mono/metadata/threadpool-worker-default.c

   1 /**
   2  * \file
   3  * native threadpool worker
   4  *
   5  * Author:
   6  *      Ludovic Henry (ludovic.henry@xamarin.com)
   7  *
   8  * Licensed under the MIT license. See LICENSE file in the project root for full license information.
   9  */
  10
  11 #include <stdlib.h>
  12 #define _USE_MATH_DEFINES // needed by MSVC to define math constants
  13 #include <math.h>
  14 #include <config.h>
  15 #include <glib.h>
  16
  17 #include <mono/metadata/class-internals.h>
  18 #include <mono/metadata/exception.h>
  19 #include <mono/metadata/gc-internals.h>
  20 #include <mono/metadata/object.h>
  21 #include <mono/metadata/object-internals.h>
  22 #include <mono/metadata/threadpool.h>
  23 #include <mono/metadata/threadpool-worker.h>
  24 #include <mono/metadata/threadpool-io.h>
  25 #include <mono/metadata/w32event.h>
  26 #include <mono/utils/atomic.h>
  27 #include <mono/utils/mono-compiler.h>
  28 #include <mono/utils/mono-complex.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 #include <mono/utils/w32api.h>
  37
  38 #define CPU_USAGE_LOW 80
  39 #define CPU_USAGE_HIGH 95
  40
  41 #define MONITOR_INTERVAL 500 // ms
  42 #define MONITOR_MINIMAL_LIFETIME 60 * 1000 // ms
  43
  44 #define WORKER_CREATION_MAX_PER_SEC 10
  45
  46 /* The exponent to apply to the gain. 1.0 means to use linear gain,
  47  * higher values will enhance large moves and damp small ones.
  48  * default: 2.0 */
  49 #define HILL_CLIMBING_GAIN_EXPONENT 2.0
  50
  51 /* The 'cost' of a thread. 0 means drive for increased throughput regardless
  52  * of thread count, higher values bias more against higher thread counts.
  53  * default: 0.15 */
  54 #define HILL_CLIMBING_BIAS 0.15
  55
  56 #define HILL_CLIMBING_WAVE_PERIOD 4
  57 #define HILL_CLIMBING_MAX_WAVE_MAGNITUDE 20
  58 #define HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER 1.0
  59 #define HILL_CLIMBING_WAVE_HISTORY_SIZE 8
  60 #define HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO 3.0
  61 #define HILL_CLIMBING_MAX_CHANGE_PER_SECOND 4
  62 #define HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE 20
  63 #define HILL_CLIMBING_SAMPLE_INTERVAL_LOW 10
  64 #define HILL_CLIMBING_SAMPLE_INTERVAL_HIGH 200
  65 #define HILL_CLIMBING_ERROR_SMOOTHING_FACTOR 0.01
  66 #define HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT 0.15
  67
  68 typedef enum {
  69         TRANSITION_WARMUP,
  70         TRANSITION_INITIALIZING,
  71         TRANSITION_RANDOM_MOVE,
  72         TRANSITION_CLIMBING_MOVE,
  73         TRANSITION_CHANGE_POINT,
  74         TRANSITION_STABILIZING,
  75         TRANSITION_STARVATION,
  76         TRANSITION_THREAD_TIMED_OUT,
  77         TRANSITION_UNDEFINED,
  78 } ThreadPoolHeuristicStateTransition;
  79
  80 typedef struct {
  81         gint32 wave_period;
  82         gint32 samples_to_measure;
  83         gdouble target_throughput_ratio;
  84         gdouble target_signal_to_noise_ratio;
  85         gdouble max_change_per_second;
  86         gdouble max_change_per_sample;
  87         gint32 max_thread_wave_magnitude;
  88         gint32 sample_interval_low;
  89         gdouble thread_magnitude_multiplier;
  90         gint32 sample_interval_high;
  91         gdouble throughput_error_smoothing_factor;
  92         gdouble gain_exponent;
  93         gdouble max_sample_error;
  94
  95         gdouble current_control_setting;
  96         gint64 total_samples;
  97         gint16 last_thread_count;
  98         gdouble elapsed_since_last_change;
  99         gdouble completions_since_last_change;
 100
 101         gdouble average_throughput_noise;
 102
 103         gdouble *samples;
 104         gdouble *thread_counts;
 105
 106         guint32 current_sample_interval;
 107         gpointer random_interval_generator;
 108
 109         gint32 accumulated_completion_count;
 110         gdouble accumulated_sample_duration;
 111 } ThreadPoolHillClimbing;
 112
 113 typedef union {
 114         struct {
 115                 gint16 max_working; /* determined by heuristic */
 116                 gint16 starting; /* starting, but not yet in worker_thread */
 117                 gint16 working; /* executing worker_thread */
 118                 gint16 parked; /* parked */
 119         } _;
 120         gint64 as_gint64;
 121 } ThreadPoolWorkerCounter
 122 #ifdef __GNUC__
 123 __attribute__((aligned(64)))
 124 #endif
 125 ;
 126
 127 typedef struct {
 128         MonoRefCount ref;
 129
 130         MonoThreadPoolWorkerCallback callback;
 131
 132         ThreadPoolWorkerCounter counters;
 133
 134         MonoCoopMutex parked_threads_lock;
 135         gint32 parked_threads_count;
 136         MonoCoopCond parked_threads_cond;
 137
 138         volatile gint32 work_items_count;
 139
 140         guint32 worker_creation_current_second;
 141         guint32 worker_creation_current_count;
 142         MonoCoopMutex worker_creation_lock;
 143
 144         gint32 heuristic_completions;
 145         gint64 heuristic_sample_start;
 146         gint64 heuristic_last_dequeue; // ms
 147         gint64 heuristic_last_adjustment; // ms
 148         gint64 heuristic_adjustment_interval; // ms
 149         ThreadPoolHillClimbing heuristic_hill_climbing;
 150         MonoCoopMutex heuristic_lock;
 151
 152         gint32 limit_worker_min;
 153         gint32 limit_worker_max;
 154
 155         MonoCpuUsageState *cpu_usage_state;
 156         gint32 cpu_usage;
 157
 158         /* suspended by the debugger */
 159         gboolean suspended;
 160
 161         gint32 monitor_status;
 162 } ThreadPoolWorker;
 163
 164 enum {
 165         MONITOR_STATUS_REQUESTED,
 166         MONITOR_STATUS_WAITING_FOR_REQUEST,
 167         MONITOR_STATUS_NOT_RUNNING,
 168 };
 169
 170 static ThreadPoolWorker worker;
 171
 172 #define COUNTER_CHECK(counter) \
 173         do { \
 174                 g_assert (counter._.max_working > 0); \
 175                 g_assert (counter._.starting >= 0); \
 176                 g_assert (counter._.working >= 0); \
