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  * Licensed under the MIT license. See LICENSE file in the project root for full license information.
   9  */
  10
  11 //
  12 // Copyright (c) Microsoft. All rights reserved.
  13 // Licensed under the MIT license. See LICENSE file in the project root for full license information.
  14 //
  15 // Files:
  16 //  - src/vm/comthreadpool.cpp
  17 //  - src/vm/win32threadpoolcpp
  18 //  - src/vm/threadpoolrequest.cpp
  19 //  - src/vm/hillclimbing.cpp
  20 //
  21 // Ported from C++ to C and adjusted to Mono runtime
  22
  23 #include <stdlib.h>
  24 #define _USE_MATH_DEFINES // needed by MSVC to define math constants
  25 #include <math.h>
  26 #include <config.h>
  27 #include <glib.h>
  28
  29 #include <mono/metadata/class-internals.h>
  30 #include <mono/metadata/exception.h>
  31 #include <mono/metadata/gc-internals.h>
  32 #include <mono/metadata/object.h>
  33 #include <mono/metadata/object-internals.h>
  34 #include <mono/metadata/threadpool-ms.h>
  35 #include <mono/metadata/threadpool-ms-io.h>
  36 #include <mono/metadata/w32event.h>
  37 #include <mono/utils/atomic.h>
  38 #include <mono/utils/mono-compiler.h>
  39 #include <mono/utils/mono-complex.h>
  40 #include <mono/utils/mono-lazy-init.h>
  41 #include <mono/utils/mono-logger.h>
  42 #include <mono/utils/mono-logger-internals.h>
  43 #include <mono/utils/mono-proclib.h>
  44 #include <mono/utils/mono-threads.h>
  45 #include <mono/utils/mono-time.h>
  46 #include <mono/utils/mono-rand.h>
  47
  48 #define CPU_USAGE_LOW 80
  49 #define CPU_USAGE_HIGH 95
  50
  51 #define MONITOR_INTERVAL 500 // ms
  52 #define MONITOR_MINIMAL_LIFETIME 60 * 1000 // ms
  53
  54 #define WORKER_CREATION_MAX_PER_SEC 10
  55
  56 /* The exponent to apply to the gain. 1.0 means to use linear gain,
  57  * higher values will enhance large moves and damp small ones.
  58  * default: 2.0 */
  59 #define HILL_CLIMBING_GAIN_EXPONENT 2.0
  60
  61 /* The 'cost' of a thread. 0 means drive for increased throughput regardless
  62  * of thread count, higher values bias more against higher thread counts.
  63  * default: 0.15 */
  64 #define HILL_CLIMBING_BIAS 0.15
  65
  66 #define HILL_CLIMBING_WAVE_PERIOD 4
  67 #define HILL_CLIMBING_MAX_WAVE_MAGNITUDE 20
  68 #define HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER 1.0
  69 #define HILL_CLIMBING_WAVE_HISTORY_SIZE 8
  70 #define HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO 3.0
  71 #define HILL_CLIMBING_MAX_CHANGE_PER_SECOND 4
  72 #define HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE 20
  73 #define HILL_CLIMBING_SAMPLE_INTERVAL_LOW 10
  74 #define HILL_CLIMBING_SAMPLE_INTERVAL_HIGH 200
  75 #define HILL_CLIMBING_ERROR_SMOOTHING_FACTOR 0.01
  76 #define HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT 0.15
  77
  78 typedef union {
  79         struct {
  80                 gint16 max_working; /* determined by heuristic */
  81                 gint16 active; /* executing worker_thread */
  82                 gint16 working; /* actively executing worker_thread, not parked */
  83                 gint16 parked; /* parked */
  84         } _;
  85         gint64 as_gint64;
  86 } ThreadPoolCounter;
  87
  88 typedef struct {
  89         MonoDomain *domain;
  90         /* Number of outstanding jobs */
  91         gint32 outstanding_request;
  92         /* Number of currently executing jobs */
  93         int     threadpool_jobs;
  94         /* Signalled when threadpool_jobs + outstanding_request is 0 */
  95         /* Protected by threadpool->domains_lock */
  96         MonoCoopCond cleanup_cond;
  97 } ThreadPoolDomain;
  98
  99 typedef MonoInternalThread ThreadPoolWorkingThread;
 100
 101 typedef struct {
 102         gint32 wave_period;
 103         gint32 samples_to_measure;
 104         gdouble target_throughput_ratio;
 105         gdouble target_signal_to_noise_ratio;
 106         gdouble max_change_per_second;
 107         gdouble max_change_per_sample;
 108         gint32 max_thread_wave_magnitude;
 109         gint32 sample_interval_low;
 110         gdouble thread_magnitude_multiplier;
 111         gint32 sample_interval_high;
 112         gdouble throughput_error_smoothing_factor;
 113         gdouble gain_exponent;
 114         gdouble max_sample_error;
 115
 116         gdouble current_control_setting;
 117         gint64 total_samples;
 118         gint16 last_thread_count;
 119         gdouble elapsed_since_last_change;
 120         gdouble completions_since_last_change;
 121
 122         gdouble average_throughput_noise;
 123
 124         gdouble *samples;
 125         gdouble *thread_counts;
 126
 127         guint32 current_sample_interval;
 128         gpointer random_interval_generator;
 129
 130         gint32 accumulated_completion_count;
 131         gdouble accumulated_sample_duration;
 132 } ThreadPoolHillClimbing;
 133
 134 typedef struct {
 135         ThreadPoolCounter counters;
 136
 137         GPtrArray *domains; // ThreadPoolDomain* []
 138         MonoCoopMutex domains_lock;
 139
 140         GPtrArray *working_threads; // ThreadPoolWorkingThread* []
 141         gint32 parked_threads_count;
 142         MonoCoopCond parked_threads_cond;
 143         MonoCoopMutex active_threads_lock; /* protect access to working_threads and parked_threads */
 144
 145         guint32 worker_creation_current_second;
 146         guint32 worker_creation_current_count;
 147         MonoCoopMutex worker_creation_lock;
 148
 149         gint32 heuristic_completions;
 150         gint64 heuristic_sample_start;
 151         gint64 heuristic_last_dequeue; // ms
 152         gint64 heuristic_last_adjustment; // ms
 153         gint64 heuristic_adjustment_interval; // ms
 154         ThreadPoolHillClimbing heuristic_hill_climbing;
 155         MonoCoopMutex heuristic_lock;
 156
 157         gint32 limit_worker_min;
 158         gint32 limit_worker_max;
 159         gint32 limit_io_min;
 160         gint32 limit_io_max;
 161
 162         MonoCpuUsageState *cpu_usage_state;
 163         gint32 cpu_usage;
 164
 165         /* suspended by the debugger */
 166         gboolean suspended;
 167 } ThreadPool;
 168
 169 typedef enum {
 170         TRANSITION_WARMUP,
 171         TRANSITION_INITIALIZING,
 172         TRANSITION_RANDOM_MOVE,
 173         TRANSITION_CLIMBING_MOVE,
 174         TRANSITION_CHANGE_POINT,
 175         TRANSITION_STABILIZING,
 176         TRANSITION_STARVATION,
 177         TRANSITION_THREAD_TIMED_OUT,
 178         TRANSITION_UNDEFINED,
 179 } ThreadPoolHeuristicStateTransition;
 180
 181 static mono_lazy_init_t status = MONO_LAZY_INIT_STATUS_NOT_INITIALIZED;
 182
 183 enum {
 184         MONITOR_STATUS_REQUESTED,
 185         MONITOR_STATUS_WAITING_FOR_REQUEST,
 186         MONITOR_STATUS_NOT_RUNNING,
 187 };
 188
 189 static gint32 monitor_status = MONITOR_STATUS_NOT_RUNNING;
 190
 191 static ThreadPool* threadpool;
 192
 193 #define COUNTER_CHECK(counter) \
 194         do { \
 195                 g_assert (counter._.max_working > 0); \
 196                 g_assert (counter._.working >= 0); \
 197                 g_assert (counter._.active >= 0); \
 198         } while (0)
 199
 200 #define COUNTER_READ() (InterlockedRead64 (&threadpool->counters.as_gint64))
 201
 202 #define COUNTER_ATOMIC(var,block) \
 203         do { \
 204                 ThreadPoolCounter __old; \
 205                 do { \
 206                         g_assert (threadpool); \
 207                         __old.as_gint64 = COUNTER_READ (); \
 208                         (var) = __old; \
 209                         { block; } \
 210                         COUNTER_CHECK (var); \
 211                 } while (InterlockedCompareExchange64 (&threadpool->counters.as_gint64, (var).as_gint64, __old.as_gint64) != __old.as_gint64); \
 212         } while (0)
 213
 214 #define COUNTER_TRY_ATOMIC(res,var,block) \
 215         do { \
 216                 ThreadPoolCounter __old; \
 217                 do { \
 218                         g_assert (threadpool); \
 219                         __old.as_gint64 = COUNTER_READ (); \
 220                         (var) = __old; \
 221                         (res) = FALSE; \
 222                         { block; } \
 223                         COUNTER_CHECK (var); \
 224                         (res) = InterlockedCompareExchange64 (&threadpool->counters.as_gint64, (var).as_gint64, __old.as_gint64) == __old.as_gint64; \
 225                 } while (0); \
 226         } while (0)
 227
 228 static inline void
 229 domains_lock (void)
 230 {
 231         mono_coop_mutex_lock (&threadpool->domains_lock);
 232 }
 233
 234 static inline void
 235 domains_unlock (void)
 236 {
 237         mono_coop_mutex_unlock (&threadpool->domains_lock);
 238 }
 239
 240 static gpointer
 241 rand_create (void)
 242 {
 243         mono_rand_open ();
 244         return mono_rand_init (NULL, 0);
 245 }
 246
 247 static guint32
 248 rand_next (gpointer *handle, guint32 min, guint32 max)
 249 {
 250         MonoError error;
 251         guint32 val;
 252         mono_rand_try_get_uint32 (handle, &val, min, max, &error);
