2 * threadpool-ms.c: Microsoft threadpool runtime support
5 * Ludovic Henry (ludovic.henry@xamarin.com)
7 * Copyright 2015 Xamarin, Inc (http://www.xamarin.com)
11 // Copyright (c) Microsoft. All rights reserved.
12 // Licensed under the MIT license. See LICENSE file in the project root for full license information.
15 // - src/vm/comthreadpool.cpp
16 // - src/vm/win32threadpoolcpp
17 // - src/vm/threadpoolrequest.cpp
18 // - src/vm/hillclimbing.cpp
20 // Ported from C++ to C and adjusted to Mono runtime
23 #define _USE_MATH_DEFINES // needed by MSVC to define math constants
28 #include <mono/metadata/class-internals.h>
29 #include <mono/metadata/exception.h>
30 #include <mono/metadata/gc-internals.h>
31 #include <mono/metadata/object.h>
32 #include <mono/metadata/object-internals.h>
33 #include <mono/metadata/threadpool-ms.h>
34 #include <mono/metadata/threadpool-ms-io.h>
35 #include <mono/utils/atomic.h>
36 #include <mono/utils/mono-compiler.h>
37 #include <mono/utils/mono-complex.h>
38 #include <mono/utils/mono-lazy-init.h>
39 #include <mono/utils/mono-logger.h>
40 #include <mono/utils/mono-logger-internals.h>
41 #include <mono/utils/mono-proclib.h>
42 #include <mono/utils/mono-threads.h>
43 #include <mono/utils/mono-time.h>
44 #include <mono/utils/mono-rand.h>
46 #define CPU_USAGE_LOW 80
47 #define CPU_USAGE_HIGH 95
49 #define MONITOR_INTERVAL 500 // ms
50 #define MONITOR_MINIMAL_LIFETIME 60 * 1000 // ms
52 #define WORKER_CREATION_MAX_PER_SEC 10
54 /* The exponent to apply to the gain. 1.0 means to use linear gain,
55 * higher values will enhance large moves and damp small ones.
57 #define HILL_CLIMBING_GAIN_EXPONENT 2.0
59 /* The 'cost' of a thread. 0 means drive for increased throughput regardless
60 * of thread count, higher values bias more against higher thread counts.
62 #define HILL_CLIMBING_BIAS 0.15
64 #define HILL_CLIMBING_WAVE_PERIOD 4
65 #define HILL_CLIMBING_MAX_WAVE_MAGNITUDE 20
66 #define HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER 1.0
67 #define HILL_CLIMBING_WAVE_HISTORY_SIZE 8
68 #define HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO 3.0
69 #define HILL_CLIMBING_MAX_CHANGE_PER_SECOND 4
70 #define HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE 20
71 #define HILL_CLIMBING_SAMPLE_INTERVAL_LOW 10
72 #define HILL_CLIMBING_SAMPLE_INTERVAL_HIGH 200
73 #define HILL_CLIMBING_ERROR_SMOOTHING_FACTOR 0.01
74 #define HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT 0.15
78 gint16 max_working; /* determined by heuristic */
79 gint16 active; /* executing worker_thread */
80 gint16 working; /* actively executing worker_thread, not parked */
81 gint16 parked; /* parked */
88 gint32 outstanding_request;
91 typedef MonoInternalThread ThreadPoolWorkingThread;
95 gint32 samples_to_measure;
96 gdouble target_throughput_ratio;
97 gdouble target_signal_to_noise_ratio;
98 gdouble max_change_per_second;
99 gdouble max_change_per_sample;
100 gint32 max_thread_wave_magnitude;
101 gint32 sample_interval_low;
102 gdouble thread_magnitude_multiplier;
103 gint32 sample_interval_high;
104 gdouble throughput_error_smoothing_factor;
105 gdouble gain_exponent;
106 gdouble max_sample_error;
108 gdouble current_control_setting;
109 gint64 total_samples;
110 gint16 last_thread_count;
111 gdouble elapsed_since_last_change;
112 gdouble completions_since_last_change;
114 gdouble average_throughput_noise;
117 gdouble *thread_counts;
119 guint32 current_sample_interval;
120 gpointer random_interval_generator;
122 gint32 accumulated_completion_count;
123 gdouble accumulated_sample_duration;
124 } ThreadPoolHillClimbing;
127 ThreadPoolCounter counters;
129 GPtrArray *domains; // ThreadPoolDomain* []
130 MonoCoopMutex domains_lock;
132 GPtrArray *working_threads; // ThreadPoolWorkingThread* []
133 gint32 parked_threads_count;
134 MonoCoopCond parked_threads_cond;
135 MonoCoopMutex active_threads_lock; /* protect access to working_threads and parked_threads */
137 guint32 worker_creation_current_second;
138 guint32 worker_creation_current_count;
139 MonoCoopMutex worker_creation_lock;
141 gint32 heuristic_completions;
142 guint32 heuristic_sample_start;
143 guint32 heuristic_last_dequeue; // ms
144 guint32 heuristic_last_adjustment; // ms
145 guint32 heuristic_adjustment_interval; // ms
146 ThreadPoolHillClimbing heuristic_hill_climbing;
147 MonoCoopMutex heuristic_lock;
149 gint32 limit_worker_min;
150 gint32 limit_worker_max;
154 MonoCpuUsageState *cpu_usage_state;
157 /* suspended by the debugger */
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;
173 static mono_lazy_init_t status = MONO_LAZY_INIT_STATUS_NOT_INITIALIZED;
176 MONITOR_STATUS_REQUESTED,
177 MONITOR_STATUS_WAITING_FOR_REQUEST,
178 MONITOR_STATUS_NOT_RUNNING,
181 static gint32 monitor_status = MONITOR_STATUS_NOT_RUNNING;
183 static ThreadPool* threadpool;
185 #define COUNTER_CHECK(counter) \
187 g_assert (counter._.max_working > 0); \
188 g_assert (counter._.working >= 0); \
189 g_assert (counter._.active >= 0); \
192 #define COUNTER_READ() (InterlockedRead64 (&threadpool->counters.as_gint64))
194 #define COUNTER_ATOMIC(var,block) \
196 ThreadPoolCounter __old; \
198 g_assert (threadpool); \
199 __old.as_gint64 = COUNTER_READ (); \
202 COUNTER_CHECK (var); \
203 } while (InterlockedCompareExchange64 (&threadpool->counters.as_gint64, (var).as_gint64, __old.as_gint64) != __old.as_gint64); \
206 #define COUNTER_TRY_ATOMIC(res,var,block) \
208 ThreadPoolCounter __old; \
210 g_assert (threadpool); \
211 __old.as_gint64 = COUNTER_READ (); \
215 COUNTER_CHECK (var); \
216 (res) = InterlockedCompareExchange64 (&threadpool->counters.as_gint64, (var).as_gint64, __old.as_gint64) == __old.as_gint64; \
224 return mono_rand_init (NULL, 0);
228 rand_next (gpointer *handle, guint32 min, guint32 max)
232 mono_rand_try_get_uint32 (handle, &val, min, max, &error);
233 // FIXME handle error
234 mono_error_assert_ok (&error);
