mono/metadata/threadpool-ms.c

   1 /*
   2  * threadpool-ms.c: Microsoft threadpool runtime support
   3  *
   4  * Author:
   5  *      Ludovic Henry (ludovic.henry@xamarin.com)
   6  *
   7  * Copyright 2015 Xamarin, Inc (http://www.xamarin.com)
   8  * Licensed under the MIT license. See LICENSE file in the project root for full license information.
   9  */
  10
  11 //
  12 // Copyright (c) Microsoft. All rights reserved.
  13 // Licensed under the MIT license. See LICENSE file in the project root for full license information.
  14 //
  15 // Files:
  16 //  - src/vm/comthreadpool.cpp
  17 //  - src/vm/win32threadpoolcpp
  18 //  - src/vm/threadpoolrequest.cpp
  19 //  - src/vm/hillclimbing.cpp
  20 //
  21 // Ported from C++ to C and adjusted to Mono runtime
  22
  23 #include <stdlib.h>
  24 #define _USE_MATH_DEFINES // needed by MSVC to define math constants
  25 #include <math.h>
  26 #include <config.h>
  27 #include <glib.h>
  28
  29 #include <mono/metadata/class-internals.h>
  30 #include <mono/metadata/exception.h>
  31 #include <mono/metadata/gc-internals.h>
  32 #include <mono/metadata/object.h>
  33 #include <mono/metadata/object-internals.h>
  34 #include <mono/metadata/threadpool-ms.h>
  35 #include <mono/metadata/threadpool-ms-io.h>
  36 #include <mono/utils/atomic.h>
  37 #include <mono/utils/mono-compiler.h>
  38 #include <mono/utils/mono-complex.h>
  39 #include <mono/utils/mono-lazy-init.h>
  40 #include <mono/utils/mono-logger.h>
  41 #include <mono/utils/mono-logger-internals.h>
  42 #include <mono/utils/mono-proclib.h>
  43 #include <mono/utils/mono-threads.h>
  44 #include <mono/utils/mono-time.h>
  45 #include <mono/utils/mono-rand.h>
  46
  47 #define CPU_USAGE_LOW 80
  48 #define CPU_USAGE_HIGH 95
  49
  50 #define MONITOR_INTERVAL 500 // ms
  51 #define MONITOR_MINIMAL_LIFETIME 60 * 1000 // ms
  52
  53 #define WORKER_CREATION_MAX_PER_SEC 10
  54
  55 /* The exponent to apply to the gain. 1.0 means to use linear gain,
  56  * higher values will enhance large moves and damp small ones.
  57  * default: 2.0 */
  58 #define HILL_CLIMBING_GAIN_EXPONENT 2.0
  59
  60 /* The 'cost' of a thread. 0 means drive for increased throughput regardless
  61  * of thread count, higher values bias more against higher thread counts.
  62  * default: 0.15 */
  63 #define HILL_CLIMBING_BIAS 0.15
  64
  65 #define HILL_CLIMBING_WAVE_PERIOD 4
  66 #define HILL_CLIMBING_MAX_WAVE_MAGNITUDE 20
  67 #define HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER 1.0
  68 #define HILL_CLIMBING_WAVE_HISTORY_SIZE 8
  69 #define HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO 3.0
  70 #define HILL_CLIMBING_MAX_CHANGE_PER_SECOND 4
  71 #define HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE 20
  72 #define HILL_CLIMBING_SAMPLE_INTERVAL_LOW 10
  73 #define HILL_CLIMBING_SAMPLE_INTERVAL_HIGH 200
  74 #define HILL_CLIMBING_ERROR_SMOOTHING_FACTOR 0.01
  75 #define HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT 0.15
  76
  77 typedef union {
  78         struct {
  79                 gint16 max_working; /* determined by heuristic */
  80                 gint16 active; /* executing worker_thread */
  81                 gint16 working; /* actively executing worker_thread, not parked */
  82                 gint16 parked; /* parked */
  83         } _;
  84         gint64 as_gint64;
  85 } ThreadPoolCounter;
  86
  87 typedef struct {
  88         MonoDomain *domain;
  89         gint32 outstanding_request;
  90 } ThreadPoolDomain;
  91
  92 typedef MonoInternalThread ThreadPoolWorkingThread;
  93
  94 typedef struct {
  95         gint32 wave_period;
  96         gint32 samples_to_measure;
  97         gdouble target_throughput_ratio;
  98         gdouble target_signal_to_noise_ratio;
  99         gdouble max_change_per_second;
 100         gdouble max_change_per_sample;
 101         gint32 max_thread_wave_magnitude;
 102         gint32 sample_interval_low;
 103         gdouble thread_magnitude_multiplier;
 104         gint32 sample_interval_high;
 105         gdouble throughput_error_smoothing_factor;
 106         gdouble gain_exponent;
 107         gdouble max_sample_error;
 108
 109         gdouble current_control_setting;
 110         gint64 total_samples;
 111         gint16 last_thread_count;
 112         gdouble elapsed_since_last_change;
 113         gdouble completions_since_last_change;
 114
 115         gdouble average_throughput_noise;
 116
 117         gdouble *samples;
 118         gdouble *thread_counts;
 119
 120         guint32 current_sample_interval;
 121         gpointer random_interval_generator;
 122
 123         gint32 accumulated_completion_count;
 124         gdouble accumulated_sample_duration;
 125 } ThreadPoolHillClimbing;
 126
 127 typedef struct {
 128         ThreadPoolCounter counters;
 129
 130         GPtrArray *domains; // ThreadPoolDomain* []
 131         MonoCoopMutex domains_lock;
 132
 133         GPtrArray *working_threads; // ThreadPoolWorkingThread* []
 134         gint32 parked_threads_count;
 135         MonoCoopCond parked_threads_cond;
 136         MonoCoopMutex active_threads_lock; /* protect access to working_threads and parked_threads */
 137
 138         guint32 worker_creation_current_second;
 139         guint32 worker_creation_current_count;
 140         MonoCoopMutex worker_creation_lock;
 141
 142         gint32 heuristic_completions;
 143         guint32 heuristic_sample_start;
 144         guint32 heuristic_last_dequeue; // ms
 145         guint32 heuristic_last_adjustment; // ms
 146         guint32 heuristic_adjustment_interval; // ms
 147         ThreadPoolHillClimbing heuristic_hill_climbing;
 148         MonoCoopMutex heuristic_lock;
 149
 150         gint32 limit_worker_min;
 151         gint32 limit_worker_max;
 152         gint32 limit_io_min;
 153         gint32 limit_io_max;
 154
 155         MonoCpuUsageState *cpu_usage_state;
 156         gint32 cpu_usage;
 157
 158         /* suspended by the debugger */
 159         gboolean suspended;
 160 } ThreadPool;
 161
 162 typedef enum {
 163         TRANSITION_WARMUP,
 164         TRANSITION_INITIALIZING,
 165         TRANSITION_RANDOM_MOVE,
 166         TRANSITION_CLIMBING_MOVE,
 167         TRANSITION_CHANGE_POINT,
 168         TRANSITION_STABILIZING,
 169         TRANSITION_STARVATION,
 170         TRANSITION_THREAD_TIMED_OUT,
 171         TRANSITION_UNDEFINED,
 172 } ThreadPoolHeuristicStateTransition;
 173
 174 static mono_lazy_init_t status = MONO_LAZY_INIT_STATUS_NOT_INITIALIZED;
 175
 176 enum {
 177         MONITOR_STATUS_REQUESTED,
 178         MONITOR_STATUS_WAITING_FOR_REQUEST,
 179         MONITOR_STATUS_NOT_RUNNING,
 180 };
 181
 182 static gint32 monitor_status = MONITOR_STATUS_NOT_RUNNING;
 183
 184 static ThreadPool* threadpool;
 185
 186 #define COUNTER_CHECK(counter) \
 187         do { \
 188                 g_assert (counter._.max_working > 0); \
 189                 g_assert (counter._.working >= 0); \
 190                 g_assert (counter._.active >= 0); \
 191         } while (0)
 192
 193 #define COUNTER_READ() (InterlockedRead64 (&threadpool->counters.as_gint64))
 194
 195 #define COUNTER_ATOMIC(var,block) \
 196         do { \
 197                 ThreadPoolCounter __old; \
 198                 do { \
 199                         g_assert (threadpool); \
 200                         __old.as_gint64 = COUNTER_READ (); \
 201                         (var) = __old; \
 202                         { block; } \
 203                         COUNTER_CHECK (var); \
 204                 } while (InterlockedCompareExchange64 (&threadpool->counters.as_gint64, (var).as_gint64, __old.as_gint64) != __old.as_gint64); \
 205         } while (0)
 206
 207 #define COUNTER_TRY_ATOMIC(res,var,block) \
 208         do { \
 209                 ThreadPoolCounter __old; \
 210                 do { \
 211                         g_assert (threadpool); \
 212                         __old.as_gint64 = COUNTER_READ (); \
 213                         (var) = __old; \
 214                         (res) = FALSE; \
 215                         { block; } \
 216                         COUNTER_CHECK (var); \
 217                         (res) = InterlockedCompareExchange64 (&threadpool->counters.as_gint64, (var).as_gint64, __old.as_gint64) == __old.as_gint64; \
 218                 } while (0); \
 219         } while (0)
 220
 221 static gpointer
 222 rand_create (void)
 223 {
 224         mono_rand_open ();
 225         return mono_rand_init (NULL, 0);
 226 }
 227
 228 static guint32
 229 rand_next (gpointer *handle, guint32 min, guint32 max)
 230 {
 231         MonoError error;
 232         guint32 val;
 233         mono_rand_try_get_uint32 (handle, &val, min, max, &error);
 234         // FIXME handle error
 235         mono_error_assert_ok (&error);
 236         return val;
 237 }
 238
 239 static void
 240 rand_free (gpointer handle)
 241 {
 242         mono_rand_close (handle);
 243 }
 244
 245 static void
 246 initialize (void)
