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 #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         guint32 val;
 231         if (!mono_rand_try_get_uint32 (handle, &val, min, max)) {
 232                 // FIXME handle error
 233                 g_assert_not_reached ();
 234         }
 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         MonoDomain *current_domain;
 353         MonoBoolean f;
 354         gpointer args [2];
 355
 356         g_assert (work_item);
 357
 358         if (!threadpool_class)
 359                 threadpool_class = mono_class_from_name (mono_defaults.corlib, "System.Threading", "ThreadPool");
 360         g_assert (threadpool_class);
 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 (unsafe_queue_custom_work_item_method, NULL, args, NULL);
 374         } else {
 375                 mono_thread_push_appdomain_ref (domain);
 376                 if (mono_domain_set (domain, FALSE)) {
 377                         mono_runtime_invoke (unsafe_queue_custom_work_item_method, NULL, args, NULL);
 378                         mono_domain_set (current_domain, TRUE);
 379                 }
 380                 mono_thread_pop_appdomain_ref ();
 381         }
 382 }
 383
 384 /* LOCKING: threadpool->domains_lock must be held */
 385 static void
 386 domain_add (ThreadPoolDomain *tpdomain)
 387 {
 388         guint i, len;
 389
 390         g_assert (tpdomain);
 391
 392         len = threadpool->domains->len;
 393         for (i = 0; i < len; ++i) {
 394                 if (g_ptr_array_index (threadpool->domains, i) == tpdomain)
 395                         break;
 396         }
 397
 398         if (i == len)
 399                 g_ptr_array_add (threadpool->domains, tpdomain);
 400 }
 401
 402 /* LOCKING: threadpool->domains_lock must be held */
 403 static gboolean
 404 domain_remove (ThreadPoolDomain *tpdomain)
 405 {
 406         g_assert (tpdomain);
 407         return g_ptr_array_remove (threadpool->domains, tpdomain);
 408 }
 409
 410 /* LOCKING: threadpool->domains_lock must be held */
 411 static ThreadPoolDomain *
 412 domain_get (MonoDomain *domain, gboolean create)
 413 {
 414         ThreadPoolDomain *tpdomain = NULL;
 415         guint i;
 416
 417         g_assert (domain);
 418
 419         for (i = 0; i < threadpool->domains->len; ++i) {
 420                 tpdomain = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i);
 421                 if (tpdomain->domain == domain)
 422                         return tpdomain;
 423         }
 424
 425         if (create) {
 426                 tpdomain = g_new0 (ThreadPoolDomain, 1);
 427                 tpdomain->domain = domain;
 428                 domain_add (tpdomain);
 429         }
 430
 431         return tpdomain;
 432 }
 433
 434 static void
 435 domain_free (ThreadPoolDomain *tpdomain)
 436 {
 437         g_free (tpdomain);
 438 }
 439
 440 /* LOCKING: threadpool->domains_lock must be held */
 441 static gboolean
 442 domain_any_has_request (void)
 443 {
 444         guint i;
 445
 446         for (i = 0; i < threadpool->domains->len; ++i) {
 447                 ThreadPoolDomain *tmp = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i);
 448                 if (tmp->outstanding_request > 0)
 449                         return TRUE;
 450         }
 451
 452         return FALSE;
 453 }
 454
 455 /* LOCKING: threadpool->domains_lock must be held */
 456 static ThreadPoolDomain *
 457 domain_get_next (ThreadPoolDomain *current)
 458 {
 459         ThreadPoolDomain *tpdomain = NULL;
 460         guint len;
 461
 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 = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i % len);
 476                         if (tmp->outstanding_request > 0) {
 477                                 tpdomain = tmp;
 478                                 break;
 479                         }
 480                 }
 481         }
 482
 483         return tpdomain;
 484 }
 485
 486 static void
 487 worker_wait_interrupt (gpointer data)
 488 {
 489         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 490         mono_coop_cond_signal (&threadpool->parked_threads_cond);
 491         mono_coop_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_coop_mutex_lock (&threadpool->active_threads_lock);
 505
 506         if (!mono_runtime_is_shutting_down ()) {
 507                 static gpointer rand_handle = NULL;
 508                 MonoInternalThread *thread_internal;
 509                 gboolean interrupted = FALSE;
 510
 511                 if (!rand_handle)
 512                         rand_handle = rand_create ();
 513                 g_assert (rand_handle);
 514
 515                 thread_internal = mono_thread_internal_current ();
 516                 g_assert (thread_internal);
 517
 518                 threadpool->parked_threads_count += 1;
 519                 g_ptr_array_remove_fast (threadpool->working_threads, thread_internal);
 520
 521                 mono_thread_info_install_interrupt (worker_wait_interrupt, NULL, &interrupted);
 522                 if (interrupted)
 523                         goto done;
 524
 525                 if (mono_coop_cond_timedwait (&threadpool->parked_threads_cond, &threadpool->active_threads_lock, rand_next ((void **)rand_handle, 5 * 1000, 60 * 1000)) != 0)
 526                         timeout = TRUE;
 527
 528                 mono_thread_info_uninstall_interrupt (&interrupted);
 529
 530 done:
 531                 g_ptr_array_add (threadpool->working_threads, thread_internal);
 532                 threadpool->parked_threads_count -= 1;
 533         }
 534
 535         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 536
 537         mono_gc_set_skip_thread (FALSE);
 538
 539         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] current worker unparking, timeout? %s", mono_native_thread_id_get (), timeout ? "yes" : "no");
 540
 541         return timeout;
 542 }
 543
 544 static gboolean
 545 worker_try_unpark (void)
 546 {
 547         gboolean res = FALSE;
 548
 549         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker", mono_native_thread_id_get ());
 550
 551         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 552         if (threadpool->parked_threads_count > 0) {
 553                 mono_coop_cond_signal (&threadpool->parked_threads_cond);
 554                 res = TRUE;
 555         }
 556         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 557
 558         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker, success? %s", mono_native_thread_id_get (), res ? "yes" : "no");
 559
 560         return res;
 561 }
 562
 563 static void
 564 worker_kill (ThreadPoolWorkingThread *thread)
 565 {
 566         if (thread == mono_thread_internal_current ())
 567                 return;
 568
 569         mono_thread_internal_stop ((MonoInternalThread*) thread);
 570 }
 571
 572 static void
 573 worker_thread (gpointer data)
 574 {
 575         MonoInternalThread *thread;
 576         ThreadPoolDomain *tpdomain, *previous_tpdomain;
 577         ThreadPoolCounter counter;
 578         gboolean retire = FALSE;
 579
 580         mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker starting", mono_native_thread_id_get ());
 581
 582         g_assert (threadpool);
 583
 584         thread = mono_thread_internal_current ();
