2 * threadpool-ms.c: Microsoft threadpool runtime support
5 * Ludovic Henry (ludovic.henry@xamarin.com)
7 * Copyright 2015 Xamarin, Inc (http://www.xamarin.com)
8 * Licensed under the MIT license. See LICENSE file in the project root for full license information.
12 // Copyright (c) Microsoft. All rights reserved.
13 // Licensed under the MIT license. See LICENSE file in the project root for full license information.
16 // - src/vm/comthreadpool.cpp
17 // - src/vm/win32threadpoolcpp
18 // - src/vm/threadpoolrequest.cpp
19 // - src/vm/hillclimbing.cpp
21 // Ported from C++ to C and adjusted to Mono runtime
24 #define _USE_MATH_DEFINES // needed by MSVC to define math constants
29 #include <mono/metadata/class-internals.h>
30 #include <mono/metadata/exception.h>
31 #include <mono/metadata/gc-internals.h>
32 #include <mono/metadata/object.h>
33 #include <mono/metadata/object-internals.h>
34 #include <mono/metadata/threadpool-ms.h>
35 #include <mono/metadata/threadpool-ms-io.h>
36 #include <mono/utils/atomic.h>
37 #include <mono/utils/mono-compiler.h>
38 #include <mono/utils/mono-complex.h>
39 #include <mono/utils/mono-lazy-init.h>
40 #include <mono/utils/mono-logger.h>
41 #include <mono/utils/mono-logger-internals.h>
42 #include <mono/utils/mono-proclib.h>
43 #include <mono/utils/mono-threads.h>
44 #include <mono/utils/mono-time.h>
45 #include <mono/utils/mono-rand.h>
47 #define CPU_USAGE_LOW 80
48 #define CPU_USAGE_HIGH 95
50 #define MONITOR_INTERVAL 500 // ms
51 #define MONITOR_MINIMAL_LIFETIME 60 * 1000 // ms
53 #define WORKER_CREATION_MAX_PER_SEC 10
55 /* The exponent to apply to the gain. 1.0 means to use linear gain,
56 * higher values will enhance large moves and damp small ones.
58 #define HILL_CLIMBING_GAIN_EXPONENT 2.0
60 /* The 'cost' of a thread. 0 means drive for increased throughput regardless
61 * of thread count, higher values bias more against higher thread counts.
63 #define HILL_CLIMBING_BIAS 0.15
65 #define HILL_CLIMBING_WAVE_PERIOD 4
66 #define HILL_CLIMBING_MAX_WAVE_MAGNITUDE 20
67 #define HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER 1.0
68 #define HILL_CLIMBING_WAVE_HISTORY_SIZE 8
69 #define HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO 3.0
70 #define HILL_CLIMBING_MAX_CHANGE_PER_SECOND 4
71 #define HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE 20
72 #define HILL_CLIMBING_SAMPLE_INTERVAL_LOW 10
73 #define HILL_CLIMBING_SAMPLE_INTERVAL_HIGH 200
74 #define HILL_CLIMBING_ERROR_SMOOTHING_FACTOR 0.01
75 #define HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT 0.15
79 gint16 max_working; /* determined by heuristic */
80 gint16 active; /* executing worker_thread */
81 gint16 working; /* actively executing worker_thread, not parked */
82 gint16 parked; /* parked */
89 gint32 outstanding_request;
92 typedef MonoInternalThread ThreadPoolWorkingThread;
96 gint32 samples_to_measure;
97 gdouble target_throughput_ratio;
98 gdouble target_signal_to_noise_ratio;
99 gdouble max_change_per_second;
100 gdouble max_change_per_sample;
101 gint32 max_thread_wave_magnitude;
102 gint32 sample_interval_low;
103 gdouble thread_magnitude_multiplier;
104 gint32 sample_interval_high;
105 gdouble throughput_error_smoothing_factor;
106 gdouble gain_exponent;
107 gdouble max_sample_error;
109 gdouble current_control_setting;
110 gint64 total_samples;
111 gint16 last_thread_count;
112 gdouble elapsed_since_last_change;
113 gdouble completions_since_last_change;
115 gdouble average_throughput_noise;
118 gdouble *thread_counts;
120 guint32 current_sample_interval;
121 gpointer random_interval_generator;
123 gint32 accumulated_completion_count;
124 gdouble accumulated_sample_duration;
125 } ThreadPoolHillClimbing;
128 ThreadPoolCounter counters;
130 GPtrArray *domains; // ThreadPoolDomain* []
131 MonoCoopMutex domains_lock;
133 GPtrArray *working_threads; // ThreadPoolWorkingThread* []
134 gint32 parked_threads_count;
135 MonoCoopCond parked_threads_cond;
136 MonoCoopMutex active_threads_lock; /* protect access to working_threads and parked_threads */
138 guint32 worker_creation_current_second;
139 guint32 worker_creation_current_count;
140 MonoCoopMutex worker_creation_lock;
142 gint32 heuristic_completions;
143 guint32 heuristic_sample_start;
144 guint32 heuristic_last_dequeue; // ms
145 guint32 heuristic_last_adjustment; // ms
146 guint32 heuristic_adjustment_interval; // ms
147 ThreadPoolHillClimbing heuristic_hill_climbing;
148 MonoCoopMutex heuristic_lock;
150 gint32 limit_worker_min;
151 gint32 limit_worker_max;
155 MonoCpuUsageState *cpu_usage_state;
158 /* suspended by the debugger */
164 TRANSITION_INITIALIZING,
165 TRANSITION_RANDOM_MOVE,
166 TRANSITION_CLIMBING_MOVE,
167 TRANSITION_CHANGE_POINT,
168 TRANSITION_STABILIZING,
169 TRANSITION_STARVATION,
170 TRANSITION_THREAD_TIMED_OUT,
171 TRANSITION_UNDEFINED,
172 } ThreadPoolHeuristicStateTransition;
174 static mono_lazy_init_t status = MONO_LAZY_INIT_STATUS_NOT_INITIALIZED;
177 MONITOR_STATUS_REQUESTED,
178 MONITOR_STATUS_WAITING_FOR_REQUEST,
179 MONITOR_STATUS_NOT_RUNNING,
182 static gint32 monitor_status = MONITOR_STATUS_NOT_RUNNING;
184 static ThreadPool* threadpool;
186 #define COUNTER_CHECK(counter) \
188 g_assert (counter._.max_working > 0); \
189 g_assert (counter._.working >= 0); \
190 g_assert (counter._.active >= 0); \
193 #define COUNTER_READ() (InterlockedRead64 (&threadpool->counters.as_gint64))
195 #define COUNTER_ATOMIC(var,block) \
197 ThreadPoolCounter __old; \
199 g_assert (threadpool); \
200 __old.as_gint64 = COUNTER_READ (); \
203 COUNTER_CHECK (var); \
204 } while (InterlockedCompareExchange64 (&threadpool->counters.as_gint64, (var).as_gint64, __old.as_gint64) != __old.as_gint64); \
207 #define COUNTER_TRY_ATOMIC(res,var,block) \
209 ThreadPoolCounter __old; \
211 g_assert (threadpool); \
212 __old.as_gint64 = COUNTER_READ (); \
216 COUNTER_CHECK (var); \