 177         } while (0)
 178
 179 #define COUNTER_ATOMIC(var,block) \
 180         do { \
 181                 ThreadPoolWorkerCounter __old; \
 182                 do { \
 183                         __old = COUNTER_READ (); \
 184                         (var) = __old; \
 185                         { block; } \
 186                         COUNTER_CHECK (var); \
 187                 } while (InterlockedCompareExchange64 (&worker.counters.as_gint64, (var).as_gint64, __old.as_gint64) != __old.as_gint64); \
 188         } while (0)
 189
 190 static inline ThreadPoolWorkerCounter
 191 COUNTER_READ (void)
 192 {
 193         ThreadPoolWorkerCounter counter;
 194         counter.as_gint64 = InterlockedRead64 (&worker.counters.as_gint64);
 195         return counter;
 196 }
 197
 198 static gpointer
 199 rand_create (void)
 200 {
 201         mono_rand_open ();
 202         return mono_rand_init (NULL, 0);
 203 }
 204
 205 static guint32
 206 rand_next (gpointer *handle, guint32 min, guint32 max)
 207 {
 208         MonoError error;
 209         guint32 val;
 210         mono_rand_try_get_uint32 (handle, &val, min, max, &error);
 211         // FIXME handle error
 212         mono_error_assert_ok (&error);
 213         return val;
 214 }
 215
 216 static void
 217 destroy (gpointer data)
 218 {
 219         mono_coop_mutex_destroy (&worker.parked_threads_lock);
 220         mono_coop_cond_destroy (&worker.parked_threads_cond);
 221
 222         mono_coop_mutex_destroy (&worker.worker_creation_lock);
 223
 224         mono_coop_mutex_destroy (&worker.heuristic_lock);
 225
 226         g_free (worker.cpu_usage_state);
 227 }
 228
 229 void
 230 mono_threadpool_worker_init (MonoThreadPoolWorkerCallback callback)
 231 {
 232         ThreadPoolHillClimbing *hc;
 233         const char *threads_per_cpu_env;
 234         gint threads_per_cpu;
 235         gint threads_count;
 236
 237         mono_refcount_init (&worker, destroy);
 238
 239         worker.callback = callback;
 240
 241         mono_coop_mutex_init (&worker.parked_threads_lock);
 242         worker.parked_threads_count = 0;
 243         mono_coop_cond_init (&worker.parked_threads_cond);
 244
 245         worker.worker_creation_current_second = -1;
 246         mono_coop_mutex_init (&worker.worker_creation_lock);
 247
 248         worker.heuristic_adjustment_interval = 10;
 249         mono_coop_mutex_init (&worker.heuristic_lock);
 250
 251         mono_rand_open ();
 252
 253         hc = &worker.heuristic_hill_climbing;
 254
 255         hc->wave_period = HILL_CLIMBING_WAVE_PERIOD;
 256         hc->max_thread_wave_magnitude = HILL_CLIMBING_MAX_WAVE_MAGNITUDE;
 257         hc->thread_magnitude_multiplier = (gdouble) HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER;
 258         hc->samples_to_measure = hc->wave_period * HILL_CLIMBING_WAVE_HISTORY_SIZE;
 259         hc->target_throughput_ratio = (gdouble) HILL_CLIMBING_BIAS;
 260         hc->target_signal_to_noise_ratio = (gdouble) HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO;
 261         hc->max_change_per_second = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SECOND;
 262         hc->max_change_per_sample = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE;
 263         hc->sample_interval_low = HILL_CLIMBING_SAMPLE_INTERVAL_LOW;
 264         hc->sample_interval_high = HILL_CLIMBING_SAMPLE_INTERVAL_HIGH;
 265         hc->throughput_error_smoothing_factor = (gdouble) HILL_CLIMBING_ERROR_SMOOTHING_FACTOR;
 266         hc->gain_exponent = (gdouble) HILL_CLIMBING_GAIN_EXPONENT;
 267         hc->max_sample_error = (gdouble) HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT;
 268         hc->current_control_setting = 0;
 269         hc->total_samples = 0;
 270         hc->last_thread_count = 0;
 271         hc->average_throughput_noise = 0;
 272         hc->elapsed_since_last_change = 0;
 273         hc->accumulated_completion_count = 0;
 274         hc->accumulated_sample_duration = 0;
 275         hc->samples = g_new0 (gdouble, hc->samples_to_measure);
 276         hc->thread_counts = g_new0 (gdouble, hc->samples_to_measure);
 277         hc->random_interval_generator = rand_create ();
 278         hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
 279
 280         if (!(threads_per_cpu_env = g_getenv ("MONO_THREADS_PER_CPU")))
 281                 threads_per_cpu = 1;
 282         else
 283                 threads_per_cpu = CLAMP (atoi (threads_per_cpu_env), 1, 50);
 284
 285         threads_count = mono_cpu_count () * threads_per_cpu;
 286
 287         worker.limit_worker_min = threads_count;
 288
 289 #if defined (HOST_ANDROID) || defined (HOST_IOS)
 290         worker.limit_worker_max = CLAMP (threads_count * 100, MIN (threads_count, 200), MAX (threads_count, 200));
 291 #else
 292         worker.limit_worker_max = threads_count * 100;
 293 #endif
 294
 295         worker.counters._.max_working = worker.limit_worker_min;
 296
 297         worker.cpu_usage_state = g_new0 (MonoCpuUsageState, 1);
 298
 299         worker.suspended = FALSE;
 300
 301         worker.monitor_status = MONITOR_STATUS_NOT_RUNNING;
 302 }
 303
 304 void
 305 mono_threadpool_worker_cleanup (void)
 306 {
 307         mono_refcount_dec (&worker);
 308 }
 309
 310 static void
 311 work_item_push (void)
 312 {
 313         gint32 old, new;
 314
 315         do {
 316                 old = InterlockedRead (&worker.work_items_count);
 317                 g_assert (old >= 0);
 318
 319                 new = old + 1;
 320         } while (InterlockedCompareExchange (&worker.work_items_count, new, old) != old);
 321 }
 322
 323 static gboolean
 324 work_item_try_pop (void)
 325 {
 326         gint32 old, new;
 327
 328         do {
 329                 old = InterlockedRead (&worker.work_items_count);
 330                 g_assert (old >= 0);
 331
 332                 if (old == 0)
 333                         return FALSE;
 334
 335                 new = old - 1;
 336         } while (InterlockedCompareExchange (&worker.work_items_count, new, old) != old);