 253         // FIXME handle error
 254         mono_error_assert_ok (&error);
 255         return val;
 256 }
 257
 258 static void
 259 rand_free (gpointer handle)
 260 {
 261         mono_rand_close (handle);
 262 }
 263
 264 static void
 265 initialize (void)
 266 {
 267         ThreadPoolHillClimbing *hc;
 268         const char *threads_per_cpu_env;
 269         gint threads_per_cpu;
 270         gint threads_count;
 271
 272         g_assert (!threadpool);
 273         threadpool = g_new0 (ThreadPool, 1);
 274         g_assert (threadpool);
 275
 276         threadpool->domains = g_ptr_array_new ();
 277         mono_coop_mutex_init (&threadpool->domains_lock);
 278
 279         threadpool->parked_threads_count = 0;
 280         mono_coop_cond_init (&threadpool->parked_threads_cond);
 281         threadpool->working_threads = g_ptr_array_new ();
 282         mono_coop_mutex_init (&threadpool->active_threads_lock);
 283
 284         threadpool->worker_creation_current_second = -1;
 285         mono_coop_mutex_init (&threadpool->worker_creation_lock);
 286
 287         threadpool->heuristic_adjustment_interval = 10;
 288         mono_coop_mutex_init (&threadpool->heuristic_lock);
 289
 290         mono_rand_open ();
 291
 292         hc = &threadpool->heuristic_hill_climbing;
 293
 294         hc->wave_period = HILL_CLIMBING_WAVE_PERIOD;
 295         hc->max_thread_wave_magnitude = HILL_CLIMBING_MAX_WAVE_MAGNITUDE;
 296         hc->thread_magnitude_multiplier = (gdouble) HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER;
 297         hc->samples_to_measure = hc->wave_period * HILL_CLIMBING_WAVE_HISTORY_SIZE;
 298         hc->target_throughput_ratio = (gdouble) HILL_CLIMBING_BIAS;
 299         hc->target_signal_to_noise_ratio = (gdouble) HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO;
 300         hc->max_change_per_second = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SECOND;
 301         hc->max_change_per_sample = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE;
 302         hc->sample_interval_low = HILL_CLIMBING_SAMPLE_INTERVAL_LOW;
 303         hc->sample_interval_high = HILL_CLIMBING_SAMPLE_INTERVAL_HIGH;
 304         hc->throughput_error_smoothing_factor = (gdouble) HILL_CLIMBING_ERROR_SMOOTHING_FACTOR;
 305         hc->gain_exponent = (gdouble) HILL_CLIMBING_GAIN_EXPONENT;
 306         hc->max_sample_error = (gdouble) HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT;
 307         hc->current_control_setting = 0;
 308         hc->total_samples = 0;
 309         hc->last_thread_count = 0;
 310         hc->average_throughput_noise = 0;
 311         hc->elapsed_since_last_change = 0;
 312         hc->accumulated_completion_count = 0;
 313         hc->accumulated_sample_duration = 0;
 314         hc->samples = g_new0 (gdouble, hc->samples_to_measure);
 315         hc->thread_counts = g_new0 (gdouble, hc->samples_to_measure);
 316         hc->random_interval_generator = rand_create ();
 317         hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
 318
 319         if (!(threads_per_cpu_env = g_getenv ("MONO_THREADS_PER_CPU")))
 320                 threads_per_cpu = 1;
 321         else
 322                 threads_per_cpu = CLAMP (atoi (threads_per_cpu_env), 1, 50);
 323
 324         threads_count = mono_cpu_count () * threads_per_cpu;
 325
 326         threadpool->limit_worker_min = threadpool->limit_io_min = threads_count;
 327
 328 #if defined (PLATFORM_ANDROID) || defined (HOST_IOS)
 329         threadpool->limit_worker_max = threadpool->limit_io_max = CLAMP (threads_count * 100, MIN (threads_count, 200), MAX (threads_count, 200));
 330 #else
 331         threadpool->limit_worker_max = threadpool->limit_io_max = threads_count * 100;
 332 #endif
 333
 334         threadpool->counters._.max_working = threadpool->limit_worker_min;
 335
 336         threadpool->cpu_usage_state = g_new0 (MonoCpuUsageState, 1);
 337
 338         threadpool->suspended = FALSE;
 339 }
 340
 341 static void worker_kill (ThreadPoolWorkingThread *thread);
 342
 343 static void
 344 cleanup (void)
 345 {
 346         guint i;
 347
 348         /* we make the assumption along the code that we are
 349          * cleaning up only if the runtime is shutting down */
 350         g_assert (mono_runtime_is_shutting_down ());
 351
 352         while (monitor_status != MONITOR_STATUS_NOT_RUNNING)
 353                 mono_thread_info_sleep (1, NULL);
 354
 355         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 356
 357         /* stop all threadpool->working_threads */
 358         for (i = 0; i < threadpool->working_threads->len; ++i)
 359                 worker_kill ((ThreadPoolWorkingThread*) g_ptr_array_index (threadpool->working_threads, i));
 360
 361         /* unpark all threadpool->parked_threads */
 362         mono_coop_cond_broadcast (&threadpool->parked_threads_cond);
 363
 364         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 365 }
 366
 367 gboolean
 368 mono_threadpool_ms_enqueue_work_item (MonoDomain *domain, MonoObject *work_item, MonoError *error)
 369 {
 370         static MonoClass *threadpool_class = NULL;
 371         static MonoMethod *unsafe_queue_custom_work_item_method = NULL;
 372         MonoDomain *current_domain;
 373         MonoBoolean f;
 374         gpointer args [2];
 375
 376         mono_error_init (error);
 377         g_assert (work_item);
 378
 379         if (!threadpool_class)
 380                 threadpool_class = mono_class_load_from_name (mono_defaults.corlib, "System.Threading", "ThreadPool");
 381
 382         if (!unsafe_queue_custom_work_item_method)
 383                 unsafe_queue_custom_work_item_method = mono_class_get_method_from_name (threadpool_class, "UnsafeQueueCustomWorkItem", 2);
 384         g_assert (unsafe_queue_custom_work_item_method);
 385
 386         f = FALSE;
 387
 388         args [0] = (gpointer) work_item;
 389         args [1] = (gpointer) &f;
 390
 391         current_domain = mono_domain_get ();
 392         if (current_domain == domain) {
 393                 mono_runtime_invoke_checked (unsafe_queue_custom_work_item_method, NULL, args, error);
 394                 return_val_if_nok (error, FALSE);
 395         } else {
 396                 mono_thread_push_appdomain_ref (domain);
 397                 if (mono_domain_set (domain, FALSE)) {
 398                         mono_runtime_invoke_checked (unsafe_queue_custom_work_item_method, NULL, args, error);
 399                         if (!is_ok (error)) {
 400                                 mono_thread_pop_appdomain_ref ();
 401                                 return FALSE;
 402                         }
 403                         mono_domain_set (current_domain, TRUE);
 404                 }
 405                 mono_thread_pop_appdomain_ref ();
 406         }
 407         return TRUE;
 408 }
 409
 410 /* LOCKING: domains_lock must be held */
 411 static void
 412 tpdomain_add (ThreadPoolDomain *tpdomain)
 413 {
 414         guint i, len;
 415
 416         g_assert (tpdomain);
 417
 418         len = threadpool->domains->len;
 419         for (i = 0; i < len; ++i) {
 420                 if (g_ptr_array_index (threadpool->domains, i) == tpdomain)
 421                         break;
 422         }
 423
 424         if (i == len)
 425                 g_ptr_array_add (threadpool->domains, tpdomain);
 426 }
 427
 428 /* LOCKING: domains_lock must be held. */
 429 static gboolean
 430 tpdomain_remove (ThreadPoolDomain *tpdomain)
 431 {
 432         g_assert (tpdomain);
 433         return g_ptr_array_remove (threadpool->domains, tpdomain);
 434 }
 435
 436 /* LOCKING: domains_lock must be held */
 437 static ThreadPoolDomain *
 438 tpdomain_get (MonoDomain *domain, gboolean create)
 439 {
 440         guint i;
 441         ThreadPoolDomain *tpdomain;
 442
 443         g_assert (domain);
 444
 445         for (i = 0; i < threadpool->domains->len; ++i) {
 446                 ThreadPoolDomain *tpdomain;
 447
 448                 tpdomain = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i);
 449                 if (tpdomain->domain == domain)
 450                         return tpdomain;
 451         }
 452
 453         if (!create)
 454                 return NULL;
 455
 456         tpdomain = g_new0 (ThreadPoolDomain, 1);
 457         tpdomain->domain = domain;
 458         mono_coop_cond_init (&tpdomain->cleanup_cond);
 459
 460         tpdomain_add (tpdomain);
 461
 462         return tpdomain;
 463 }
 464
 465 static void
 466 tpdomain_free (ThreadPoolDomain *tpdomain)
 467 {
 468         g_free (tpdomain);
 469 }
 470
 471 /* LOCKING: domains_lock must be held */
 472 static gboolean
 473 domain_any_has_request (void)
 474 {
 475         guint i;
 476
 477         for (i = 0; i < threadpool->domains->len; ++i) {
 478                 ThreadPoolDomain *tmp = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i);
 479                 if (tmp->outstanding_request > 0)
 480                         return TRUE;
 481         }
 482
 483         return FALSE;
 484 }
 485
 486 /* LOCKING: domains_lock must be held */