239 rand_free (gpointer handle)
241 mono_rand_close (handle);
247 ThreadPoolHillClimbing *hc;
248 const char *threads_per_cpu_env;
249 gint threads_per_cpu;
252 g_assert (!threadpool);
253 threadpool = g_new0 (ThreadPool, 1);
254 g_assert (threadpool);
256 threadpool->domains = g_ptr_array_new ();
257 mono_coop_mutex_init (&threadpool->domains_lock);
259 threadpool->parked_threads_count = 0;
260 mono_coop_cond_init (&threadpool->parked_threads_cond);
261 threadpool->working_threads = g_ptr_array_new ();
262 mono_coop_mutex_init (&threadpool->active_threads_lock);
264 threadpool->worker_creation_current_second = -1;
265 mono_coop_mutex_init (&threadpool->worker_creation_lock);
267 threadpool->heuristic_adjustment_interval = 10;
268 mono_coop_mutex_init (&threadpool->heuristic_lock);
272 hc = &threadpool->heuristic_hill_climbing;
274 hc->wave_period = HILL_CLIMBING_WAVE_PERIOD;
275 hc->max_thread_wave_magnitude = HILL_CLIMBING_MAX_WAVE_MAGNITUDE;
276 hc->thread_magnitude_multiplier = (gdouble) HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER;
277 hc->samples_to_measure = hc->wave_period * HILL_CLIMBING_WAVE_HISTORY_SIZE;
278 hc->target_throughput_ratio = (gdouble) HILL_CLIMBING_BIAS;
279 hc->target_signal_to_noise_ratio = (gdouble) HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO;
280 hc->max_change_per_second = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SECOND;
281 hc->max_change_per_sample = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE;
282 hc->sample_interval_low = HILL_CLIMBING_SAMPLE_INTERVAL_LOW;
283 hc->sample_interval_high = HILL_CLIMBING_SAMPLE_INTERVAL_HIGH;
284 hc->throughput_error_smoothing_factor = (gdouble) HILL_CLIMBING_ERROR_SMOOTHING_FACTOR;
285 hc->gain_exponent = (gdouble) HILL_CLIMBING_GAIN_EXPONENT;
286 hc->max_sample_error = (gdouble) HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT;
287 hc->current_control_setting = 0;
288 hc->total_samples = 0;
289 hc->last_thread_count = 0;
290 hc->average_throughput_noise = 0;
291 hc->elapsed_since_last_change = 0;
292 hc->accumulated_completion_count = 0;
293 hc->accumulated_sample_duration = 0;
294 hc->samples = g_new0 (gdouble, hc->samples_to_measure);
295 hc->thread_counts = g_new0 (gdouble, hc->samples_to_measure);
296 hc->random_interval_generator = rand_create ();
297 hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
299 if (!(threads_per_cpu_env = g_getenv ("MONO_THREADS_PER_CPU")))
302 threads_per_cpu = CLAMP (atoi (threads_per_cpu_env), 1, 50);
304 threads_count = mono_cpu_count () * threads_per_cpu;
306 threadpool->limit_worker_min = threadpool->limit_io_min = threads_count;
308 #if defined (PLATFORM_ANDROID) || defined (HOST_IOS)
309 threadpool->limit_worker_max = threadpool->limit_io_max = CLAMP (threads_count * 100, MIN (threads_count, 200), MAX (threads_count, 200));
311 threadpool->limit_worker_max = threadpool->limit_io_max = threads_count * 100;
314 threadpool->counters._.max_working = threadpool->limit_worker_min;
316 threadpool->cpu_usage_state = g_new0 (MonoCpuUsageState, 1);
318 threadpool->suspended = FALSE;
321 static void worker_kill (ThreadPoolWorkingThread *thread);
328 /* we make the assumption along the code that we are
329 * cleaning up only if the runtime is shutting down */
330 g_assert (mono_runtime_is_shutting_down ());
332 while (monitor_status != MONITOR_STATUS_NOT_RUNNING)
333 mono_thread_info_sleep (1, NULL);
335 mono_coop_mutex_lock (&threadpool->active_threads_lock);
337 /* stop all threadpool->working_threads */
338 for (i = 0; i < threadpool->working_threads->len; ++i)
339 worker_kill ((ThreadPoolWorkingThread*) g_ptr_array_index (threadpool->working_threads, i));
341 /* unpark all threadpool->parked_threads */
342 mono_coop_cond_broadcast (&threadpool->parked_threads_cond);
344 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
348 mono_threadpool_ms_enqueue_work_item (MonoDomain *domain, MonoObject *work_item)
350 static MonoClass *threadpool_class = NULL;
351 static MonoMethod *unsafe_queue_custom_work_item_method = NULL;
353 MonoDomain *current_domain;
357 g_assert (work_item);
359 if (!threadpool_class)
360 threadpool_class = mono_class_load_from_name (mono_defaults.corlib, "System.Threading", "ThreadPool");
362 if (!unsafe_queue_custom_work_item_method)
363 unsafe_queue_custom_work_item_method = mono_class_get_method_from_name (threadpool_class, "UnsafeQueueCustomWorkItem", 2);
364 g_assert (unsafe_queue_custom_work_item_method);
368 args [0] = (gpointer) work_item;
369 args [1] = (gpointer) &f;
371 current_domain = mono_domain_get ();
372 if (current_domain == domain) {
373 mono_runtime_invoke_checked (unsafe_queue_custom_work_item_method, NULL, args, &error);
374 mono_error_raise_exception (&error); /* FIXME don't raise here */
376 mono_thread_push_appdomain_ref (domain);
377 if (mono_domain_set (domain, FALSE)) {
378 mono_runtime_invoke_checked (unsafe_queue_custom_work_item_method, NULL, args, &error);
379 mono_error_raise_exception (&error); /* FIXME don't raise here */
380 mono_domain_set (current_domain, TRUE);
382 mono_thread_pop_appdomain_ref ();
386 /* LOCKING: threadpool->domains_lock must be held */
388 domain_add (ThreadPoolDomain *tpdomain)
394 len = threadpool->domains->len;
395 for (i = 0; i < len; ++i) {
396 if (g_ptr_array_index (threadpool->domains, i) == tpdomain)
401 g_ptr_array_add (threadpool->domains, tpdomain);
404 /* LOCKING: threadpool->domains_lock must be held */
406 domain_remove (ThreadPoolDomain *tpdomain)
409 return g_ptr_array_remove (threadpool->domains, tpdomain);
412 /* LOCKING: threadpool->domains_lock must be held */
413 static ThreadPoolDomain *
414 domain_get (MonoDomain *domain, gboolean create)
416 ThreadPoolDomain *tpdomain = NULL;
421 for (i = 0; i < threadpool->domains->len; ++i) {
422 tpdomain = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i);
423 if (tpdomain->domain == domain)
