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 ((void **)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_domain_get (), "Threadpool worker"), FALSE);
 596
 597         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 598         g_ptr_array_add (threadpool->working_threads, thread);
 599         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 600
 601         previous_tpdomain = NULL;
 602
 603         mono_coop_mutex_lock (&threadpool->domains_lock);
 604
 605         while (!mono_runtime_is_shutting_down ()) {
 606                 tpdomain = NULL;
 607
 608                 if ((thread->state & (ThreadState_StopRequested | ThreadState_SuspendRequested)) != 0) {
 609                         mono_coop_mutex_unlock (&threadpool->domains_lock);
 610                         mono_thread_interruption_checkpoint ();
 611                         mono_coop_mutex_lock (&threadpool->domains_lock);
 612                 }
 613
 614                 if (retire || !(tpdomain = domain_get_next (previous_tpdomain))) {
 615                         gboolean timeout;
 616
 617                         COUNTER_ATOMIC (counter, {
 618                                 counter._.working --;
 619                                 counter._.parked ++;
 620                         });
 621
 622                         mono_coop_mutex_unlock (&threadpool->domains_lock);
 623                         timeout = worker_park ();
 624                         mono_coop_mutex_lock (&threadpool->domains_lock);
 625
 626                         COUNTER_ATOMIC (counter, {
 627                                 counter._.working ++;
 628                                 counter._.parked --;
 629                         });
 630
 631                         if (timeout)
 632                                 break;
 633
 634                         if (retire)
 635                                 retire = FALSE;
 636
 637                         continue;
 638                 }
 639
 640                 tpdomain->outstanding_request --;
 641                 g_assert (tpdomain->outstanding_request >= 0);
 642
 643                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker running in domain %p",
 644                         mono_native_thread_id_get (), tpdomain->domain, tpdomain->outstanding_request);
 645
 646                 g_assert (tpdomain->domain);
 647                 g_assert (tpdomain->domain->threadpool_jobs >= 0);
 648                 tpdomain->domain->threadpool_jobs ++;
 649
 650                 mono_coop_mutex_unlock (&threadpool->domains_lock);
 651
 652                 mono_thread_push_appdomain_ref (tpdomain->domain);
 653                 if (mono_domain_set (tpdomain->domain, FALSE)) {
 654                         MonoObject *exc = NULL, *res;
 655
 656                         res = mono_runtime_try_invoke (mono_defaults.threadpool_perform_wait_callback_method, NULL, NULL, &exc, &error);
 657                         if (exc || !mono_error_ok(&error)) {
 658                                 if (exc == NULL)
 659                                         exc = (MonoObject *) mono_error_convert_to_exception (&error);
 660                                 else
 661                                         mono_error_cleanup (&error);
 662                                 mono_thread_internal_unhandled_exception (exc);
 663                         } else if (res && *(MonoBoolean*) mono_object_unbox (res) == FALSE)
 664                                 retire = TRUE;
 665
 666                         mono_thread_clr_state (thread, (MonoThreadState)~ThreadState_Background);
 667                         if (!mono_thread_test_state (thread , ThreadState_Background))
 668                                 ves_icall_System_Threading_Thread_SetState (thread, ThreadState_Background);
 669
 670                         mono_domain_set (mono_get_root_domain (), TRUE);
 671                 }
 672                 mono_thread_pop_appdomain_ref ();
 673
 674                 mono_coop_mutex_lock (&threadpool->domains_lock);
 675
 676                 tpdomain->domain->threadpool_jobs --;
 677                 g_assert (tpdomain->domain->threadpool_jobs >= 0);
 678
 679                 if (tpdomain->domain->threadpool_jobs == 0 && mono_domain_is_unloading (tpdomain->domain)) {
 680                         gboolean removed = domain_remove (tpdomain);
 681                         g_assert (removed);
 682                         if (tpdomain->domain->cleanup_semaphore)
 683                                 ReleaseSemaphore (tpdomain->domain->cleanup_semaphore, 1, NULL);
 684                         domain_free (tpdomain);
 685                         tpdomain = NULL;
 686                 }
 687
 688                 previous_tpdomain = tpdomain;
 689         }
 690
 691         mono_coop_mutex_unlock (&threadpool->domains_lock);
 692
 693         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 694         g_ptr_array_remove_fast (threadpool->working_threads, thread);
 695         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 696
 697         COUNTER_ATOMIC (counter, {
 698                 counter._.working--;
 699                 counter._.active --;
 700         });
 701
 702         mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker finishing", mono_native_thread_id_get ());
 703 }
 704
 705 static gboolean
 706 worker_try_create (void)
 707 {
 708         ThreadPoolCounter counter;
 709         MonoInternalThread *thread;
 710         gint32 now;
 711
 712         mono_coop_mutex_lock (&threadpool->worker_creation_lock);
 713
 714         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker", mono_native_thread_id_get ());
 715
 716         if ((now = mono_100ns_ticks () / 10 / 1000 / 1000) == 0) {
 717                 g_warning ("failed to get 100ns ticks");
 718         } else {
 719                 if (threadpool->worker_creation_current_second != now) {
 720                         threadpool->worker_creation_current_second = now;
 721                         threadpool->worker_creation_current_count = 0;