 487 static ThreadPoolDomain *
 488 tpdomain_get_next (ThreadPoolDomain *current)
 489 {
 490         ThreadPoolDomain *tpdomain = NULL;
 491         guint len;
 492
 493         len = threadpool->domains->len;
 494         if (len > 0) {
 495                 guint i, current_idx = -1;
 496                 if (current) {
 497                         for (i = 0; i < len; ++i) {
 498                                 if (current == g_ptr_array_index (threadpool->domains, i)) {
 499                                         current_idx = i;
 500                                         break;
 501                                 }
 502                         }
 503                         g_assert (current_idx != (guint)-1);
 504                 }
 505                 for (i = current_idx + 1; i < len + current_idx + 1; ++i) {
 506                         ThreadPoolDomain *tmp = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i % len);
 507                         if (tmp->outstanding_request > 0) {
 508                                 tpdomain = tmp;
 509                                 break;
 510                         }
 511                 }
 512         }
 513
 514         return tpdomain;
 515 }
 516
 517 static void
 518 worker_wait_interrupt (gpointer data)
 519 {
 520         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 521         mono_coop_cond_signal (&threadpool->parked_threads_cond);
 522         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 523 }
 524
 525 /* return TRUE if timeout, FALSE otherwise (worker unpark or interrupt) */
 526 static gboolean
 527 worker_park (void)
 528 {
 529         gboolean timeout = FALSE;
 530
 531         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] current worker parking", mono_native_thread_id_get ());
 532
 533         mono_gc_set_skip_thread (TRUE);
 534
 535         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 536
 537         if (!mono_runtime_is_shutting_down ()) {
 538                 static gpointer rand_handle = NULL;
 539                 MonoInternalThread *thread_internal;
 540                 gboolean interrupted = FALSE;
 541
 542                 if (!rand_handle)
 543                         rand_handle = rand_create ();
 544                 g_assert (rand_handle);
 545
 546                 thread_internal = mono_thread_internal_current ();
 547                 g_assert (thread_internal);
 548
 549                 threadpool->parked_threads_count += 1;
 550                 g_ptr_array_remove_fast (threadpool->working_threads, thread_internal);
 551
 552                 mono_thread_info_install_interrupt (worker_wait_interrupt, NULL, &interrupted);
 553                 if (interrupted)
 554                         goto done;
 555
 556                 if (mono_coop_cond_timedwait (&threadpool->parked_threads_cond, &threadpool->active_threads_lock, rand_next (&rand_handle, 5 * 1000, 60 * 1000)) != 0)
 557                         timeout = TRUE;
 558
 559                 mono_thread_info_uninstall_interrupt (&interrupted);
 560
 561 done:
 562                 g_ptr_array_add (threadpool->working_threads, thread_internal);
 563                 threadpool->parked_threads_count -= 1;
 564         }
 565
 566         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 567
 568         mono_gc_set_skip_thread (FALSE);
 569
 570         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] current worker unparking, timeout? %s", mono_native_thread_id_get (), timeout ? "yes" : "no");
 571
 572         return timeout;
 573 }
 574
 575 static gboolean
 576 worker_try_unpark (void)
 577 {
 578         gboolean res = FALSE;
 579
 580         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker", mono_native_thread_id_get ());
 581
 582         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 583         if (threadpool->parked_threads_count > 0) {
 584                 mono_coop_cond_signal (&threadpool->parked_threads_cond);
 585                 res = TRUE;
 586         }
 587         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 588
 589         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker, success? %s", mono_native_thread_id_get (), res ? "yes" : "no");
 590
 591         return res;
 592 }
 593
 594 static void
 595 worker_kill (ThreadPoolWorkingThread *thread)
 596 {
 597         if (thread == mono_thread_internal_current ())
 598                 return;
 599
 600         mono_thread_internal_stop ((MonoInternalThread*) thread);
 601 }
 602
 603 static void
 604 worker_thread (gpointer data)
 605 {
 606         MonoError error;
 607         MonoInternalThread *thread;
 608         ThreadPoolDomain *tpdomain, *previous_tpdomain;
 609         ThreadPoolCounter counter;
 610         gboolean retire = FALSE;
 611
 612         mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker starting", mono_native_thread_id_get ());
 613
 614         g_assert (threadpool);
 615
 616         thread = mono_thread_internal_current ();
 617         g_assert (thread);
 618
 619         mono_thread_set_name_internal (thread, mono_string_new (mono_get_root_domain (), "Threadpool worker"), FALSE, &error);
 620         mono_error_assert_ok (&error);
 621
 622         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 623         g_ptr_array_add (threadpool->working_threads, thread);
 624         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 625
 626         previous_tpdomain = NULL;
 627
 628         domains_lock ();
 629
 630         while (!mono_runtime_is_shutting_down ()) {
 631                 tpdomain = NULL;
 632
 633                 if ((thread->state & (ThreadState_StopRequested | ThreadState_SuspendRequested)) != 0) {
 634                         domains_unlock ();
 635                         mono_thread_interruption_checkpoint ();
 636                         domains_lock ();
 637                 }
 638
 639                 if (retire || !(tpdomain = tpdomain_get_next (previous_tpdomain))) {
 640                         gboolean timeout;
 641
 642                         COUNTER_ATOMIC (counter, {
 643                                 counter._.working --;
 644                                 counter._.parked ++;
 645                         });
 646
 647                         domains_unlock ();
 648                         timeout = worker_park ();
 649                         domains_lock ();
 650
 651                         COUNTER_ATOMIC (counter, {
 652                                 counter._.working ++;
 653                                 counter._.parked --;
 654                         });
 655
 656                         if (timeout)
 657                                 break;
 658
 659                         if (retire)
 660                                 retire = FALSE;
 661
 662                         /* The tpdomain->domain might have unloaded, while this thread was parked */
 663                         previous_tpdomain = NULL;
 664
 665                         continue;
 666                 }
 667
 668                 tpdomain->outstanding_request --;
 669                 g_assert (tpdomain->outstanding_request >= 0);
 670
 671                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker running in domain %p (outstanding requests %d) ",
 672                         mono_native_thread_id_get (), tpdomain->domain, tpdomain->outstanding_request);
 673
 674                 g_assert (tpdomain->domain);
 675                 g_assert (tpdomain->threadpool_jobs >= 0);
 676                 tpdomain->threadpool_jobs ++;
 677
 678                 domains_unlock ();
 679
 680                 mono_thread_push_appdomain_ref (tpdomain->domain);
 681                 if (mono_domain_set (tpdomain->domain, FALSE)) {
 682                         MonoObject *exc = NULL, *res;
 683
 684                         res = mono_runtime_try_invoke (mono_defaults.threadpool_perform_wait_callback_method, NULL, NULL, &exc, &error);
 685                         if (exc || !mono_error_ok(&error)) {
 686                                 if (exc == NULL)
 687                                         exc = (MonoObject *) mono_error_convert_to_exception (&error);
 688                                 else
 689                                         mono_error_cleanup (&error);
 690                                 mono_thread_internal_unhandled_exception (exc);
 691                         } else if (res && *(MonoBoolean*) mono_object_unbox (res) == FALSE)
 692                                 retire = TRUE;
 693
 694                         mono_thread_clr_state (thread, (MonoThreadState)~ThreadState_Background);
 695                         if (!mono_thread_test_state (thread , ThreadState_Background))
 696                                 ves_icall_System_Threading_Thread_SetState (thread, ThreadState_Background);
 697
 698                         mono_domain_set (mono_get_root_domain (), TRUE);
 699                 }
 700                 mono_thread_pop_appdomain_ref ();
 701
 702                 domains_lock ();
 703
 704                 tpdomain->threadpool_jobs --;