 337
 338         return TRUE;
 339 }
 340
 341 static gint32
 342 work_item_count (void)
 343 {
 344         return InterlockedRead (&worker.work_items_count);
 345 }
 346
 347 static void worker_request (void);
 348
 349 void
 350 mono_threadpool_worker_request (void)
 351 {
 352         if (!mono_refcount_tryinc (&worker))
 353                 return;
 354
 355         work_item_push ();
 356
 357         worker_request ();
 358
 359         mono_refcount_dec (&worker);
 360 }
 361
 362 static void
 363 worker_wait_interrupt (gpointer unused)
 364 {
 365         /* If the runtime is not shutting down, we are not using this mechanism to wake up a unparked thread, and if the
 366          * runtime is shutting down, then we need to wake up ALL the threads.
 367          * It might be a bit wasteful, but I witnessed shutdown hang where the main thread would abort and then wait for all
 368          * background threads to exit (see mono_thread_manage). This would go wrong because not all threadpool threads would
 369          * be unparked. It would end up getting unstucked because of the timeout, but that would delay shutdown by 5-60s. */
 370         if (!mono_runtime_is_shutting_down ())
 371                 return;
 372
 373         if (!mono_refcount_tryinc (&worker))
 374                 return;
 375
 376         mono_coop_mutex_lock (&worker.parked_threads_lock);
 377         mono_coop_cond_broadcast (&worker.parked_threads_cond);
 378         mono_coop_mutex_unlock (&worker.parked_threads_lock);
 379
 380         mono_refcount_dec (&worker);
 381 }
 382
 383 /* return TRUE if timeout, FALSE otherwise (worker unpark or interrupt) */
 384 static gboolean
 385 worker_park (void)
 386 {
 387         gboolean timeout = FALSE;
 388         gboolean interrupted = FALSE;
 389
 390         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker parking",
 391                 GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())));
 392
 393         mono_coop_mutex_lock (&worker.parked_threads_lock);
 394
 395         if (!mono_runtime_is_shutting_down ()) {
 396                 static gpointer rand_handle = NULL;
 397                 MonoInternalThread *thread;
 398                 ThreadPoolWorkerCounter counter;
 399
 400                 if (!rand_handle)
 401                         rand_handle = rand_create ();
 402                 g_assert (rand_handle);
 403
 404                 thread = mono_thread_internal_current ();
 405                 g_assert (thread);
 406
 407                 COUNTER_ATOMIC (counter, {
 408                         counter._.working --;
 409                         counter._.parked ++;
 410                 });
 411
 412                 worker.parked_threads_count += 1;
 413
 414                 mono_thread_info_install_interrupt (worker_wait_interrupt, NULL, &interrupted);
 415                 if (interrupted)
 416                         goto done;
 417
 418                 if (mono_coop_cond_timedwait (&worker.parked_threads_cond, &worker.parked_threads_lock, rand_next (&rand_handle, 5 * 1000, 60 * 1000)) != 0)
 419                         timeout = TRUE;
 420
 421                 mono_thread_info_uninstall_interrupt (&interrupted);
 422
 423 done:
 424                 worker.parked_threads_count -= 1;
 425
 426                 COUNTER_ATOMIC (counter, {
 427                         counter._.working ++;
 428                         counter._.parked --;
 429                 });
 430         }
 431
 432         mono_coop_mutex_unlock (&worker.parked_threads_lock);
 433
 434         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker unparking, timeout? %s interrupted? %s",
 435                 GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())), timeout ? "yes" : "no", interrupted ? "yes" : "no");
 436
 437         return timeout;
 438 }
 439
 440 static gboolean
 441 worker_try_unpark (void)
 442 {
 443         gboolean res = FALSE;
 444
 445         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker",
 446                 GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())));
 447
 448         mono_coop_mutex_lock (&worker.parked_threads_lock);
 449         if (worker.parked_threads_count > 0) {
 450                 mono_coop_cond_signal (&worker.parked_threads_cond);
 451                 res = TRUE;
 452         }
 453         mono_coop_mutex_unlock (&worker.parked_threads_lock);
 454
 455         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker, success? %s",
 456                 GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())), res ? "yes" : "no");
 457
 458         return res;
 459 }
 460
 461 static gsize WINAPI
 462 worker_thread (gpointer unused)
 463 {
 464         MonoInternalThread *thread;
 465         ThreadPoolWorkerCounter counter;
 466
 467         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker starting",
 468                 GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())));
 469
 470         if (!mono_refcount_tryinc (&worker))
 471                 return 0;
 472
 473         COUNTER_ATOMIC (counter, {
 474                 counter._.starting --;
 475                 counter._.working ++;
 476         });
 477
 478         thread = mono_thread_internal_current ();
 479         g_assert (thread);
 480
 481         while (!mono_runtime_is_shutting_down ()) {
 482                 if (mono_thread_interruption_checkpoint ())
 483                         continue;
 484
 485                 if (!work_item_try_pop ()) {
 486                         gboolean timeout;
 487
 488                         timeout = worker_park ();
 489                         if (timeout)
 490                                 break;
 491
 492                         continue;
 493                 }
 494
 495                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker executing",
 496                         GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())));
 497
 498                 worker.callback ();