 722                 } else {
 723                         g_assert (threadpool->worker_creation_current_count <= WORKER_CREATION_MAX_PER_SEC);
 724                         if (threadpool->worker_creation_current_count == WORKER_CREATION_MAX_PER_SEC) {
 725                                 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",
 726                                         mono_native_thread_id_get (), threadpool->worker_creation_current_count);
 727                                 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
 728                                 return FALSE;
 729                         }
 730                 }
 731         }
 732
 733         COUNTER_ATOMIC (counter, {
 734                 if (counter._.working >= counter._.max_working) {
 735                         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: maximum number of working threads reached",
 736                                 mono_native_thread_id_get ());
 737                         mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
 738                         return FALSE;
 739                 }
 740                 counter._.working ++;
 741                 counter._.active ++;
 742         });
 743
 744         if ((thread = mono_thread_create_internal (mono_get_root_domain (), worker_thread, NULL, TRUE, 0)) != NULL) {
 745                 threadpool->worker_creation_current_count += 1;
 746
 747                 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);
 748                 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
 749                 return TRUE;
 750         }
 751
 752         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: could not create thread", mono_native_thread_id_get ());
 753
 754         COUNTER_ATOMIC (counter, {
 755                 counter._.working --;
 756                 counter._.active --;
 757         });
 758
 759         mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
 760         return FALSE;
 761 }
 762
 763 static void monitor_ensure_running (void);
 764
 765 static gboolean
 766 worker_request (MonoDomain *domain)
 767 {
 768         ThreadPoolDomain *tpdomain;
 769
 770         g_assert (domain);
 771         g_assert (threadpool);
 772
 773         if (mono_runtime_is_shutting_down ())
 774                 return FALSE;
 775
 776         mono_coop_mutex_lock (&threadpool->domains_lock);
 777
 778         /* synchronize check with worker_thread */
 779         if (mono_domain_is_unloading (domain)) {
 780                 mono_coop_mutex_unlock (&threadpool->domains_lock);
 781                 return FALSE;
 782         }
 783
 784         tpdomain = domain_get (domain, TRUE);
 785         g_assert (tpdomain);
 786         tpdomain->outstanding_request ++;
 787
 788         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, domain = %p, outstanding_request = %d",
 789                 mono_native_thread_id_get (), tpdomain->domain, tpdomain->outstanding_request);
 790
 791         mono_coop_mutex_unlock (&threadpool->domains_lock);
 792
 793         if (threadpool->suspended)
 794                 return FALSE;
 795
 796         monitor_ensure_running ();
 797
 798         if (worker_try_unpark ()) {
 799                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, unparked", mono_native_thread_id_get ());
 800                 return TRUE;
 801         }
 802
 803         if (worker_try_create ()) {
 804                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, created", mono_native_thread_id_get ());
 805                 return TRUE;
 806         }
 807
 808         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, failed", mono_native_thread_id_get ());
 809         return FALSE;
 810 }
 811
 812 static gboolean
 813 monitor_should_keep_running (void)
 814 {
 815         static gint64 last_should_keep_running = -1;
 816
 817         g_assert (monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || monitor_status == MONITOR_STATUS_REQUESTED);
 818
 819         if (InterlockedExchange (&monitor_status, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST) {
 820                 gboolean should_keep_running = TRUE, force_should_keep_running = FALSE;
 821
 822                 if (mono_runtime_is_shutting_down ()) {
 823                         should_keep_running = FALSE;
 824                 } else {
 825                         mono_coop_mutex_lock (&threadpool->domains_lock);
 826                         if (!domain_any_has_request ())
 827                                 should_keep_running = FALSE;
 828                         mono_coop_mutex_unlock (&threadpool->domains_lock);
 829
 830                         if (!should_keep_running) {
 831                                 if (last_should_keep_running == -1 || mono_100ns_ticks () - last_should_keep_running < MONITOR_MINIMAL_LIFETIME * 1000 * 10) {
 832                                         should_keep_running = force_should_keep_running = TRUE;
 833                                 }
 834                         }
 835                 }
 836
 837                 if (should_keep_running) {
 838                         if (last_should_keep_running == -1 || !force_should_keep_running)
 839                                 last_should_keep_running = mono_100ns_ticks ();
 840                 } else {
 841                         last_should_keep_running = -1;
 842                         if (InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_NOT_RUNNING, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST)
 843                                 return FALSE;
 844                 }
 845         }
 846
 847         g_assert (monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || monitor_status == MONITOR_STATUS_REQUESTED);
 848
 849         return TRUE;
 850 }
 851
 852 static gboolean
 853 monitor_sufficient_delay_since_last_dequeue (void)
 854 {
 855         guint32 threshold;
 856
 857         g_assert (threadpool);
 858
 859         if (threadpool->cpu_usage < CPU_USAGE_LOW) {