 705                 g_assert (tpdomain->threadpool_jobs >= 0);
 706
 707                 if (tpdomain->outstanding_request + tpdomain->threadpool_jobs == 0 && mono_domain_is_unloading (tpdomain->domain)) {
 708                         gboolean removed;
 709
 710                         removed = tpdomain_remove (tpdomain);
 711                         g_assert (removed);
 712
 713                         mono_coop_cond_signal (&tpdomain->cleanup_cond);
 714                         tpdomain = NULL;
 715                 }
 716
 717                 previous_tpdomain = tpdomain;
 718         }
 719
 720         domains_unlock ();
 721
 722         mono_coop_mutex_lock (&threadpool->active_threads_lock);
 723         g_ptr_array_remove_fast (threadpool->working_threads, thread);
 724         mono_coop_mutex_unlock (&threadpool->active_threads_lock);
 725
 726         COUNTER_ATOMIC (counter, {
 727                 counter._.working--;
 728                 counter._.active --;
 729         });
 730
 731         mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker finishing", mono_native_thread_id_get ());
 732 }
 733
 734 static gboolean
 735 worker_try_create (void)
 736 {
 737         ThreadPoolCounter counter;
 738         MonoInternalThread *thread;
 739         gint64 current_ticks;
 740         gint32 now;
 741
 742         mono_coop_mutex_lock (&threadpool->worker_creation_lock);
 743
 744         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker", mono_native_thread_id_get ());
 745         current_ticks = mono_100ns_ticks ();
 746         now = current_ticks / (10 * 1000 * 1000);
 747         if (0 == current_ticks) {
 748                 g_warning ("failed to get 100ns ticks");
 749         } else {
 750                 if (threadpool->worker_creation_current_second != now) {
 751                         threadpool->worker_creation_current_second = now;
 752                         threadpool->worker_creation_current_count = 0;
 753                 } else {
 754                         g_assert (threadpool->worker_creation_current_count <= WORKER_CREATION_MAX_PER_SEC);
 755                         if (threadpool->worker_creation_current_count == WORKER_CREATION_MAX_PER_SEC) {
 756                                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: maximum number of worker created per second reached, current count = %d",
 757                                         mono_native_thread_id_get (), threadpool->worker_creation_current_count);
 758                                 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
 759                                 return FALSE;
 760                         }
 761                 }
 762         }
 763
 764         COUNTER_ATOMIC (counter, {
 765                 if (counter._.working >= counter._.max_working) {
 766                         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: maximum number of working threads reached",
 767                                 mono_native_thread_id_get ());
 768                         mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
 769                         return FALSE;
 770                 }
 771                 counter._.working ++;
 772                 counter._.active ++;
 773         });
 774
 775         MonoError error;
 776         if ((thread = mono_thread_create_internal (mono_get_root_domain (), worker_thread, NULL, TRUE, 0, &error)) != NULL) {
 777                 threadpool->worker_creation_current_count += 1;
 778
 779                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, created %p, now = %d count = %d", mono_native_thread_id_get (), GUINT_TO_POINTER(thread->tid), now, threadpool->worker_creation_current_count);
 780                 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
 781                 return TRUE;
 782         }
 783
 784         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: could not create thread due to %s", mono_native_thread_id_get (), mono_error_get_message (&error));
 785         mono_error_cleanup (&error);
 786
 787         COUNTER_ATOMIC (counter, {
 788                 counter._.working --;
 789                 counter._.active --;
 790         });
 791
 792         mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
 793         return FALSE;
 794 }
 795
 796 static void monitor_ensure_running (void);
 797
 798 static gboolean
 799 worker_request (MonoDomain *domain)
 800 {
 801         ThreadPoolDomain *tpdomain;
 802
 803         g_assert (domain);
 804         g_assert (threadpool);
 805
 806         if (mono_runtime_is_shutting_down ())
 807                 return FALSE;
 808
 809         domains_lock ();
 810
 811         /* synchronize check with worker_thread */
 812         if (mono_domain_is_unloading (domain)) {
 813                 domains_unlock ();
 814                 return FALSE;
 815         }
 816
 817         tpdomain = tpdomain_get (domain, TRUE);
 818         g_assert (tpdomain);
 819         tpdomain->outstanding_request ++;
 820
 821         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, domain = %p, outstanding_request = %d",
 822                 mono_native_thread_id_get (), tpdomain->domain, tpdomain->outstanding_request);
 823
 824         domains_unlock ();
 825
 826         if (threadpool->suspended)
 827                 return FALSE;
 828
 829         monitor_ensure_running ();
 830
 831         if (worker_try_unpark ()) {
 832                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, unparked", mono_native_thread_id_get ());
 833                 return TRUE;
 834         }
 835
 836         if (worker_try_create ()) {
 837                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, created", mono_native_thread_id_get ());
 838                 return TRUE;
 839         }
 840
 841         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, failed", mono_native_thread_id_get ());
 842         return FALSE;
 843 }
 844
 845 static gboolean
 846 monitor_should_keep_running (void)
 847 {
 848         static gint64 last_should_keep_running = -1;
 849
 850         g_assert (monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || monitor_status == MONITOR_STATUS_REQUESTED);
 851
 852         if (InterlockedExchange (&monitor_status, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST) {
 853                 gboolean should_keep_running = TRUE, force_should_keep_running = FALSE;
 854
 855                 if (mono_runtime_is_shutting_down ()) {
 856                         should_keep_running = FALSE;
 857                 } else {
 858                         domains_lock ();
 859                         if (!domain_any_has_request ())
 860                                 should_keep_running = FALSE;
 861                         domains_unlock ();
 862
 863                         if (!should_keep_running) {
 864                                 if (last_should_keep_running == -1 || mono_100ns_ticks () - last_should_keep_running < MONITOR_MINIMAL_LIFETIME * 1000 * 10) {
 865                                         should_keep_running = force_should_keep_running = TRUE;
 866                                 }
 867                         }
 868                 }
 869
 870                 if (should_keep_running) {
 871                         if (last_should_keep_running == -1 || !force_should_keep_running)
 872                                 last_should_keep_running = mono_100ns_ticks ();
 873                 } else {
 874                         last_should_keep_running = -1;
 875                         if (InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_NOT_RUNNING, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST)
 876                                 return FALSE;
 877                 }
 878         }
 879
 880         g_assert (monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || monitor_status == MONITOR_STATUS_REQUESTED);
 881
 882         return TRUE;
 883 }
 884
 885 static gboolean
 886 monitor_sufficient_delay_since_last_dequeue (void)
 887 {
 888         gint64 threshold;
 889
 890         g_assert (threadpool);
 891
 892         if (threadpool->cpu_usage < CPU_USAGE_LOW) {
 893                 threshold = MONITOR_INTERVAL;
 894         } else {
 895                 ThreadPoolCounter counter;
 896                 counter.as_gint64 = COUNTER_READ();
 897                 threshold = counter._.max_working * MONITOR_INTERVAL * 2;
 898         }
 899
 900         return mono_msec_ticks () >= threadpool->heuristic_last_dequeue + threshold;
 901 }
 902
 903 static void hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition);
 904
 905 static void
 906 monitor_thread (void)
 907 {
 908         MonoInternalThread *current_thread = mono_thread_internal_current ();
 909         guint i;
 910
 911         mono_cpu_usage (threadpool->cpu_usage_state);
 912
 913         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, started", mono_native_thread_id_get ());
 914
 915         do {
 916                 ThreadPoolCounter counter;
 917                 gboolean limit_worker_max_reached;