 499         }
 500
 501         COUNTER_ATOMIC (counter, {
 502                 counter._.working --;
 503         });
 504
 505         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker finishing",
 506                 GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())));
 507
 508         mono_refcount_dec (&worker);
 509
 510         return 0;
 511 }
 512
 513 static gboolean
 514 worker_try_create (void)
 515 {
 516         MonoError error;
 517         MonoInternalThread *thread;
 518         gint64 current_ticks;
 519         gint32 now;
 520         ThreadPoolWorkerCounter counter;
 521
 522         if (mono_runtime_is_shutting_down ())
 523                 return FALSE;
 524
 525         mono_coop_mutex_lock (&worker.worker_creation_lock);
 526
 527         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker",
 528                 GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())));
 529
 530         current_ticks = mono_100ns_ticks ();
 531         if (0 == current_ticks) {
 532                 g_warning ("failed to get 100ns ticks");
 533         } else {
 534                 now = current_ticks / (10 * 1000 * 1000);
 535                 if (worker.worker_creation_current_second != now) {
 536                         worker.worker_creation_current_second = now;
 537                         worker.worker_creation_current_count = 0;
 538                 } else {
 539                         g_assert (worker.worker_creation_current_count <= WORKER_CREATION_MAX_PER_SEC);
 540                         if (worker.worker_creation_current_count == WORKER_CREATION_MAX_PER_SEC) {
 541                                 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",
 542                                         GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())), worker.worker_creation_current_count);
 543                                 mono_coop_mutex_unlock (&worker.worker_creation_lock);
 544                                 return FALSE;
 545                         }
 546                 }
 547         }
 548
 549         COUNTER_ATOMIC (counter, {
 550                 if (counter._.working >= counter._.max_working) {
 551                         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: maximum number of working threads reached",
 552                                 GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())));
 553                         mono_coop_mutex_unlock (&worker.worker_creation_lock);
 554                         return FALSE;
 555                 }
 556                 counter._.starting ++;
 557         });
 558
 559         thread = mono_thread_create_internal (mono_get_root_domain (), worker_thread, NULL, MONO_THREAD_CREATE_FLAGS_THREADPOOL, &error);
 560         if (!thread) {
 561                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: could not create thread due to %s",
 562                         GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())), mono_error_get_message (&error));
 563                 mono_error_cleanup (&error);
 564
 565                 COUNTER_ATOMIC (counter, {
 566                         counter._.starting --;
 567                 });
 568
 569                 mono_coop_mutex_unlock (&worker.worker_creation_lock);
 570
 571                 return FALSE;
 572         }
 573
 574         worker.worker_creation_current_count += 1;
 575
 576         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, created %p, now = %d count = %d",
 577                 GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())), (gpointer) thread->tid, now, worker.worker_creation_current_count);
 578
 579         mono_coop_mutex_unlock (&worker.worker_creation_lock);
 580         return TRUE;
 581 }
 582
 583 static void monitor_ensure_running (void);
 584
 585 static void
 586 worker_request (void)
 587 {
 588         if (worker.suspended)
 589                 return;
 590
 591         monitor_ensure_running ();
 592
 593         if (worker_try_unpark ()) {
 594                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, unparked",
 595                         GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())));
 596                 return;
 597         }
 598
 599         if (worker_try_create ()) {
 600                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, created",
 601                         GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())));
 602                 return;
 603         }
 604
 605         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, failed",
 606                 GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())));
 607 }
 608
 609 static gboolean
 610 monitor_should_keep_running (void)
 611 {
 612         static gint64 last_should_keep_running = -1;
 613
 614         g_assert (worker.monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || worker.monitor_status == MONITOR_STATUS_REQUESTED);
 615
 616         if (InterlockedExchange (&worker.monitor_status, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST) {
 617                 gboolean should_keep_running = TRUE, force_should_keep_running = FALSE;
 618
 619                 if (mono_runtime_is_shutting_down ()) {
 620                         should_keep_running = FALSE;
 621                 } else {
 622                         if (work_item_count () == 0)
 623                                 should_keep_running = FALSE;
 624
 625                         if (!should_keep_running) {
 626                                 if (last_should_keep_running == -1 || mono_100ns_ticks () - last_should_keep_running < MONITOR_MINIMAL_LIFETIME * 1000 * 10) {
 627                                         should_keep_running = force_should_keep_running = TRUE;
 628                                 }
 629                         }
 630                 }
 631
 632                 if (should_keep_running) {
 633                         if (last_should_keep_running == -1 || !force_should_keep_running)