 860                 threshold = MONITOR_INTERVAL;
 861         } else {
 862                 ThreadPoolCounter counter;
 863                 counter.as_gint64 = COUNTER_READ();
 864                 threshold = counter._.max_working * MONITOR_INTERVAL * 2;
 865         }
 866
 867         return mono_msec_ticks () >= threadpool->heuristic_last_dequeue + threshold;
 868 }
 869
 870 static void hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition);
 871
 872 static void
 873 monitor_thread (void)
 874 {
 875         MonoInternalThread *current_thread = mono_thread_internal_current ();
 876         guint i;
 877
 878         mono_cpu_usage (threadpool->cpu_usage_state);
 879
 880         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, started", mono_native_thread_id_get ());
 881
 882         do {
 883                 ThreadPoolCounter counter;
 884                 gboolean limit_worker_max_reached;
 885                 gint32 interval_left = MONITOR_INTERVAL;
 886                 gint32 awake = 0; /* number of spurious awakes we tolerate before doing a round of rebalancing */
 887
 888                 g_assert (monitor_status != MONITOR_STATUS_NOT_RUNNING);
 889
 890                 mono_gc_set_skip_thread (TRUE);
 891
 892                 do {
 893                         guint32 ts;
 894                         gboolean alerted = FALSE;
 895
 896                         if (mono_runtime_is_shutting_down ())
 897                                 break;
 898
 899                         ts = mono_msec_ticks ();
 900                         if (mono_thread_info_sleep (interval_left, &alerted) == 0)
 901                                 break;
 902                         interval_left -= mono_msec_ticks () - ts;
 903
 904                         mono_gc_set_skip_thread (FALSE);
 905                         if ((current_thread->state & (ThreadState_StopRequested | ThreadState_SuspendRequested)) != 0)
 906                                 mono_thread_interruption_checkpoint ();
 907                         mono_gc_set_skip_thread (TRUE);
 908                 } while (interval_left > 0 && ++awake < 10);
 909
 910                 mono_gc_set_skip_thread (FALSE);
 911
 912                 if (threadpool->suspended)
 913                         continue;
 914
 915                 if (mono_runtime_is_shutting_down ())
 916                         continue;
 917
 918                 mono_coop_mutex_lock (&threadpool->domains_lock);
 919                 if (!domain_any_has_request ()) {
 920                         mono_coop_mutex_unlock (&threadpool->domains_lock);
 921                         continue;
 922                 }
 923                 mono_coop_mutex_unlock (&threadpool->domains_lock);
 924
 925                 threadpool->cpu_usage = mono_cpu_usage (threadpool->cpu_usage_state);
 926
 927                 if (!monitor_sufficient_delay_since_last_dequeue ())
 928                         continue;
 929
 930                 limit_worker_max_reached = FALSE;
 931
 932                 COUNTER_ATOMIC (counter, {
 933                         if (counter._.max_working >= threadpool->limit_worker_max) {
 934                                 limit_worker_max_reached = TRUE;
 935                                 break;
 936                         }
 937                         counter._.max_working ++;
 938                 });
 939
 940                 if (limit_worker_max_reached)
 941                         continue;
 942
 943                 hill_climbing_force_change (counter._.max_working, TRANSITION_STARVATION);
 944
 945                 for (i = 0; i < 5; ++i) {
 946                         if (mono_runtime_is_shutting_down ())
 947                                 break;
 948
 949                         if (worker_try_unpark ()) {
 950                                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, unparked", mono_native_thread_id_get ());
 951                                 break;
 952                         }
 953
 954                         if (worker_try_create ()) {
 955                                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, created", mono_native_thread_id_get ());
 956                                 break;
 957                         }
 958                 }
 959         } while (monitor_should_keep_running ());
 960
 961         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, finished", mono_native_thread_id_get ());
 962 }
 963
 964 static void
 965 monitor_ensure_running (void)
 966 {
 967         for (;;) {
 968                 switch (monitor_status) {
 969                 case MONITOR_STATUS_REQUESTED:
 970                         return;
 971                 case MONITOR_STATUS_WAITING_FOR_REQUEST:
 972                         InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_WAITING_FOR_REQUEST);
 973                         break;
 974                 case MONITOR_STATUS_NOT_RUNNING:
 975                         if (mono_runtime_is_shutting_down ())
 976                                 return;
 977                         if (InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_NOT_RUNNING) == MONITOR_STATUS_NOT_RUNNING) {
 978                                 if (!mono_thread_create_internal (mono_get_root_domain (), monitor_thread, NULL, TRUE, SMALL_STACK))
 979                                         monitor_status = MONITOR_STATUS_NOT_RUNNING;
 980                                 return;
 981                         }
 982                         break;
 983                 default: g_assert_not_reached ();
 984                 }
 985         }
 986 }
 987
 988 static void
 989 hill_climbing_change_thread_count (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
 990 {
 991         ThreadPoolHillClimbing *hc;
 992
 993         g_assert (threadpool);
 994
 995         hc = &threadpool->heuristic_hill_climbing;
 996
 997         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);
 998
 999         hc->last_thread_count = new_thread_count;