 918                 gint32 interval_left = MONITOR_INTERVAL;
 919                 gint32 awake = 0; /* number of spurious awakes we tolerate before doing a round of rebalancing */
 920
 921                 g_assert (monitor_status != MONITOR_STATUS_NOT_RUNNING);
 922
 923                 mono_gc_set_skip_thread (TRUE);
 924
 925                 do {
 926                         gint64 ts;
 927                         gboolean alerted = FALSE;
 928
 929                         if (mono_runtime_is_shutting_down ())
 930                                 break;
 931
 932                         ts = mono_msec_ticks ();
 933                         if (mono_thread_info_sleep (interval_left, &alerted) == 0)
 934                                 break;
 935                         interval_left -= mono_msec_ticks () - ts;
 936
 937                         mono_gc_set_skip_thread (FALSE);
 938                         if ((current_thread->state & (ThreadState_StopRequested | ThreadState_SuspendRequested)) != 0)
 939                                 mono_thread_interruption_checkpoint ();
 940                         mono_gc_set_skip_thread (TRUE);
 941                 } while (interval_left > 0 && ++awake < 10);
 942
 943                 mono_gc_set_skip_thread (FALSE);
 944
 945                 if (threadpool->suspended)
 946                         continue;
 947
 948                 if (mono_runtime_is_shutting_down ())
 949                         continue;
 950
 951                 domains_lock ();
 952                 if (!domain_any_has_request ()) {
 953                         domains_unlock ();
 954                         continue;
 955                 }
 956                 domains_unlock ();
 957
 958                 threadpool->cpu_usage = mono_cpu_usage (threadpool->cpu_usage_state);
 959
 960                 if (!monitor_sufficient_delay_since_last_dequeue ())
 961                         continue;
 962
 963                 limit_worker_max_reached = FALSE;
 964
 965                 COUNTER_ATOMIC (counter, {
 966                         if (counter._.max_working >= threadpool->limit_worker_max) {
 967                                 limit_worker_max_reached = TRUE;
 968                                 break;
 969                         }
 970                         counter._.max_working ++;
 971                 });
 972
 973                 if (limit_worker_max_reached)
 974                         continue;
 975
 976                 hill_climbing_force_change (counter._.max_working, TRANSITION_STARVATION);
 977
 978                 for (i = 0; i < 5; ++i) {
 979                         if (mono_runtime_is_shutting_down ())
 980                                 break;
 981
 982                         if (worker_try_unpark ()) {
 983                                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, unparked", mono_native_thread_id_get ());
 984                                 break;
 985                         }
 986
 987                         if (worker_try_create ()) {
 988                                 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, created", mono_native_thread_id_get ());
 989                                 break;
 990                         }
 991                 }
 992         } while (monitor_should_keep_running ());
 993
 994         mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, finished", mono_native_thread_id_get ());
 995 }
 996
 997 static void
 998 monitor_ensure_running (void)
 999 {
1000         MonoError error;
1001         for (;;) {
1002                 switch (monitor_status) {
1003                 case MONITOR_STATUS_REQUESTED:
1004                         return;
1005                 case MONITOR_STATUS_WAITING_FOR_REQUEST:
1006                         InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_WAITING_FOR_REQUEST);
1007                         break;
1008                 case MONITOR_STATUS_NOT_RUNNING:
1009                         if (mono_runtime_is_shutting_down ())
1010                                 return;
1011                         if (InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_NOT_RUNNING) == MONITOR_STATUS_NOT_RUNNING) {
1012                                 if (!mono_thread_create_internal (mono_get_root_domain (), monitor_thread, NULL, TRUE, SMALL_STACK, &error)) {
1013                                         monitor_status = MONITOR_STATUS_NOT_RUNNING;
1014                                         mono_error_cleanup (&error);
1015                                 }
1016                                 return;
1017                         }
1018                         break;
1019                 default: g_assert_not_reached ();
1020                 }
1021         }
1022 }
1023
1024 static void
1025 hill_climbing_change_thread_count (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
1026 {
1027         ThreadPoolHillClimbing *hc;
1028
1029         g_assert (threadpool);
1030
1031         hc = &threadpool->heuristic_hill_climbing;
1032
1033         mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] hill climbing, change max number of threads %d", mono_native_thread_id_get (), new_thread_count);
1034
1035         hc->last_thread_count = new_thread_count;
1036         hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
1037         hc->elapsed_since_last_change = 0;
1038         hc->completions_since_last_change = 0;
1039 }
1040
1041 static void
1042 hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
1043 {
1044         ThreadPoolHillClimbing *hc;
1045
1046         g_assert (threadpool);
1047
1048         hc = &threadpool->heuristic_hill_climbing;
1049
1050         if (new_thread_count != hc->last_thread_count) {
1051                 hc->current_control_setting += new_thread_count - hc->last_thread_count;
1052                 hill_climbing_change_thread_count (new_thread_count, transition);
1053         }
1054 }
1055
1056 static double_complex
1057 hill_climbing_get_wave_component (gdouble *samples, guint sample_count, gdouble period)
1058 {
1059         ThreadPoolHillClimbing *hc;
1060         gdouble w, cosine, sine, coeff, q0, q1, q2;
1061         guint i;
1062
1063         g_assert (threadpool);
1064         g_assert (sample_count >= period);
1065         g_assert (period >= 2);
1066
1067         hc = &threadpool->heuristic_hill_climbing;
1068
1069         w = 2.0 * M_PI / period;
1070         cosine = cos (w);
1071         sine = sin (w);
1072         coeff = 2.0 * cosine;
1073         q0 = q1 = q2 = 0;
1074
1075         for (i = 0; i < sample_count; ++i) {
1076                 q0 = coeff * q1 - q2 + samples [(hc->total_samples - sample_count + i) % hc->samples_to_measure];
1077                 q2 = q1;
1078                 q1 = q0;
1079         }
1080
1081         return mono_double_complex_scalar_div (mono_double_complex_make (q1 - q2 * cosine, (q2 * sine)), ((gdouble)sample_count));
1082 }
1083
1084 static gint16
1085 hill_climbing_update (gint16 current_thread_count, guint32 sample_duration, gint32 completions, gint64 *adjustment_interval)
1086 {
1087         ThreadPoolHillClimbing *hc;
1088         ThreadPoolHeuristicStateTransition transition;
1089         gdouble throughput;
1090         gdouble throughput_error_estimate;
1091         gdouble confidence;
1092         gdouble move;
1093         gdouble gain;
1094         gint sample_index;
1095         gint sample_count;
1096         gint new_thread_wave_magnitude;
1097         gint new_thread_count;
1098         double_complex thread_wave_component;
1099         double_complex throughput_wave_component;
1100         double_complex ratio;
1101
1102         g_assert (threadpool);
1103         g_assert (adjustment_interval);
1104
1105         hc = &threadpool->heuristic_hill_climbing;
1106
1107         /* If someone changed the thread count without telling us, update our records accordingly. */
1108         if (current_thread_count != hc->last_thread_count)
1109                 hill_climbing_force_change (current_thread_count, TRANSITION_INITIALIZING);
1110
1111         /* Update the cumulative stats for this thread count */
1112         hc->elapsed_since_last_change += sample_duration;
1113         hc->completions_since_last_change += completions;
1114
1115         /* Add in any data we've already collected about this sample */
1116         sample_duration += hc->accumulated_sample_duration;
1117         completions += hc->accumulated_completion_count;
1118
1119         /* We need to make sure we're collecting reasonably accurate data. Since we're just counting the end
1120          * of each work item, we are goinng to be missing some data about what really happened during the
1121          * sample interval. The count produced by each thread includes an initial work item that may have
1122          * started well before the start of the interval, and each thread may have been running some new
1123          * work item for some time before the end of the interval, which did not yet get counted. So
1124          * our count is going to be off by +/- threadCount workitems.