428 tpdomain = g_new0 (ThreadPoolDomain, 1);
429 tpdomain->domain = domain;
430 domain_add (tpdomain);
437 domain_free (ThreadPoolDomain *tpdomain)
442 /* LOCKING: threadpool->domains_lock must be held */
444 domain_any_has_request (void)
448 for (i = 0; i < threadpool->domains->len; ++i) {
449 ThreadPoolDomain *tmp = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i);
450 if (tmp->outstanding_request > 0)
457 /* LOCKING: threadpool->domains_lock must be held */
458 static ThreadPoolDomain *
459 domain_get_next (ThreadPoolDomain *current)
461 ThreadPoolDomain *tpdomain = NULL;
464 len = threadpool->domains->len;
466 guint i, current_idx = -1;
468 for (i = 0; i < len; ++i) {
469 if (current == g_ptr_array_index (threadpool->domains, i)) {
474 g_assert (current_idx >= 0);
476 for (i = current_idx + 1; i < len + current_idx + 1; ++i) {
477 ThreadPoolDomain *tmp = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i % len);
478 if (tmp->outstanding_request > 0) {
489 worker_wait_interrupt (gpointer data)
491 mono_coop_mutex_lock (&threadpool->active_threads_lock);
492 mono_coop_cond_signal (&threadpool->parked_threads_cond);
493 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
496 /* return TRUE if timeout, FALSE otherwise (worker unpark or interrupt) */
500 gboolean timeout = FALSE;
502 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] current worker parking", mono_native_thread_id_get ());
504 mono_gc_set_skip_thread (TRUE);
506 mono_coop_mutex_lock (&threadpool->active_threads_lock);
508 if (!mono_runtime_is_shutting_down ()) {
509 static gpointer rand_handle = NULL;
510 MonoInternalThread *thread_internal;
511 gboolean interrupted = FALSE;
514 rand_handle = rand_create ();
515 g_assert (rand_handle);
517 thread_internal = mono_thread_internal_current ();
518 g_assert (thread_internal);
520 threadpool->parked_threads_count += 1;
521 g_ptr_array_remove_fast (threadpool->working_threads, thread_internal);
523 mono_thread_info_install_interrupt (worker_wait_interrupt, NULL, &interrupted);
527 if (mono_coop_cond_timedwait (&threadpool->parked_threads_cond, &threadpool->active_threads_lock, rand_next ((void **)rand_handle, 5 * 1000, 60 * 1000)) != 0)
530 mono_thread_info_uninstall_interrupt (&interrupted);
533 g_ptr_array_add (threadpool->working_threads, thread_internal);
534 threadpool->parked_threads_count -= 1;
537 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
539 mono_gc_set_skip_thread (FALSE);
541 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] current worker unparking, timeout? %s", mono_native_thread_id_get (), timeout ? "yes" : "no");
547 worker_try_unpark (void)
549 gboolean res = FALSE;
551 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker", mono_native_thread_id_get ());
553 mono_coop_mutex_lock (&threadpool->active_threads_lock);
554 if (threadpool->parked_threads_count > 0) {
555 mono_coop_cond_signal (&threadpool->parked_threads_cond);
558 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
560 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker, success? %s", mono_native_thread_id_get (), res ? "yes" : "no");
566 worker_kill (ThreadPoolWorkingThread *thread)
568 if (thread == mono_thread_internal_current ())
571 mono_thread_internal_stop ((MonoInternalThread*) thread);
575 worker_thread (gpointer data)
578 MonoInternalThread *thread;
579 ThreadPoolDomain *tpdomain, *previous_tpdomain;
580 ThreadPoolCounter counter;
581 gboolean retire = FALSE;
583 mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker starting", mono_native_thread_id_get ());
585 g_assert (threadpool);
587 thread = mono_thread_internal_current ();
590 mono_thread_set_name_internal (thread, mono_string_new (mono_domain_get (), "Threadpool worker"), FALSE);
592 mono_coop_mutex_lock (&threadpool->active_threads_lock);
593 g_ptr_array_add (threadpool->working_threads, thread);
594 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
596 previous_tpdomain = NULL;
598 mono_coop_mutex_lock (&threadpool->domains_lock);
600 while (!mono_runtime_is_shutting_down ()) {
603 if ((thread->state & (ThreadState_StopRequested | ThreadState_SuspendRequested)) != 0) {
604 mono_coop_mutex_unlock (&threadpool->domains_lock);
605 mono_thread_interruption_checkpoint ();
606 mono_coop_mutex_lock (&threadpool->domains_lock);
609 if (retire || !(tpdomain = domain_get_next (previous_tpdomain))) {
612 COUNTER_ATOMIC (counter, {
613 counter._.working --;
617 mono_coop_mutex_unlock (&threadpool->domains_lock);
618 timeout = worker_park ();
619 mono_coop_mutex_lock (&threadpool->domains_lock);
621 COUNTER_ATOMIC (counter, {
622 counter._.working ++;
635 tpdomain->outstanding_request --;
636 g_assert (tpdomain->outstanding_request >= 0);
638 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker running in domain %p",
639 mono_native_thread_id_get (), tpdomain->domain, tpdomain->outstanding_request);
641 g_assert (tpdomain->domain);
642 g_assert (tpdomain->domain->threadpool_jobs >= 0);
643 tpdomain->domain->threadpool_jobs ++;
645 mono_coop_mutex_unlock (&threadpool->domains_lock);
647 mono_thread_push_appdomain_ref (tpdomain->domain);
648 if (mono_domain_set (tpdomain->domain, FALSE)) {
649 MonoObject *exc = NULL, *res;
651 res = mono_runtime_try_invoke (mono_defaults.threadpool_perform_wait_callback_method, NULL, NULL, &exc, &error);
652 if (exc || !mono_error_ok(&error)) {
654 exc = (MonoObject *) mono_error_convert_to_exception (&error);
656 mono_error_cleanup (&error);
657 mono_thread_internal_unhandled_exception (exc);
658 } else if (res && *(MonoBoolean*) mono_object_unbox (res) == FALSE)
661 mono_thread_clr_state (thread, (MonoThreadState)~ThreadState_Background);
662 if (!mono_thread_test_state (thread , ThreadState_Background))
663 ves_icall_System_Threading_Thread_SetState (thread, ThreadState_Background);
665 mono_domain_set (mono_get_root_domain (), TRUE);
667 mono_thread_pop_appdomain_ref ();