 585         g_assert (thread);
 586
 587         mono_thread_set_name_internal (thread, mono_string_new (mono_domain_get (), "Threadpool worker"), FALSE);
 588
 589         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 590         g_ptr_array_add (threadpool->working_threads, thread);
 591         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 592
 593         previous_tpdomain = NULL;
 594
 595         mono_coop_mutex_lock (&threadpool->domains_lock);
 596
 597         while (!mono_runtime_is_shutting_down ()) {
 598                 tpdomain = NULL;
 599
 600                 if ((thread->state & (ThreadState_StopRequested | ThreadState_SuspendRequested)) != 0) {
 601                         mono_coop_mutex_unlock (&threadpool->domains_lock);
 602                         mono_thread_interruption_checkpoint ();
 603                         mono_coop_mutex_lock (&threadpool->domains_lock);
 604                 }
 605
 606                 if (retire || !(tpdomain = domain_get_next (previous_tpdomain))) {
 607                         gboolean timeout;
 608
 609                         COUNTER_ATOMIC (counter, {
 610                                 counter._.working --;
 611                                 counter._.parked ++;
 612                         });
 613
 614                         mono_coop_mutex_unlock (&threadpool->domains_lock);
 615                         timeout = worker_park ();
 616                         mono_coop_mutex_lock (&threadpool->domains_lock);
 617
 618                         COUNTER_ATOMIC (counter, {
 619                                 counter._.working ++;
 620                                 counter._.parked --;
 621                         });
 622
 623                         if (timeout)
 624                                 break;
 625
 626                         if (retire)
 627                                 retire = FALSE;
 628
 629                         continue;
 630                 }
 631
 632                 tpdomain->outstanding_request --;
 633                 g_assert (tpdomain->outstanding_request >= 0);
 634
 635                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker running in domain %p",
 636                         mono_native_thread_id_get (), tpdomain->domain, tpdomain->outstanding_request);
 637
 638                 g_assert (tpdomain->domain);
 639                 g_assert (tpdomain->domain->threadpool_jobs >= 0);
 640                 tpdomain->domain->threadpool_jobs ++;
 641
 642                 mono_coop_mutex_unlock (&threadpool->domains_lock);
 643
 644                 mono_thread_push_appdomain_ref (tpdomain->domain);
 645                 if (mono_domain_set (tpdomain->domain, FALSE)) {
 646                         MonoObject *exc = NULL;
 647                         MonoObject *res = mono_runtime_invoke (mono_defaults.threadpool_perform_wait_callback_method, NULL, NULL, &exc);
 648                         if (exc)
 649                                 mono_thread_internal_unhandled_exception (exc);
 650                         else if (res && *(MonoBoolean*) mono_object_unbox (res) == FALSE)
 651                                 retire = TRUE;
 652
 653                         mono_thread_clr_state (thread, (MonoThreadState)~ThreadState_Background);
 654                         if (!mono_thread_test_state (thread , ThreadState_Background))
 655                                 ves_icall_System_Threading_Thread_SetState (thread, ThreadState_Background);
 656
 657                         mono_domain_set (mono_get_root_domain (), TRUE);
 658                 }
 659                 mono_thread_pop_appdomain_ref ();
 660
 661                 mono_coop_mutex_lock (&threadpool->domains_lock);
 662
 663                 tpdomain->domain->threadpool_jobs --;
 664                 g_assert (tpdomain->domain->threadpool_jobs >= 0);
 665
 666                 if (tpdomain->domain->threadpool_jobs == 0 && mono_domain_is_unloading (tpdomain->domain)) {
 667                         gboolean removed = domain_remove (tpdomain);
 668                         g_assert (removed);
 669                         if (tpdomain->domain->cleanup_semaphore)
 670                                 ReleaseSemaphore (tpdomain->domain->cleanup_semaphore, 1, NULL);
 671                         domain_free (tpdomain);
 672                         tpdomain = NULL;
 673                 }
 674
 675                 previous_tpdomain = tpdomain;
 676         }
 677
 678         mono_coop_mutex_unlock (&threadpool->domains_lock);
 679
 680         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 681         g_ptr_array_remove_fast (threadpool->working_threads, thread);
 682         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 683
 684         COUNTER_ATOMIC (counter, {
 685                 counter._.working--;
 686                 counter._.active --;
 687         });
 688
 689         mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker finishing", mono_native_thread_id_get ());
 690 }
 691
 692 static gboolean
 693 worker_try_create (void)
 694 {
 695         ThreadPoolCounter counter;
 696         MonoInternalThread *thread;
 697         gint32 now;
 698
 699         mono_coop_mutex_lock (&threadpool->worker_creation_lock);
 700
 701         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker", mono_native_thread_id_get ());
 702
 703         if ((now = mono_100ns_ticks () / 10 / 1000 / 1000) == 0) {
 704                 g_warning ("failed to get 100ns ticks");
 705         } else {
 706                 if (threadpool->worker_creation_current_second != now) {
 707                         threadpool->worker_creation_current_second = now;
 708                         threadpool->worker_creation_current_count = 0;
 709                 } else {
 710                         g_assert (threadpool->worker_creation_current_count <= WORKER_CREATION_MAX_PER_SEC);
 711                         if (threadpool->worker_creation_current_count == WORKER_CREATION_MAX_PER_SEC) {
 712                                 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",
 713                                         mono_native_thread_id_get (), threadpool->worker_creation_current_count);
 714                                 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
 715                                 return FALSE;
 716                         }
 717                 }
 718         }
 719
 720         COUNTER_ATOMIC (counter, {
 721                 if (counter._.working >= counter._.max_working) {
 722                         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: maximum number of working threads reached",
 723                                 mono_native_thread_id_get ());
 724                         mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
 725                         return FALSE;
 726                 }
 727                 counter._.working ++;
 728                 counter._.active ++;
 729         });
 730
 731         if ((thread = mono_thread_create_internal (mono_get_root_domain (), worker_thread, NULL, TRUE, 0)) != NULL) {
 732                 threadpool->worker_creation_current_count += 1;
 733
 734                 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);
 735                 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