1000         hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
1001         hc->elapsed_since_last_change = 0;
1002         hc->completions_since_last_change = 0;
1003 }
1004
1005 static void
1006 hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
1007 {
1008         ThreadPoolHillClimbing *hc;
1009
1010         g_assert (threadpool);
1011
1012         hc = &threadpool->heuristic_hill_climbing;
1013
1014         if (new_thread_count != hc->last_thread_count) {
1015                 hc->current_control_setting += new_thread_count - hc->last_thread_count;
1016                 hill_climbing_change_thread_count (new_thread_count, transition);
1017         }
1018 }
1019
1020 static double_complex
1021 hill_climbing_get_wave_component (gdouble *samples, guint sample_count, gdouble period)
1022 {
1023         ThreadPoolHillClimbing *hc;
1024         gdouble w, cosine, sine, coeff, q0, q1, q2;
1025         guint i;
1026
1027         g_assert (threadpool);
1028         g_assert (sample_count >= period);
1029         g_assert (period >= 2);
1030
1031         hc = &threadpool->heuristic_hill_climbing;
1032
1033         w = 2.0 * M_PI / period;
1034         cosine = cos (w);
1035         sine = sin (w);
1036         coeff = 2.0 * cosine;
1037         q0 = q1 = q2 = 0;
1038
1039         for (i = 0; i < sample_count; ++i) {
1040                 q0 = coeff * q1 - q2 + samples [(hc->total_samples - sample_count + i) % hc->samples_to_measure];
1041                 q2 = q1;
1042                 q1 = q0;
1043         }
1044
1045         return mono_double_complex_scalar_div (mono_double_complex_make (q1 - q2 * cosine, (q2 * sine)), ((gdouble)sample_count));
1046 }
1047
1048 static gint16
1049 hill_climbing_update (gint16 current_thread_count, guint32 sample_duration, gint32 completions, guint32 *adjustment_interval)
1050 {
1051         ThreadPoolHillClimbing *hc;
1052         ThreadPoolHeuristicStateTransition transition;
1053         gdouble throughput;
1054         gdouble throughput_error_estimate;
1055         gdouble confidence;
1056         gdouble move;
1057         gdouble gain;
1058         gint sample_index;
1059         gint sample_count;
1060         gint new_thread_wave_magnitude;
1061         gint new_thread_count;
1062         double_complex thread_wave_component;
1063         double_complex throughput_wave_component;
1064         double_complex ratio;
1065
1066         g_assert (threadpool);
1067         g_assert (adjustment_interval);
1068
1069         hc = &threadpool->heuristic_hill_climbing;
1070
1071         /* If someone changed the thread count without telling us, update our records accordingly. */
1072         if (current_thread_count != hc->last_thread_count)
1073                 hill_climbing_force_change (current_thread_count, TRANSITION_INITIALIZING);
1074
1075         /* Update the cumulative stats for this thread count */
1076         hc->elapsed_since_last_change += sample_duration;
1077         hc->completions_since_last_change += completions;
1078
1079         /* Add in any data we've already collected about this sample */
1080         sample_duration += hc->accumulated_sample_duration;
1081         completions += hc->accumulated_completion_count;
1082
1083         /* We need to make sure we're collecting reasonably accurate data. Since we're just counting the end
1084          * of each work item, we are goinng to be missing some data about what really happened during the
1085          * sample interval. The count produced by each thread includes an initial work item that may have
1086          * started well before the start of the interval, and each thread may have been running some new
1087          * work item for some time before the end of the interval, which did not yet get counted. So
1088          * our count is going to be off by +/- threadCount workitems.
1089          *
1090          * The exception is that the thread that reported to us last time definitely wasn't running any work
1091          * at that time, and the thread that's reporting now definitely isn't running a work item now. So
1092          * we really only need to consider threadCount-1 threads.
1093          *
1094          * Thus the percent error in our count is +/- (threadCount-1)/numCompletions.
1095          *
1096          * We cannot rely on the frequency-domain analysis we'll be doing later to filter out this error, because
1097          * of the way it accumulates over time. If this sample is off by, say, 33% in the negative direction,
1098          * then the next one likely will be too. The one after that will include the sum of the completions
1099          * we missed in the previous samples, and so will be 33% positive. So every three samples we'll have
1100          * two "low" samples and one "high" sample. This will appear as periodic variation right in the frequency
1101          * range we're targeting, which will not be filtered by the frequency-domain translation. */
1102         if (hc->total_samples > 0 && ((current_thread_count - 1.0) / completions) >= hc->max_sample_error) {
1103                 /* Not accurate enough yet. Let's accumulate the data so
1104                  * far, and tell the ThreadPool to collect a little more. */
1105                 hc->accumulated_sample_duration = sample_duration;
1106                 hc->accumulated_completion_count = completions;
1107                 *adjustment_interval = 10;
1108                 return current_thread_count;
1109         }
1110
1111         /* We've got enouugh data for our sample; reset our accumulators for next time. */
1112         hc->accumulated_sample_duration = 0;
1113         hc->accumulated_completion_count = 0;
1114
1115         /* Add the current thread count and throughput sample to our history. */
1116         throughput = ((gdouble) completions) / sample_duration;
1117
1118         sample_index = hc->total_samples % hc->samples_to_measure;
1119         hc->samples [sample_index] = throughput;
1120         hc->thread_counts [sample_index] = current_thread_count;
1121         hc->total_samples ++;
1122