669 mono_coop_mutex_lock (&threadpool->domains_lock);
671 tpdomain->domain->threadpool_jobs --;
672 g_assert (tpdomain->domain->threadpool_jobs >= 0);
674 if (tpdomain->domain->threadpool_jobs == 0 && mono_domain_is_unloading (tpdomain->domain)) {
675 gboolean removed = domain_remove (tpdomain);
677 if (tpdomain->domain->cleanup_semaphore)
678 ReleaseSemaphore (tpdomain->domain->cleanup_semaphore, 1, NULL);
679 domain_free (tpdomain);
683 previous_tpdomain = tpdomain;
686 mono_coop_mutex_unlock (&threadpool->domains_lock);
688 mono_coop_mutex_lock (&threadpool->active_threads_lock);
689 g_ptr_array_remove_fast (threadpool->working_threads, thread);
690 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
692 COUNTER_ATOMIC (counter, {
697 mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker finishing", mono_native_thread_id_get ());
701 worker_try_create (void)
703 ThreadPoolCounter counter;
704 MonoInternalThread *thread;
707 mono_coop_mutex_lock (&threadpool->worker_creation_lock);
709 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker", mono_native_thread_id_get ());
711 if ((now = mono_100ns_ticks () / 10 / 1000 / 1000) == 0) {
712 g_warning ("failed to get 100ns ticks");
714 if (threadpool->worker_creation_current_second != now) {
715 threadpool->worker_creation_current_second = now;
716 threadpool->worker_creation_current_count = 0;
718 g_assert (threadpool->worker_creation_current_count <= WORKER_CREATION_MAX_PER_SEC);
719 if (threadpool->worker_creation_current_count == WORKER_CREATION_MAX_PER_SEC) {
720 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",
721 mono_native_thread_id_get (), threadpool->worker_creation_current_count);
722 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
728 COUNTER_ATOMIC (counter, {
729 if (counter._.working >= counter._.max_working) {
730 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: maximum number of working threads reached",
731 mono_native_thread_id_get ());
732 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
735 counter._.working ++;
739 if ((thread = mono_thread_create_internal (mono_get_root_domain (), worker_thread, NULL, TRUE, 0)) != NULL) {
740 threadpool->worker_creation_current_count += 1;
742 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, created %p, now = %d count = %d", mono_native_thread_id_get (), thread->tid, now, threadpool->worker_creation_current_count);
743 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
747 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: could not create thread", mono_native_thread_id_get ());
749 COUNTER_ATOMIC (counter, {
750 counter._.working --;
754 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
758 static void monitor_ensure_running (void);
761 worker_request (MonoDomain *domain)
763 ThreadPoolDomain *tpdomain;
766 g_assert (threadpool);
768 if (mono_runtime_is_shutting_down ())
771 mono_coop_mutex_lock (&threadpool->domains_lock);
773 /* synchronize check with worker_thread */
774 if (mono_domain_is_unloading (domain)) {
775 mono_coop_mutex_unlock (&threadpool->domains_lock);
779 tpdomain = domain_get (domain, TRUE);
781 tpdomain->outstanding_request ++;
783 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, domain = %p, outstanding_request = %d",
784 mono_native_thread_id_get (), tpdomain->domain, tpdomain->outstanding_request);
786 mono_coop_mutex_unlock (&threadpool->domains_lock);
788 if (threadpool->suspended)
791 monitor_ensure_running ();
793 if (worker_try_unpark ()) {
794 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, unparked", mono_native_thread_id_get ());
798 if (worker_try_create ()) {
799 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, created", mono_native_thread_id_get ());
803 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, failed", mono_native_thread_id_get ());
808 monitor_should_keep_running (void)
810 static gint64 last_should_keep_running = -1;
812 g_assert (monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || monitor_status == MONITOR_STATUS_REQUESTED);
814 if (InterlockedExchange (&monitor_status, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST) {
815 gboolean should_keep_running = TRUE, force_should_keep_running = FALSE;
817 if (mono_runtime_is_shutting_down ()) {
818 should_keep_running = FALSE;
820 mono_coop_mutex_lock (&threadpool->domains_lock);
821 if (!domain_any_has_request ())
822 should_keep_running = FALSE;
823 mono_coop_mutex_unlock (&threadpool->domains_lock);
825 if (!should_keep_running) {
826 if (last_should_keep_running == -1 || mono_100ns_ticks () - last_should_keep_running < MONITOR_MINIMAL_LIFETIME * 1000 * 10) {
827 should_keep_running = force_should_keep_running = TRUE;
832 if (should_keep_running) {
833 if (last_should_keep_running == -1 || !force_should_keep_running)
834 last_should_keep_running = mono_100ns_ticks ();
836 last_should_keep_running = -1;
837 if (InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_NOT_RUNNING, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST)
842 g_assert (monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || monitor_status == MONITOR_STATUS_REQUESTED);
848 monitor_sufficient_delay_since_last_dequeue (void)
852 g_assert (threadpool);
854 if (threadpool->cpu_usage < CPU_USAGE_LOW) {
855 threshold = MONITOR_INTERVAL;
857 ThreadPoolCounter counter;
858 counter.as_gint64 = COUNTER_READ();
859 threshold = counter._.max_working * MONITOR_INTERVAL * 2;
862 return mono_msec_ticks () >= threadpool->heuristic_last_dequeue + threshold;
865 static void hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition);
868 monitor_thread (void)
870 MonoInternalThread *current_thread = mono_thread_internal_current ();
873 mono_cpu_usage (threadpool->cpu_usage_state);
875 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, started", mono_native_thread_id_get ());
878 ThreadPoolCounter counter;
