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