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