1125          *
1126          * The exception is that the thread that reported to us last time definitely wasn't running any work
1127          * at that time, and the thread that's reporting now definitely isn't running a work item now. So
1128          * we really only need to consider threadCount-1 threads.
1129          *
1130          * Thus the percent error in our count is +/- (threadCount-1)/numCompletions.
1131          *
1132          * We cannot rely on the frequency-domain analysis we'll be doing later to filter out this error, because
1133          * of the way it accumulates over time. If this sample is off by, say, 33% in the negative direction,
1134          * then the next one likely will be too. The one after that will include the sum of the completions
1135          * we missed in the previous samples, and so will be 33% positive. So every three samples we'll have
1136          * two "low" samples and one "high" sample. This will appear as periodic variation right in the frequency
1137          * range we're targeting, which will not be filtered by the frequency-domain translation. */
1138         if (hc->total_samples > 0 && ((current_thread_count - 1.0) / completions) >= hc->max_sample_error) {
1139                 /* Not accurate enough yet. Let's accumulate the data so
1140                  * far, and tell the ThreadPool to collect a little more. */
1141                 hc->accumulated_sample_duration = sample_duration;
1142                 hc->accumulated_completion_count = completions;
1143                 *adjustment_interval = 10;
1144                 return current_thread_count;
1145         }
1146
1147         /* We've got enouugh data for our sample; reset our accumulators for next time. */
1148         hc->accumulated_sample_duration = 0;
1149         hc->accumulated_completion_count = 0;
1150
1151         /* Add the current thread count and throughput sample to our history. */
1152         throughput = ((gdouble) completions) / sample_duration;
1153
1154         sample_index = hc->total_samples % hc->samples_to_measure;
1155         hc->samples [sample_index] = throughput;
1156         hc->thread_counts [sample_index] = current_thread_count;
1157         hc->total_samples ++;
1158
1159         /* Set up defaults for our metrics. */
1160         thread_wave_component = mono_double_complex_make(0, 0);
1161         throughput_wave_component = mono_double_complex_make(0, 0);
1162         throughput_error_estimate = 0;
1163         ratio = mono_double_complex_make(0, 0);
1164         confidence = 0;
1165
1166         transition = TRANSITION_WARMUP;
1167
1168         /* How many samples will we use? It must be at least the three wave periods we're looking for, and it must also
1169          * be a whole multiple of the primary wave's period; otherwise the frequency we're looking for will fall between
1170          * two frequency bands in the Fourier analysis, and we won't be able to measure it accurately. */
1171         sample_count = ((gint) MIN (hc->total_samples - 1, hc->samples_to_measure) / hc->wave_period) * hc->wave_period;
1172
1173         if (sample_count > hc->wave_period) {
1174                 guint i;
1175                 gdouble average_throughput;
1176                 gdouble average_thread_count;
1177                 gdouble sample_sum = 0;
1178                 gdouble thread_sum = 0;
1179
1180                 /* Average the throughput and thread count samples, so we can scale the wave magnitudes later. */
1181                 for (i = 0; i < sample_count; ++i) {
1182                         guint j = (hc->total_samples - sample_count + i) % hc->samples_to_measure;
1183                         sample_sum += hc->samples [j];
1184                         thread_sum += hc->thread_counts [j];
1185                 }
1186
1187                 average_throughput = sample_sum / sample_count;
1188                 average_thread_count = thread_sum / sample_count;
1189
1190                 if (average_throughput > 0 && average_thread_count > 0) {
1191                         gdouble noise_for_confidence, adjacent_period_1, adjacent_period_2;
1192
1193                         /* Calculate the periods of the adjacent frequency bands we'll be using to
1194                          * measure noise levels. We want the two adjacent Fourier frequency bands. */
1195                         adjacent_period_1 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) + 1);
1196                         adjacent_period_2 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) - 1);
1197
1198                         /* Get the the three different frequency components of the throughput (scaled by average
1199                          * throughput). Our "error" estimate (the amount of noise that might be present in the
1200                          * frequency band we're really interested in) is the average of the adjacent bands. */
1201                         throughput_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, hc->wave_period), average_throughput);
1202                         throughput_error_estimate = cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, adjacent_period_1), average_throughput));
1203
1204                         if (adjacent_period_2 <= sample_count) {
1205                                 throughput_error_estimate = MAX (throughput_error_estimate, cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (
1206                                         hc->samples, sample_count, adjacent_period_2), average_throughput)));
1207                         }
1208
1209                         /* Do the same for the thread counts, so we have something to compare to. We don't
1210                          * measure thread count noise, because there is none; these are exact measurements. */
1211                         thread_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->thread_counts, sample_count, hc->wave_period), average_thread_count);
1212
1213                         /* Update our moving average of the throughput noise. We'll use this
1214                          * later as feedback to determine the new size of the thread wave. */
1215                         if (hc->average_throughput_noise == 0) {
1216                                 hc->average_throughput_noise = throughput_error_estimate;
1217                         } else {
1218                                 hc->average_throughput_noise = (hc->throughput_error_smoothing_factor * throughput_error_estimate)
1219                                         + ((1.0 + hc->throughput_error_smoothing_factor) * hc->average_throughput_noise);
1220                         }
1221
1222                         if (cabs (thread_wave_component) > 0) {
1223                                 /* Adjust the throughput wave so it's centered around the target wave,
1224                                  * and then calculate the adjusted throughput/thread ratio. */
1225                                 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);
1226                                 transition = TRANSITION_CLIMBING_MOVE;
1227                         } else {
1228                                 ratio = mono_double_complex_make (0, 0);
1229                                 transition = TRANSITION_STABILIZING;
1230                         }
1231
1232                         noise_for_confidence = MAX (hc->average_throughput_noise, throughput_error_estimate);
1233                         if (noise_for_confidence > 0) {
1234                                 confidence = cabs (thread_wave_component) / noise_for_confidence / hc->target_signal_to_noise_ratio;
1235                         } else {
1236                                 /* there is no noise! */
1237                                 confidence = 1.0;
1238                         }
1239                 }
1240         }
1241
1242         /* We use just the real part of the complex ratio we just calculated. If the throughput signal
1243          * is exactly in phase with the thread signal, this will be the same as taking the magnitude of
1244          * the complex move and moving that far up. If they're 180 degrees out of phase, we'll move
1245          * backward (because this indicates that our changes are having the opposite of the intended effect).
