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