1123         /* Set up defaults for our metrics. */
1124         thread_wave_component = mono_double_complex_make(0, 0);
1125         throughput_wave_component = mono_double_complex_make(0, 0);
1126         throughput_error_estimate = 0;
1127         ratio = mono_double_complex_make(0, 0);
1128         confidence = 0;
1129
1130         transition = TRANSITION_WARMUP;
1131
1132         /* How many samples will we use? It must be at least the three wave periods we're looking for, and it must also
1133          * be a whole multiple of the primary wave's period; otherwise the frequency we're looking for will fall between
1134          * two frequency bands in the Fourier analysis, and we won't be able to measure it accurately. */
1135         sample_count = ((gint) MIN (hc->total_samples - 1, hc->samples_to_measure) / hc->wave_period) * hc->wave_period;
1136
1137         if (sample_count > hc->wave_period) {
1138                 guint i;
1139                 gdouble average_throughput;
1140                 gdouble average_thread_count;
1141                 gdouble sample_sum = 0;
1142                 gdouble thread_sum = 0;
1143
1144                 /* Average the throughput and thread count samples, so we can scale the wave magnitudes later. */
1145                 for (i = 0; i < sample_count; ++i) {
1146                         guint j = (hc->total_samples - sample_count + i) % hc->samples_to_measure;
1147                         sample_sum += hc->samples [j];
1148                         thread_sum += hc->thread_counts [j];
1149                 }
1150
1151                 average_throughput = sample_sum / sample_count;
1152                 average_thread_count = thread_sum / sample_count;
1153
1154                 if (average_throughput > 0 && average_thread_count > 0) {
1155                         gdouble noise_for_confidence, adjacent_period_1, adjacent_period_2;
1156
1157                         /* Calculate the periods of the adjacent frequency bands we'll be using to
1158                          * measure noise levels. We want the two adjacent Fourier frequency bands. */
1159                         adjacent_period_1 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) + 1);
1160                         adjacent_period_2 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) - 1);
1161
1162                         /* Get the the three different frequency components of the throughput (scaled by average
1163                          * throughput). Our "error" estimate (the amount of noise that might be present in the
1164                          * frequency band we're really interested in) is the average of the adjacent bands. */
1165                         throughput_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, hc->wave_period), average_throughput);
1166                         throughput_error_estimate = cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, adjacent_period_1), average_throughput));
1167
1168                         if (adjacent_period_2 <= sample_count) {
1169                                 throughput_error_estimate = MAX (throughput_error_estimate, cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (
1170                                         hc->samples, sample_count, adjacent_period_2), average_throughput)));
1171                         }
1172
1173                         /* Do the same for the thread counts, so we have something to compare to. We don't
1174                          * measure thread count noise, because there is none; these are exact measurements. */
1175                         thread_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->thread_counts, sample_count, hc->wave_period), average_thread_count);
1176
1177                         /* Update our moving average of the throughput noise. We'll use this
1178                          * later as feedback to determine the new size of the thread wave. */
1179                         if (hc->average_throughput_noise == 0) {
1180                                 hc->average_throughput_noise = throughput_error_estimate;
1181                         } else {
1182                                 hc->average_throughput_noise = (hc->throughput_error_smoothing_factor * throughput_error_estimate)
1183                                         + ((1.0 + hc->throughput_error_smoothing_factor) * hc->average_throughput_noise);
1184                         }
1185
1186                         if (cabs (thread_wave_component) > 0) {
1187                                 /* Adjust the throughput wave so it's centered around the target wave,
1188                                  * and then calculate the adjusted throughput/thread ratio. */
1189                                 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);
1190                                 transition = TRANSITION_CLIMBING_MOVE;
1191                         } else {
1192                                 ratio = mono_double_complex_make (0, 0);
1193                                 transition = TRANSITION_STABILIZING;
1194                         }
1195
1196                         noise_for_confidence = MAX (hc->average_throughput_noise, throughput_error_estimate);
1197                         if (noise_for_confidence > 0) {
1198                                 confidence = cabs (thread_wave_component) / noise_for_confidence / hc->target_signal_to_noise_ratio;
1199                         } else {
1200                                 /* there is no noise! */
1201                                 confidence = 1.0;
1202                         }
1203                 }
1204         }
1205
1206         /* We use just the real part of the complex ratio we just calculated. If the throughput signal
1207          * is exactly in phase with the thread signal, this will be the same as taking the magnitude of
1208          * the complex move and moving that far up. If they're 180 degrees out of phase, we'll move
1209          * backward (because this indicates that our changes are having the opposite of the intended effect).