 634                                 last_should_keep_running = mono_100ns_ticks ();
 635                 } else {
 636                         last_should_keep_running = -1;
 637                         if (InterlockedCompareExchange (&worker.monitor_status, MONITOR_STATUS_NOT_RUNNING, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST)
 638                                 return FALSE;
 639                 }
 640         }
 641
 642         g_assert (worker.monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || worker.monitor_status == MONITOR_STATUS_REQUESTED);
 643
 644         return TRUE;
 645 }
 646
 647 static gboolean
 648 monitor_sufficient_delay_since_last_dequeue (void)
 649 {
 650         gint64 threshold;
 651
 652         if (worker.cpu_usage < CPU_USAGE_LOW) {
 653                 threshold = MONITOR_INTERVAL;
 654         } else {
 655                 ThreadPoolWorkerCounter counter;
 656                 counter = COUNTER_READ ();
 657                 threshold = counter._.max_working * MONITOR_INTERVAL * 2;
 658         }
 659
 660         return mono_msec_ticks () >= worker.heuristic_last_dequeue + threshold;
 661 }
 662
 663 static void hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition);
 664
 665 static gsize WINAPI
 666 monitor_thread (gpointer unused)
 667 {
 668         MonoInternalThread *internal;
 669         guint i;
 670
 671         if (!mono_refcount_tryinc (&worker))
 672                 return 0;
 673
 674         internal = mono_thread_internal_current ();
 675         g_assert (internal);
 676
 677         mono_cpu_usage (worker.cpu_usage_state);
 678
 679         // printf ("monitor_thread: start\n");
 680
 681         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, started",
 682                 GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())));
 683
 684         do {
 685                 ThreadPoolWorkerCounter counter;
 686                 gboolean limit_worker_max_reached;
 687                 gint32 interval_left = MONITOR_INTERVAL;
 688                 gint32 awake = 0; /* number of spurious awakes we tolerate before doing a round of rebalancing */
 689
 690                 g_assert (worker.monitor_status != MONITOR_STATUS_NOT_RUNNING);
 691
 692                 // counter = COUNTER_READ ();
 693                 // printf ("monitor_thread: starting = %d working = %d parked = %d max_working = %d\n",
 694                 //      counter._.starting, counter._.working, counter._.parked, counter._.max_working);
 695
 696                 do {
 697                         gint64 ts;
 698                         gboolean alerted = FALSE;
 699
 700                         if (mono_runtime_is_shutting_down ())
 701                                 break;
 702
 703                         ts = mono_msec_ticks ();
 704                         if (mono_thread_info_sleep (interval_left, &alerted) == 0)
 705                                 break;
 706                         interval_left -= mono_msec_ticks () - ts;
 707
 708                         mono_thread_interruption_checkpoint ();
 709                 } while (interval_left > 0 && ++awake < 10);
 710
 711                 if (mono_runtime_is_shutting_down ())
 712                         continue;
 713
 714                 if (worker.suspended)
 715                         continue;
 716
 717                 if (work_item_count () == 0)
 718                         continue;
 719
 720                 worker.cpu_usage = mono_cpu_usage (worker.cpu_usage_state);
 721
 722                 if (!monitor_sufficient_delay_since_last_dequeue ())
 723                         continue;
 724
 725                 limit_worker_max_reached = FALSE;
 726
 727                 COUNTER_ATOMIC (counter, {
 728                         if (counter._.max_working >= worker.limit_worker_max) {
 729                                 limit_worker_max_reached = TRUE;
 730                                 break;
 731                         }
 732                         counter._.max_working ++;
 733                 });
 734
 735                 if (limit_worker_max_reached)
 736                         continue;
 737
 738                 hill_climbing_force_change (counter._.max_working, TRANSITION_STARVATION);
 739
 740                 for (i = 0; i < 5; ++i) {
 741                         if (mono_runtime_is_shutting_down ())
 742                                 break;
 743
 744                         if (worker_try_unpark ()) {
 745                                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, unparked",
 746                                         GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())));
 747                                 break;
 748                         }
 749
 750                         if (worker_try_create ()) {
 751                                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, created",
 752                                         GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())));
 753                                 break;
 754                         }
 755                 }
 756         } while (monitor_should_keep_running ());
 757
 758         // printf ("monitor_thread: stop\n");
 759
 760         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, finished",
 761                 GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())));
 762
 763         mono_refcount_dec (&worker);
 764         return 0;
 765 }
 766
 767 static void
 768 monitor_ensure_running (void)
 769 {
 770         MonoError error;
 771         for (;;) {
 772                 switch (worker.monitor_status) {
 773                 case MONITOR_STATUS_REQUESTED:
 774                         // printf ("monitor_thread: requested\n");
 775                         return;
 776                 case MONITOR_STATUS_WAITING_FOR_REQUEST:
 777                         // printf ("monitor_thread: waiting for request\n");
 778                         InterlockedCompareExchange (&worker.monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_WAITING_FOR_REQUEST);