217 (res) = InterlockedCompareExchange64 (&threadpool->counters.as_gint64, (var).as_gint64, __old.as_gint64) == __old.as_gint64; \
225 return mono_rand_init (NULL, 0);
229 rand_next (gpointer *handle, guint32 min, guint32 max)
233 mono_rand_try_get_uint32 (handle, &val, min, max, &error);
234 // FIXME handle error
235 mono_error_assert_ok (&error);
240 rand_free (gpointer handle)
242 mono_rand_close (handle);
248 ThreadPoolHillClimbing *hc;
249 const char *threads_per_cpu_env;
250 gint threads_per_cpu;
253 g_assert (!threadpool);
254 threadpool = g_new0 (ThreadPool, 1);
255 g_assert (threadpool);
257 threadpool->domains = g_ptr_array_new ();
258 mono_coop_mutex_init (&threadpool->domains_lock);
260 threadpool->parked_threads_count = 0;
261 mono_coop_cond_init (&threadpool->parked_threads_cond);
262 threadpool->working_threads = g_ptr_array_new ();
263 mono_coop_mutex_init (&threadpool->active_threads_lock);
265 threadpool->worker_creation_current_second = -1;
266 mono_coop_mutex_init (&threadpool->worker_creation_lock);
268 threadpool->heuristic_adjustment_interval = 10;
269 mono_coop_mutex_init (&threadpool->heuristic_lock);
273 hc = &threadpool->heuristic_hill_climbing;
275 hc->wave_period = HILL_CLIMBING_WAVE_PERIOD;
276 hc->max_thread_wave_magnitude = HILL_CLIMBING_MAX_WAVE_MAGNITUDE;
277 hc->thread_magnitude_multiplier = (gdouble) HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER;
278 hc->samples_to_measure = hc->wave_period * HILL_CLIMBING_WAVE_HISTORY_SIZE;
279 hc->target_throughput_ratio = (gdouble) HILL_CLIMBING_BIAS;
280 hc->target_signal_to_noise_ratio = (gdouble) HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO;
281 hc->max_change_per_second = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SECOND;
282 hc->max_change_per_sample = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE;
283 hc->sample_interval_low = HILL_CLIMBING_SAMPLE_INTERVAL_LOW;
284 hc->sample_interval_high = HILL_CLIMBING_SAMPLE_INTERVAL_HIGH;
285 hc->throughput_error_smoothing_factor = (gdouble) HILL_CLIMBING_ERROR_SMOOTHING_FACTOR;
286 hc->gain_exponent = (gdouble) HILL_CLIMBING_GAIN_EXPONENT;
287 hc->max_sample_error = (gdouble) HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT;
288 hc->current_control_setting = 0;
289 hc->total_samples = 0;
290 hc->last_thread_count = 0;
291 hc->average_throughput_noise = 0;
292 hc->elapsed_since_last_change = 0;
293 hc->accumulated_completion_count = 0;
294 hc->accumulated_sample_duration = 0;
295 hc->samples = g_new0 (gdouble, hc->samples_to_measure);
296 hc->thread_counts = g_new0 (gdouble, hc->samples_to_measure);
297 hc->random_interval_generator = rand_create ();
298 hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
300 if (!(threads_per_cpu_env = g_getenv ("MONO_THREADS_PER_CPU")))
303 threads_per_cpu = CLAMP (atoi (threads_per_cpu_env), 1, 50);
305 threads_count = mono_cpu_count () * threads_per_cpu;
307 threadpool->limit_worker_min = threadpool->limit_io_min = threads_count;
309 #if defined (PLATFORM_ANDROID) || defined (HOST_IOS)
310 threadpool->limit_worker_max = threadpool->limit_io_max = CLAMP (threads_count * 100, MIN (threads_count, 200), MAX (threads_count, 200));
312 threadpool->limit_worker_max = threadpool->limit_io_max = threads_count * 100;
315 threadpool->counters._.max_working = threadpool->limit_worker_min;
317 threadpool->cpu_usage_state = g_new0 (MonoCpuUsageState, 1);
319 threadpool->suspended = FALSE;
322 static void worker_kill (ThreadPoolWorkingThread *thread);
329 /* we make the assumption along the code that we are
330 * cleaning up only if the runtime is shutting down */
331 g_assert (mono_runtime_is_shutting_down ());
333 while (monitor_status != MONITOR_STATUS_NOT_RUNNING)
334 mono_thread_info_sleep (1, NULL);
336 mono_coop_mutex_lock (&threadpool->active_threads_lock);
338 /* stop all threadpool->working_threads */
339 for (i = 0; i < threadpool->working_threads->len; ++i)
340 worker_kill ((ThreadPoolWorkingThread*) g_ptr_array_index (threadpool->working_threads, i));
342 /* unpark all threadpool->parked_threads */
343 mono_coop_cond_broadcast (&threadpool->parked_threads_cond);
345 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
349 mono_threadpool_ms_enqueue_work_item (MonoDomain *domain, MonoObject *work_item, MonoError *error)
351 static MonoClass *threadpool_class = NULL;
352 static MonoMethod *unsafe_queue_custom_work_item_method = NULL;
353 MonoDomain *current_domain;
357 mono_error_init (error);
358 g_assert (work_item);
360 if (!threadpool_class)
361 threadpool_class = mono_class_load_from_name (mono_defaults.corlib, "System.Threading", "ThreadPool");
363 if (!unsafe_queue_custom_work_item_method)
364 unsafe_queue_custom_work_item_method = mono_class_get_method_from_name (threadpool_class, "UnsafeQueueCustomWorkItem", 2);
365 g_assert (unsafe_queue_custom_work_item_method);
369 args [0] = (gpointer) work_item;
370 args [1] = (gpointer) &f;
372 current_domain = mono_domain_get ();
373 if (current_domain == domain) {
374 mono_runtime_invoke_checked (unsafe_queue_custom_work_item_method, NULL, args, error);
375 return_val_if_nok (error, FALSE);
377 mono_thread_push_appdomain_ref (domain);
378 if (mono_domain_set (domain, FALSE)) {
379 mono_runtime_invoke_checked (unsafe_queue_custom_work_item_method, NULL, args, error);
380 if (!is_ok (error)) {
381 mono_thread_pop_appdomain_ref ();
384 mono_domain_set (current_domain, TRUE);
386 mono_thread_pop_appdomain_ref ();
391 /* LOCKING: threadpool->domains_lock must be held */
393 domain_add (ThreadPoolDomain *tpdomain)
399 len = threadpool->domains->len;
400 for (i = 0; i < len; ++i) {
401 if (g_ptr_array_index (threadpool->domains, i) == tpdomain)
406 g_ptr_array_add (threadpool->domains, tpdomain);
409 /* LOCKING: threadpool->domains_lock must be held */
411 domain_remove (ThreadPoolDomain *tpdomain)
414 return g_ptr_array_remove (threadpool->domains, tpdomain);
417 /* LOCKING: threadpool->domains_lock must be held */