 247 {
 248         ThreadPoolHillClimbing *hc;
 249         const char *threads_per_cpu_env;
 250         gint threads_per_cpu;
 251         gint threads_count;
 252
 253         g_assert (!threadpool);
 254         threadpool = g_new0 (ThreadPool, 1);
 255         g_assert (threadpool);
 256
 257         threadpool->domains = g_ptr_array_new ();
 258         mono_coop_mutex_init (&threadpool->domains_lock);
 259
 260         threadpool->parked_threads_count = 0;
 261         mono_coop_cond_init (&threadpool->parked_threads_cond);
 262         threadpool->working_threads = g_ptr_array_new ();
 263         mono_coop_mutex_init (&threadpool->active_threads_lock);
 264
 265         threadpool->worker_creation_current_second = -1;
 266         mono_coop_mutex_init (&threadpool->worker_creation_lock);
 267
 268         threadpool->heuristic_adjustment_interval = 10;
 269         mono_coop_mutex_init (&threadpool->heuristic_lock);
 270
 271         mono_rand_open ();
 272
 273         hc = &threadpool->heuristic_hill_climbing;
 274
 275         hc->wave_period = HILL_CLIMBING_WAVE_PERIOD;
 276         hc->max_thread_wave_magnitude = HILL_CLIMBING_MAX_WAVE_MAGNITUDE;
 277         hc->thread_magnitude_multiplier = (gdouble) HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER;
 278         hc->samples_to_measure = hc->wave_period * HILL_CLIMBING_WAVE_HISTORY_SIZE;
 279         hc->target_throughput_ratio = (gdouble) HILL_CLIMBING_BIAS;
 280         hc->target_signal_to_noise_ratio = (gdouble) HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO;
 281         hc->max_change_per_second = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SECOND;
 282         hc->max_change_per_sample = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE;
 283         hc->sample_interval_low = HILL_CLIMBING_SAMPLE_INTERVAL_LOW;
 284         hc->sample_interval_high = HILL_CLIMBING_SAMPLE_INTERVAL_HIGH;
 285         hc->throughput_error_smoothing_factor = (gdouble) HILL_CLIMBING_ERROR_SMOOTHING_FACTOR;
 286         hc->gain_exponent = (gdouble) HILL_CLIMBING_GAIN_EXPONENT;
 287         hc->max_sample_error = (gdouble) HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT;
 288         hc->current_control_setting = 0;
 289         hc->total_samples = 0;
 290         hc->last_thread_count = 0;
 291         hc->average_throughput_noise = 0;
 292         hc->elapsed_since_last_change = 0;
 293         hc->accumulated_completion_count = 0;
 294         hc->accumulated_sample_duration = 0;
 295         hc->samples = g_new0 (gdouble, hc->samples_to_measure);
 296         hc->thread_counts = g_new0 (gdouble, hc->samples_to_measure);
 297         hc->random_interval_generator = rand_create ();
 298         hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
 299
 300         if (!(threads_per_cpu_env = g_getenv ("MONO_THREADS_PER_CPU")))
 301                 threads_per_cpu = 1;
 302         else
 303                 threads_per_cpu = CLAMP (atoi (threads_per_cpu_env), 1, 50);
 304
 305         threads_count = mono_cpu_count () * threads_per_cpu;
 306
 307         threadpool->limit_worker_min = threadpool->limit_io_min = threads_count;
 308
 309 #if defined (PLATFORM_ANDROID) || defined (HOST_IOS)
 310         threadpool->limit_worker_max = threadpool->limit_io_max = CLAMP (threads_count * 100, MIN (threads_count, 200), MAX (threads_count, 200));
 311 #else
 312         threadpool->limit_worker_max = threadpool->limit_io_max = threads_count * 100;
 313 #endif
 314
 315         threadpool->counters._.max_working = threadpool->limit_worker_min;
 316
 317         threadpool->cpu_usage_state = g_new0 (MonoCpuUsageState, 1);
 318
 319         threadpool->suspended = FALSE;
 320 }
 321
 322 static void worker_kill (ThreadPoolWorkingThread *thread);
 323
 324 static void
 325 cleanup (void)
 326 {
 327         guint i;
 328
 329         /* we make the assumption along the code that we are
 330          * cleaning up only if the runtime is shutting down */
 331         g_assert (mono_runtime_is_shutting_down ());
 332
 333         while (monitor_status != MONITOR_STATUS_NOT_RUNNING)
 334                 mono_thread_info_sleep (1, NULL);
 335
 336         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 337
 338         /* stop all threadpool->working_threads */
 339         for (i = 0; i < threadpool->working_threads->len; ++i)
 340                 worker_kill ((ThreadPoolWorkingThread*) g_ptr_array_index (threadpool->working_threads, i));
 341
 342         /* unpark all threadpool->parked_threads */
 343         mono_coop_cond_broadcast (&threadpool->parked_threads_cond);
 344
 345         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 346 }
 347
 348 gboolean
 349 mono_threadpool_ms_enqueue_work_item (MonoDomain *domain, MonoObject *work_item, MonoError *error)
 350 {
 351         static MonoClass *threadpool_class = NULL;
 352         static MonoMethod *unsafe_queue_custom_work_item_method = NULL;
 353         MonoDomain *current_domain;
 354         MonoBoolean f;
 355         gpointer args [2];
 356
 357         mono_error_init (error);
 358         g_assert (work_item);
 359
 360         if (!threadpool_class)
 361                 threadpool_class = mono_class_load_from_name (mono_defaults.corlib, "System.Threading", "ThreadPool");
 362
 363         if (!unsafe_queue_custom_work_item_method)
 364                 unsafe_queue_custom_work_item_method = mono_class_get_method_from_name (threadpool_class, "UnsafeQueueCustomWorkItem", 2);
 365         g_assert (unsafe_queue_custom_work_item_method);
 366
 367         f = FALSE;
 368
 369         args [0] = (gpointer) work_item;
 370         args [1] = (gpointer) &f;
 371
 372         current_domain = mono_domain_get ();
 373         if (current_domain == domain) {
 374                 mono_runtime_invoke_checked (unsafe_queue_custom_work_item_method, NULL, args, error);
 375                 return_val_if_nok (error, FALSE);
 376         } else {
 377                 mono_thread_push_appdomain_ref (domain);
 378                 if (mono_domain_set (domain, FALSE)) {
 379                         mono_runtime_invoke_checked (unsafe_queue_custom_work_item_method, NULL, args, error);
 380                         if (!is_ok (error)) {
 381                                 mono_thread_pop_appdomain_ref ();
 382                                 return FALSE;
 383                         }
 384                         mono_domain_set (current_domain, TRUE);
 385                 }
 386                 mono_thread_pop_appdomain_ref ();
 387         }
 388         return TRUE;
 389 }
 390
 391 /* LOCKING: threadpool->domains_lock must be held */
 392 static void
 393 domain_add (ThreadPoolDomain *tpdomain)
 394 {
 395         guint i, len;
 396
 397         g_assert (tpdomain);
 398
 399         len = threadpool->domains->len;
 400         for (i = 0; i < len; ++i) {
 401                 if (g_ptr_array_index (threadpool->domains, i) == tpdomain)
 402                         break;
 403         }
 404
 405         if (i == len)
 406                 g_ptr_array_add (threadpool->domains, tpdomain);
 407 }
 408
 409 /* LOCKING: threadpool->domains_lock must be held */
 410 static gboolean
 411 domain_remove (ThreadPoolDomain *tpdomain)
 412 {
 413         g_assert (tpdomain);
 414         return g_ptr_array_remove (threadpool->domains, tpdomain);
 415 }
 416
 417 /* LOCKING: threadpool->domains_lock must be held */
 418 static ThreadPoolDomain *
 419 domain_get (MonoDomain *domain, gboolean create)
 420 {
 421         ThreadPoolDomain *tpdomain = NULL;
 422         guint i;
 423
 424         g_assert (domain);
 425
 426         for (i = 0; i < threadpool->domains->len; ++i) {
 427                 tpdomain = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i);
 428                 if (tpdomain->domain == domain)
 429                         return tpdomain;
 430         }
 431
 432         if (create) {
 433                 tpdomain = g_new0 (ThreadPoolDomain, 1);
 434                 tpdomain->domain = domain;
 435                 domain_add (tpdomain);
 436         }
 437
 438         return tpdomain;
 439 }
 440
 441 static void
 442 domain_free (ThreadPoolDomain *tpdomain)
 443 {
 444         g_free (tpdomain);
 445 }
 446
 447 /* LOCKING: threadpool->domains_lock must be held */
 448 static gboolean
 449 domain_any_has_request (void)
 450 {
 451         guint i;
 452
 453         for (i = 0; i < threadpool->domains->len; ++i) {
 454                 ThreadPoolDomain *tmp = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i);
 455                 if (tmp->outstanding_request > 0)
 456                         return TRUE;
 457         }
 458
 459         return FALSE;
 460 }
 461
 462 /* LOCKING: threadpool->domains_lock must be held */
 463 static ThreadPoolDomain *
 464 domain_get_next (ThreadPoolDomain *current)
 465 {
 466         ThreadPoolDomain *tpdomain = NULL;
 467         guint len;
 468
 469         len = threadpool->domains->len;
