                 heuristic_adjust (worker);
1218 }
1219
1220 gboolean
1221 mono_threadpool_worker_notify_completed (MonoThreadPoolWorker *worker)
1222 {
1223         ThreadPoolWorkerCounter counter;
1224
1225         heuristic_notify_work_completed (worker);
1226
1227         counter = COUNTER_READ (worker);
1228         return counter._.working <= counter._.max_working;
1229 }
1230
1231 gint32
1232 mono_threadpool_worker_get_min (MonoThreadPoolWorker *worker)
1233 {
1234         return worker->limit_worker_min;
1235 }
1236
1237 gboolean
1238 mono_threadpool_worker_set_min (MonoThreadPoolWorker *worker, gint32 value)
1239 {
1240         if (value <= 0 || value > worker->limit_worker_max)
1241                 return FALSE;
1242
1243         worker->limit_worker_min = value;
1244         return TRUE;
1245 }
1246
1247 gint32
1248 mono_threadpool_worker_get_max (MonoThreadPoolWorker *worker)
1249 {
1250         return worker->limit_worker_max;
1251 }
1252
1253 gboolean
1254 mono_threadpool_worker_set_max (MonoThreadPoolWorker *worker, gint32 value)
1255 {
1256         gint32 cpu_count = mono_cpu_count ();
1257
1258         if (value < worker->limit_worker_min || value < cpu_count)
1259                 return FALSE;
1260
1261         worker->limit_worker_max = value;
1262         return TRUE;
1263 }
1264
1265 void
1266 mono_threadpool_worker_set_suspended (MonoThreadPoolWorker *worker, gboolean suspended)
1267 {
1268         worker->suspended = suspended;
1269         if (!suspended)
1270                 worker_request (worker);
1271 }