 779                         break;
 780                 case MONITOR_STATUS_NOT_RUNNING:
 781                         // printf ("monitor_thread: not running\n");
 782                         if (mono_runtime_is_shutting_down ())
 783                                 return;
 784                         if (InterlockedCompareExchange (&worker.monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_NOT_RUNNING) == MONITOR_STATUS_NOT_RUNNING) {
 785                                 // printf ("monitor_thread: creating\n");
 786                                 if (!mono_thread_create_internal (mono_get_root_domain (), monitor_thread, NULL, MONO_THREAD_CREATE_FLAGS_THREADPOOL | MONO_THREAD_CREATE_FLAGS_SMALL_STACK, &error)) {
 787                                         // printf ("monitor_thread: creating failed\n");
 788                                         worker.monitor_status = MONITOR_STATUS_NOT_RUNNING;
 789                                         mono_error_cleanup (&error);
 790                                         mono_refcount_dec (&worker);
 791                                 }
 792                                 return;
 793                         }
 794                         break;
 795                 default: g_assert_not_reached ();
 796                 }
 797         }
 798 }
 799
 800 static void
 801 hill_climbing_change_thread_count (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
 802 {
 803         ThreadPoolHillClimbing *hc;
 804
 805         hc = &worker.heuristic_hill_climbing;
 806
 807         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] hill climbing, change max number of threads %d",
 808                 GUINT_TO_POINTER (MONO_NATIVE_THREAD_ID_TO_UINT (mono_native_thread_id_get ())), new_thread_count);
 809
 810         hc->last_thread_count = new_thread_count;
 811         hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
 812         hc->elapsed_since_last_change = 0;
 813         hc->completions_since_last_change = 0;
 814 }
 815
 816 static void
 817 hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
 818 {
 819         ThreadPoolHillClimbing *hc;
 820
 821         hc = &worker.heuristic_hill_climbing;
 822
 823         if (new_thread_count != hc->last_thread_count) {
 824                 hc->current_control_setting += new_thread_count - hc->last_thread_count;
 825                 hill_climbing_change_thread_count (new_thread_count, transition);
 826         }
 827 }
 828
 829 static double_complex
 830 hill_climbing_get_wave_component (gdouble *samples, guint sample_count, gdouble period)
 831 {
 832         ThreadPoolHillClimbing *hc;
 833         gdouble w, cosine, sine, coeff, q0, q1, q2;
 834         guint i;
 835
 836         g_assert (sample_count >= period);
 837         g_assert (period >= 2);
 838
 839         hc = &worker.heuristic_hill_climbing;
 840
 841         w = 2.0 * M_PI / period;
 842         cosine = cos (w);
 843         sine = sin (w);
 844         coeff = 2.0 * cosine;
 845         q0 = q1 = q2 = 0;
 846
 847         for (i = 0; i < sample_count; ++i) {
 848                 q0 = coeff * q1 - q2 + samples [(hc->total_samples - sample_count + i) % hc->samples_to_measure];
 849                 q2 = q1;
 850                 q1 = q0;
 851         }
 852
 853         return mono_double_complex_scalar_div (mono_double_complex_make (q1 - q2 * cosine, (q2 * sine)), ((gdouble)sample_count));
 854 }
 855
 856 static gint16
 857 hill_climbing_update (gint16 current_thread_count, guint32 sample_duration, gint32 completions, gint64 *adjustment_interval)
 858 {
 859         ThreadPoolHillClimbing *hc;
 860         ThreadPoolHeuristicStateTransition transition;
 861         gdouble throughput;
 862         gdouble throughput_error_estimate;
 863         gdouble confidence;
 864         gdouble move;
 865         gdouble gain;
 866         gint sample_index;
 867         gint sample_count;
 868         gint new_thread_wave_magnitude;
 869         gint new_thread_count;
 870         double_complex thread_wave_component;
 871         double_complex throughput_wave_component;
 872         double_complex ratio;
 873
 874         g_assert (adjustment_interval);
 875
 876         hc = &worker.heuristic_hill_climbing;
 877
 878         /* If someone changed the thread count without telling us, update our records accordingly. */
 879         if (current_thread_count != hc->last_thread_count)
 880                 hill_climbing_force_change (current_thread_count, TRANSITION_INITIALIZING);
 881
 882         /* Update the cumulative stats for this thread count */
 883         hc->elapsed_since_last_change += sample_duration;
 884         hc->completions_since_last_change += completions;
 885
 886         /* Add in any data we've already collected about this sample */
 887         sample_duration += hc->accumulated_sample_duration;
 888         completions += hc->accumulated_completion_count;
 889
 890         /* We need to make sure we're collecting reasonably accurate data. Since we're just counting the end
 891          * of each work item, we are goinng to be missing some data about what really happened during the
 892          * sample interval. The count produced by each thread includes an initial work item that may have
 893          * started well before the start of the interval, and each thread may have been running some new
 894          * work item for some time before the end of the interval, which did not yet get counted. So
 895          * our count is going to be off by +/- threadCount workitems.
 896          *
 897          * The exception is that the thread that reported to us last time definitely wasn't running any work
 898          * at that time, and the thread that's reporting now definitely isn't running a work item now. So
 899          * we really only need to consider threadCount-1 threads.
 900          *
 901          * Thus the percent error in our count is +/- (threadCount-1)/numCompletions.