879 gboolean limit_worker_max_reached;
880 gint32 interval_left = MONITOR_INTERVAL;
881 gint32 awake = 0; /* number of spurious awakes we tolerate before doing a round of rebalancing */
883 g_assert (monitor_status != MONITOR_STATUS_NOT_RUNNING);
885 mono_gc_set_skip_thread (TRUE);
889 gboolean alerted = FALSE;
891 if (mono_runtime_is_shutting_down ())
894 ts = mono_msec_ticks ();
895 if (mono_thread_info_sleep (interval_left, &alerted) == 0)
897 interval_left -= mono_msec_ticks () - ts;
899 mono_gc_set_skip_thread (FALSE);
900 if ((current_thread->state & (ThreadState_StopRequested | ThreadState_SuspendRequested)) != 0)
901 mono_thread_interruption_checkpoint ();
902 mono_gc_set_skip_thread (TRUE);
903 } while (interval_left > 0 && ++awake < 10);
905 mono_gc_set_skip_thread (FALSE);
907 if (threadpool->suspended)
910 if (mono_runtime_is_shutting_down ())
913 mono_coop_mutex_lock (&threadpool->domains_lock);
914 if (!domain_any_has_request ()) {
915 mono_coop_mutex_unlock (&threadpool->domains_lock);
918 mono_coop_mutex_unlock (&threadpool->domains_lock);
920 threadpool->cpu_usage = mono_cpu_usage (threadpool->cpu_usage_state);
922 if (!monitor_sufficient_delay_since_last_dequeue ())
925 limit_worker_max_reached = FALSE;
927 COUNTER_ATOMIC (counter, {
928 if (counter._.max_working >= threadpool->limit_worker_max) {
929 limit_worker_max_reached = TRUE;
932 counter._.max_working ++;
935 if (limit_worker_max_reached)
938 hill_climbing_force_change (counter._.max_working, TRANSITION_STARVATION);
940 for (i = 0; i < 5; ++i) {
941 if (mono_runtime_is_shutting_down ())
944 if (worker_try_unpark ()) {
945 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, unparked", mono_native_thread_id_get ());
949 if (worker_try_create ()) {
950 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, created", mono_native_thread_id_get ());
954 } while (monitor_should_keep_running ());
956 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, finished", mono_native_thread_id_get ());
960 monitor_ensure_running (void)
963 switch (monitor_status) {
964 case MONITOR_STATUS_REQUESTED:
966 case MONITOR_STATUS_WAITING_FOR_REQUEST:
967 InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_WAITING_FOR_REQUEST);
969 case MONITOR_STATUS_NOT_RUNNING:
970 if (mono_runtime_is_shutting_down ())
972 if (InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_NOT_RUNNING) == MONITOR_STATUS_NOT_RUNNING) {
973 if (!mono_thread_create_internal (mono_get_root_domain (), monitor_thread, NULL, TRUE, SMALL_STACK))
974 monitor_status = MONITOR_STATUS_NOT_RUNNING;
978 default: g_assert_not_reached ();
984 hill_climbing_change_thread_count (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
986 ThreadPoolHillClimbing *hc;
988 g_assert (threadpool);
990 hc = &threadpool->heuristic_hill_climbing;
992 mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] hill climbing, change max number of threads %d", mono_native_thread_id_get (), new_thread_count);
994 hc->last_thread_count = new_thread_count;
995 hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
996 hc->elapsed_since_last_change = 0;
997 hc->completions_since_last_change = 0;
1001 hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
1003 ThreadPoolHillClimbing *hc;
1005 g_assert (threadpool);
1007 hc = &threadpool->heuristic_hill_climbing;
1009 if (new_thread_count != hc->last_thread_count) {
1010 hc->current_control_setting += new_thread_count - hc->last_thread_count;
1011 hill_climbing_change_thread_count (new_thread_count, transition);
1015 static double_complex
1016 hill_climbing_get_wave_component (gdouble *samples, guint sample_count, gdouble period)
1018 ThreadPoolHillClimbing *hc;
1019 gdouble w, cosine, sine, coeff, q0, q1, q2;
1022 g_assert (threadpool);
1023 g_assert (sample_count >= period);
1024 g_assert (period >= 2);
1026 hc = &threadpool->heuristic_hill_climbing;
1028 w = 2.0 * M_PI / period;
1031 coeff = 2.0 * cosine;
1034 for (i = 0; i < sample_count; ++i) {
1035 q0 = coeff * q1 - q2 + samples [(hc->total_samples - sample_count + i) % hc->samples_to_measure];
1040 return mono_double_complex_scalar_div (mono_double_complex_make (q1 - q2 * cosine, (q2 * sine)), ((gdouble)sample_count));
1044 hill_climbing_update (gint16 current_thread_count, guint32 sample_duration, gint32 completions, guint32 *adjustment_interval)
1046 ThreadPoolHillClimbing *hc;
1047 ThreadPoolHeuristicStateTransition transition;
1049 gdouble throughput_error_estimate;
1055 gint new_thread_wave_magnitude;
1056 gint new_thread_count;
1057 double_complex thread_wave_component;
1058 double_complex throughput_wave_component;
1059 double_complex ratio;
1061 g_assert (threadpool);
1062 g_assert (adjustment_interval);
1064 hc = &threadpool->heuristic_hill_climbing;
1066 /* If someone changed the thread count without telling us, update our records accordingly. */
1067 if (current_thread_count != hc->last_thread_count)
1068 hill_climbing_force_change (current_thread_count, TRANSITION_INITIALIZING);
1070 /* Update the cumulative stats for this thread count */
1071 hc->elapsed_since_last_change += sample_duration;
1072 hc->completions_since_last_change += completions;
1074 /* Add in any data we've already collected about this sample */
1075 sample_duration += hc->accumulated_sample_duration;
1076 completions += hc->accumulated_completion_count;
1078 /* We need to make sure we're collecting reasonably accurate data. Since we're just counting the end
1079 * of each work item, we are goinng to be missing some data about what really happened during the
1080 * sample interval. The count produced by each thread includes an initial work item that may have
1081 * started well before the start of the interval, and each thread may have been running some new
1082 * work item for some time before the end of the interval, which did not yet get counted. So
1083 * our count is going to be off by +/- threadCount workitems.
