mono/metadata/threadpool-ms.c

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