 736                 return TRUE;
 737         }
 738
 739         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: could not create thread", mono_native_thread_id_get ());
 740
 741         COUNTER_ATOMIC (counter, {
 742                 counter._.working --;
 743                 counter._.active --;
 744         });
 745
 746         mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
 747         return FALSE;
 748 }
 749
 750 static void monitor_ensure_running (void);
 751
 752 static gboolean
 753 worker_request (MonoDomain *domain)
 754 {
 755         ThreadPoolDomain *tpdomain;
 756
 757         g_assert (domain);
 758         g_assert (threadpool);
 759
 760         if (mono_runtime_is_shutting_down ())
 761                 return FALSE;
 762
 763         mono_coop_mutex_lock (&threadpool->domains_lock);
 764
 765         /* synchronize check with worker_thread */
 766         if (mono_domain_is_unloading (domain)) {
 767                 mono_coop_mutex_unlock (&threadpool->domains_lock);
 768                 return FALSE;
 769         }
 770
 771         tpdomain = domain_get (domain, TRUE);
 772         g_assert (tpdomain);
 773         tpdomain->outstanding_request ++;
 774
 775         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, domain = %p, outstanding_request = %d",
 776                 mono_native_thread_id_get (), tpdomain->domain, tpdomain->outstanding_request);
 777
 778         mono_coop_mutex_unlock (&threadpool->domains_lock);
 779
 780         if (threadpool->suspended)
 781                 return FALSE;
 782
 783         monitor_ensure_running ();
 784
 785         if (worker_try_unpark ()) {
 786                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, unparked", mono_native_thread_id_get ());
 787                 return TRUE;
 788         }
 789
 790         if (worker_try_create ()) {
 791                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, created", mono_native_thread_id_get ());
 792                 return TRUE;
 793         }
 794
 795         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, failed", mono_native_thread_id_get ());
 796         return FALSE;
 797 }
 798
 799 static gboolean
 800 monitor_should_keep_running (void)
 801 {
 802         static gint64 last_should_keep_running = -1;
 803
 804         g_assert (monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || monitor_status == MONITOR_STATUS_REQUESTED);
 805
 806         if (InterlockedExchange (&monitor_status, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST) {
 807                 gboolean should_keep_running = TRUE, force_should_keep_running = FALSE;
 808
 809                 if (mono_runtime_is_shutting_down ()) {
 810                         should_keep_running = FALSE;
 811                 } else {
 812                         mono_coop_mutex_lock (&threadpool->domains_lock);
 813                         if (!domain_any_has_request ())
 814                                 should_keep_running = FALSE;
 815                         mono_coop_mutex_unlock (&threadpool->domains_lock);
 816
 817                         if (!should_keep_running) {
 818                                 if (last_should_keep_running == -1 || mono_100ns_ticks () - last_should_keep_running < MONITOR_MINIMAL_LIFETIME * 1000 * 10) {
 819                                         should_keep_running = force_should_keep_running = TRUE;
 820                                 }
 821                         }
 822                 }
 823
 824                 if (should_keep_running) {
 825                         if (last_should_keep_running == -1 || !force_should_keep_running)
 826                                 last_should_keep_running = mono_100ns_ticks ();
 827                 } else {
 828                         last_should_keep_running = -1;
 829                         if (InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_NOT_RUNNING, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST)
 830                                 return FALSE;
 831                 }
 832         }
 833
 834         g_assert (monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || monitor_status == MONITOR_STATUS_REQUESTED);
 835
 836         return TRUE;
 837 }
 838
 839 static gboolean
 840 monitor_sufficient_delay_since_last_dequeue (void)
 841 {
 842         guint32 threshold;
 843
 844         g_assert (threadpool);
 845
 846         if (threadpool->cpu_usage < CPU_USAGE_LOW) {
 847                 threshold = MONITOR_INTERVAL;
 848         } else {
 849                 ThreadPoolCounter counter;
 850                 counter.as_gint64 = COUNTER_READ();
 851                 threshold = counter._.max_working * MONITOR_INTERVAL * 2;
 852         }
 853
 854         return mono_msec_ticks () >= threadpool->heuristic_last_dequeue + threshold;
 855 }
 856
 857 static void hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition);
 858
 859 static void
 860 monitor_thread (void)
 861 {
 862         MonoInternalThread *current_thread = mono_thread_internal_current ();
 863         guint i;
 864
 865         mono_cpu_usage (threadpool->cpu_usage_state);
 866
 867         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, started", mono_native_thread_id_get ());
 868
 869         do {
 870                 MonoInternalThread *thread;
 871                 gboolean all_waitsleepjoin = TRUE;
 872                 gint32 interval_left = MONITOR_INTERVAL;
 873                 gint32 awake = 0; /* number of spurious awakes we tolerate before doing a round of rebalancing */
 874
 875                 g_assert (monitor_status != MONITOR_STATUS_NOT_RUNNING);
 876
 877                 mono_gc_set_skip_thread (TRUE);
 878
 879                 do {
 880                         guint32 ts;
 881                         gboolean alerted = FALSE;
 882
 883                         if (mono_runtime_is_shutting_down ())
 884                                 break;
 885
 886                         ts = mono_msec_ticks ();
 887                         if (mono_thread_info_sleep (interval_left, &alerted) == 0)
 888                                 break;
 889                         interval_left -= mono_msec_ticks () - ts;
 890
 891                         mono_gc_set_skip_thread (FALSE);
 892                         if ((current_thread->state & (ThreadState_StopRequested | ThreadState_SuspendRequested)) != 0)
 893                                 mono_thread_interruption_checkpoint ();
 894                         mono_gc_set_skip_thread (TRUE);
 895                 } while (interval_left > 0 && ++awake < 10);
 896
 897                 mono_gc_set_skip_thread (FALSE);
 898
 899                 if (threadpool->suspended)
 900                         continue;
 901
 902                 if (mono_runtime_is_shutting_down ())
 903                         continue;
 904
 905                 mono_coop_mutex_lock (&threadpool->domains_lock);
 906                 if (!domain_any_has_request ()) {
 907                         mono_coop_mutex_unlock (&threadpool->domains_lock);
 908                         continue;
 909                 }
 910                 mono_coop_mutex_unlock (&threadpool->domains_lock);
 911
 912
 913                 mono_coop_mutex_lock (&threadpool->active_threads_lock);
 914                 for (i = 0; i < threadpool->working_threads->len; ++i) {