 902          *
 903          * We cannot rely on the frequency-domain analysis we'll be doing later to filter out this error, because
 904          * of the way it accumulates over time. If this sample is off by, say, 33% in the negative direction,
 905          * then the next one likely will be too. The one after that will include the sum of the completions
 906          * we missed in the previous samples, and so will be 33% positive. So every three samples we'll have
 907          * two "low" samples and one "high" sample. This will appear as periodic variation right in the frequency
 908          * range we're targeting, which will not be filtered by the frequency-domain translation. */
 909         if (hc->total_samples > 0 && ((current_thread_count - 1.0) / completions) >= hc->max_sample_error) {
 910                 /* Not accurate enough yet. Let's accumulate the data so
 911                  * far, and tell the ThreadPoolWorker to collect a little more. */
 912                 hc->accumulated_sample_duration = sample_duration;
 913                 hc->accumulated_completion_count = completions;
 914                 *adjustment_interval = 10;
 915                 return current_thread_count;
 916         }
 917
 918         /* We've got enouugh data for our sample; reset our accumulators for next time. */
 919         hc->accumulated_sample_duration = 0;
 920         hc->accumulated_completion_count = 0;
 921
 922         /* Add the current thread count and throughput sample to our history. */
 923         throughput = ((gdouble) completions) / sample_duration;
 924
 925         sample_index = hc->total_samples % hc->samples_to_measure;
 926         hc->samples [sample_index] = throughput;
 927         hc->thread_counts [sample_index] = current_thread_count;
 928         hc->total_samples ++;
 929
 930         /* Set up defaults for our metrics. */
 931         thread_wave_component = mono_double_complex_make(0, 0);
 932         throughput_wave_component = mono_double_complex_make(0, 0);
 933         throughput_error_estimate = 0;
 934         ratio = mono_double_complex_make(0, 0);
 935         confidence = 0;
 936
 937         transition = TRANSITION_WARMUP;
 938
 939         /* How many samples will we use? It must be at least the three wave periods we're looking for, and it must also
 940          * be a whole multiple of the primary wave's period; otherwise the frequency we're looking for will fall between
 941          * two frequency bands in the Fourier analysis, and we won't be able to measure it accurately. */
 942         sample_count = ((gint) MIN (hc->total_samples - 1, hc->samples_to_measure) / hc->wave_period) * hc->wave_period;
 943
 944         if (sample_count > hc->wave_period) {
 945                 guint i;
 946                 gdouble average_throughput;
 947                 gdouble average_thread_count;
 948                 gdouble sample_sum = 0;
 949                 gdouble thread_sum = 0;
 950
 951                 /* Average the throughput and thread count samples, so we can scale the wave magnitudes later. */
 952                 for (i = 0; i < sample_count; ++i) {
 953                         guint j = (hc->total_samples - sample_count + i) % hc->samples_to_measure;
 954                         sample_sum += hc->samples [j];
 955                         thread_sum += hc->thread_counts [j];
 956                 }
 957
 958                 average_throughput = sample_sum / sample_count;
 959                 average_thread_count = thread_sum / sample_count;
 960
 961                 if (average_throughput > 0 && average_thread_count > 0) {
 962                         gdouble noise_for_confidence, adjacent_period_1, adjacent_period_2;
 963
 964                         /* Calculate the periods of the adjacent frequency bands we'll be using to
 965                          * measure noise levels. We want the two adjacent Fourier frequency bands. */
 966                         adjacent_period_1 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) + 1);
 967                         adjacent_period_2 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) - 1);
 968
 969                         /* Get the the three different frequency components of the throughput (scaled by average
 970                          * throughput). Our "error" estimate (the amount of noise that might be present in the
 971                          * frequency band we're really interested in) is the average of the adjacent bands. */
 972                         throughput_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, hc->wave_period), average_throughput);
 973                         throughput_error_estimate = cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, adjacent_period_1), average_throughput));
 974
 975                         if (adjacent_period_2 <= sample_count) {
 976                                 throughput_error_estimate = MAX (throughput_error_estimate, cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (
 977                                         hc->samples, sample_count, adjacent_period_2), average_throughput)));
 978                         }
 979
 980                         /* Do the same for the thread counts, so we have something to compare to. We don't
 981                          * measure thread count noise, because there is none; these are exact measurements. */
 982                         thread_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->thread_counts, sample_count, hc->wave_period), average_thread_count);
 983
 984                         /* Update our moving average of the throughput noise. We'll use this
 985                          * later as feedback to determine the new size of the thread wave. */
 986                         if (hc->average_throughput_noise == 0) {
 987                                 hc->average_throughput_noise = throughput_error_estimate;
 988                         } else {
 989                                 hc->average_throughput_noise = (hc->throughput_error_smoothing_factor * throughput_error_estimate)
 990                                         + ((1.0 + hc->throughput_error_smoothing_factor) * hc->average_throughput_noise);
 991                         }
 992
 993                         if (cabs (thread_wave_component) > 0) {
 994                                 /* Adjust the throughput wave so it's centered around the target wave,
 995                                  * and then calculate the adjusted throughput/thread ratio. */
 996                                 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);
 997                                 transition = TRANSITION_CLIMBING_MOVE;
 998                         } else {
 999                                 ratio = mono_double_complex_make (0, 0);
1000                                 transition = TRANSITION_STABILIZING;
1001                         }
1002
1003                         noise_for_confidence = MAX (hc->average_throughput_noise, throughput_error_estimate);
1004                         if (noise_for_confidence > 0) {
1005                                 confidence = cabs (thread_wave_component) / noise_for_confidence / hc->target_signal_to_noise_ratio;
1006                         } else {
1007                                 /* there is no noise! */
1008                                 confidence = 1.0;
1009                         }
1010                 }
1011         }
1012
1013         /* We use just the real part of the complex ratio we just calculated. If the throughput signal
1014          * is exactly in phase with the thread signal, this will be the same as taking the magnitude of
1015          * the complex move and moving that far up. If they're 180 degrees out of phase, we'll move
1016          * backward (because this indicates that our changes are having the opposite of the intended effect).