418 static ThreadPoolDomain *
419 domain_get (MonoDomain *domain, gboolean create)
421 ThreadPoolDomain *tpdomain = NULL;
426 for (i = 0; i < threadpool->domains->len; ++i) {
427 tpdomain = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i);
428 if (tpdomain->domain == domain)
433 tpdomain = g_new0 (ThreadPoolDomain, 1);
434 tpdomain->domain = domain;
435 domain_add (tpdomain);
442 domain_free (ThreadPoolDomain *tpdomain)
447 /* LOCKING: threadpool->domains_lock must be held */
449 domain_any_has_request (void)
453 for (i = 0; i < threadpool->domains->len; ++i) {
454 ThreadPoolDomain *tmp = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i);
455 if (tmp->outstanding_request > 0)
462 /* LOCKING: threadpool->domains_lock must be held */
463 static ThreadPoolDomain *
464 domain_get_next (ThreadPoolDomain *current)
466 ThreadPoolDomain *tpdomain = NULL;
469 len = threadpool->domains->len;
471 guint i, current_idx = -1;
473 for (i = 0; i < len; ++i) {
474 if (current == g_ptr_array_index (threadpool->domains, i)) {
479 g_assert (current_idx >= 0);
481 for (i = current_idx + 1; i < len + current_idx + 1; ++i) {
482 ThreadPoolDomain *tmp = (ThreadPoolDomain *)g_ptr_array_index (threadpool->domains, i % len);
483 if (tmp->outstanding_request > 0) {
494 worker_wait_interrupt (gpointer data)
496 mono_coop_mutex_lock (&threadpool->active_threads_lock);
497 mono_coop_cond_signal (&threadpool->parked_threads_cond);
498 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
501 /* return TRUE if timeout, FALSE otherwise (worker unpark or interrupt) */
505 gboolean timeout = FALSE;
507 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] current worker parking", mono_native_thread_id_get ());
509 mono_gc_set_skip_thread (TRUE);
511 mono_coop_mutex_lock (&threadpool->active_threads_lock);
513 if (!mono_runtime_is_shutting_down ()) {
514 static gpointer rand_handle = NULL;
515 MonoInternalThread *thread_internal;
516 gboolean interrupted = FALSE;
519 rand_handle = rand_create ();
520 g_assert (rand_handle);
522 thread_internal = mono_thread_internal_current ();
523 g_assert (thread_internal);
525 threadpool->parked_threads_count += 1;
526 g_ptr_array_remove_fast (threadpool->working_threads, thread_internal);
528 mono_thread_info_install_interrupt (worker_wait_interrupt, NULL, &interrupted);
532 if (mono_coop_cond_timedwait (&threadpool->parked_threads_cond, &threadpool->active_threads_lock, rand_next (&rand_handle, 5 * 1000, 60 * 1000)) != 0)
535 mono_thread_info_uninstall_interrupt (&interrupted);
538 g_ptr_array_add (threadpool->working_threads, thread_internal);
539 threadpool->parked_threads_count -= 1;
542 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
544 mono_gc_set_skip_thread (FALSE);
546 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] current worker unparking, timeout? %s", mono_native_thread_id_get (), timeout ? "yes" : "no");
552 worker_try_unpark (void)
554 gboolean res = FALSE;
556 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker", mono_native_thread_id_get ());
558 mono_coop_mutex_lock (&threadpool->active_threads_lock);
559 if (threadpool->parked_threads_count > 0) {
560 mono_coop_cond_signal (&threadpool->parked_threads_cond);
563 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
565 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker, success? %s", mono_native_thread_id_get (), res ? "yes" : "no");
571 worker_kill (ThreadPoolWorkingThread *thread)
573 if (thread == mono_thread_internal_current ())
576 mono_thread_internal_stop ((MonoInternalThread*) thread);
580 worker_thread (gpointer data)
583 MonoInternalThread *thread;
584 ThreadPoolDomain *tpdomain, *previous_tpdomain;
585 ThreadPoolCounter counter;
586 gboolean retire = FALSE;
588 mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker starting", mono_native_thread_id_get ());
590 g_assert (threadpool);
592 thread = mono_thread_internal_current ();
595 mono_thread_set_name_internal (thread, mono_string_new (mono_domain_get (), "Threadpool worker"), FALSE);
597 mono_coop_mutex_lock (&threadpool->active_threads_lock);
598 g_ptr_array_add (threadpool->working_threads, thread);
599 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
601 previous_tpdomain = NULL;
603 mono_coop_mutex_lock (&threadpool->domains_lock);
605 while (!mono_runtime_is_shutting_down ()) {
608 if ((thread->state & (ThreadState_StopRequested | ThreadState_SuspendRequested)) != 0) {
609 mono_coop_mutex_unlock (&threadpool->domains_lock);
610 mono_thread_interruption_checkpoint ();
611 mono_coop_mutex_lock (&threadpool->domains_lock);
614 if (retire || !(tpdomain = domain_get_next (previous_tpdomain))) {
617 COUNTER_ATOMIC (counter, {
618 counter._.working --;
622 mono_coop_mutex_unlock (&threadpool->domains_lock);
623 timeout = worker_park ();
624 mono_coop_mutex_lock (&threadpool->domains_lock);
626 COUNTER_ATOMIC (counter, {
627 counter._.working ++;
640 tpdomain->outstanding_request --;
641 g_assert (tpdomain->outstanding_request >= 0);
643 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker running in domain %p",
644 mono_native_thread_id_get (), tpdomain->domain, tpdomain->outstanding_request);
646 g_assert (tpdomain->domain);
647 g_assert (tpdomain->domain->threadpool_jobs >= 0);
648 tpdomain->domain->threadpool_jobs ++;
650 mono_coop_mutex_unlock (&threadpool->domains_lock);
652 mono_thread_push_appdomain_ref (tpdomain->domain);
653 if (mono_domain_set (tpdomain->domain, FALSE)) {
654 MonoObject *exc = NULL, *res;
656 res = mono_runtime_try_invoke (mono_defaults.threadpool_perform_wait_callback_method, NULL, NULL, &exc, &error);
657 if (exc || !mono_error_ok(&error)) {
659 exc = (MonoObject *) mono_error_convert_to_exception (&error);
661 mono_error_cleanup (&error);
662 mono_thread_internal_unhandled_exception (exc);
663 } else if (res && *(MonoBoolean*) mono_object_unbox (res) == FALSE)
666 mono_thread_clr_state (thread, (MonoThreadState)~ThreadState_Background);