 915                         thread = (MonoInternalThread *)g_ptr_array_index (threadpool->working_threads, i);
 916                         if ((thread->state & ThreadState_WaitSleepJoin) == 0) {
 917                                 all_waitsleepjoin = FALSE;
 918                                 break;
 919                         }
 920                 }
 921                 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 922
 923                 if (all_waitsleepjoin) {
 924                         ThreadPoolCounter counter;
 925                         gboolean 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                                 hill_climbing_force_change (counter._.max_working, TRANSITION_STARVATION);
 937                 }
 938
 939                 threadpool->cpu_usage = mono_cpu_usage (threadpool->cpu_usage_state);
 940
 941                 if (monitor_sufficient_delay_since_last_dequeue ()) {
 942                         for (i = 0; i < 5; ++i) {
 943                                 if (mono_runtime_is_shutting_down ())
 944                                         break;
 945
 946                                 if (worker_try_unpark ()) {
 947                                         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, unparked", mono_native_thread_id_get ());
 948                                         break;
 949                                 }
 950
 951                                 if (worker_try_create ()) {
 952                                         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, created", mono_native_thread_id_get ());
 953                                         break;
 954                                 }
 955                         }
 956                 }
 957         } while (monitor_should_keep_running ());
 958
 959         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, finished", mono_native_thread_id_get ());
 960 }
 961
 962 static void
 963 monitor_ensure_running (void)
 964 {
 965         for (;;) {
 966                 switch (monitor_status) {
 967                 case MONITOR_STATUS_REQUESTED:
 968                         return;
 969                 case MONITOR_STATUS_WAITING_FOR_REQUEST:
 970                         InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_WAITING_FOR_REQUEST);
 971                         break;
 972                 case MONITOR_STATUS_NOT_RUNNING:
 973                         if (mono_runtime_is_shutting_down ())
 974                                 return;
 975                         if (InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_NOT_RUNNING) == MONITOR_STATUS_NOT_RUNNING) {
 976                                 if (!mono_thread_create_internal (mono_get_root_domain (), monitor_thread, NULL, TRUE, SMALL_STACK))
 977                                         monitor_status = MONITOR_STATUS_NOT_RUNNING;
 978                                 return;
 979                         }
 980                         break;
 981                 default: g_assert_not_reached ();
 982                 }
 983         }
 984 }
 985
 986 static void
 987 hill_climbing_change_thread_count (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
 988 {
 989         ThreadPoolHillClimbing *hc;
 990
 991         g_assert (threadpool);
 992
 993         hc = &threadpool->heuristic_hill_climbing;
 994
 995         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);
 996
 997         hc->last_thread_count = new_thread_count;
 998         hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
 999         hc->elapsed_since_last_change = 0;
1000         hc->completions_since_last_change = 0;
1001 }
1002
1003 static void
1004 hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
1005 {
1006         ThreadPoolHillClimbing *hc;
1007
1008         g_assert (threadpool);
1009
1010         hc = &threadpool->heuristic_hill_climbing;
1011
1012         if (new_thread_count != hc->last_thread_count) {
1013                 hc->current_control_setting += new_thread_count - hc->last_thread_count;
1014                 hill_climbing_change_thread_count (new_thread_count, transition);
1015         }
1016 }
1017
1018 static double_complex
1019 hill_climbing_get_wave_component (gdouble *samples, guint sample_count, gdouble period)
1020 {
1021         ThreadPoolHillClimbing *hc;
1022         gdouble w, cosine, sine, coeff, q0, q1, q2;
1023         guint i;
1024
1025         g_assert (threadpool);
1026         g_assert (sample_count >= period);
1027         g_assert (period >= 2);
1028
1029         hc = &threadpool->heuristic_hill_climbing;
1030
1031         w = 2.0 * M_PI / period;
1032         cosine = cos (w);
1033         sine = sin (w);
1034         coeff = 2.0 * cosine;
1035         q0 = q1 = q2 = 0;
1036
1037         for (i = 0; i < sample_count; ++i) {
1038                 q0 = coeff * q1 - q2 + samples [(hc->total_samples - sample_count + i) % hc->samples_to_measure];
1039                 q2 = q1;
1040                 q1 = q0;
1041         }
1042
1043         return mono_double_complex_scalar_div (mono_double_complex_make (q1 - q2 * cosine, (q2 * sine)), ((gdouble)sample_count));
1044 }
1045
1046 static gint16
1047 hill_climbing_update (gint16 current_thread_count, guint32 sample_duration, gint32 completions, guint32 *adjustment_interval)
1048 {
1049         ThreadPoolHillClimbing *hc;
1050         ThreadPoolHeuristicStateTransition transition;
1051         gdouble throughput;
1052         gdouble throughput_error_estimate;
1053         gdouble confidence;
1054         gdouble move;
1055         gdouble gain;
1056         gint sample_index;
1057         gint sample_count;
1058         gint new_thread_wave_magnitude;
1059         gint new_thread_count;
1060         double_complex thread_wave_component;
1061         double_complex throughput_wave_component;
1062         double_complex ratio;
1063
1064         g_assert (threadpool);
1065         g_assert (adjustment_interval);
1066
1067         hc = &threadpool->heuristic_hill_climbing;
1068
1069         /* If someone changed the thread count without telling us, update our records accordingly. */
1070         if (current_thread_count != hc->last_thread_count)
1071                 hill_climbing_force_change (current_thread_count, TRANSITION_INITIALIZING);
1072
1073         /* Update the cumulative stats for this thread count */
1074         hc->elapsed_since_last_change += sample_duration;
1075         hc->completions_since_last_change += completions;
1076
1077         /* Add in any data we've already collected about this sample */
1078         sample_duration += hc->accumulated_sample_duration;
1079         completions += hc->accumulated_completion_count;
1080
1081         /* We need to make sure we're collecting reasonably accurate data. Since we're just counting the end
1082          * of each work item, we are goinng to be missing some data about what really happened during the
1083          * sample interval. The count produced by each thread includes an initial work item that may have
1084          * started well before the start of the interval, and each thread may have been running some new
1085          * work item for some time before the end of the interval, which did not yet get counted. So
1086          * our count is going to be off by +/- threadCount workitems.