 470         if (len > 0) {
 471                 guint i, current_idx = -1;
 472                 if (current) {
 473                         for (i = 0; i < len; ++i) {
 474                                 if (current == g_ptr_array_index (threadpool->domains, i)) {
 475                                         current_idx = i;
 476                                         break;
 477                                 }
 478                         }
 479                         g_assert (current_idx >= 0);
 480                 }
 481                 for (i = current_idx + 1; i < len + current_idx + 1; ++i) {
 482                         ThreadPoolDomain *tmp = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i % len);
 483                         if (tmp->outstanding_request > 0) {
 484                                 tpdomain = tmp;
 485                                 break;
 486                         }
 487                 }
 488         }
 489
 490         return tpdomain;
 491 }
 492
 493 static void
 494 worker_wait_interrupt (gpointer data)
 495 {
 496         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 497         mono_coop_cond_signal (&threadpool->parked_threads_cond);
 498         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 499 }
 500
 501 /* return TRUE if timeout, FALSE otherwise (worker unpark or interrupt) */
 502 static gboolean
 503 worker_park (void)
 504 {
 505         gboolean timeout = FALSE;
 506
 507         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] current worker parking", mono_native_thread_id_get ());
 508
 509         mono_gc_set_skip_thread (TRUE);
 510
 511         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 512
 513         if (!mono_runtime_is_shutting_down ()) {
 514                 static gpointer rand_handle = NULL;
 515                 MonoInternalThread *thread_internal;
 516                 gboolean interrupted = FALSE;
 517
 518                 if (!rand_handle)
 519                         rand_handle = rand_create ();
 520                 g_assert (rand_handle);
 521
 522                 thread_internal = mono_thread_internal_current ();
 523                 g_assert (thread_internal);
 524
 525                 threadpool->parked_threads_count += 1;
 526                 g_ptr_array_remove_fast (threadpool->working_threads, thread_internal);
 527
 528                 mono_thread_info_install_interrupt (worker_wait_interrupt, NULL, &interrupted);
 529                 if (interrupted)
 530                         goto done;
 531
 532                 if (mono_coop_cond_timedwait (&threadpool->parked_threads_cond, &threadpool->active_threads_lock, rand_next (&rand_handle, 5 * 1000, 60 * 1000)) != 0)
 533                         timeout = TRUE;
 534
 535                 mono_thread_info_uninstall_interrupt (&interrupted);
 536
 537 done:
 538                 g_ptr_array_add (threadpool->working_threads, thread_internal);
 539                 threadpool->parked_threads_count -= 1;
 540         }
 541
 542         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 543
 544         mono_gc_set_skip_thread (FALSE);
 545
 546         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] current worker unparking, timeout? %s", mono_native_thread_id_get (), timeout ? "yes" : "no");
 547
 548         return timeout;
 549 }
 550
 551 static gboolean
 552 worker_try_unpark (void)
 553 {
 554         gboolean res = FALSE;
 555
 556         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker", mono_native_thread_id_get ());
 557
 558         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 559         if (threadpool->parked_threads_count > 0) {
 560                 mono_coop_cond_signal (&threadpool->parked_threads_cond);
 561                 res = TRUE;
 562         }
 563         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 564
 565         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker, success? %s", mono_native_thread_id_get (), res ? "yes" : "no");
 566
 567         return res;
 568 }
 569
 570 static void
 571 worker_kill (ThreadPoolWorkingThread *thread)
 572 {
 573         if (thread == mono_thread_internal_current ())
 574                 return;
 575
 576         mono_thread_internal_stop ((MonoInternalThread*) thread);
 577 }
 578
 579 static void
 580 worker_thread (gpointer data)
 581 {
 582         MonoError error;
 583         MonoInternalThread *thread;
 584         ThreadPoolDomain *tpdomain, *previous_tpdomain;
 585         ThreadPoolCounter counter;
 586         gboolean retire = FALSE;
 587
 588         mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker starting", mono_native_thread_id_get ());
 589
 590         g_assert (threadpool);
 591
 592         thread = mono_thread_internal_current ();
 593         g_assert (thread);
 594
 595         mono_thread_set_name_internal (thread, mono_string_new (mono_get_root_domain (), "Threadpool worker"), FALSE, &error);
 596         mono_error_assert_ok (&error);
 597
 598         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 599         g_ptr_array_add (threadpool->working_threads, thread);
 600         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 601
 602         previous_tpdomain = NULL;
 603
 604         mono_coop_mutex_lock (&threadpool->domains_lock);
 605
 606         while (!mono_runtime_is_shutting_down ()) {
 607                 tpdomain = NULL;
 608
 609                 if ((thread->state & (ThreadState_StopRequested | ThreadState_SuspendRequested)) != 0) {
 610                         mono_coop_mutex_unlock (&threadpool->domains_lock);
 611                         mono_thread_interruption_checkpoint ();
 612                         mono_coop_mutex_lock (&threadpool->domains_lock);
 613                 }
 614
 615                 if (retire || !(tpdomain = domain_get_next (previous_tpdomain))) {
 616                         gboolean timeout;
 617
 618                         COUNTER_ATOMIC (counter, {
 619                                 counter._.working --;
 620                                 counter._.parked ++;
 621                         });
 622
 623                         mono_coop_mutex_unlock (&threadpool->domains_lock);
 624                         timeout = worker_park ();
 625                         mono_coop_mutex_lock (&threadpool->domains_lock);
 626
 627                         COUNTER_ATOMIC (counter, {
 628                                 counter._.working ++;
 629                                 counter._.parked --;
 630                         });
 631
 632                         if (timeout)
 633                                 break;
 634
 635                         if (retire)
 636                                 retire = FALSE;
 637
 638                         continue;
 639                 }
 640
 641                 tpdomain->outstanding_request --;
 642                 g_assert (tpdomain->outstanding_request >= 0);
 643
 644                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker running in domain %p",
 645                         mono_native_thread_id_get (), tpdomain->domain, tpdomain->outstanding_request);
 646
 647                 g_assert (tpdomain->domain);
 648                 g_assert (tpdomain->domain->threadpool_jobs >= 0);
 649                 tpdomain->domain->threadpool_jobs ++;
 650
 651                 mono_coop_mutex_unlock (&threadpool->domains_lock);
 652
 653                 mono_thread_push_appdomain_ref (tpdomain->domain);
 654                 if (mono_domain_set (tpdomain->domain, FALSE)) {
 655                         MonoObject *exc = NULL, *res;
 656
 657                         res = mono_runtime_try_invoke (mono_defaults.threadpool_perform_wait_callback_method, NULL, NULL, &exc, &error);
 658                         if (exc || !mono_error_ok(&error)) {
 659                                 if (exc == NULL)
 660                                         exc = (MonoObject *) mono_error_convert_to_exception (&error);
 661                                 else
 662                                         mono_error_cleanup (&error);
 663                                 mono_thread_internal_unhandled_exception (exc);
 664                         } else if (res && *(MonoBoolean*) mono_object_unbox (res) == FALSE)
 665                                 retire = TRUE;
 666
 667                         mono_thread_clr_state (thread, (MonoThreadState)~ThreadState_Background);
 668                         if (!mono_thread_test_state (thread , ThreadState_Background))
 669                                 ves_icall_System_Threading_Thread_SetState (thread, ThreadState_Background);
 670
 671                         mono_domain_set (mono_get_root_domain (), TRUE);
 672                 }
 673                 mono_thread_pop_appdomain_ref ();
 674
 675                 mono_coop_mutex_lock (&threadpool->domains_lock);
 676
 677                 tpdomain->domain->threadpool_jobs --;