1210          * If they're 90 degrees out of phase, we won't move at all, because we can't tell wether we're
1211          * having a negative or positive effect on throughput. */
1212         move = creal (ratio);
1213         move = CLAMP (move, -1.0, 1.0);
1214
1215         /* Apply our confidence multiplier. */
1216         move *= CLAMP (confidence, -1.0, 1.0);
1217
1218         /* Now apply non-linear gain, such that values around zero are attenuated, while higher values
1219          * are enhanced. This allows us to move quickly if we're far away from the target, but more slowly
1220         * if we're getting close, giving us rapid ramp-up without wild oscillations around the target. */
1221         gain = hc->max_change_per_second * sample_duration;
1222         move = pow (fabs (move), hc->gain_exponent) * (move >= 0.0 ? 1 : -1) * gain;
1223         move = MIN (move, hc->max_change_per_sample);
1224
1225         /* If the result was positive, and CPU is > 95%, refuse the move. */
1226         if (move > 0.0 && threadpool->cpu_usage > CPU_USAGE_HIGH)
1227                 move = 0.0;
1228
1229         /* Apply the move to our control setting. */
1230         hc->current_control_setting += move;
1231
1232         /* Calculate the new thread wave magnitude, which is based on the moving average we've been keeping of the
1233          * throughput error.  This average starts at zero, so we'll start with a nice safe little wave at first. */
1234         new_thread_wave_magnitude = (gint)(0.5 + (hc->current_control_setting * hc->average_throughput_noise
1235                 * hc->target_signal_to_noise_ratio * hc->thread_magnitude_multiplier * 2.0));
1236         new_thread_wave_magnitude = CLAMP (new_thread_wave_magnitude, 1, hc->max_thread_wave_magnitude);
1237
1238         /* Make sure our control setting is within the ThreadPool's limits. */
1239         hc->current_control_setting = CLAMP (hc->current_control_setting, threadpool->limit_worker_min, threadpool->limit_worker_max - new_thread_wave_magnitude);
1240
1241         /* Calculate the new thread count (control setting + square wave). */
1242         new_thread_count = (gint)(hc->current_control_setting + new_thread_wave_magnitude * ((hc->total_samples / (hc->wave_period / 2)) % 2));
1243
1244         /* Make sure the new thread count doesn't exceed the ThreadPool's limits. */
1245         new_thread_count = CLAMP (new_thread_count, threadpool->limit_worker_min, threadpool->limit_worker_max);
1246
1247         if (new_thread_count != current_thread_count)
1248                 hill_climbing_change_thread_count (new_thread_count, transition);
1249
1250         if (creal (ratio) < 0.0 && new_thread_count == threadpool->limit_worker_min)
1251                 *adjustment_interval = (gint)(0.5 + hc->current_sample_interval * (10.0 * MAX (-1.0 * creal (ratio), 1.0)));
1252         else
1253                 *adjustment_interval = hc->current_sample_interval;
1254
1255         return new_thread_count;
1256 }
1257
1258 static void
1259 heuristic_notify_work_completed (void)
1260 {
1261         g_assert (threadpool);
1262
1263         InterlockedIncrement (&threadpool->heuristic_completions);
1264         threadpool->heuristic_last_dequeue = mono_msec_ticks ();
1265 }
1266
1267 static gboolean
1268 heuristic_should_adjust (void)
1269 {
1270         g_assert (threadpool);
1271
1272         if (threadpool->heuristic_last_dequeue > threadpool->heuristic_last_adjustment + threadpool->heuristic_adjustment_interval) {
1273                 ThreadPoolCounter counter;
1274                 counter.as_gint64 = COUNTER_READ();
1275                 if (counter._.working <= counter._.max_working)
1276                         return TRUE;
1277         }
1278
1279         return FALSE;
1280 }
1281
1282 static void
1283 heuristic_adjust (void)
1284 {
1285         g_assert (threadpool);
1286
1287         if (mono_coop_mutex_trylock (&threadpool->heuristic_lock) == 0) {
1288                 gint32 completions = InterlockedExchange (&threadpool->heuristic_completions, 0);
1289                 guint32 sample_end = mono_msec_ticks ();
1290                 guint32 sample_duration = sample_end - threadpool->heuristic_sample_start;
1291
1292                 if (sample_duration >= threadpool->heuristic_adjustment_interval / 2) {
1293                         ThreadPoolCounter counter;
1294                         gint16 new_thread_count;
1295
1296                         counter.as_gint64 = COUNTER_READ ();
1297                         new_thread_count = hill_climbing_update (counter._.max_working, sample_duration, completions, &threadpool->heuristic_adjustment_interval);
1298
1299                         COUNTER_ATOMIC (counter, { counter._.max_working = new_thread_count; });
1300
1301                         if (new_thread_count > counter._.max_working)
1302                                 worker_request (mono_domain_get ());
1303
1304                         threadpool->heuristic_sample_start = sample_end;
1305                         threadpool->heuristic_last_adjustment = mono_msec_ticks ();
1306                 }
1307
1308                 mono_coop_mutex_unlock (&threadpool->heuristic_lock);
1309         }
1310 }
1311