1087          *
1088          * The exception is that the thread that reported to us last time definitely wasn't running any work
1089          * at that time, and the thread that's reporting now definitely isn't running a work item now. So
1090          * we really only need to consider threadCount-1 threads.
1091          *
1092          * Thus the percent error in our count is +/- (threadCount-1)/numCompletions.
1093          *
1094          * We cannot rely on the frequency-domain analysis we'll be doing later to filter out this error, because
1095          * of the way it accumulates over time. If this sample is off by, say, 33% in the negative direction,
1096          * then the next one likely will be too. The one after that will include the sum of the completions
1097          * we missed in the previous samples, and so will be 33% positive. So every three samples we'll have
1098          * two "low" samples and one "high" sample. This will appear as periodic variation right in the frequency
1099          * range we're targeting, which will not be filtered by the frequency-domain translation. */
1100         if (hc->total_samples > 0 && ((current_thread_count - 1.0) / completions) >= hc->max_sample_error) {
1101                 /* Not accurate enough yet. Let's accumulate the data so
1102                  * far, and tell the ThreadPool to collect a little more. */
1103                 hc->accumulated_sample_duration = sample_duration;
1104                 hc->accumulated_completion_count = completions;
1105                 *adjustment_interval = 10;
1106                 return current_thread_count;
1107         }
1108
1109         /* We've got enouugh data for our sample; reset our accumulators for next time. */
1110         hc->accumulated_sample_duration = 0;
1111         hc->accumulated_completion_count = 0;
1112
1113         /* Add the current thread count and throughput sample to our history. */
1114         throughput = ((gdouble) completions) / sample_duration;
1115
1116         sample_index = hc->total_samples % hc->samples_to_measure;
1117         hc->samples [sample_index] = throughput;
1118         hc->thread_counts [sample_index] = current_thread_count;
1119         hc->total_samples ++;
1120
1121         /* Set up defaults for our metrics. */
1122         thread_wave_component = mono_double_complex_make(0, 0);
1123         throughput_wave_component = mono_double_complex_make(0, 0);
1124         throughput_error_estimate = 0;
1125         ratio = mono_double_complex_make(0, 0);
1126         confidence = 0;
1127
1128         transition = TRANSITION_WARMUP;
1129
1130         /* How many samples will we use? It must be at least the three wave periods we're looking for, and it must also
1131          * be a whole multiple of the primary wave's period; otherwise the frequency we're looking for will fall between
1132          * two frequency bands in the Fourier analysis, and we won't be able to measure it accurately. */
1133         sample_count = ((gint) MIN (hc->total_samples - 1, hc->samples_to_measure) / hc->wave_period) * hc->wave_period;
1134
1135         if (sample_count > hc->wave_period) {
1136                 guint i;
1137                 gdouble average_throughput;
1138                 gdouble average_thread_count;
1139                 gdouble sample_sum = 0;
1140                 gdouble thread_sum = 0;
1141
1142                 /* Average the throughput and thread count samples, so we can scale the wave magnitudes later. */
1143                 for (i = 0; i < sample_count; ++i) {
1144                         guint j = (hc->total_samples - sample_count + i) % hc->samples_to_measure;
1145                         sample_sum += hc->samples [j];
1146                         thread_sum += hc->thread_counts [j];
1147                 }
1148
1149                 average_throughput = sample_sum / sample_count;
1150                 average_thread_count = thread_sum / sample_count;
1151
1152                 if (average_throughput > 0 && average_thread_count > 0) {
1153                         gdouble noise_for_confidence, adjacent_period_1, adjacent_period_2;
1154
1155                         /* Calculate the periods of the adjacent frequency bands we'll be using to
1156                          * measure noise levels. We want the two adjacent Fourier frequency bands. */
1157                         adjacent_period_1 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) + 1);
1158                         adjacent_period_2 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) - 1);
1159
1160                         /* Get the the three different frequency components of the throughput (scaled by average
1161                          * throughput). Our "error" estimate (the amount of noise that might be present in the
1162                          * frequency band we're really interested in) is the average of the adjacent bands. */
1163                         throughput_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, hc->wave_period), average_throughput);
1164                         throughput_error_estimate = cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, adjacent_period_1), average_throughput));
1165
1166                         if (adjacent_period_2 <= sample_count) {
1167                                 throughput_error_estimate = MAX (throughput_error_estimate, cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (
1168                                         hc->samples, sample_count, adjacent_period_2), average_throughput)));
1169                         }
1170
1171                         /* Do the same for the thread counts, so we have something to compare to. We don't
1172                          * measure thread count noise, because there is none; these are exact measurements. */
1173                         thread_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->thread_counts, sample_count, hc->wave_period), average_thread_count);
1174
1175                         /* Update our moving average of the throughput noise. We'll use this
1176                          * later as feedback to determine the new size of the thread wave. */
1177                         if (hc->average_throughput_noise == 0) {
1178                                 hc->average_throughput_noise = throughput_error_estimate;
1179                         } else {
1180                                 hc->average_throughput_noise = (hc->throughput_error_smoothing_factor * throughput_error_estimate)
1181                                         + ((1.0 + hc->throughput_error_smoothing_factor) * hc->average_throughput_noise);
1182                         }
1183
1184                         if (cabs (thread_wave_component) > 0) {
1185                                 /* Adjust the throughput wave so it's centered around the target wave,
1186                                  * and then calculate the adjusted throughput/thread ratio. */
1187                                 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);
1188                                 transition = TRANSITION_CLIMBING_MOVE;
1189                         } else {
1190                                 ratio = mono_double_complex_make (0, 0);
1191                                 transition = TRANSITION_STABILIZING;
1192                         }
1193
1194                         noise_for_confidence = MAX (hc->average_throughput_noise, throughput_error_estimate);
1195                         if (noise_for_confidence > 0) {
1196                                 confidence = cabs (thread_wave_component) / noise_for_confidence / hc->target_signal_to_noise_ratio;
1197                         } else {
1198                                 /* there is no noise! */
1199                                 confidence = 1.0;
1200                         }
1201                 }
1202         }
1203
1204         /* We use just the real part of the complex ratio we just calculated. If the throughput signal
1205          * is exactly in phase with the thread signal, this will be the same as taking the magnitude of
1206          * the complex move and moving that far up. If they're 180 degrees out of phase, we'll move
1207          * backward (because this indicates that our changes are having the opposite of the intended effect).