 678                 g_assert (tpdomain->domain->threadpool_jobs >= 0);
 679
 680                 if (tpdomain->domain->threadpool_jobs == 0 && mono_domain_is_unloading (tpdomain->domain)) {
 681                         gboolean removed = domain_remove (tpdomain);
 682                         g_assert (removed);
 683                         if (tpdomain->domain->cleanup_semaphore)
 684                                 ReleaseSemaphore (tpdomain->domain->cleanup_semaphore, 1, NULL);
 685                         domain_free (tpdomain);
 686                         tpdomain = NULL;
 687                 }
 688
 689                 previous_tpdomain = tpdomain;
 690         }
 691
 692         mono_coop_mutex_unlock (&threadpool->domains_lock);
 693
 694         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 695         g_ptr_array_remove_fast (threadpool->working_threads, thread);
 696         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 697
 698         COUNTER_ATOMIC (counter, {
 699                 counter._.working--;
 700                 counter._.active --;
 701         });
 702
 703         mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker finishing", mono_native_thread_id_get ());
 704 }
 705
 706 static gboolean
 707 worker_try_create (void)
 708 {
 709         ThreadPoolCounter counter;
 710         MonoInternalThread *thread;
 711         gint32 now;
 712
 713         mono_coop_mutex_lock (&threadpool->worker_creation_lock);
 714
 715         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker", mono_native_thread_id_get ());
 716
 717         if ((now = mono_100ns_ticks () / 10 / 1000 / 1000) == 0) {
 718                 g_warning ("failed to get 100ns ticks");
 719         } else {
 720                 if (threadpool->worker_creation_current_second != now) {
 721                         threadpool->worker_creation_current_second = now;
 722                         threadpool->worker_creation_current_count = 0;
 723                 } else {
 724                         g_assert (threadpool->worker_creation_current_count <= WORKER_CREATION_MAX_PER_SEC);
 725                         if (threadpool->worker_creation_current_count == WORKER_CREATION_MAX_PER_SEC) {
 726                                 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",
 727                                         mono_native_thread_id_get (), threadpool->worker_creation_current_count);
 728                                 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
 729                                 return FALSE;
 730                         }
 731                 }
 732         }
 733
 734         COUNTER_ATOMIC (counter, {
 735                 if (counter._.working >= counter._.max_working) {
 736                         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: maximum number of working threads reached",
 737                                 mono_native_thread_id_get ());
 738                         mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
 739                         return FALSE;
 740                 }
 741                 counter._.working ++;
 742                 counter._.active ++;
 743         });
 744
 745         if ((thread = mono_thread_create_internal (mono_get_root_domain (), worker_thread, NULL, TRUE, 0)) != NULL) {
 746                 threadpool->worker_creation_current_count += 1;
 747
 748                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, created %p, now = %d count = %d", mono_native_thread_id_get (), thread->tid, now, threadpool->worker_creation_current_count);
 749                 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
 750                 return TRUE;
 751         }
 752
 753         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: could not create thread", mono_native_thread_id_get ());
 754
 755         COUNTER_ATOMIC (counter, {
 756                 counter._.working --;
 757                 counter._.active --;
 758         });
 759
 760         mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
 761         return FALSE;
 762 }
 763
 764 static void monitor_ensure_running (void);
 765
 766 static gboolean
 767 worker_request (MonoDomain *domain)
 768 {
 769         ThreadPoolDomain *tpdomain;
 770
 771         g_assert (domain);
 772         g_assert (threadpool);
 773
 774         if (mono_runtime_is_shutting_down ())
 775                 return FALSE;
 776
 777         mono_coop_mutex_lock (&threadpool->domains_lock);
 778
 779         /* synchronize check with worker_thread */
 780         if (mono_domain_is_unloading (domain)) {
 781                 mono_coop_mutex_unlock (&threadpool->domains_lock);
 782                 return FALSE;
 783         }
 784
 785         tpdomain = domain_get (domain, TRUE);
 786         g_assert (tpdomain);
 787         tpdomain->outstanding_request ++;
 788
 789         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, domain = %p, outstanding_request = %d",
 790                 mono_native_thread_id_get (), tpdomain->domain, tpdomain->outstanding_request);
 791
 792         mono_coop_mutex_unlock (&threadpool->domains_lock);
 793
 794         if (threadpool->suspended)
 795                 return FALSE;
 796
 797         monitor_ensure_running ();
 798
 799         if (worker_try_unpark ()) {
 800                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, unparked", mono_native_thread_id_get ());
 801                 return TRUE;
 802         }
 803
 804         if (worker_try_create ()) {
 805                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, created", mono_native_thread_id_get ());
 806                 return TRUE;
 807         }
 808
 809         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, failed", mono_native_thread_id_get ());
 810         return FALSE;
 811 }
 812
 813 static gboolean
 814 monitor_should_keep_running (void)
 815 {
 816         static gint64 last_should_keep_running = -1;
 817
 818         g_assert (monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || monitor_status == MONITOR_STATUS_REQUESTED);
 819
 820         if (InterlockedExchange (&monitor_status, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST) {
 821                 gboolean should_keep_running = TRUE, force_should_keep_running = FALSE;
 822
 823                 if (mono_runtime_is_shutting_down ()) {
 824                         should_keep_running = FALSE;
 825                 } else {
 826                         mono_coop_mutex_lock (&threadpool->domains_lock);
 827                         if (!domain_any_has_request ())
 828                                 should_keep_running = FALSE;
 829                         mono_coop_mutex_unlock (&threadpool->domains_lock);
 830
 831                         if (!should_keep_running) {
 832                                 if (last_should_keep_running == -1 || mono_100ns_ticks () - last_should_keep_running < MONITOR_MINIMAL_LIFETIME * 1000 * 10) {
 833                                         should_keep_running = force_should_keep_running = TRUE;
 834                                 }
 835                         }
 836                 }
 837
 838                 if (should_keep_running) {
 839                         if (last_should_keep_running == -1 || !force_should_keep_running)
 840                                 last_should_keep_running = mono_100ns_ticks ();
 841                 } else {
 842                         last_should_keep_running = -1;
 843                         if (InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_NOT_RUNNING, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST)
 844                                 return FALSE;
 845                 }
 846         }
 847
 848         g_assert (monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || monitor_status == MONITOR_STATUS_REQUESTED);
 849
 850         return TRUE;
 851 }
 852
 853 static gboolean
 854 monitor_sufficient_delay_since_last_dequeue (void)
 855 {
 856         guint32 threshold;
 857
 858         g_assert (threadpool);
 859
 860         if (threadpool->cpu_usage < CPU_USAGE_LOW) {
 861                 threshold = MONITOR_INTERVAL;
 862         } else {
 863                 ThreadPoolCounter counter;
 864                 counter.as_gint64 = COUNTER_READ();
 865                 threshold = counter._.max_working * MONITOR_INTERVAL * 2;
 866         }
 867
 868         return mono_msec_ticks () >= threadpool->heuristic_last_dequeue + threshold;
 869 }
 870
 871 static void hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition);
 872
 873 static void
 874 monitor_thread (void)
 875 {
 876         MonoInternalThread *current_thread = mono_thread_internal_current ();
 877         guint i;
 878
 879         mono_cpu_usage (threadpool->cpu_usage_state);
 880
 881         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, started", mono_native_thread_id_get ());
 882
 883         do {
 884                 ThreadPoolCounter counter;