1312 void
1313 mono_threadpool_ms_cleanup (void)
1314 {
1315         #ifndef DISABLE_SOCKETS
1316                 mono_threadpool_ms_io_cleanup ();
1317         #endif
1318         mono_lazy_cleanup (&status, cleanup);
1319 }
1320
1321 MonoAsyncResult *
1322 mono_threadpool_ms_begin_invoke (MonoDomain *domain, MonoObject *target, MonoMethod *method, gpointer *params, MonoError *error)
1323 {
1324         static MonoClass *async_call_klass = NULL;
1325         MonoMethodMessage *message;
1326         MonoAsyncResult *async_result;
1327         MonoAsyncCall *async_call;
1328         MonoDelegate *async_callback = NULL;
1329         MonoObject *state = NULL;
1330
1331         if (!async_call_klass)
1332                 async_call_klass = mono_class_load_from_name (mono_defaults.corlib, "System", "MonoAsyncCall");
1333
1334         mono_lazy_initialize (&status, initialize);
1335
1336         mono_error_init (error);
1337
1338         message = mono_method_call_message_new (method, params, mono_get_delegate_invoke (method->klass), (params != NULL) ? (&async_callback) : NULL, (params != NULL) ? (&state) : NULL);
1339
1340         async_call = (MonoAsyncCall*) mono_object_new_checked (domain, async_call_klass, error);
1341         return_val_if_nok (error, NULL);
1342
1343         MONO_OBJECT_SETREF (async_call, msg, message);
1344         MONO_OBJECT_SETREF (async_call, state, state);
1345
1346         if (async_callback) {
1347                 MONO_OBJECT_SETREF (async_call, cb_method, mono_get_delegate_invoke (((MonoObject*) async_callback)->vtable->klass));
1348                 MONO_OBJECT_SETREF (async_call, cb_target, async_callback);
1349         }
1350
1351         async_result = mono_async_result_new (domain, NULL, async_call->state, NULL, (MonoObject*) async_call);
1352         MONO_OBJECT_SETREF (async_result, async_delegate, target);
1353
1354         mono_threadpool_ms_enqueue_work_item (domain, (MonoObject*) async_result, error);
1355         return_val_if_nok (error, NULL);
1356
1357         return async_result;
1358 }
1359
1360 MonoObject *
1361 mono_threadpool_ms_end_invoke (MonoAsyncResult *ares, MonoArray **out_args, MonoObject **exc)
1362 {
1363         MonoAsyncCall *ac;
1364
1365         g_assert (exc);
1366         g_assert (out_args);
1367
1368         *exc = NULL;
1369         *out_args = NULL;
1370
1371         /* check if already finished */
1372         mono_monitor_enter ((MonoObject*) ares);
1373
1374         if (ares->endinvoke_called) {
1375                 *exc = (MonoObject*) mono_get_exception_invalid_operation (NULL);
1376                 mono_monitor_exit ((MonoObject*) ares);
1377                 return NULL;
1378         }
1379
1380         ares->endinvoke_called = 1;
1381
1382         /* wait until we are really finished */
1383         if (ares->completed) {
1384                 mono_monitor_exit ((MonoObject *) ares);
1385         } else {
1386                 gpointer wait_event;
1387                 if (ares->handle) {
1388                         wait_event = mono_wait_handle_get_handle ((MonoWaitHandle*) ares->handle);
1389                 } else {
1390                         wait_event = CreateEvent (NULL, TRUE, FALSE, NULL);
1391                         g_assert(wait_event);
1392                         MONO_OBJECT_SETREF (ares, handle, (MonoObject*) mono_wait_handle_new (mono_object_domain (ares), wait_event));
1393                 }
1394                 mono_monitor_exit ((MonoObject*) ares);
1395                 MONO_PREPARE_BLOCKING;
1396                 WaitForSingleObjectEx (wait_event, INFINITE, TRUE);
1397                 MONO_FINISH_BLOCKING;
1398         }
1399
1400         ac = (MonoAsyncCall*) ares->object_data;
1401         g_assert (ac);
1402
1403         *exc = ac->msg->exc; /* FIXME: GC add write barrier */
1404         *out_args = ac->out_args;
1405         return ac->res;
1406 }
1407
1408 gboolean
1409 mono_threadpool_ms_remove_domain_jobs (MonoDomain *domain, int timeout)
1410 {
1411         gboolean res = TRUE;
1412         guint32 start;
1413         gpointer sem;
1414
1415         g_assert (domain);
1416         g_assert (timeout >= -1);
1417
1418         g_assert (mono_domain_is_unloading (domain));
1419
1420         if (timeout != -1)
1421                 start = mono_msec_ticks ();
1422
1423 #ifndef DISABLE_SOCKETS
1424         mono_threadpool_ms_io_remove_domain_jobs (domain);
1425         if (timeout != -1) {
1426                 timeout -= mono_msec_ticks () - start;
1427                 if (timeout < 0)
1428                         return FALSE;
1429         }
1430 #endif
1431
1432         /*
1433          * There might be some threads out that could be about to execute stuff from the given domain.
1434          * We avoid that by setting up a semaphore to be pulsed by the thread that reaches zero.
1435          */
1436         sem = domain->cleanup_semaphore = CreateSemaphore (NULL, 0, 1, NULL);
1437
1438         /*
1439          * The memory barrier here is required to have global ordering between assigning to cleanup_semaphone
1440          * and reading threadpool_jobs. Otherwise this thread could read a stale version of threadpool_jobs
1441          * and wait forever.