1208          * If they're 90 degrees out of phase, we won't move at all, because we can't tell wether we're
1209          * having a negative or positive effect on throughput. */
1210         move = creal (ratio);
1211         move = CLAMP (move, -1.0, 1.0);
1212
1213         /* Apply our confidence multiplier. */
1214         move *= CLAMP (confidence, -1.0, 1.0);
1215
1216         /* Now apply non-linear gain, such that values around zero are attenuated, while higher values
1217          * are enhanced. This allows us to move quickly if we're far away from the target, but more slowly
1218         * if we're getting close, giving us rapid ramp-up without wild oscillations around the target. */
1219         gain = hc->max_change_per_second * sample_duration;
1220         move = pow (fabs (move), hc->gain_exponent) * (move >= 0.0 ? 1 : -1) * gain;
1221         move = MIN (move, hc->max_change_per_sample);
1222
1223         /* If the result was positive, and CPU is > 95%, refuse the move. */
1224         if (move > 0.0 && threadpool->cpu_usage > CPU_USAGE_HIGH)
1225                 move = 0.0;
1226
1227         /* Apply the move to our control setting. */
1228         hc->current_control_setting += move;
1229
1230         /* Calculate the new thread wave magnitude, which is based on the moving average we've been keeping of the
1231          * throughput error.  This average starts at zero, so we'll start with a nice safe little wave at first. */
1232         new_thread_wave_magnitude = (gint)(0.5 + (hc->current_control_setting * hc->average_throughput_noise
1233                 * hc->target_signal_to_noise_ratio * hc->thread_magnitude_multiplier * 2.0));
1234         new_thread_wave_magnitude = CLAMP (new_thread_wave_magnitude, 1, hc->max_thread_wave_magnitude);
1235
1236         /* Make sure our control setting is within the ThreadPool's limits. */
1237         hc->current_control_setting = CLAMP (hc->current_control_setting, threadpool->limit_worker_min, threadpool->limit_worker_max - new_thread_wave_magnitude);
1238
1239         /* Calculate the new thread count (control setting + square wave). */
1240         new_thread_count = (gint)(hc->current_control_setting + new_thread_wave_magnitude * ((hc->total_samples / (hc->wave_period / 2)) % 2));
1241
1242         /* Make sure the new thread count doesn't exceed the ThreadPool's limits. */
1243         new_thread_count = CLAMP (new_thread_count, threadpool->limit_worker_min, threadpool->limit_worker_max);
1244
1245         if (new_thread_count != current_thread_count)
1246                 hill_climbing_change_thread_count (new_thread_count, transition);
1247
1248         if (creal (ratio) < 0.0 && new_thread_count == threadpool->limit_worker_min)
1249                 *adjustment_interval = (gint)(0.5 + hc->current_sample_interval * (10.0 * MAX (-1.0 * creal (ratio), 1.0)));
1250         else
1251                 *adjustment_interval = hc->current_sample_interval;
1252
1253         return new_thread_count;
1254 }
1255
1256 static void
1257 heuristic_notify_work_completed (void)
1258 {
1259         g_assert (threadpool);
1260
1261         InterlockedIncrement (&threadpool->heuristic_completions);
1262         threadpool->heuristic_last_dequeue = mono_msec_ticks ();
1263 }
1264
1265 static gboolean
1266 heuristic_should_adjust (void)
1267 {
1268         g_assert (threadpool);
1269
1270         if (threadpool->heuristic_last_dequeue > threadpool->heuristic_last_adjustment + threadpool->heuristic_adjustment_interval) {
1271                 ThreadPoolCounter counter;
1272                 counter.as_gint64 = COUNTER_READ();
1273                 if (counter._.working <= counter._.max_working)
1274                         return TRUE;
1275         }
1276
1277         return FALSE;
1278 }
1279
1280 static void
1281 heuristic_adjust (void)
1282 {
1283         g_assert (threadpool);
1284
1285         if (mono_coop_mutex_trylock (&threadpool->heuristic_lock) == 0) {
1286                 gint32 completions = InterlockedExchange (&threadpool->heuristic_completions, 0);
1287                 guint32 sample_end = mono_msec_ticks ();
1288                 guint32 sample_duration = sample_end - threadpool->heuristic_sample_start;
1289
1290                 if (sample_duration >= threadpool->heuristic_adjustment_interval / 2) {
1291                         ThreadPoolCounter counter;
1292                         gint16 new_thread_count;
1293
1294                         counter.as_gint64 = COUNTER_READ ();
1295                         new_thread_count = hill_climbing_update (counter._.max_working, sample_duration, completions, &threadpool->heuristic_adjustment_interval);
1296
1297                         COUNTER_ATOMIC (counter, { counter._.max_working = new_thread_count; });
1298
1299                         if (new_thread_count > counter._.max_working)
1300                                 worker_request (mono_domain_get ());
1301
1302                         threadpool->heuristic_sample_start = sample_end;
1303                         threadpool->heuristic_last_adjustment = mono_msec_ticks ();
1304                 }
1305