 885                 gboolean limit_worker_max_reached;
 886                 gint32 interval_left = MONITOR_INTERVAL;
 887                 gint32 awake = 0; /* number of spurious awakes we tolerate before doing a round of rebalancing */
 888
 889                 g_assert (monitor_status != MONITOR_STATUS_NOT_RUNNING);
 890
 891                 mono_gc_set_skip_thread (TRUE);
 892
 893                 do {
 894                         guint32 ts;
 895                         gboolean alerted = FALSE;
 896
 897                         if (mono_runtime_is_shutting_down ())
 898                                 break;
 899
 900                         ts = mono_msec_ticks ();
 901                         if (mono_thread_info_sleep (interval_left, &alerted) == 0)
 902                                 break;
 903                         interval_left -= mono_msec_ticks () - ts;
 904
 905                         mono_gc_set_skip_thread (FALSE);
 906                         if ((current_thread->state & (ThreadState_StopRequested | ThreadState_SuspendRequested)) != 0)
 907                                 mono_thread_interruption_checkpoint ();
 908                         mono_gc_set_skip_thread (TRUE);
 909                 } while (interval_left > 0 && ++awake < 10);
 910
 911                 mono_gc_set_skip_thread (FALSE);
 912
 913                 if (threadpool->suspended)
 914                         continue;
 915
 916                 if (mono_runtime_is_shutting_down ())
 917                         continue;
 918
 919                 mono_coop_mutex_lock (&threadpool->domains_lock);
 920                 if (!domain_any_has_request ()) {
 921                         mono_coop_mutex_unlock (&threadpool->domains_lock);
 922                         continue;
 923                 }
 924                 mono_coop_mutex_unlock (&threadpool->domains_lock);
 925
 926                 threadpool->cpu_usage = mono_cpu_usage (threadpool->cpu_usage_state);
 927
 928                 if (!monitor_sufficient_delay_since_last_dequeue ())
 929                         continue;
 930
 931                 limit_worker_max_reached = FALSE;
 932
 933                 COUNTER_ATOMIC (counter, {
 934                         if (counter._.max_working >= threadpool->limit_worker_max) {
 935                                 limit_worker_max_reached = TRUE;
 936                                 break;
 937                         }
 938                         counter._.max_working ++;
 939                 });
 940
 941                 if (limit_worker_max_reached)
 942                         continue;
 943
 944                 hill_climbing_force_change (counter._.max_working, TRANSITION_STARVATION);
 945
 946                 for (i = 0; i < 5; ++i) {
 947                         if (mono_runtime_is_shutting_down ())
 948                                 break;
 949
 950                         if (worker_try_unpark ()) {
 951                                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, unparked", mono_native_thread_id_get ());
 952                                 break;
 953                         }
 954
 955                         if (worker_try_create ()) {
 956                                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, created", mono_native_thread_id_get ());
 957                                 break;
 958                         }
 959                 }
 960         } while (monitor_should_keep_running ());
 961
 962         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, finished", mono_native_thread_id_get ());
 963 }
 964
 965 static void
 966 monitor_ensure_running (void)
 967 {
 968         for (;;) {
 969                 switch (monitor_status) {
 970                 case MONITOR_STATUS_REQUESTED:
 971                         return;
 972                 case MONITOR_STATUS_WAITING_FOR_REQUEST:
 973                         InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_WAITING_FOR_REQUEST);
 974                         break;
 975                 case MONITOR_STATUS_NOT_RUNNING:
 976                         if (mono_runtime_is_shutting_down ())
 977                                 return;
 978                         if (InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_NOT_RUNNING) == MONITOR_STATUS_NOT_RUNNING) {
 979                                 if (!mono_thread_create_internal (mono_get_root_domain (), monitor_thread, NULL, TRUE, SMALL_STACK))
 980                                         monitor_status = MONITOR_STATUS_NOT_RUNNING;
 981                                 return;
 982                         }
 983                         break;
 984                 default: g_assert_not_reached ();
 985                 }
 986         }
 987 }
 988
 989 static void
 990 hill_climbing_change_thread_count (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
 991 {
 992         ThreadPoolHillClimbing *hc;
 993
 994         g_assert (threadpool);
 995
 996         hc = &threadpool->heuristic_hill_climbing;
 997
 998         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);
 999
1000         hc->last_thread_count = new_thread_count;
1001         hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
1002         hc->elapsed_since_last_change = 0;
1003         hc->completions_since_last_change = 0;
1004 }
1005
1006 static void
1007 hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
1008 {
1009         ThreadPoolHillClimbing *hc;
1010
1011         g_assert (threadpool);
1012
1013         hc = &threadpool->heuristic_hill_climbing;
1014
1015         if (new_thread_count != hc->last_thread_count) {
1016                 hc->current_control_setting += new_thread_count - hc->last_thread_count;
1017                 hill_climbing_change_thread_count (new_thread_count, transition);
1018         }
1019 }
1020
1021 static double_complex
1022 hill_climbing_get_wave_component (gdouble *samples, guint sample_count, gdouble period)
1023 {
1024         ThreadPoolHillClimbing *hc;
1025         gdouble w, cosine, sine, coeff, q0, q1, q2;
1026         guint i;
1027
1028         g_assert (threadpool);
1029         g_assert (sample_count >= period);
1030         g_assert (period >= 2);
1031
1032         hc = &threadpool->heuristic_hill_climbing;
1033
1034         w = 2.0 * M_PI / period;
1035         cosine = cos (w);
1036         sine = sin (w);
1037         coeff = 2.0 * cosine;
1038         q0 = q1 = q2 = 0;
1039
1040         for (i = 0; i < sample_count; ++i) {
1041                 q0 = coeff * q1 - q2 + samples [(hc->total_samples - sample_count + i) % hc->samples_to_measure];
1042                 q2 = q1;
1043                 q1 = q0;
1044         }
1045
1046         return mono_double_complex_scalar_div (mono_double_complex_make (q1 - q2 * cosine, (q2 * sine)), ((gdouble)sample_count));
1047 }
1048
1049 static gint16
1050 hill_climbing_update (gint16 current_thread_count, guint32 sample_duration, gint32 completions, guint32 *adjustment_interval)
1051 {
1052         ThreadPoolHillClimbing *hc;
1053         ThreadPoolHeuristicStateTransition transition;
1054         gdouble throughput;
1055         gdouble throughput_error_estimate;
1056         gdouble confidence;
1057         gdouble move;
1058         gdouble gain;
1059         gint sample_index;
1060         gint sample_count;
1061         gint new_thread_wave_magnitude;
1062         gint new_thread_count;
1063         double_complex thread_wave_component;
1064         double_complex throughput_wave_component;
1065         double_complex ratio;
1066
1067         g_assert (threadpool);
1068         g_assert (adjustment_interval);
1069
1070         hc = &threadpool->heuristic_hill_climbing;
1071
1072         /* If someone changed the thread count without telling us, update our records accordingly. */
1073         if (current_thread_count != hc->last_thread_count)
1074                 hill_climbing_force_change (current_thread_count, TRANSITION_INITIALIZING);
1075
1076         /* Update the cumulative stats for this thread count */
1077         hc->elapsed_since_last_change += sample_duration;
1078         hc->completions_since_last_change += completions;
1079
1080         /* Add in any data we've already collected about this sample */
1081         sample_duration += hc->accumulated_sample_duration;
1082         completions += hc->accumulated_completion_count;
1083
1084         /* We need to make sure we're collecting reasonably accurate data. Since we're just counting the end
1085          * of each work item, we are goinng to be missing some data about what really happened during the
1086          * sample interval. The count produced by each thread includes an initial work item that may have
1087          * started well before the start of the interval, and each thread may have been running some new
1088          * work item for some time before the end of the interval, which did not yet get counted. So
1089          * our count is going to be off by +/- threadCount workitems.