1442          */
1443         mono_memory_write_barrier ();
1444
1445         while (domain->threadpool_jobs) {
1446                 MONO_PREPARE_BLOCKING;
1447                 WaitForSingleObject (sem, timeout);
1448                 MONO_FINISH_BLOCKING;
1449                 if (timeout != -1) {
1450                         timeout -= mono_msec_ticks () - start;
1451                         if (timeout <= 0) {
1452                                 res = FALSE;
1453                                 break;
1454                         }
1455                 }
1456         }
1457
1458         domain->cleanup_semaphore = NULL;
1459         CloseHandle (sem);
1460
1461         return res;
1462 }
1463
1464 void
1465 mono_threadpool_ms_suspend (void)
1466 {
1467         if (threadpool)
1468                 threadpool->suspended = TRUE;
1469 }
1470
1471 void
1472 mono_threadpool_ms_resume (void)
1473 {
1474         if (threadpool)
1475                 threadpool->suspended = FALSE;
1476 }
1477
1478 void
1479 ves_icall_System_Threading_ThreadPool_GetAvailableThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1480 {
1481         ThreadPoolCounter counter;
1482
1483         if (!worker_threads || !completion_port_threads)
1484                 return;
1485
1486         mono_lazy_initialize (&status, initialize);
1487
1488         counter.as_gint64 = COUNTER_READ ();
1489
1490         *worker_threads = MAX (0, threadpool->limit_worker_max - counter._.active);
1491         *completion_port_threads = threadpool->limit_io_max;
1492 }
1493
1494 void
1495 ves_icall_System_Threading_ThreadPool_GetMinThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1496 {
1497         if (!worker_threads || !completion_port_threads)
1498                 return;
1499
1500         mono_lazy_initialize (&status, initialize);
1501
1502         *worker_threads = threadpool->limit_worker_min;
1503         *completion_port_threads = threadpool->limit_io_min;
1504 }
1505
1506 void
1507 ves_icall_System_Threading_ThreadPool_GetMaxThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1508 {
1509         if (!worker_threads || !completion_port_threads)
1510                 return;
1511
1512         mono_lazy_initialize (&status, initialize);
1513
1514         *worker_threads = threadpool->limit_worker_max;
1515         *completion_port_threads = threadpool->limit_io_max;
1516 }
1517
1518 MonoBoolean
1519 ves_icall_System_Threading_ThreadPool_SetMinThreadsNative (gint32 worker_threads, gint32 completion_port_threads)
1520 {
1521         mono_lazy_initialize (&status, initialize);
1522
1523         if (worker_threads <= 0 || worker_threads > threadpool->limit_worker_max)
1524                 return FALSE;
1525         if (completion_port_threads <= 0 || completion_port_threads > threadpool->limit_io_max)
1526                 return FALSE;
1527
1528         threadpool->limit_worker_min = worker_threads;
1529         threadpool->limit_io_min = completion_port_threads;
1530
1531         return TRUE;
1532 }
1533
1534 MonoBoolean
1535 ves_icall_System_Threading_ThreadPool_SetMaxThreadsNative (gint32 worker_threads, gint32 completion_port_threads)
1536 {
1537         gint cpu_count = mono_cpu_count ();
1538
1539         mono_lazy_initialize (&status, initialize);
1540
1541         if (worker_threads < threadpool->limit_worker_min || worker_threads < cpu_count)
1542                 return FALSE;
1543         if (completion_port_threads < threadpool->limit_io_min || completion_port_threads < cpu_count)
1544                 return FALSE;
1545
1546         threadpool->limit_worker_max = worker_threads;
1547         threadpool->limit_io_max = completion_port_threads;
1548
1549         return TRUE;
1550 }
1551
1552 void
1553 ves_icall_System_Threading_ThreadPool_InitializeVMTp (MonoBoolean *enable_worker_tracking)
1554 {
1555         if (enable_worker_tracking) {
1556                 // TODO implement some kind of switch to have the possibily to use it
1557                 *enable_worker_tracking = FALSE;
1558         }
1559
1560         mono_lazy_initialize (&status, initialize);
1561 }
1562
1563 MonoBoolean
1564 ves_icall_System_Threading_ThreadPool_NotifyWorkItemComplete (void)
1565 {
1566         ThreadPoolCounter counter;
1567
1568         if (mono_domain_is_unloading (mono_domain_get ()) || mono_runtime_is_shutting_down ())
1569                 return FALSE;
1570
1571         heuristic_notify_work_completed ();
1572
1573         if (heuristic_should_adjust ())
1574                 heuristic_adjust ();
1575
1576         counter.as_gint64 = COUNTER_READ ();
1577         return counter._.working <= counter._.max_working;
1578 }
1579
1580 void
1581 ves_icall_System_Threading_ThreadPool_NotifyWorkItemProgressNative (void)
1582 {
1583         heuristic_notify_work_completed ();
1584
1585         if (heuristic_should_adjust ())
1586                 heuristic_adjust ();
1587 }
1588
1589 void
1590 ves_icall_System_Threading_ThreadPool_ReportThreadStatus (MonoBoolean is_working)
1591 {
1592         // TODO
1593         MonoError error;
1594         mono_error_set_not_implemented (&error, "");
1595         mono_error_set_pending_exception (&error);
1596 }
1597
1598 MonoBoolean
1599 ves_icall_System_Threading_ThreadPool_RequestWorkerThread (void)
1600 {
1601         return worker_request (mono_domain_get ());
1602 }
1603
1604 MonoBoolean G_GNUC_UNUSED
1605 ves_icall_System_Threading_ThreadPool_PostQueuedCompletionStatus (MonoNativeOverlapped *native_overlapped)
1606 {
1607         /* This copy the behavior of the current Mono implementation */
1608         MonoError error;
1609         mono_error_set_not_implemented (&error, "");
1610         mono_error_set_pending_exception (&error);
1611         return FALSE;
1612 }
1613
1614 MonoBoolean G_GNUC_UNUSED
1615 ves_icall_System_Threading_ThreadPool_BindIOCompletionCallbackNative (gpointer file_handle)
1616 {
1617         /* This copy the behavior of the current Mono implementation */
1618         return TRUE;
1619 }
1620
1621 MonoBoolean G_GNUC_UNUSED
1622 ves_icall_System_Threading_ThreadPool_IsThreadPoolHosted (void)
1623 {
1624         return FALSE;
1625 }