1306                 mono_coop_mutex_unlock (&threadpool->heuristic_lock);
1307         }
1308 }
1309
1310 void
1311 mono_threadpool_ms_cleanup (void)
1312 {
1313         #ifndef DISABLE_SOCKETS
1314                 mono_threadpool_ms_io_cleanup ();
1315         #endif
1316         mono_lazy_cleanup (&status, cleanup);
1317 }
1318
1319 MonoAsyncResult *
1320 mono_threadpool_ms_begin_invoke (MonoDomain *domain, MonoObject *target, MonoMethod *method, gpointer *params)
1321 {
1322         static MonoClass *async_call_klass = NULL;
1323         MonoMethodMessage *message;
1324         MonoAsyncResult *async_result;
1325         MonoAsyncCall *async_call;
1326         MonoDelegate *async_callback = NULL;
1327         MonoObject *state = NULL;
1328
1329         if (!async_call_klass)
1330                 async_call_klass = mono_class_from_name (mono_defaults.corlib, "System", "MonoAsyncCall");
1331         g_assert (async_call_klass);
1332
1333         mono_lazy_initialize (&status, initialize);
1334
1335         message = mono_method_call_message_new (method, params, mono_get_delegate_invoke (method->klass), (params != NULL) ? (&async_callback) : NULL, (params != NULL) ? (&state) : NULL);
1336
1337         async_call = (MonoAsyncCall*) mono_object_new (domain, async_call_klass);
1338         MONO_OBJECT_SETREF (async_call, msg, message);
1339         MONO_OBJECT_SETREF (async_call, state, state);
1340
1341         if (async_callback) {
1342                 MONO_OBJECT_SETREF (async_call, cb_method, mono_get_delegate_invoke (((MonoObject*) async_callback)->vtable->klass));
1343                 MONO_OBJECT_SETREF (async_call, cb_target, async_callback);
1344         }
1345
1346         async_result = mono_async_result_new (domain, NULL, async_call->state, NULL, (MonoObject*) async_call);
1347         MONO_OBJECT_SETREF (async_result, async_delegate, target);
1348
1349         mono_threadpool_ms_enqueue_work_item (domain, (MonoObject*) async_result);
1350
1351         return async_result;
1352 }
1353
1354 MonoObject *
1355 mono_threadpool_ms_end_invoke (MonoAsyncResult *ares, MonoArray **out_args, MonoObject **exc)
1356 {
1357         MonoAsyncCall *ac;
1358
1359         g_assert (exc);
1360         g_assert (out_args);
1361
1362         *exc = NULL;
1363         *out_args = NULL;
1364
1365         /* check if already finished */
1366         mono_monitor_enter ((MonoObject*) ares);
1367
1368         if (ares->endinvoke_called) {
1369                 *exc = (MonoObject*) mono_get_exception_invalid_operation (NULL);
1370                 mono_monitor_exit ((MonoObject*) ares);
1371                 return NULL;
1372         }
1373
1374         ares->endinvoke_called = 1;
1375
1376         /* wait until we are really finished */
1377         if (ares->completed) {
1378                 mono_monitor_exit ((MonoObject *) ares);
1379         } else {
1380                 gpointer wait_event;
1381                 if (ares->handle) {
1382                         wait_event = mono_wait_handle_get_handle ((MonoWaitHandle*) ares->handle);
1383                 } else {
1384                         wait_event = CreateEvent (NULL, TRUE, FALSE, NULL);
1385                         g_assert(wait_event);
1386                         MONO_OBJECT_SETREF (ares, handle, (MonoObject*) mono_wait_handle_new (mono_object_domain (ares), wait_event));
1387                 }
1388                 mono_monitor_exit ((MonoObject*) ares);
1389                 MONO_PREPARE_BLOCKING;
1390                 WaitForSingleObjectEx (wait_event, INFINITE, TRUE);
1391                 MONO_FINISH_BLOCKING;
1392         }
1393
1394         ac = (MonoAsyncCall*) ares->object_data;
1395         g_assert (ac);
1396
1397         *exc = ac->msg->exc; /* FIXME: GC add write barrier */
1398         *out_args = ac->out_args;
1399         return ac->res;
1400 }
1401
1402 gboolean
1403 mono_threadpool_ms_remove_domain_jobs (MonoDomain *domain, int timeout)
1404 {
1405         gboolean res = TRUE;
1406         guint32 start;
1407         gpointer sem;
1408
1409         g_assert (domain);
1410         g_assert (timeout >= -1);
1411
1412         g_assert (mono_domain_is_unloading (domain));
1413
1414         if (timeout != -1)
1415                 start = mono_msec_ticks ();
1416
1417 #ifndef DISABLE_SOCKETS
1418         mono_threadpool_ms_io_remove_domain_jobs (domain);
1419         if (timeout != -1) {
1420                 timeout -= mono_msec_ticks () - start;
1421                 if (timeout < 0)
1422                         return FALSE;
1423         }
1424 #endif
1425
1426         /*
1427          * There might be some threads out that could be about to execute stuff from the given domain.
1428          * We avoid that by setting up a semaphore to be pulsed by the thread that reaches zero.
1429          */
1430         sem = domain->cleanup_semaphore = CreateSemaphore (NULL, 0, 1, NULL);
1431
1432         /*
1433          * The memory barrier here is required to have global ordering between assigning to cleanup_semaphone
1434          * and reading threadpool_jobs. Otherwise this thread could read a stale version of threadpool_jobs
1435          * and wait forever.