1085 * The exception is that the thread that reported to us last time definitely wasn't running any work
1086 * at that time, and the thread that's reporting now definitely isn't running a work item now. So
1087 * we really only need to consider threadCount-1 threads.
1089 * Thus the percent error in our count is +/- (threadCount-1)/numCompletions.
1091 * We cannot rely on the frequency-domain analysis we'll be doing later to filter out this error, because
1092 * of the way it accumulates over time. If this sample is off by, say, 33% in the negative direction,
1093 * then the next one likely will be too. The one after that will include the sum of the completions
1094 * we missed in the previous samples, and so will be 33% positive. So every three samples we'll have
1095 * two "low" samples and one "high" sample. This will appear as periodic variation right in the frequency
1096 * range we're targeting, which will not be filtered by the frequency-domain translation. */
1097 if (hc->total_samples > 0 && ((current_thread_count - 1.0) / completions) >= hc->max_sample_error) {
1098 /* Not accurate enough yet. Let's accumulate the data so
1099 * far, and tell the ThreadPool to collect a little more. */
1100 hc->accumulated_sample_duration = sample_duration;
1101 hc->accumulated_completion_count = completions;
1102 *adjustment_interval = 10;
1103 return current_thread_count;
1106 /* We've got enouugh data for our sample; reset our accumulators for next time. */
1107 hc->accumulated_sample_duration = 0;
1108 hc->accumulated_completion_count = 0;
1110 /* Add the current thread count and throughput sample to our history. */
1111 throughput = ((gdouble) completions) / sample_duration;
1113 sample_index = hc->total_samples % hc->samples_to_measure;
1114 hc->samples [sample_index] = throughput;
1115 hc->thread_counts [sample_index] = current_thread_count;
1116 hc->total_samples ++;
1118 /* Set up defaults for our metrics. */
1119 thread_wave_component = mono_double_complex_make(0, 0);
1120 throughput_wave_component = mono_double_complex_make(0, 0);
1121 throughput_error_estimate = 0;
1122 ratio = mono_double_complex_make(0, 0);
1125 transition = TRANSITION_WARMUP;
1127 /* How many samples will we use? It must be at least the three wave periods we're looking for, and it must also
1128 * be a whole multiple of the primary wave's period; otherwise the frequency we're looking for will fall between
1129 * two frequency bands in the Fourier analysis, and we won't be able to measure it accurately. */
1130 sample_count = ((gint) MIN (hc->total_samples - 1, hc->samples_to_measure) / hc->wave_period) * hc->wave_period;
1132 if (sample_count > hc->wave_period) {
1134 gdouble average_throughput;
1135 gdouble average_thread_count;
1136 gdouble sample_sum = 0;
1137 gdouble thread_sum = 0;
1139 /* Average the throughput and thread count samples, so we can scale the wave magnitudes later. */
1140 for (i = 0; i < sample_count; ++i) {
1141 guint j = (hc->total_samples - sample_count + i) % hc->samples_to_measure;
1142 sample_sum += hc->samples [j];
1143 thread_sum += hc->thread_counts [j];
1146 average_throughput = sample_sum / sample_count;
1147 average_thread_count = thread_sum / sample_count;
1149 if (average_throughput > 0 && average_thread_count > 0) {
1150 gdouble noise_for_confidence, adjacent_period_1, adjacent_period_2;
1152 /* Calculate the periods of the adjacent frequency bands we'll be using to
1153 * measure noise levels. We want the two adjacent Fourier frequency bands. */
1154 adjacent_period_1 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) + 1);
1155 adjacent_period_2 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) - 1);
1157 /* Get the the three different frequency components of the throughput (scaled by average
1158 * throughput). Our "error" estimate (the amount of noise that might be present in the
1159 * frequency band we're really interested in) is the average of the adjacent bands. */
1160 throughput_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, hc->wave_period), average_throughput);
1161 throughput_error_estimate = cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, adjacent_period_1), average_throughput));
1163 if (adjacent_period_2 <= sample_count) {
1164 throughput_error_estimate = MAX (throughput_error_estimate, cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (
1165 hc->samples, sample_count, adjacent_period_2), average_throughput)));
1168 /* Do the same for the thread counts, so we have something to compare to. We don't
1169 * measure thread count noise, because there is none; these are exact measurements. */
1170 thread_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->thread_counts, sample_count, hc->wave_period), average_thread_count);
1172 /* Update our moving average of the throughput noise. We'll use this
1173 * later as feedback to determine the new size of the thread wave. */
1174 if (hc->average_throughput_noise == 0) {
1175 hc->average_throughput_noise = throughput_error_estimate;
1177 hc->average_throughput_noise = (hc->throughput_error_smoothing_factor * throughput_error_estimate)
1178 + ((1.0 + hc->throughput_error_smoothing_factor) * hc->average_throughput_noise);
1181 if (cabs (thread_wave_component) > 0) {
1182 /* Adjust the throughput wave so it's centered around the target wave,
1183 * and then calculate the adjusted throughput/thread ratio. */
1184 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);
1185 transition = TRANSITION_CLIMBING_MOVE;
1187 ratio = mono_double_complex_make (0, 0);
1188 transition = TRANSITION_STABILIZING;
1191 noise_for_confidence = MAX (hc->average_throughput_noise, throughput_error_estimate);
1192 if (noise_for_confidence > 0) {
1193 confidence = cabs (thread_wave_component) / noise_for_confidence / hc->target_signal_to_noise_ratio;
1195 /* there is no noise! */
1201 /* We use just the real part of the complex ratio we just calculated. If the throughput signal
1202 * is exactly in phase with the thread signal, this will be the same as taking the magnitude of
1203 * the complex move and moving that far up. If they're 180 degrees out of phase, we'll move
1204 * backward (because this indicates that our changes are having the opposite of the intended effect).