1246          * If they're 90 degrees out of phase, we won't move at all, because we can't tell wether we're
1247          * having a negative or positive effect on throughput. */
1248         move = creal (ratio);
1249         move = CLAMP (move, -1.0, 1.0);
1250
1251         /* Apply our confidence multiplier. */
1252         move *= CLAMP (confidence, -1.0, 1.0);
1253
1254         /* Now apply non-linear gain, such that values around zero are attenuated, while higher values
1255          * are enhanced. This allows us to move quickly if we're far away from the target, but more slowly
1256         * if we're getting close, giving us rapid ramp-up without wild oscillations around the target. */
1257         gain = hc->max_change_per_second * sample_duration;
1258         move = pow (fabs (move), hc->gain_exponent) * (move >= 0.0 ? 1 : -1) * gain;
1259         move = MIN (move, hc->max_change_per_sample);
1260
1261         /* If the result was positive, and CPU is > 95%, refuse the move. */
1262         if (move > 0.0 && threadpool->cpu_usage > CPU_USAGE_HIGH)
1263                 move = 0.0;
1264
1265         /* Apply the move to our control setting. */
1266         hc->current_control_setting += move;
1267
1268         /* Calculate the new thread wave magnitude, which is based on the moving average we've been keeping of the
1269          * throughput error.  This average starts at zero, so we'll start with a nice safe little wave at first. */
1270         new_thread_wave_magnitude = (gint)(0.5 + (hc->current_control_setting * hc->average_throughput_noise
1271                 * hc->target_signal_to_noise_ratio * hc->thread_magnitude_multiplier * 2.0));
1272         new_thread_wave_magnitude = CLAMP (new_thread_wave_magnitude, 1, hc->max_thread_wave_magnitude);
1273
1274         /* Make sure our control setting is within the ThreadPool's limits. */
1275         hc->current_control_setting = CLAMP (hc->current_control_setting, threadpool->limit_worker_min, threadpool->limit_worker_max - new_thread_wave_magnitude);
1276
1277         /* Calculate the new thread count (control setting + square wave). */
1278         new_thread_count = (gint)(hc->current_control_setting + new_thread_wave_magnitude * ((hc->total_samples / (hc->wave_period / 2)) % 2));
1279
1280         /* Make sure the new thread count doesn't exceed the ThreadPool's limits. */
1281         new_thread_count = CLAMP (new_thread_count, threadpool->limit_worker_min, threadpool->limit_worker_max);
1282
1283         if (new_thread_count != current_thread_count)
1284                 hill_climbing_change_thread_count (new_thread_count, transition);
1285
1286         if (creal (ratio) < 0.0 && new_thread_count == threadpool->limit_worker_min)
1287                 *adjustment_interval = (gint)(0.5 + hc->current_sample_interval * (10.0 * MAX (-1.0 * creal (ratio), 1.0)));
1288         else
1289                 *adjustment_interval = hc->current_sample_interval;
1290
1291         return new_thread_count;
1292 }
1293
1294 static void
1295 heuristic_notify_work_completed (void)
1296 {
1297         g_assert (threadpool);
1298
1299         InterlockedIncrement (&threadpool->heuristic_completions);
1300         threadpool->heuristic_last_dequeue = mono_msec_ticks ();
1301 }
1302
1303 static gboolean
1304 heuristic_should_adjust (void)
1305 {
1306         g_assert (threadpool);
1307
1308         if (threadpool->heuristic_last_dequeue > threadpool->heuristic_last_adjustment + threadpool->heuristic_adjustment_interval) {
1309                 ThreadPoolCounter counter;
1310                 counter.as_gint64 = COUNTER_READ();
1311                 if (counter._.working <= counter._.max_working)
1312                         return TRUE;
1313         }
1314
1315         return FALSE;
1316 }
1317
1318 static void
1319 heuristic_adjust (void)
1320 {
1321         g_assert (threadpool);
1322
1323         if (mono_coop_mutex_trylock (&threadpool->heuristic_lock) == 0) {
1324                 gint32 completions = InterlockedExchange (&threadpool->heuristic_completions, 0);
1325                 gint64 sample_end = mono_msec_ticks ();
1326                 gint64 sample_duration = sample_end - threadpool->heuristic_sample_start;
1327
1328                 if (sample_duration >= threadpool->heuristic_adjustment_interval / 2) {
1329                         ThreadPoolCounter counter;
1330                         gint16 new_thread_count;
1331
1332                         counter.as_gint64 = COUNTER_READ ();
1333                         new_thread_count = hill_climbing_update (counter._.max_working, sample_duration, completions, &threadpool->heuristic_adjustment_interval);
1334
1335                         COUNTER_ATOMIC (counter, { counter._.max_working = new_thread_count; });
1336
1337                         if (new_thread_count > counter._.max_working)
1338                                 worker_request (mono_domain_get ());
1339
1340                         threadpool->heuristic_sample_start = sample_end;
1341                         threadpool->heuristic_last_adjustment = mono_msec_ticks ();
1342                 }
1343
1344                 mono_coop_mutex_unlock (&threadpool->heuristic_lock);
1345         }
1346 }
1347
1348 void
1349 mono_threadpool_ms_cleanup (void)
1350 {
1351 #ifndef DISABLE_SOCKETS
1352         mono_threadpool_ms_io_cleanup ();
1353 #endif
1354         mono_lazy_cleanup (&status, cleanup);
1355 }
1356
1357 MonoAsyncResult *
1358 mono_threadpool_ms_begin_invoke (MonoDomain *domain, MonoObject *target, MonoMethod *method, gpointer *params, MonoError *error)
1359 {
1360         static MonoClass *async_call_klass = NULL;
1361         MonoMethodMessage *message;
1362         MonoAsyncResult *async_result;
1363         MonoAsyncCall *async_call;
1364         MonoDelegate *async_callback = NULL;
1365         MonoObject *state = NULL;
1366
1367         if (!async_call_klass)
1368                 async_call_klass = mono_class_load_from_name (mono_defaults.corlib, "System", "MonoAsyncCall");
1369
1370         mono_lazy_initialize (&status, initialize);
1371
1372         mono_error_init (error);
1373
1374         message = mono_method_call_message_new (method, params, mono_get_delegate_invoke (method->klass), (params != NULL) ? (&async_callback) : NULL, (params != NULL) ? (&state) : NULL, error);
1375         return_val_if_nok (error, NULL);
1376
1377         async_call = (MonoAsyncCall*) mono_object_new_checked (domain, async_call_klass, error);
1378         return_val_if_nok (error, NULL);
1379
1380         MONO_OBJECT_SETREF (async_call, msg, message);
1381         MONO_OBJECT_SETREF (async_call, state, state);
1382
1383         if (async_callback) {
1384                 MONO_OBJECT_SETREF (async_call, cb_method, mono_get_delegate_invoke (((MonoObject*) async_callback)->vtable->klass));
1385                 MONO_OBJECT_SETREF (async_call, cb_target, async_callback);
1386         }
1387
1388         async_result = mono_async_result_new (domain, NULL, async_call->state, NULL, (MonoObject*) async_call, error);
1389         return_val_if_nok (error, NULL);
1390         MONO_OBJECT_SETREF (async_result, async_delegate, target);
1391
1392         mono_threadpool_ms_enqueue_work_item (domain, (MonoObject*) async_result, error);
1393         return_val_if_nok (error, NULL);
1394
1395         return async_result;
1396 }
1397
1398 MonoObject *
1399 mono_threadpool_ms_end_invoke (MonoAsyncResult *ares, MonoArray **out_args, MonoObject **exc, MonoError *error)
1400 {
1401         MonoAsyncCall *ac;
1402
1403         mono_error_init (error);
1404         g_assert (exc);
1405         g_assert (out_args);
1406
1407         *exc = NULL;
1408         *out_args = NULL;
1409
1410         /* check if already finished */
1411         mono_monitor_enter ((MonoObject*) ares);
1412
1413         if (ares->endinvoke_called) {
1414                 mono_error_set_invalid_operation(error, "Delegate EndInvoke method called more than once");
1415                 mono_monitor_exit ((MonoObject*) ares);
1416                 return NULL;
1417         }
1418
1419         ares->endinvoke_called = 1;
1420
1421         /* wait until we are really finished */
1422         if (ares->completed) {
1423                 mono_monitor_exit ((MonoObject *) ares);
1424         } else {
1425                 gpointer wait_event;
1426                 if (ares->handle) {
1427                         wait_event = mono_wait_handle_get_handle ((MonoWaitHandle*) ares->handle);
1428                 } else {
1429                         wait_event = mono_w32event_create (TRUE, FALSE);
1430                         g_assert(wait_event);
1431                         MonoWaitHandle *wait_handle = mono_wait_handle_new (mono_object_domain (ares), wait_event, error);
1432                         if (!is_ok (error)) {
1433                                 CloseHandle (wait_event);
1434                                 return NULL;
1435                         }
1436                         MONO_OBJECT_SETREF (ares, handle, (MonoObject*) wait_handle);
1437                 }
1438                 mono_monitor_exit ((MonoObject*) ares);
1439                 MONO_ENTER_GC_SAFE;
1440                 WaitForSingleObjectEx (wait_event, INFINITE, TRUE);
1441                 MONO_EXIT_GC_SAFE;
1442         }
1443
1444         ac = (MonoAsyncCall*) ares->object_data;
1445         g_assert (ac);
1446
1447         *exc = ac->msg->exc; /* FIXME: GC add write barrier */
1448         *out_args = ac->out_args;
1449         return ac->res;
1450 }
1451
1452 gboolean
1453 mono_threadpool_ms_remove_domain_jobs (MonoDomain *domain, int timeout)
1454 {
1455         gint64 end;
1456         ThreadPoolDomain *tpdomain;
1457         gboolean ret;
1458
1459         g_assert (domain);
1460         g_assert (timeout >= -1);
1461
1462         g_assert (mono_domain_is_unloading (domain));
1463
1464         if (timeout != -1)
1465                 end = mono_msec_ticks () + timeout;
1466
1467 #ifndef DISABLE_SOCKETS
1468         mono_threadpool_ms_io_remove_domain_jobs (domain);
1469         if (timeout != -1) {
1470                 if (mono_msec_ticks () > end)
1471                         return FALSE;
1472         }
1473 #endif
1474
1475         /*
1476          * Wait for all threads which execute jobs in the domain to exit.