1436          */
1437         mono_memory_write_barrier ();
1438
1439         while (domain->threadpool_jobs) {
1440                 MONO_PREPARE_BLOCKING;
1441                 WaitForSingleObject (sem, timeout);
1442                 MONO_FINISH_BLOCKING;
1443                 if (timeout != -1) {
1444                         timeout -= mono_msec_ticks () - start;
1445                         if (timeout <= 0) {
1446                                 res = FALSE;
1447                                 break;
1448                         }
1449                 }
1450         }
1451
1452         domain->cleanup_semaphore = NULL;
1453         CloseHandle (sem);
1454
1455         return res;
1456 }
1457
1458 void
1459 mono_threadpool_ms_suspend (void)
1460 {
1461         if (threadpool)
1462                 threadpool->suspended = TRUE;
1463 }
1464
1465 void
1466 mono_threadpool_ms_resume (void)
1467 {
1468         if (threadpool)
1469                 threadpool->suspended = FALSE;
1470 }
1471
1472 void
1473 ves_icall_System_Threading_ThreadPool_GetAvailableThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1474 {
1475         ThreadPoolCounter counter;
1476
1477         if (!worker_threads || !completion_port_threads)
1478                 return;
1479
1480         mono_lazy_initialize (&status, initialize);
1481
1482         counter.as_gint64 = COUNTER_READ ();
1483
1484         *worker_threads = MAX (0, threadpool->limit_worker_max - counter._.active);
1485         *completion_port_threads = threadpool->limit_io_max;
1486 }
1487
1488 void
1489 ves_icall_System_Threading_ThreadPool_GetMinThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1490 {
1491         if (!worker_threads || !completion_port_threads)
1492                 return;
1493
1494         mono_lazy_initialize (&status, initialize);
1495
1496         *worker_threads = threadpool->limit_worker_min;
1497         *completion_port_threads = threadpool->limit_io_min;
1498 }
1499
1500 void
1501 ves_icall_System_Threading_ThreadPool_GetMaxThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1502 {
1503         if (!worker_threads || !completion_port_threads)
1504                 return;
1505
1506         mono_lazy_initialize (&status, initialize);
1507
1508         *worker_threads = threadpool->limit_worker_max;
1509         *completion_port_threads = threadpool->limit_io_max;
1510 }
1511
1512 MonoBoolean
1513 ves_icall_System_Threading_ThreadPool_SetMinThreadsNative (gint32 worker_threads, gint32 completion_port_threads)
1514 {
1515         mono_lazy_initialize (&status, initialize);
1516
1517         if (worker_threads <= 0 || worker_threads > threadpool->limit_worker_max)
1518                 return FALSE;
1519         if (completion_port_threads <= 0 || completion_port_threads > threadpool->limit_io_max)
1520                 return FALSE;
1521
1522         threadpool->limit_worker_min = worker_threads;
1523         threadpool->limit_io_min = completion_port_threads;
1524
1525         return TRUE;
1526 }
1527
1528 MonoBoolean
1529 ves_icall_System_Threading_ThreadPool_SetMaxThreadsNative (gint32 worker_threads, gint32 completion_port_threads)
1530 {
1531         gint cpu_count = mono_cpu_count ();
1532
1533         mono_lazy_initialize (&status, initialize);
1534
1535         if (worker_threads < threadpool->limit_worker_min || worker_threads < cpu_count)
1536                 return FALSE;
1537         if (completion_port_threads < threadpool->limit_io_min || completion_port_threads < cpu_count)
1538                 return FALSE;
1539
1540         threadpool->limit_worker_max = worker_threads;
1541         threadpool->limit_io_max = completion_port_threads;
1542
1543         return TRUE;
1544 }
1545
1546 void
1547 ves_icall_System_Threading_ThreadPool_InitializeVMTp (MonoBoolean *enable_worker_tracking)
1548 {
1549         if (enable_worker_tracking) {
1550                 // TODO implement some kind of switch to have the possibily to use it
1551                 *enable_worker_tracking = FALSE;
1552         }
1553
1554         mono_lazy_initialize (&status, initialize);
1555 }
1556
1557 MonoBoolean
1558 ves_icall_System_Threading_ThreadPool_NotifyWorkItemComplete (void)
1559 {
1560         ThreadPoolCounter counter;
1561
1562         if (mono_domain_is_unloading (mono_domain_get ()) || mono_runtime_is_shutting_down ())
1563                 return FALSE;
1564
1565         heuristic_notify_work_completed ();
1566
1567         if (heuristic_should_adjust ())
1568                 heuristic_adjust ();
1569
1570         counter.as_gint64 = COUNTER_READ ();
1571         return counter._.working <= counter._.max_working;
1572 }
1573
1574 void
1575 ves_icall_System_Threading_ThreadPool_NotifyWorkItemProgressNative (void)
1576 {
1577         heuristic_notify_work_completed ();
1578
1579         if (heuristic_should_adjust ())
1580                 heuristic_adjust ();
1581 }
1582
1583 void
1584 ves_icall_System_Threading_ThreadPool_ReportThreadStatus (MonoBoolean is_working)
1585 {
1586         // TODO
1587         mono_raise_exception (mono_get_exception_not_implemented (NULL));
1588 }
1589
1590 MonoBoolean
1591 ves_icall_System_Threading_ThreadPool_RequestWorkerThread (void)
1592 {
1593         return worker_request (mono_domain_get ());
1594 }
1595
1596 MonoBoolean G_GNUC_UNUSED
1597 ves_icall_System_Threading_ThreadPool_PostQueuedCompletionStatus (MonoNativeOverlapped *native_overlapped)
1598 {
1599         /* This copy the behavior of the current Mono implementation */
1600         mono_raise_exception (mono_get_exception_not_implemented (NULL));
1601         return FALSE;
1602 }
1603
1604 MonoBoolean G_GNUC_UNUSED
1605 ves_icall_System_Threading_ThreadPool_BindIOCompletionCallbackNative (gpointer file_handle)
1606 {
1607         /* This copy the behavior of the current Mono implementation */
1608         return TRUE;
1609 }
1610
1611 MonoBoolean G_GNUC_UNUSED
1612 ves_icall_System_Threading_ThreadPool_IsThreadPoolHosted (void)
1613 {
1614         return FALSE;
1615 }