667 if (!mono_thread_test_state (thread , ThreadState_Background))
668 ves_icall_System_Threading_Thread_SetState (thread, ThreadState_Background);
670 mono_domain_set (mono_get_root_domain (), TRUE);
672 mono_thread_pop_appdomain_ref ();
674 mono_coop_mutex_lock (&threadpool->domains_lock);
676 tpdomain->domain->threadpool_jobs --;
677 g_assert (tpdomain->domain->threadpool_jobs >= 0);
679 if (tpdomain->domain->threadpool_jobs == 0 && mono_domain_is_unloading (tpdomain->domain)) {
680 gboolean removed = domain_remove (tpdomain);
682 if (tpdomain->domain->cleanup_semaphore)
683 ReleaseSemaphore (tpdomain->domain->cleanup_semaphore, 1, NULL);
684 domain_free (tpdomain);
688 previous_tpdomain = tpdomain;
691 mono_coop_mutex_unlock (&threadpool->domains_lock);
693 mono_coop_mutex_lock (&threadpool->active_threads_lock);
694 g_ptr_array_remove_fast (threadpool->working_threads, thread);
695 mono_coop_mutex_unlock (&threadpool->active_threads_lock);
697 COUNTER_ATOMIC (counter, {
702 mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker finishing", mono_native_thread_id_get ());
706 worker_try_create (void)
708 ThreadPoolCounter counter;
709 MonoInternalThread *thread;
712 mono_coop_mutex_lock (&threadpool->worker_creation_lock);
714 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker", mono_native_thread_id_get ());
716 if ((now = mono_100ns_ticks () / 10 / 1000 / 1000) == 0) {
717 g_warning ("failed to get 100ns ticks");
719 if (threadpool->worker_creation_current_second != now) {
720 threadpool->worker_creation_current_second = now;
721 threadpool->worker_creation_current_count = 0;
723 g_assert (threadpool->worker_creation_current_count <= WORKER_CREATION_MAX_PER_SEC);
724 if (threadpool->worker_creation_current_count == WORKER_CREATION_MAX_PER_SEC) {
725 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",
726 mono_native_thread_id_get (), threadpool->worker_creation_current_count);
727 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
733 COUNTER_ATOMIC (counter, {
734 if (counter._.working >= counter._.max_working) {
735 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: maximum number of working threads reached",
736 mono_native_thread_id_get ());
737 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
740 counter._.working ++;
744 if ((thread = mono_thread_create_internal (mono_get_root_domain (), worker_thread, NULL, TRUE, 0)) != NULL) {
745 threadpool->worker_creation_current_count += 1;
747 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);
748 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
752 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: could not create thread", mono_native_thread_id_get ());
754 COUNTER_ATOMIC (counter, {
755 counter._.working --;
759 mono_coop_mutex_unlock (&threadpool->worker_creation_lock);
763 static void monitor_ensure_running (void);
766 worker_request (MonoDomain *domain)
768 ThreadPoolDomain *tpdomain;
771 g_assert (threadpool);
773 if (mono_runtime_is_shutting_down ())
776 mono_coop_mutex_lock (&threadpool->domains_lock);
778 /* synchronize check with worker_thread */
779 if (mono_domain_is_unloading (domain)) {
780 mono_coop_mutex_unlock (&threadpool->domains_lock);
784 tpdomain = domain_get (domain, TRUE);
786 tpdomain->outstanding_request ++;
788 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, domain = %p, outstanding_request = %d",
789 mono_native_thread_id_get (), tpdomain->domain, tpdomain->outstanding_request);
791 mono_coop_mutex_unlock (&threadpool->domains_lock);
793 if (threadpool->suspended)
796 monitor_ensure_running ();
798 if (worker_try_unpark ()) {
799 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, unparked", mono_native_thread_id_get ());
803 if (worker_try_create ()) {
804 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, created", mono_native_thread_id_get ());
808 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, failed", mono_native_thread_id_get ());
813 monitor_should_keep_running (void)
815 static gint64 last_should_keep_running = -1;
817 g_assert (monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || monitor_status == MONITOR_STATUS_REQUESTED);
819 if (InterlockedExchange (&monitor_status, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST) {
820 gboolean should_keep_running = TRUE, force_should_keep_running = FALSE;
822 if (mono_runtime_is_shutting_down ()) {
823 should_keep_running = FALSE;
825 mono_coop_mutex_lock (&threadpool->domains_lock);
826 if (!domain_any_has_request ())
827 should_keep_running = FALSE;
828 mono_coop_mutex_unlock (&threadpool->domains_lock);
830 if (!should_keep_running) {
831 if (last_should_keep_running == -1 || mono_100ns_ticks () - last_should_keep_running < MONITOR_MINIMAL_LIFETIME * 1000 * 10) {
832 should_keep_running = force_should_keep_running = TRUE;
837 if (should_keep_running) {
838 if (last_should_keep_running == -1 || !force_should_keep_running)
839 last_should_keep_running = mono_100ns_ticks ();
841 last_should_keep_running = -1;
842 if (InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_NOT_RUNNING, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST)
847 g_assert (monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || monitor_status == MONITOR_STATUS_REQUESTED);
853 monitor_sufficient_delay_since_last_dequeue (void)
857 g_assert (threadpool);
859 if (threadpool->cpu_usage < CPU_USAGE_LOW) {
860 threshold = MONITOR_INTERVAL;
862 ThreadPoolCounter counter;
863 counter.as_gint64 = COUNTER_READ();
864 threshold = counter._.max_working * MONITOR_INTERVAL * 2;
867 return mono_msec_ticks () >= threadpool->heuristic_last_dequeue + threshold;
870 static void hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition);
873 monitor_thread (void)
875 MonoInternalThread *current_thread = mono_thread_internal_current ();
878 mono_cpu_usage (threadpool->cpu_usage_state);
880 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, started", mono_native_thread_id_get ());