1090          *
1091          * The exception is that the thread that reported to us last time definitely wasn't running any work
1092          * at that time, and the thread that's reporting now definitely isn't running a work item now. So
1093          * we really only need to consider threadCount-1 threads.
1094          *
1095          * Thus the percent error in our count is +/- (threadCount-1)/numCompletions.
1096          *
1097          * We cannot rely on the frequency-domain analysis we'll be doing later to filter out this error, because
1098          * of the way it accumulates over time. If this sample is off by, say, 33% in the negative direction,
1099          * then the next one likely will be too. The one after that will include the sum of the completions
1100          * we missed in the previous samples, and so will be 33% positive. So every three samples we'll have
1101          * two "low" samples and one "high" sample. This will appear as periodic variation right in the frequency
1102          * range we're targeting, which will not be filtered by the frequency-domain translation. */
1103         if (hc->total_samples > 0 && ((current_thread_count - 1.0) / completions) >= hc->max_sample_error) {
1104                 /* Not accurate enough yet. Let's accumulate the data so
1105                  * far, and tell the ThreadPool to collect a little more. */
1106                 hc->accumulated_sample_duration = sample_duration;
1107                 hc->accumulated_completion_count = completions;
1108                 *adjustment_interval = 10;
1109                 return current_thread_count;
1110         }
1111
1112         /* We've got enouugh data for our sample; reset our accumulators for next time. */
1113         hc->accumulated_sample_duration = 0;
1114         hc->accumulated_completion_count = 0;
1115
1116         /* Add the current thread count and throughput sample to our history. */
1117         throughput = ((gdouble) completions) / sample_duration;
1118
1119         sample_index = hc->total_samples % hc->samples_to_measure;
1120         hc->samples [sample_index] = throughput;
1121         hc->thread_counts [sample_index] = current_thread_count;
1122         hc->total_samples ++;
1123
1124         /* Set up defaults for our metrics. */
1125         thread_wave_component = mono_double_complex_make(0, 0);
1126         throughput_wave_component = mono_double_complex_make(0, 0);
1127         throughput_error_estimate = 0;
1128         ratio = mono_double_complex_make(0, 0);
1129         confidence = 0;
1130
1131         transition = TRANSITION_WARMUP;
1132
1133         /* How many samples will we use? It must be at least the three wave periods we're looking for, and it must also
1134          * be a whole multiple of the primary wave's period; otherwise the frequency we're looking for will fall between
1135          * two frequency bands in the Fourier analysis, and we won't be able to measure it accurately. */
1136         sample_count = ((gint) MIN (hc->total_samples - 1, hc->samples_to_measure) / hc->wave_period) * hc->wave_period;
1137
1138         if (sample_count > hc->wave_period) {
1139                 guint i;
1140                 gdouble average_throughput;
1141                 gdouble average_thread_count;
1142                 gdouble sample_sum = 0;
1143                 gdouble thread_sum = 0;
1144
1145                 /* Average the throughput and thread count samples, so we can scale the wave magnitudes later. */
1146                 for (i = 0; i < sample_count; ++i) {
1147                         guint j = (hc->total_samples - sample_count + i) % hc->samples_to_measure;
1148                         sample_sum += hc->samples [j];
1149                         thread_sum += hc->thread_counts [j];
1150                 }
1151
1152                 average_throughput = sample_sum / sample_count;
1153                 average_thread_count = thread_sum / sample_count;
1154
1155                 if (average_throughput > 0 && average_thread_count > 0) {
1156                         gdouble noise_for_confidence, adjacent_period_1, adjacent_period_2;
1157
1158                         /* Calculate the periods of the adjacent frequency bands we'll be using to
1159                          * measure noise levels. We want the two adjacent Fourier frequency bands. */
1160                         adjacent_period_1 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) + 1);
1161                         adjacent_period_2 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) - 1);
1162
1163                         /* Get the the three different frequency components of the throughput (scaled by average
1164                          * throughput). Our "error" estimate (the amount of noise that might be present in the
1165                          * frequency band we're really interested in) is the average of the adjacent bands. */
1166                         throughput_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, hc->wave_period), average_throughput);
1167                         throughput_error_estimate = cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, adjacent_period_1), average_throughput));
1168
1169                         if (adjacent_period_2 <= sample_count) {
1170                                 throughput_error_estimate = MAX (throughput_error_estimate, cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (
1171                                         hc->samples, sample_count, adjacent_period_2), average_throughput)));
1172                         }
1173
1174                         /* Do the same for the thread counts, so we have something to compare to. We don't
1175                          * measure thread count noise, because there is none; these are exact measurements. */
1176                         thread_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->thread_counts, sample_count, hc->wave_period), average_thread_count);
1177
1178                         /* Update our moving average of the throughput noise. We'll use this
1179                          * later as feedback to determine the new size of the thread wave. */
1180                         if (hc->average_throughput_noise == 0) {
1181                                 hc->average_throughput_noise = throughput_error_estimate;
1182                         } else {
1183                                 hc->average_throughput_noise = (hc->throughput_error_smoothing_factor * throughput_error_estimate)
1184                                         + ((1.0 + hc->throughput_error_smoothing_factor) * hc->average_throughput_noise);
1185                         }
1186
1187                         if (cabs (thread_wave_component) > 0) {
1188                                 /* Adjust the throughput wave so it's centered around the target wave,
1189                                  * and then calculate the adjusted throughput/thread ratio. */
1190                                 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);
1191                                 transition = TRANSITION_CLIMBING_MOVE;
1192                         } else {
1193                                 ratio = mono_double_complex_make (0, 0);
1194                                 transition = TRANSITION_STABILIZING;
1195                         }
1196
1197                         noise_for_confidence = MAX (hc->average_throughput_noise, throughput_error_estimate);
1198                         if (noise_for_confidence > 0) {
1199                                 confidence = cabs (thread_wave_component) / noise_for_confidence / hc->target_signal_to_noise_ratio;
1200                         } else {
1201                                 /* there is no noise! */
1202                                 confidence = 1.0;
1203                         }
1204                 }
1205         }
1206
1207         /* We use just the real part of the complex ratio we just calculated. If the throughput signal
1208          * is exactly in phase with the thread signal, this will be the same as taking the magnitude of
1209          * the complex move and moving that far up. If they're 180 degrees out of phase, we'll move
1210          * backward (because this indicates that our changes are having the opposite of the intended effect).
