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