883 ThreadPoolCounter counter;
884 gboolean limit_worker_max_reached;
885 gint32 interval_left = MONITOR_INTERVAL;
886 gint32 awake = 0; /* number of spurious awakes we tolerate before doing a round of rebalancing */
888 g_assert (monitor_status != MONITOR_STATUS_NOT_RUNNING);
890 mono_gc_set_skip_thread (TRUE);
894 gboolean alerted = FALSE;
896 if (mono_runtime_is_shutting_down ())
899 ts = mono_msec_ticks ();
900 if (mono_thread_info_sleep (interval_left, &alerted) == 0)
902 interval_left -= mono_msec_ticks () - ts;
904 mono_gc_set_skip_thread (FALSE);
905 if ((current_thread->state & (ThreadState_StopRequested | ThreadState_SuspendRequested)) != 0)
906 mono_thread_interruption_checkpoint ();
907 mono_gc_set_skip_thread (TRUE);
908 } while (interval_left > 0 && ++awake < 10);
910 mono_gc_set_skip_thread (FALSE);
912 if (threadpool->suspended)
915 if (mono_runtime_is_shutting_down ())
918 mono_coop_mutex_lock (&threadpool->domains_lock);
919 if (!domain_any_has_request ()) {
920 mono_coop_mutex_unlock (&threadpool->domains_lock);
923 mono_coop_mutex_unlock (&threadpool->domains_lock);
925 threadpool->cpu_usage = mono_cpu_usage (threadpool->cpu_usage_state);
927 if (!monitor_sufficient_delay_since_last_dequeue ())
930 limit_worker_max_reached = FALSE;
932 COUNTER_ATOMIC (counter, {
933 if (counter._.max_working >= threadpool->limit_worker_max) {
934 limit_worker_max_reached = TRUE;
937 counter._.max_working ++;
940 if (limit_worker_max_reached)
943 hill_climbing_force_change (counter._.max_working, TRANSITION_STARVATION);
945 for (i = 0; i < 5; ++i) {
946 if (mono_runtime_is_shutting_down ())
949 if (worker_try_unpark ()) {
950 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, unparked", mono_native_thread_id_get ());
954 if (worker_try_create ()) {
955 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, created", mono_native_thread_id_get ());
959 } while (monitor_should_keep_running ());
961 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, finished", mono_native_thread_id_get ());
965 monitor_ensure_running (void)
968 switch (monitor_status) {
969 case MONITOR_STATUS_REQUESTED:
971 case MONITOR_STATUS_WAITING_FOR_REQUEST:
972 InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_WAITING_FOR_REQUEST);
974 case MONITOR_STATUS_NOT_RUNNING:
975 if (mono_runtime_is_shutting_down ())
977 if (InterlockedCompareExchange (&monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_NOT_RUNNING) == MONITOR_STATUS_NOT_RUNNING) {
978 if (!mono_thread_create_internal (mono_get_root_domain (), monitor_thread, NULL, TRUE, SMALL_STACK))
979 monitor_status = MONITOR_STATUS_NOT_RUNNING;
983 default: g_assert_not_reached ();
989 hill_climbing_change_thread_count (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
991 ThreadPoolHillClimbing *hc;
993 g_assert (threadpool);
995 hc = &threadpool->heuristic_hill_climbing;
997 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);
999 hc->last_thread_count = new_thread_count;
1000 hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
1001 hc->elapsed_since_last_change = 0;
1002 hc->completions_since_last_change = 0;
1006 hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
1008 ThreadPoolHillClimbing *hc;
1010 g_assert (threadpool);
1012 hc = &threadpool->heuristic_hill_climbing;
1014 if (new_thread_count != hc->last_thread_count) {
1015 hc->current_control_setting += new_thread_count - hc->last_thread_count;
1016 hill_climbing_change_thread_count (new_thread_count, transition);
1020 static double_complex
1021 hill_climbing_get_wave_component (gdouble *samples, guint sample_count, gdouble period)
1023 ThreadPoolHillClimbing *hc;
1024 gdouble w, cosine, sine, coeff, q0, q1, q2;
1027 g_assert (threadpool);
1028 g_assert (sample_count >= period);
1029 g_assert (period >= 2);
1031 hc = &threadpool->heuristic_hill_climbing;
1033 w = 2.0 * M_PI / period;
1036 coeff = 2.0 * cosine;
1039 for (i = 0; i < sample_count; ++i) {
1040 q0 = coeff * q1 - q2 + samples [(hc->total_samples - sample_count + i) % hc->samples_to_measure];
1045 return mono_double_complex_scalar_div (mono_double_complex_make (q1 - q2 * cosine, (q2 * sine)), ((gdouble)sample_count));
1049 hill_climbing_update (gint16 current_thread_count, guint32 sample_duration, gint32 completions, guint32 *adjustment_interval)
1051 ThreadPoolHillClimbing *hc;
1052 ThreadPoolHeuristicStateTransition transition;
1054 gdouble throughput_error_estimate;
1060 gint new_thread_wave_magnitude;
1061 gint new_thread_count;
1062 double_complex thread_wave_component;
1063 double_complex throughput_wave_component;
1064 double_complex ratio;
1066 g_assert (threadpool);
1067 g_assert (adjustment_interval);
1069 hc = &threadpool->heuristic_hill_climbing;
1071 /* If someone changed the thread count without telling us, update our records accordingly. */
1072 if (current_thread_count != hc->last_thread_count)
1073 hill_climbing_force_change (current_thread_count, TRANSITION_INITIALIZING);
1075 /* Update the cumulative stats for this thread count */
1076 hc->elapsed_since_last_change += sample_duration;
1077 hc->completions_since_last_change += completions;
1079 /* Add in any data we've already collected about this sample */
1080 sample_duration += hc->accumulated_sample_duration;
1081 completions += hc->accumulated_completion_count;
1083 /* We need to make sure we're collecting reasonably accurate data. Since we're just counting the end
1084 * of each work item, we are goinng to be missing some data about what really happened during the
1085 * sample interval. The count produced by each thread includes an initial work item that may have
1086 * started well before the start of the interval, and each thread may have been running some new
1087 * work item for some time before the end of the interval, which did not yet get counted. So
1088 * our count is going to be off by +/- threadCount workitems.