1205 * If they're 90 degrees out of phase, we won't move at all, because we can't tell wether we're
1206 * having a negative or positive effect on throughput. */
1207 move = creal (ratio);
1208 move = CLAMP (move, -1.0, 1.0);
1210 /* Apply our confidence multiplier. */
1211 move *= CLAMP (confidence, -1.0, 1.0);
1213 /* Now apply non-linear gain, such that values around zero are attenuated, while higher values
1214 * are enhanced. This allows us to move quickly if we're far away from the target, but more slowly
1215 * if we're getting close, giving us rapid ramp-up without wild oscillations around the target. */
1216 gain = hc->max_change_per_second * sample_duration;
1217 move = pow (fabs (move), hc->gain_exponent) * (move >= 0.0 ? 1 : -1) * gain;
1218 move = MIN (move, hc->max_change_per_sample);
1220 /* If the result was positive, and CPU is > 95%, refuse the move. */
1221 if (move > 0.0 && threadpool->cpu_usage > CPU_USAGE_HIGH)
1224 /* Apply the move to our control setting. */
1225 hc->current_control_setting += move;
1227 /* Calculate the new thread wave magnitude, which is based on the moving average we've been keeping of the
1228 * throughput error. This average starts at zero, so we'll start with a nice safe little wave at first. */
1229 new_thread_wave_magnitude = (gint)(0.5 + (hc->current_control_setting * hc->average_throughput_noise
1230 * hc->target_signal_to_noise_ratio * hc->thread_magnitude_multiplier * 2.0));
1231 new_thread_wave_magnitude = CLAMP (new_thread_wave_magnitude, 1, hc->max_thread_wave_magnitude);
1233 /* Make sure our control setting is within the ThreadPool's limits. */
1234 hc->current_control_setting = CLAMP (hc->current_control_setting, threadpool->limit_worker_min, threadpool->limit_worker_max - new_thread_wave_magnitude);
1236 /* Calculate the new thread count (control setting + square wave). */
1237 new_thread_count = (gint)(hc->current_control_setting + new_thread_wave_magnitude * ((hc->total_samples / (hc->wave_period / 2)) % 2));
1239 /* Make sure the new thread count doesn't exceed the ThreadPool's limits. */
1240 new_thread_count = CLAMP (new_thread_count, threadpool->limit_worker_min, threadpool->limit_worker_max);
1242 if (new_thread_count != current_thread_count)
1243 hill_climbing_change_thread_count (new_thread_count, transition);
1245 if (creal (ratio) < 0.0 && new_thread_count == threadpool->limit_worker_min)
1246 *adjustment_interval = (gint)(0.5 + hc->current_sample_interval * (10.0 * MAX (-1.0 * creal (ratio), 1.0)));
1248 *adjustment_interval = hc->current_sample_interval;
1250 return new_thread_count;
1254 heuristic_notify_work_completed (void)
1256 g_assert (threadpool);
1258 InterlockedIncrement (&threadpool->heuristic_completions);
1259 threadpool->heuristic_last_dequeue = mono_msec_ticks ();
1263 heuristic_should_adjust (void)
1265 g_assert (threadpool);
1267 if (threadpool->heuristic_last_dequeue > threadpool->heuristic_last_adjustment + threadpool->heuristic_adjustment_interval) {
1268 ThreadPoolCounter counter;
1269 counter.as_gint64 = COUNTER_READ();
1270 if (counter._.working <= counter._.max_working)
1278 heuristic_adjust (void)
1280 g_assert (threadpool);
1282 if (mono_coop_mutex_trylock (&threadpool->heuristic_lock) == 0) {
1283 gint32 completions = InterlockedExchange (&threadpool->heuristic_completions, 0);
1284 guint32 sample_end = mono_msec_ticks ();
1285 guint32 sample_duration = sample_end - threadpool->heuristic_sample_start;
1287 if (sample_duration >= threadpool->heuristic_adjustment_interval / 2) {
1288 ThreadPoolCounter counter;
1289 gint16 new_thread_count;
1291 counter.as_gint64 = COUNTER_READ ();
1292 new_thread_count = hill_climbing_update (counter._.max_working, sample_duration, completions, &threadpool->heuristic_adjustment_interval);
1294 COUNTER_ATOMIC (counter, { counter._.max_working = new_thread_count; });
1296 if (new_thread_count > counter._.max_working)
1297 worker_request (mono_domain_get ());
1299 threadpool->heuristic_sample_start = sample_end;
1300 threadpool->heuristic_last_adjustment = mono_msec_ticks ();
1303 mono_coop_mutex_unlock (&threadpool->heuristic_lock);
1308 mono_threadpool_ms_cleanup (void)
1310 #ifndef DISABLE_SOCKETS
1311 mono_threadpool_ms_io_cleanup ();
1313 mono_lazy_cleanup (&status, cleanup);
1317 mono_threadpool_ms_begin_invoke (MonoDomain *domain, MonoObject *target, MonoMethod *method, gpointer *params)
1319 static MonoClass *async_call_klass = NULL;
1321 MonoMethodMessage *message;
1322 MonoAsyncResult *async_result;
1323 MonoAsyncCall *async_call;
1324 MonoDelegate *async_callback = NULL;
1325 MonoObject *state = NULL;
1327 if (!async_call_klass)
1328 async_call_klass = mono_class_load_from_name (mono_defaults.corlib, "System", "MonoAsyncCall");
1330 mono_lazy_initialize (&status, initialize);
1332 message = mono_method_call_message_new (method, params, mono_get_delegate_invoke (method->klass), (params != NULL) ? (&async_callback) : NULL, (params != NULL) ? (&state) : NULL);
1334 async_call = (MonoAsyncCall*) mono_object_new_checked (domain, async_call_klass, &error);
1335 mono_error_raise_exception (&error); /* FIXME don't raise here */
1337 MONO_OBJECT_SETREF (async_call, msg, message);
1338 MONO_OBJECT_SETREF (async_call, state, state);
1340 if (async_callback) {
1341 MONO_OBJECT_SETREF (async_call, cb_method, mono_get_delegate_invoke (((MonoObject*) async_callback)->vtable->klass));
1342 MONO_OBJECT_SETREF (async_call, cb_target, async_callback);
1345 async_result = mono_async_result_new (domain, NULL, async_call->state, NULL, (MonoObject*) async_call);
1346 MONO_OBJECT_SETREF (async_result, async_delegate, target);
1348 mono_threadpool_ms_enqueue_work_item (domain, (MonoObject*) async_result);
1350 return async_result;
1354 mono_threadpool_ms_end_invoke (MonoAsyncResult *ares, MonoArray **out_args, MonoObject **exc)
1359 g_assert (out_args);
1364 /* check if already finished */
1365 mono_monitor_enter ((MonoObject*) ares);
1367 if (ares->endinvoke_called) {
1368 *exc = (MonoObject*) mono_get_exception_invalid_operation (NULL);
1369 mono_monitor_exit ((MonoObject*) ares);
1373 ares->endinvoke_called = 1;
1375 /* wait until we are really finished */
1376 if (ares->completed) {
1377 mono_monitor_exit ((MonoObject *) ares);
1379 gpointer wait_event;
1381 wait_event = mono_wait_handle_get_handle ((MonoWaitHandle*) ares->handle);
1383 wait_event = CreateEvent (NULL, TRUE, FALSE, NULL);
1384 g_assert(wait_event);
1385 MONO_OBJECT_SETREF (ares, handle, (MonoObject*) mono_wait_handle_new (mono_object_domain (ares), wait_event));
1387 mono_monitor_exit ((MonoObject*) ares);
1388 MONO_PREPARE_BLOCKING;
1389 WaitForSingleObjectEx (wait_event, INFINITE, TRUE);
1390 MONO_FINISH_BLOCKING;
1393 ac = (MonoAsyncCall*) ares->object_data;
1396 *exc = ac->msg->exc; /* FIXME: GC add write barrier */
1397 *out_args = ac->out_args;
1402 mono_threadpool_ms_remove_domain_jobs (MonoDomain *domain, int timeout)
1404 gboolean res = TRUE;
1409 g_assert (timeout >= -1);
1411 g_assert (mono_domain_is_unloading (domain));
1414 start = mono_msec_ticks ();
1416 #ifndef DISABLE_SOCKETS
1417 mono_threadpool_ms_io_remove_domain_jobs (domain);
1418 if (timeout != -1) {
1419 timeout -= mono_msec_ticks () - start;
1426 * There might be some threads out that could be about to execute stuff from the given domain.