1017          * If they're 90 degrees out of phase, we won't move at all, because we can't tell wether we're
1018          * having a negative or positive effect on throughput. */
1019         move = creal (ratio);
1020         move = CLAMP (move, -1.0, 1.0);
1021
1022         /* Apply our confidence multiplier. */
1023         move *= CLAMP (confidence, -1.0, 1.0);
1024
1025         /* Now apply non-linear gain, such that values around zero are attenuated, while higher values
1026          * are enhanced. This allows us to move quickly if we're far away from the target, but more slowly
1027         * if we're getting close, giving us rapid ramp-up without wild oscillations around the target. */
1028         gain = hc->max_change_per_second * sample_duration;
1029         move = pow (fabs (move), hc->gain_exponent) * (move >= 0.0 ? 1 : -1) * gain;
1030         move = MIN (move, hc->max_change_per_sample);
1031
1032         /* If the result was positive, and CPU is > 95%, refuse the move. */
1033         if (move > 0.0 && worker.cpu_usage > CPU_USAGE_HIGH)
1034                 move = 0.0;
1035
1036         /* Apply the move to our control setting. */
1037         hc->current_control_setting += move;
1038
1039         /* Calculate the new thread wave magnitude, which is based on the moving average we've been keeping of the
1040          * throughput error.  This average starts at zero, so we'll start with a nice safe little wave at first. */
1041         new_thread_wave_magnitude = (gint)(0.5 + (hc->current_control_setting * hc->average_throughput_noise
1042                 * hc->target_signal_to_noise_ratio * hc->thread_magnitude_multiplier * 2.0));
1043         new_thread_wave_magnitude = CLAMP (new_thread_wave_magnitude, 1, hc->max_thread_wave_magnitude);
1044
1045         /* Make sure our control setting is within the ThreadPoolWorker's limits. */
1046         hc->current_control_setting = CLAMP (hc->current_control_setting, worker.limit_worker_min, worker.limit_worker_max - new_thread_wave_magnitude);
1047
1048         /* Calculate the new thread count (control setting + square wave). */
1049         new_thread_count = (gint)(hc->current_control_setting + new_thread_wave_magnitude * ((hc->total_samples / (hc->wave_period / 2)) % 2));
1050
1051         /* Make sure the new thread count doesn't exceed the ThreadPoolWorker's limits. */
1052         new_thread_count = CLAMP (new_thread_count, worker.limit_worker_min, worker.limit_worker_max);
1053
1054         if (new_thread_count != current_thread_count)
1055                 hill_climbing_change_thread_count (new_thread_count, transition);
1056
1057         if (creal (ratio) < 0.0 && new_thread_count == worker.limit_worker_min)
1058                 *adjustment_interval = (gint)(0.5 + hc->current_sample_interval * (10.0 * MAX (-1.0 * creal (ratio), 1.0)));
1059         else
1060                 *adjustment_interval = hc->current_sample_interval;
1061
1062         return new_thread_count;
1063 }
1064
1065 static gboolean
1066 heuristic_should_adjust (void)
1067 {
1068         if (worker.heuristic_last_dequeue > worker.heuristic_last_adjustment + worker.heuristic_adjustment_interval) {
1069                 ThreadPoolWorkerCounter counter;
1070                 counter = COUNTER_READ ();
1071                 if (counter._.working <= counter._.max_working)
1072                         return TRUE;
1073         }
1074
1075         return FALSE;
1076 }
1077
1078 static void
1079 heuristic_adjust (void)
1080 {
1081         if (mono_coop_mutex_trylock (&worker.heuristic_lock) == 0) {
1082                 gint32 completions = InterlockedExchange (&worker.heuristic_completions, 0);
1083                 gint64 sample_end = mono_msec_ticks ();
1084                 gint64 sample_duration = sample_end - worker.heuristic_sample_start;
1085
1086                 if (sample_duration >= worker.heuristic_adjustment_interval / 2) {
1087                         ThreadPoolWorkerCounter counter;
1088                         gint16 new_thread_count;
1089
1090                         counter = COUNTER_READ ();
1091                         new_thread_count = hill_climbing_update (counter._.max_working, sample_duration, completions, &worker.heuristic_adjustment_interval);
1092
1093                         COUNTER_ATOMIC (counter, {
1094                                 counter._.max_working = new_thread_count;
1095                         });
1096
1097                         if (new_thread_count > counter._.max_working)
1098                                 worker_request ();
1099
1100                         worker.heuristic_sample_start = sample_end;
1101                         worker.heuristic_last_adjustment = mono_msec_ticks ();
1102                 }
1103
1104                 mono_coop_mutex_unlock (&worker.heuristic_lock);
1105         }
1106 }
1107
1108 static void
1109 heuristic_notify_work_completed (void)
1110 {
1111         InterlockedIncrement (&worker.heuristic_completions);
1112         worker.heuristic_last_dequeue = mono_msec_ticks ();
1113
1114         if (heuristic_should_adjust ())
1115                 heuristic_adjust ();
1116 }
1117
1118 gboolean
1119 mono_threadpool_worker_notify_completed (void)
1120 {
1121         ThreadPoolWorkerCounter counter;
1122
1123         heuristic_notify_work_completed ();
1124
1125         counter = COUNTER_READ ();
1126         return counter._.working <= counter._.max_working;
1127 }
1128
1129 gint32
1130 mono_threadpool_worker_get_min (void)
1131 {
1132         gint32 ret;
1133
1134         if (!mono_refcount_tryinc (&worker))
1135                 return 0;
1136
1137         ret = worker.limit_worker_min;
1138
1139         mono_refcount_dec (&worker);
1140         return ret;
1141 }
1142
1143 gboolean
1144 mono_threadpool_worker_set_min (gint32 value)
1145 {
1146         if (value <= 0 || value > worker.limit_worker_max)
1147                 return FALSE;
1148
1149         if (!mono_refcount_tryinc (&worker))
1150                 return FALSE;
1151
1152         worker.limit_worker_min = value;
1153
1154         mono_refcount_dec (&worker);
1155         return TRUE;
1156 }
1157
1158 gint32
1159 mono_threadpool_worker_get_max (void)
1160 {
1161         gint32 ret;
1162
1163         if (!mono_refcount_tryinc (&worker))
1164                 return 0;
1165
1166         ret = worker.limit_worker_max;
1167
1168         mono_refcount_dec (&worker);
1169         return ret;
1170 }
1171
1172 gboolean
1173 mono_threadpool_worker_set_max (gint32 value)
1174 {
1175         gint32 cpu_count;
1176
1177         cpu_count = mono_cpu_count ();
1178         if (value < worker.limit_worker_min || value < cpu_count)
1179                 return FALSE;
1180
1181         if (!mono_refcount_tryinc (&worker))
1182                 return FALSE;
1183
1184         worker.limit_worker_max = value;
1185
1186         mono_refcount_dec (&worker);
1187         return TRUE;
1188 }
1189
1190 void
1191 mono_threadpool_worker_set_suspended (gboolean suspended)
1192 {
1193         if (!mono_refcount_tryinc (&worker))
1194                 return;
1195
1196         worker.suspended = suspended;
1197         if (!suspended)
1198                 worker_request ();
1199
1200         mono_refcount_dec (&worker);
1201 }