1090 * The exception is that the thread that reported to us last time definitely wasn't running any work
1091 * at that time, and the thread that's reporting now definitely isn't running a work item now. So
1092 * we really only need to consider threadCount-1 threads.
1094 * Thus the percent error in our count is +/- (threadCount-1)/numCompletions.
1096 * We cannot rely on the frequency-domain analysis we'll be doing later to filter out this error, because
1097 * of the way it accumulates over time. If this sample is off by, say, 33% in the negative direction,
1098 * then the next one likely will be too. The one after that will include the sum of the completions
1099 * we missed in the previous samples, and so will be 33% positive. So every three samples we'll have
1100 * two "low" samples and one "high" sample. This will appear as periodic variation right in the frequency
1101 * range we're targeting, which will not be filtered by the frequency-domain translation. */
1102 if (hc->total_samples > 0 && ((current_thread_count - 1.0) / completions) >= hc->max_sample_error) {
1103 /* Not accurate enough yet. Let's accumulate the data so
1104 * far, and tell the ThreadPool to collect a little more. */
1105 hc->accumulated_sample_duration = sample_duration;
1106 hc->accumulated_completion_count = completions;
1107 *adjustment_interval = 10;
1108 return current_thread_count;
1111 /* We've got enouugh data for our sample; reset our accumulators for next time. */
1112 hc->accumulated_sample_duration = 0;
1113 hc->accumulated_completion_count = 0;
1115 /* Add the current thread count and throughput sample to our history. */
1116 throughput = ((gdouble) completions) / sample_duration;
1118 sample_index = hc->total_samples % hc->samples_to_measure;
1119 hc->samples [sample_index] = throughput;
1120 hc->thread_counts [sample_index] = current_thread_count;
1121 hc->total_samples ++;
1123 /* Set up defaults for our metrics. */
1124 thread_wave_component = mono_double_complex_make(0, 0);
1125 throughput_wave_component = mono_double_complex_make(0, 0);
1126 throughput_error_estimate = 0;
1127 ratio = mono_double_complex_make(0, 0);
1130 transition = TRANSITION_WARMUP;
1132 /* How many samples will we use? It must be at least the three wave periods we're looking for, and it must also
1133 * be a whole multiple of the primary wave's period; otherwise the frequency we're looking for will fall between
1134 * two frequency bands in the Fourier analysis, and we won't be able to measure it accurately. */
1135 sample_count = ((gint) MIN (hc->total_samples - 1, hc->samples_to_measure) / hc->wave_period) * hc->wave_period;
1137 if (sample_count > hc->wave_period) {
1139 gdouble average_throughput;
1140 gdouble average_thread_count;
1141 gdouble sample_sum = 0;
1142 gdouble thread_sum = 0;
1144 /* Average the throughput and thread count samples, so we can scale the wave magnitudes later. */
1145 for (i = 0; i < sample_count; ++i) {
1146 guint j = (hc->total_samples - sample_count + i) % hc->samples_to_measure;
1147 sample_sum += hc->samples [j];
1148 thread_sum += hc->thread_counts [j];
1151 average_throughput = sample_sum / sample_count;
1152 average_thread_count = thread_sum / sample_count;
1154 if (average_throughput > 0 && average_thread_count > 0) {
1155 gdouble noise_for_confidence, adjacent_period_1, adjacent_period_2;
1157 /* Calculate the periods of the adjacent frequency bands we'll be using to
1158 * measure noise levels. We want the two adjacent Fourier frequency bands. */
1159 adjacent_period_1 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) + 1);
1160 adjacent_period_2 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) - 1);
1162 /* Get the the three different frequency components of the throughput (scaled by average
1163 * throughput). Our "error" estimate (the amount of noise that might be present in the
1164 * frequency band we're really interested in) is the average of the adjacent bands. */
1165 throughput_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, hc->wave_period), average_throughput);
1166 throughput_error_estimate = cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, adjacent_period_1), average_throughput));
1168 if (adjacent_period_2 <= sample_count) {
1169 throughput_error_estimate = MAX (throughput_error_estimate, cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (
1170 hc->samples, sample_count, adjacent_period_2), average_throughput)));
1173 /* Do the same for the thread counts, so we have something to compare to. We don't
1174 * measure thread count noise, because there is none; these are exact measurements. */
1175 thread_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->thread_counts, sample_count, hc->wave_period), average_thread_count);
1177 /* Update our moving average of the throughput noise. We'll use this
1178 * later as feedback to determine the new size of the thread wave. */
1179 if (hc->average_throughput_noise == 0) {
1180 hc->average_throughput_noise = throughput_error_estimate;
1182 hc->average_throughput_noise = (hc->throughput_error_smoothing_factor * throughput_error_estimate)
1183 + ((1.0 + hc->throughput_error_smoothing_factor) * hc->average_throughput_noise);
1186 if (cabs (thread_wave_component) > 0) {
1187 /* Adjust the throughput wave so it's centered around the target wave,
1188 * and then calculate the adjusted throughput/thread ratio. */
1189 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);
1190 transition = TRANSITION_CLIMBING_MOVE;
1192 ratio = mono_double_complex_make (0, 0);
1193 transition = TRANSITION_STABILIZING;
1196 noise_for_confidence = MAX (hc->average_throughput_noise, throughput_error_estimate);
1197 if (noise_for_confidence > 0) {
1198 confidence = cabs (thread_wave_component) / noise_for_confidence / hc->target_signal_to_noise_ratio;
1200 /* there is no noise! */
1206 /* We use just the real part of the complex ratio we just calculated. If the throughput signal
1207 * is exactly in phase with the thread signal, this will be the same as taking the magnitude of
1208 * the complex move and moving that far up. If they're 180 degrees out of phase, we'll move
1209 * backward (because this indicates that our changes are having the opposite of the intended effect).