1211          * If they're 90 degrees out of phase, we won't move at all, because we can't tell wether we're
1212          * having a negative or positive effect on throughput. */
1213         move = creal (ratio);
1214         move = CLAMP (move, -1.0, 1.0);
1215
1216         /* Apply our confidence multiplier. */
1217         move *= CLAMP (confidence, -1.0, 1.0);
1218
1219         /* Now apply non-linear gain, such that values around zero are attenuated, while higher values
1220          * are enhanced. This allows us to move quickly if we're far away from the target, but more slowly
1221         * if we're getting close, giving us rapid ramp-up without wild oscillations around the target. */
1222         gain = hc->max_change_per_second * sample_duration;
1223         move = pow (fabs (move), hc->gain_exponent) * (move >= 0.0 ? 1 : -1) * gain;
1224         move = MIN (move, hc->max_change_per_sample);
1225
1226         /* If the result was positive, and CPU is > 95%, refuse the move. */
1227         if (move > 0.0 && threadpool->cpu_usage > CPU_USAGE_HIGH)
1228                 move = 0.0;
1229
1230         /* Apply the move to our control setting. */
1231         hc->current_control_setting += move;
1232
1233         /* Calculate the new thread wave magnitude, which is based on the moving average we've been keeping of the
1234          * throughput error.  This average starts at zero, so we'll start with a nice safe little wave at first. */
1235         new_thread_wave_magnitude = (gint)(0.5 + (hc->current_control_setting * hc->average_throughput_noise
1236                 * hc->target_signal_to_noise_ratio * hc->thread_magnitude_multiplier * 2.0));
1237         new_thread_wave_magnitude = CLAMP (new_thread_wave_magnitude, 1, hc->max_thread_wave_magnitude);
1238
1239         /* Make sure our control setting is within the ThreadPool's limits. */
1240         hc->current_control_setting = CLAMP (hc->current_control_setting, threadpool->limit_worker_min, threadpool->limit_worker_max - new_thread_wave_magnitude);
1241
1242         /* Calculate the new thread count (control setting + square wave). */
1243         new_thread_count = (gint)(hc->current_control_setting + new_thread_wave_magnitude * ((hc->total_samples / (hc->wave_period / 2)) % 2));
1244
1245         /* Make sure the new thread count doesn't exceed the ThreadPool's limits. */
1246         new_thread_count = CLAMP (new_thread_count, threadpool->limit_worker_min, threadpool->limit_worker_max);
1247
1248         if (new_thread_count != current_thread_count)
1249                 hill_climbing_change_thread_count (new_thread_count, transition);
1250
1251         if (creal (ratio) < 0.0 && new_thread_count == threadpool->limit_worker_min)
1252                 *adjustment_interval = (gint)(0.5 + hc->current_sample_interval * (10.0 * MAX (-1.0 * creal (ratio), 1.0)));
1253         else
1254                 *adjustment_interval = hc->current_sample_interval;
1255
1256         return new_thread_count;
1257 }
1258
1259 static void
1260 heuristic_notify_work_completed (void)
1261 {
1262         g_assert (threadpool);
1263
1264         InterlockedIncrement (&threadpool->heuristic_completions);
1265         threadpool->heuristic_last_dequeue = mono_msec_ticks ();
1266 }
1267
1268 static gboolean
1269 heuristic_should_adjust (void)
1270 {
1271         g_assert (threadpool);
1272
1273         if (threadpool->heuristic_last_dequeue > threadpool->heuristic_last_adjustment + threadpool->heuristic_adjustment_interval) {
1274                 ThreadPoolCounter counter;
1275                 counter.as_gint64 = COUNTER_READ();
1276                 if (counter._.working <= counter._.max_working)
1277                         return TRUE;
1278         }
1279
1280         return FALSE;
1281 }
1282
1283 static void
1284 heuristic_adjust (void)
1285 {
1286         g_assert (threadpool);
1287
1288         if (mono_coop_mutex_trylock (&threadpool->heuristic_lock) == 0) {
1289                 gint32 completions = InterlockedExchange (&threadpool->heuristic_completions, 0);
1290                 guint32 sample_end = mono_msec_ticks ();
1291                 guint32 sample_duration = sample_end - threadpool->heuristic_sample_start;
1292
1293                 if (sample_duration >= threadpool->heuristic_adjustment_interval / 2) {
1294                         ThreadPoolCounter counter;
1295                         gint16 new_thread_count;
1296
1297                         counter.as_gint64 = COUNTER_READ ();
1298                         new_thread_count = hill_climbing_update (counter._.max_working, sample_duration, completions, &threadpool->heuristic_adjustment_interval);
1299
1300                         COUNTER_ATOMIC (counter, { counter._.max_working = new_thread_count; });
1301
1302                         if (new_thread_count > counter._.max_working)
1303                                 worker_request (mono_domain_get ());
1304
1305                         threadpool->heuristic_sample_start = sample_end;
1306                         threadpool->heuristic_last_adjustment = mono_msec_ticks ();
1307                 }
1308
1309                 mono_coop_mutex_unlock (&threadpool->heuristic_lock);
1310         }
1311 }
1312
1313 void
1314 mono_threadpool_ms_cleanup (void)
1315 {
1316         #ifndef DISABLE_SOCKETS
1317                 mono_threadpool_ms_io_cleanup ();
1318         #endif
1319         mono_lazy_cleanup (&status, cleanup);
1320 }
1321
1322 MonoAsyncResult *
1323 mono_threadpool_ms_begin_invoke (MonoDomain *domain, MonoObject *target, MonoMethod *method, gpointer *params, MonoError *error)
1324 {
1325         static MonoClass *async_call_klass = NULL;
1326         MonoMethodMessage *message;
1327         MonoAsyncResult *async_result;
1328         MonoAsyncCall *async_call;
1329         MonoDelegate *async_callback = NULL;
1330         MonoObject *state = NULL;
1331
1332         if (!async_call_klass)
1333                 async_call_klass = mono_class_load_from_name (mono_defaults.corlib, "System", "MonoAsyncCall");
1334
1335         mono_lazy_initialize (&status, initialize);
1336
1337         mono_error_init (error);
1338
1339         message = mono_method_call_message_new (method, params, mono_get_delegate_invoke (method->klass), (params != NULL) ? (&async_callback) : NULL, (params != NULL) ? (&state) : NULL);
1340
1341         async_call = (MonoAsyncCall*) mono_object_new_checked (domain, async_call_klass, error);
1342         return_val_if_nok (error, NULL);
1343
1344         MONO_OBJECT_SETREF (async_call, msg, message);
1345         MONO_OBJECT_SETREF (async_call, state, state);
1346
1347         if (async_callback) {
1348                 MONO_OBJECT_SETREF (async_call, cb_method, mono_get_delegate_invoke (((MonoObject*) async_callback)->vtable->klass));
1349                 MONO_OBJECT_SETREF (async_call, cb_target, async_callback);
1350         }
1351
1352         async_result = mono_async_result_new (domain, NULL, async_call->state, NULL, (MonoObject*) async_call);
1353         MONO_OBJECT_SETREF (async_result, async_delegate, target);
1354
1355         mono_threadpool_ms_enqueue_work_item (domain, (MonoObject*) async_result, error);
1356         return_val_if_nok (error, NULL);
1357
1358         return async_result;
1359 }
1360
1361 MonoObject *
1362 mono_threadpool_ms_end_invoke (MonoAsyncResult *ares, MonoArray **out_args, MonoObject **exc)
1363 {
1364         MonoError error;
1365         MonoAsyncCall *ac;
1366
1367         g_assert (exc);
1368         g_assert (out_args);
1369
1370         *exc = NULL;
1371         *out_args = NULL;
1372
1373         /* check if already finished */
1374         mono_monitor_enter ((MonoObject*) ares);
1375
1376         if (ares->endinvoke_called) {
1377                 *exc = (MonoObject*) mono_get_exception_invalid_operation (NULL);
1378                 mono_monitor_exit ((MonoObject*) ares);
1379                 return NULL;
1380         }
1381
1382         ares->endinvoke_called = 1;
1383
1384         /* wait until we are really finished */
1385         if (ares->completed) {
1386                 mono_monitor_exit ((MonoObject *) ares);
1387         } else {
1388                 gpointer wait_event;
1389                 if (ares->handle) {
1390                         wait_event = mono_wait_handle_get_handle ((MonoWaitHandle*) ares->handle);
1391                 } else {
1392                         wait_event = CreateEvent (NULL, TRUE, FALSE, NULL);
1393                         g_assert(wait_event);
1394                         MonoWaitHandle *wait_handle = mono_wait_handle_new (mono_object_domain (ares), wait_event, &error);
1395                         mono_error_raise_exception (&error); /* FIXME don't raise here */
1396                         MONO_OBJECT_SETREF (ares, handle, (MonoObject*) wait_handle);
1397                 }
1398                 mono_monitor_exit ((MonoObject*) ares);
1399                 MONO_PREPARE_BLOCKING;
1400                 WaitForSingleObjectEx (wait_event, INFINITE, TRUE);
1401                 MONO_FINISH_BLOCKING;
1402         }
1403
1404         ac = (MonoAsyncCall*) ares->object_data;
1405         g_assert (ac);
1406
1407         *exc = ac->msg->exc; /* FIXME: GC add write barrier */
1408         *out_args = ac->out_args;
1409         return ac->res;
1410 }
1411
1412 gboolean
1413 mono_threadpool_ms_remove_domain_jobs (MonoDomain *domain, int timeout)
1414 {
1415         gboolean res = TRUE;
1416         guint32 start;
1417         gpointer sem;
1418
1419         g_assert (domain);
1420         g_assert (timeout >= -1);
1421
1422         g_assert (mono_domain_is_unloading (domain));
1423
1424         if (timeout != -1)
1425                 start = mono_msec_ticks ();
1426
1427 #ifndef DISABLE_SOCKETS
1428         mono_threadpool_ms_io_remove_domain_jobs (domain);
1429         if (timeout != -1) {
1430                 timeout -= mono_msec_ticks () - start;
1431                 if (timeout < 0)
1432                         return FALSE;
1433         }
1434 #endif
1435
1436         /*
1437          * There might be some threads out that could be about to execute stuff from the given domain.