1477          * The is_unloading () check in worker_request () ensures that
1478          * no new jobs are added after we enter the lock below.
1479          */
1480         mono_lazy_initialize (&status, initialize);
1481         domains_lock ();
1482
1483         tpdomain = tpdomain_get (domain, FALSE);
1484         if (!tpdomain) {
1485                 domains_unlock ();
1486                 return TRUE;
1487         }
1488
1489         ret = TRUE;
1490
1491         while (tpdomain->outstanding_request + tpdomain->threadpool_jobs > 0) {
1492                 if (timeout == -1) {
1493                         mono_coop_cond_wait (&tpdomain->cleanup_cond, &threadpool->domains_lock);
1494                 } else {
1495                         gint64 now;
1496                         gint res;
1497
1498                         now = mono_msec_ticks();
1499                         if (now > end) {
1500                                 ret = FALSE;
1501                                 break;
1502                         }
1503
1504                         res = mono_coop_cond_timedwait (&tpdomain->cleanup_cond, &threadpool->domains_lock, end - now);
1505                         if (res != 0) {
1506                                 ret = FALSE;
1507                                 break;
1508                         }
1509                 }
1510         }
1511
1512         /* Remove from the list the worker threads look at */
1513         tpdomain_remove (tpdomain);
1514
1515         domains_unlock ();
1516
1517         mono_coop_cond_destroy (&tpdomain->cleanup_cond);
1518         tpdomain_free (tpdomain);
1519
1520         return ret;
1521 }
1522
1523 void
1524 mono_threadpool_ms_suspend (void)
1525 {
1526         if (threadpool)
1527                 threadpool->suspended = TRUE;
1528 }
1529
1530 void
1531 mono_threadpool_ms_resume (void)
1532 {
1533         if (threadpool)
1534                 threadpool->suspended = FALSE;
1535 }
1536
1537 void
1538 ves_icall_System_Threading_ThreadPool_GetAvailableThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1539 {
1540         ThreadPoolCounter counter;
1541
1542         if (!worker_threads || !completion_port_threads)
1543                 return;
1544
1545         mono_lazy_initialize (&status, initialize);
1546
1547         counter.as_gint64 = COUNTER_READ ();
1548
1549         *worker_threads = MAX (0, threadpool->limit_worker_max - counter._.active);
1550         *completion_port_threads = threadpool->limit_io_max;
1551 }
1552
1553 void
1554 ves_icall_System_Threading_ThreadPool_GetMinThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1555 {
1556         if (!worker_threads || !completion_port_threads)
1557                 return;
1558
1559         mono_lazy_initialize (&status, initialize);
1560
1561         *worker_threads = threadpool->limit_worker_min;
1562         *completion_port_threads = threadpool->limit_io_min;
1563 }
1564
1565 void
1566 ves_icall_System_Threading_ThreadPool_GetMaxThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1567 {
1568         if (!worker_threads || !completion_port_threads)
1569                 return;
1570
1571         mono_lazy_initialize (&status, initialize);
1572
1573         *worker_threads = threadpool->limit_worker_max;
1574         *completion_port_threads = threadpool->limit_io_max;
1575 }
1576
1577 MonoBoolean
1578 ves_icall_System_Threading_ThreadPool_SetMinThreadsNative (gint32 worker_threads, gint32 completion_port_threads)
1579 {
1580         mono_lazy_initialize (&status, initialize);
1581
1582         if (worker_threads <= 0 || worker_threads > threadpool->limit_worker_max)
1583                 return FALSE;
1584         if (completion_port_threads <= 0 || completion_port_threads > threadpool->limit_io_max)
1585                 return FALSE;
1586
1587         threadpool->limit_worker_min = worker_threads;
1588         threadpool->limit_io_min = completion_port_threads;
1589
1590         return TRUE;
1591 }
1592
1593 MonoBoolean
1594 ves_icall_System_Threading_ThreadPool_SetMaxThreadsNative (gint32 worker_threads, gint32 completion_port_threads)
1595 {
1596         gint cpu_count = mono_cpu_count ();
1597
1598         mono_lazy_initialize (&status, initialize);
1599
1600         if (worker_threads < threadpool->limit_worker_min || worker_threads < cpu_count)
1601                 return FALSE;
1602         if (completion_port_threads < threadpool->limit_io_min || completion_port_threads < cpu_count)
1603                 return FALSE;
1604
1605         threadpool->limit_worker_max = worker_threads;
1606         threadpool->limit_io_max = completion_port_threads;
1607
1608         return TRUE;
1609 }
1610
1611 void
1612 ves_icall_System_Threading_ThreadPool_InitializeVMTp (MonoBoolean *enable_worker_tracking)
1613 {
1614         if (enable_worker_tracking) {
1615                 // TODO implement some kind of switch to have the possibily to use it
1616                 *enable_worker_tracking = FALSE;
1617         }
1618
1619         mono_lazy_initialize (&status, initialize);
1620 }
1621
1622 MonoBoolean
1623 ves_icall_System_Threading_ThreadPool_NotifyWorkItemComplete (void)
1624 {
1625         ThreadPoolCounter counter;
1626
1627         if (mono_domain_is_unloading (mono_domain_get ()) || mono_runtime_is_shutting_down ())
1628                 return FALSE;
1629
1630         heuristic_notify_work_completed ();
1631
1632         if (heuristic_should_adjust ())
1633                 heuristic_adjust ();
1634
1635         counter.as_gint64 = COUNTER_READ ();
1636         return counter._.working <= counter._.max_working;
1637 }
1638
1639 void
1640 ves_icall_System_Threading_ThreadPool_NotifyWorkItemProgressNative (void)
1641 {
1642         heuristic_notify_work_completed ();
1643
1644         if (heuristic_should_adjust ())
1645                 heuristic_adjust ();
1646 }
1647
1648 void
1649 ves_icall_System_Threading_ThreadPool_ReportThreadStatus (MonoBoolean is_working)
1650 {
1651         // TODO
1652         MonoError error;
1653         mono_error_set_not_implemented (&error, "");
1654         mono_error_set_pending_exception (&error);
1655 }
1656
1657 MonoBoolean
1658 ves_icall_System_Threading_ThreadPool_RequestWorkerThread (void)
1659 {
1660         return worker_request (mono_domain_get ());
1661 }
1662
1663 MonoBoolean G_GNUC_UNUSED
1664 ves_icall_System_Threading_ThreadPool_PostQueuedCompletionStatus (MonoNativeOverlapped *native_overlapped)
1665 {
1666         /* This copy the behavior of the current Mono implementation */
1667         MonoError error;
1668         mono_error_set_not_implemented (&error, "");
1669         mono_error_set_pending_exception (&error);
1670         return FALSE;
1671 }
1672
1673 MonoBoolean G_GNUC_UNUSED
1674 ves_icall_System_Threading_ThreadPool_BindIOCompletionCallbackNative (gpointer file_handle)
1675 {
1676         /* This copy the behavior of the current Mono implementation */
1677         return TRUE;
1678 }
1679
1680 MonoBoolean G_GNUC_UNUSED
1681 ves_icall_System_Threading_ThreadPool_IsThreadPoolHosted (void)
1682 {
1683         return FALSE;
1684 }