1210 * If they're 90 degrees out of phase, we won't move at all, because we can't tell wether we're
1211 * having a negative or positive effect on throughput. */
1212 move = creal (ratio);
1213 move = CLAMP (move, -1.0, 1.0);
1215 /* Apply our confidence multiplier. */
1216 move *= CLAMP (confidence, -1.0, 1.0);
1218 /* Now apply non-linear gain, such that values around zero are attenuated, while higher values
1219 * are enhanced. This allows us to move quickly if we're far away from the target, but more slowly
1220 * if we're getting close, giving us rapid ramp-up without wild oscillations around the target. */
1221 gain = hc->max_change_per_second * sample_duration;
1222 move = pow (fabs (move), hc->gain_exponent) * (move >= 0.0 ? 1 : -1) * gain;
1223 move = MIN (move, hc->max_change_per_sample);
1225 /* If the result was positive, and CPU is > 95%, refuse the move. */
1226 if (move > 0.0 && threadpool->cpu_usage > CPU_USAGE_HIGH)
1229 /* Apply the move to our control setting. */
1230 hc->current_control_setting += move;
1232 /* Calculate the new thread wave magnitude, which is based on the moving average we've been keeping of the
1233 * throughput error. This average starts at zero, so we'll start with a nice safe little wave at first. */
1234 new_thread_wave_magnitude = (gint)(0.5 + (hc->current_control_setting * hc->average_throughput_noise
1235 * hc->target_signal_to_noise_ratio * hc->thread_magnitude_multiplier * 2.0));
1236 new_thread_wave_magnitude = CLAMP (new_thread_wave_magnitude, 1, hc->max_thread_wave_magnitude);
1238 /* Make sure our control setting is within the ThreadPool's limits. */
1239 hc->current_control_setting = CLAMP (hc->current_control_setting, threadpool->limit_worker_min, threadpool->limit_worker_max - new_thread_wave_magnitude);
1241 /* Calculate the new thread count (control setting + square wave). */
1242 new_thread_count = (gint)(hc->current_control_setting + new_thread_wave_magnitude * ((hc->total_samples / (hc->wave_period / 2)) % 2));
1244 /* Make sure the new thread count doesn't exceed the ThreadPool's limits. */
1245 new_thread_count = CLAMP (new_thread_count, threadpool->limit_worker_min, threadpool->limit_worker_max);
1247 if (new_thread_count != current_thread_count)
1248 hill_climbing_change_thread_count (new_thread_count, transition);
1250 if (creal (ratio) < 0.0 && new_thread_count == threadpool->limit_worker_min)
1251 *adjustment_interval = (gint)(0.5 + hc->current_sample_interval * (10.0 * MAX (-1.0 * creal (ratio), 1.0)));
1253 *adjustment_interval = hc->current_sample_interval;
1255 return new_thread_count;
1259 heuristic_notify_work_completed (void)
1261 g_assert (threadpool);
1263 InterlockedIncrement (&threadpool->heuristic_completions);
1264 threadpool->heuristic_last_dequeue = mono_msec_ticks ();
1268 heuristic_should_adjust (void)
1270 g_assert (threadpool);
1272 if (threadpool->heuristic_last_dequeue > threadpool->heuristic_last_adjustment + threadpool->heuristic_adjustment_interval) {
1273 ThreadPoolCounter counter;
1274 counter.as_gint64 = COUNTER_READ();
1275 if (counter._.working <= counter._.max_working)
1283 heuristic_adjust (void)
1285 g_assert (threadpool);
1287 if (mono_coop_mutex_trylock (&threadpool->heuristic_lock) == 0) {
1288 gint32 completions = InterlockedExchange (&threadpool->heuristic_completions, 0);
1289 guint32 sample_end = mono_msec_ticks ();
1290 guint32 sample_duration = sample_end - threadpool->heuristic_sample_start;
1292 if (sample_duration >= threadpool->heuristic_adjustment_interval / 2) {
1293 ThreadPoolCounter counter;
1294 gint16 new_thread_count;
1296 counter.as_gint64 = COUNTER_READ ();
1297 new_thread_count = hill_climbing_update (counter._.max_working, sample_duration, completions, &threadpool->heuristic_adjustment_interval);
1299 COUNTER_ATOMIC (counter, { counter._.max_working = new_thread_count; });
1301 if (new_thread_count > counter._.max_working)
1302 worker_request (mono_domain_get ());
1304 threadpool->heuristic_sample_start = sample_end;
1305 threadpool->heuristic_last_adjustment = mono_msec_ticks ();
1308 mono_coop_mutex_unlock (&threadpool->heuristic_lock);
1313 mono_threadpool_ms_cleanup (void)
1315 #ifndef DISABLE_SOCKETS
1316 mono_threadpool_ms_io_cleanup ();
1318 mono_lazy_cleanup (&status, cleanup);
1322 mono_threadpool_ms_begin_invoke (MonoDomain *domain, MonoObject *target, MonoMethod *method, gpointer *params, MonoError *error)
1324 static MonoClass *async_call_klass = NULL;
1325 MonoMethodMessage *message;
1326 MonoAsyncResult *async_result;
1327 MonoAsyncCall *async_call;
1328 MonoDelegate *async_callback = NULL;
1329 MonoObject *state = NULL;
1331 if (!async_call_klass)
1332 async_call_klass = mono_class_load_from_name (mono_defaults.corlib, "System", "MonoAsyncCall");
1334 mono_lazy_initialize (&status, initialize);
1336 mono_error_init (error);
1338 message = mono_method_call_message_new (method, params, mono_get_delegate_invoke (method->klass), (params != NULL) ? (&async_callback) : NULL, (params != NULL) ? (&state) : NULL);
1340 async_call = (MonoAsyncCall*) mono_object_new_checked (domain, async_call_klass, error);
1341 return_val_if_nok (error, NULL);
1343 MONO_OBJECT_SETREF (async_call, msg, message);
1344 MONO_OBJECT_SETREF (async_call, state, state);
1346 if (async_callback) {
1347 MONO_OBJECT_SETREF (async_call, cb_method, mono_get_delegate_invoke (((MonoObject*) async_callback)->vtable->klass));
1348 MONO_OBJECT_SETREF (async_call, cb_target, async_callback);