1438          * We avoid that by setting up a semaphore to be pulsed by the thread that reaches zero.
1439          */
1440         sem = domain->cleanup_semaphore = CreateSemaphore (NULL, 0, 1, NULL);
1441
1442         /*
1443          * The memory barrier here is required to have global ordering between assigning to cleanup_semaphone
1444          * and reading threadpool_jobs. Otherwise this thread could read a stale version of threadpool_jobs
1445          * and wait forever.
1446          */
1447         mono_memory_write_barrier ();
1448
1449         while (domain->threadpool_jobs) {
1450                 MONO_PREPARE_BLOCKING;
1451                 WaitForSingleObject (sem, timeout);
1452                 MONO_FINISH_BLOCKING;
1453                 if (timeout != -1) {
1454                         timeout -= mono_msec_ticks () - start;
1455                         if (timeout <= 0) {
1456                                 res = FALSE;
1457                                 break;
1458                         }
1459                 }
1460         }
1461
1462         domain->cleanup_semaphore = NULL;
1463         CloseHandle (sem);
1464
1465         return res;
1466 }
1467
1468 void
1469 mono_threadpool_ms_suspend (void)
1470 {
1471         if (threadpool)
1472                 threadpool->suspended = TRUE;
1473 }
1474
1475 void
1476 mono_threadpool_ms_resume (void)
1477 {
1478         if (threadpool)
1479                 threadpool->suspended = FALSE;
1480 }
1481
1482 void
1483 ves_icall_System_Threading_ThreadPool_GetAvailableThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1484 {
1485         ThreadPoolCounter counter;
1486
1487         if (!worker_threads || !completion_port_threads)
1488                 return;
1489
1490         mono_lazy_initialize (&status, initialize);
1491
1492         counter.as_gint64 = COUNTER_READ ();
1493
1494         *worker_threads = MAX (0, threadpool->limit_worker_max - counter._.active);
1495         *completion_port_threads = threadpool->limit_io_max;
1496 }
1497
1498 void
1499 ves_icall_System_Threading_ThreadPool_GetMinThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1500 {
1501         if (!worker_threads || !completion_port_threads)
1502                 return;
1503
1504         mono_lazy_initialize (&status, initialize);
1505
1506         *worker_threads = threadpool->limit_worker_min;
1507         *completion_port_threads = threadpool->limit_io_min;
1508 }
1509
1510 void
1511 ves_icall_System_Threading_ThreadPool_GetMaxThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1512 {
1513         if (!worker_threads || !completion_port_threads)
1514                 return;
1515
1516         mono_lazy_initialize (&status, initialize);
1517
1518         *worker_threads = threadpool->limit_worker_max;
1519         *completion_port_threads = threadpool->limit_io_max;
1520 }
1521
1522 MonoBoolean
1523 ves_icall_System_Threading_ThreadPool_SetMinThreadsNative (gint32 worker_threads, gint32 completion_port_threads)
1524 {
1525         mono_lazy_initialize (&status, initialize);
1526
1527         if (worker_threads <= 0 || worker_threads > threadpool->limit_worker_max)
1528                 return FALSE;
1529         if (completion_port_threads <= 0 || completion_port_threads > threadpool->limit_io_max)
1530                 return FALSE;
1531
1532         threadpool->limit_worker_min = worker_threads;
1533         threadpool->limit_io_min = completion_port_threads;
1534
1535         return TRUE;
1536 }
1537
1538 MonoBoolean
1539 ves_icall_System_Threading_ThreadPool_SetMaxThreadsNative (gint32 worker_threads, gint32 completion_port_threads)
1540 {
1541         gint cpu_count = mono_cpu_count ();
1542
1543         mono_lazy_initialize (&status, initialize);
1544
1545         if (worker_threads < threadpool->limit_worker_min || worker_threads < cpu_count)
1546                 return FALSE;
1547         if (completion_port_threads < threadpool->limit_io_min || completion_port_threads < cpu_count)
1548                 return FALSE;
1549
1550         threadpool->limit_worker_max = worker_threads;
1551         threadpool->limit_io_max = completion_port_threads;
1552
1553         return TRUE;
1554 }
1555
1556 void
1557 ves_icall_System_Threading_ThreadPool_InitializeVMTp (MonoBoolean *enable_worker_tracking)
1558 {
1559         if (enable_worker_tracking) {
1560                 // TODO implement some kind of switch to have the possibily to use it
1561                 *enable_worker_tracking = FALSE;
1562         }
1563
1564         mono_lazy_initialize (&status, initialize);
1565 }
1566
1567 MonoBoolean
1568 ves_icall_System_Threading_ThreadPool_NotifyWorkItemComplete (void)
1569 {
1570         ThreadPoolCounter counter;
1571
1572         if (mono_domain_is_unloading (mono_domain_get ()) || mono_runtime_is_shutting_down ())
1573                 return FALSE;
1574
1575         heuristic_notify_work_completed ();
1576
1577         if (heuristic_should_adjust ())
1578                 heuristic_adjust ();
1579
1580         counter.as_gint64 = COUNTER_READ ();
1581         return counter._.working <= counter._.max_working;
1582 }
1583
1584 void
1585 ves_icall_System_Threading_ThreadPool_NotifyWorkItemProgressNative (void)
1586 {
1587         heuristic_notify_work_completed ();
1588
1589         if (heuristic_should_adjust ())
1590                 heuristic_adjust ();
1591 }
1592
1593 void
1594 ves_icall_System_Threading_ThreadPool_ReportThreadStatus (MonoBoolean is_working)
1595 {
1596         // TODO
1597         MonoError error;
1598         mono_error_set_not_implemented (&error, "");
1599         mono_error_set_pending_exception (&error);
1600 }
1601
1602 MonoBoolean
1603 ves_icall_System_Threading_ThreadPool_RequestWorkerThread (void)
1604 {
1605         return worker_request (mono_domain_get ());
1606 }
1607
1608 MonoBoolean G_GNUC_UNUSED
1609 ves_icall_System_Threading_ThreadPool_PostQueuedCompletionStatus (MonoNativeOverlapped *native_overlapped)
1610 {
1611         /* This copy the behavior of the current Mono implementation */
1612         MonoError error;
1613         mono_error_set_not_implemented (&error, "");
1614         mono_error_set_pending_exception (&error);
1615         return FALSE;
1616 }
1617
1618 MonoBoolean G_GNUC_UNUSED
1619 ves_icall_System_Threading_ThreadPool_BindIOCompletionCallbackNative (gpointer file_handle)
1620 {
1621         /* This copy the behavior of the current Mono implementation */
1622         return TRUE;
1623 }
1624
1625 MonoBoolean G_GNUC_UNUSED
1626 ves_icall_System_Threading_ThreadPool_IsThreadPoolHosted (void)
1627 {
1628         return FALSE;
1629 }