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