1427 * We avoid that by setting up a semaphore to be pulsed by the thread that reaches zero.
1429 sem = domain->cleanup_semaphore = CreateSemaphore (NULL, 0, 1, NULL);
1432 * The memory barrier here is required to have global ordering between assigning to cleanup_semaphone
1433 * and reading threadpool_jobs. Otherwise this thread could read a stale version of threadpool_jobs
1436 mono_memory_write_barrier ();
1438 while (domain->threadpool_jobs) {
1439 MONO_PREPARE_BLOCKING;
1440 WaitForSingleObject (sem, timeout);
1441 MONO_FINISH_BLOCKING;
1442 if (timeout != -1) {
1443 timeout -= mono_msec_ticks () - start;
1451 domain->cleanup_semaphore = NULL;
1458 mono_threadpool_ms_suspend (void)
1461 threadpool->suspended = TRUE;
1465 mono_threadpool_ms_resume (void)
1468 threadpool->suspended = FALSE;
1472 ves_icall_System_Threading_ThreadPool_GetAvailableThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1474 ThreadPoolCounter counter;
1476 if (!worker_threads || !completion_port_threads)
1479 mono_lazy_initialize (&status, initialize);
1481 counter.as_gint64 = COUNTER_READ ();
1483 *worker_threads = MAX (0, threadpool->limit_worker_max - counter._.active);
1484 *completion_port_threads = threadpool->limit_io_max;
1488 ves_icall_System_Threading_ThreadPool_GetMinThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1490 if (!worker_threads || !completion_port_threads)
1493 mono_lazy_initialize (&status, initialize);
1495 *worker_threads = threadpool->limit_worker_min;
1496 *completion_port_threads = threadpool->limit_io_min;
1500 ves_icall_System_Threading_ThreadPool_GetMaxThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1502 if (!worker_threads || !completion_port_threads)
1505 mono_lazy_initialize (&status, initialize);
1507 *worker_threads = threadpool->limit_worker_max;
1508 *completion_port_threads = threadpool->limit_io_max;
1512 ves_icall_System_Threading_ThreadPool_SetMinThreadsNative (gint32 worker_threads, gint32 completion_port_threads)
1514 mono_lazy_initialize (&status, initialize);
1516 if (worker_threads <= 0 || worker_threads > threadpool->limit_worker_max)
1518 if (completion_port_threads <= 0 || completion_port_threads > threadpool->limit_io_max)
1521 threadpool->limit_worker_min = worker_threads;
1522 threadpool->limit_io_min = completion_port_threads;
1528 ves_icall_System_Threading_ThreadPool_SetMaxThreadsNative (gint32 worker_threads, gint32 completion_port_threads)
1530 gint cpu_count = mono_cpu_count ();
1532 mono_lazy_initialize (&status, initialize);
1534 if (worker_threads < threadpool->limit_worker_min || worker_threads < cpu_count)
1536 if (completion_port_threads < threadpool->limit_io_min || completion_port_threads < cpu_count)
1539 threadpool->limit_worker_max = worker_threads;
1540 threadpool->limit_io_max = completion_port_threads;
1546 ves_icall_System_Threading_ThreadPool_InitializeVMTp (MonoBoolean *enable_worker_tracking)
1548 if (enable_worker_tracking) {
1549 // TODO implement some kind of switch to have the possibily to use it
1550 *enable_worker_tracking = FALSE;
1553 mono_lazy_initialize (&status, initialize);
1557 ves_icall_System_Threading_ThreadPool_NotifyWorkItemComplete (void)
1559 ThreadPoolCounter counter;
1561 if (mono_domain_is_unloading (mono_domain_get ()) || mono_runtime_is_shutting_down ())
1564 heuristic_notify_work_completed ();
1566 if (heuristic_should_adjust ())
1567 heuristic_adjust ();
1569 counter.as_gint64 = COUNTER_READ ();
1570 return counter._.working <= counter._.max_working;
1574 ves_icall_System_Threading_ThreadPool_NotifyWorkItemProgressNative (void)
1576 heuristic_notify_work_completed ();
1578 if (heuristic_should_adjust ())
1579 heuristic_adjust ();
1583 ves_icall_System_Threading_ThreadPool_ReportThreadStatus (MonoBoolean is_working)
1586 mono_raise_exception (mono_get_exception_not_implemented (NULL));
1590 ves_icall_System_Threading_ThreadPool_RequestWorkerThread (void)
1592 return worker_request (mono_domain_get ());
1595 MonoBoolean G_GNUC_UNUSED
1596 ves_icall_System_Threading_ThreadPool_PostQueuedCompletionStatus (MonoNativeOverlapped *native_overlapped)
1598 /* This copy the behavior of the current Mono implementation */
1599 mono_raise_exception (mono_get_exception_not_implemented (NULL));
1603 MonoBoolean G_GNUC_UNUSED
1604 ves_icall_System_Threading_ThreadPool_BindIOCompletionCallbackNative (gpointer file_handle)
1606 /* This copy the behavior of the current Mono implementation */
1610 MonoBoolean G_GNUC_UNUSED
1611 ves_icall_System_Threading_ThreadPool_IsThreadPoolHosted (void)