1351 async_result = mono_async_result_new (domain, NULL, async_call->state, NULL, (MonoObject*) async_call);
1352 MONO_OBJECT_SETREF (async_result, async_delegate, target);
1354 mono_threadpool_ms_enqueue_work_item (domain, (MonoObject*) async_result, error);
1355 return_val_if_nok (error, NULL);
1357 return async_result;
1361 mono_threadpool_ms_end_invoke (MonoAsyncResult *ares, MonoArray **out_args, MonoObject **exc)
1366 g_assert (out_args);
1371 /* check if already finished */
1372 mono_monitor_enter ((MonoObject*) ares);
1374 if (ares->endinvoke_called) {
1375 *exc = (MonoObject*) mono_get_exception_invalid_operation (NULL);
1376 mono_monitor_exit ((MonoObject*) ares);
1380 ares->endinvoke_called = 1;
1382 /* wait until we are really finished */
1383 if (ares->completed) {
1384 mono_monitor_exit ((MonoObject *) ares);
1386 gpointer wait_event;
1388 wait_event = mono_wait_handle_get_handle ((MonoWaitHandle*) ares->handle);
1390 wait_event = CreateEvent (NULL, TRUE, FALSE, NULL);
1391 g_assert(wait_event);
1392 MONO_OBJECT_SETREF (ares, handle, (MonoObject*) mono_wait_handle_new (mono_object_domain (ares), wait_event));
1394 mono_monitor_exit ((MonoObject*) ares);
1395 MONO_PREPARE_BLOCKING;
1396 WaitForSingleObjectEx (wait_event, INFINITE, TRUE);
1397 MONO_FINISH_BLOCKING;
1400 ac = (MonoAsyncCall*) ares->object_data;
1403 *exc = ac->msg->exc; /* FIXME: GC add write barrier */
1404 *out_args = ac->out_args;
1409 mono_threadpool_ms_remove_domain_jobs (MonoDomain *domain, int timeout)
1411 gboolean res = TRUE;
1416 g_assert (timeout >= -1);
1418 g_assert (mono_domain_is_unloading (domain));
1421 start = mono_msec_ticks ();
1423 #ifndef DISABLE_SOCKETS
1424 mono_threadpool_ms_io_remove_domain_jobs (domain);
1425 if (timeout != -1) {
1426 timeout -= mono_msec_ticks () - start;
1433 * There might be some threads out that could be about to execute stuff from the given domain.
1434 * We avoid that by setting up a semaphore to be pulsed by the thread that reaches zero.
1436 sem = domain->cleanup_semaphore = CreateSemaphore (NULL, 0, 1, NULL);
1439 * The memory barrier here is required to have global ordering between assigning to cleanup_semaphone
1440 * and reading threadpool_jobs. Otherwise this thread could read a stale version of threadpool_jobs
1443 mono_memory_write_barrier ();
1445 while (domain->threadpool_jobs) {
1446 MONO_PREPARE_BLOCKING;
1447 WaitForSingleObject (sem, timeout);
1448 MONO_FINISH_BLOCKING;
1449 if (timeout != -1) {
1450 timeout -= mono_msec_ticks () - start;
1458 domain->cleanup_semaphore = NULL;
1465 mono_threadpool_ms_suspend (void)
1468 threadpool->suspended = TRUE;
1472 mono_threadpool_ms_resume (void)
1475 threadpool->suspended = FALSE;
1479 ves_icall_System_Threading_ThreadPool_GetAvailableThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1481 ThreadPoolCounter counter;
1483 if (!worker_threads || !completion_port_threads)
1486 mono_lazy_initialize (&status, initialize);
1488 counter.as_gint64 = COUNTER_READ ();
1490 *worker_threads = MAX (0, threadpool->limit_worker_max - counter._.active);
1491 *completion_port_threads = threadpool->limit_io_max;
1495 ves_icall_System_Threading_ThreadPool_GetMinThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1497 if (!worker_threads || !completion_port_threads)
1500 mono_lazy_initialize (&status, initialize);
1502 *worker_threads = threadpool->limit_worker_min;
1503 *completion_port_threads = threadpool->limit_io_min;
1507 ves_icall_System_Threading_ThreadPool_GetMaxThreadsNative (gint32 *worker_threads, gint32 *completion_port_threads)
1509 if (!worker_threads || !completion_port_threads)
1512 mono_lazy_initialize (&status, initialize);
1514 *worker_threads = threadpool->limit_worker_max;
1515 *completion_port_threads = threadpool->limit_io_max;
1519 ves_icall_System_Threading_ThreadPool_SetMinThreadsNative (gint32 worker_threads, gint32 completion_port_threads)
1521 mono_lazy_initialize (&status, initialize);
1523 if (worker_threads <= 0 || worker_threads > threadpool->limit_worker_max)
1525 if (completion_port_threads <= 0 || completion_port_threads > threadpool->limit_io_max)
1528 threadpool->limit_worker_min = worker_threads;
1529 threadpool->limit_io_min = completion_port_threads;
1535 ves_icall_System_Threading_ThreadPool_SetMaxThreadsNative (gint32 worker_threads, gint32 completion_port_threads)
1537 gint cpu_count = mono_cpu_count ();
1539 mono_lazy_initialize (&status, initialize);
1541 if (worker_threads < threadpool->limit_worker_min || worker_threads < cpu_count)
1543 if (completion_port_threads < threadpool->limit_io_min || completion_port_threads < cpu_count)
1546 threadpool->limit_worker_max = worker_threads;
1547 threadpool->limit_io_max = completion_port_threads;
1553 ves_icall_System_Threading_ThreadPool_InitializeVMTp (MonoBoolean *enable_worker_tracking)
1555 if (enable_worker_tracking) {
1556 // TODO implement some kind of switch to have the possibily to use it
1557 *enable_worker_tracking = FALSE;
1560 mono_lazy_initialize (&status, initialize);
1564 ves_icall_System_Threading_ThreadPool_NotifyWorkItemComplete (void)
1566 ThreadPoolCounter counter;
1568 if (mono_domain_is_unloading (mono_domain_get ()) || mono_runtime_is_shutting_down ())
1571 heuristic_notify_work_completed ();
1573 if (heuristic_should_adjust ())
1574 heuristic_adjust ();
1576 counter.as_gint64 = COUNTER_READ ();
1577 return counter._.working <= counter._.max_working;
1581 ves_icall_System_Threading_ThreadPool_NotifyWorkItemProgressNative (void)
1583 heuristic_notify_work_completed ();
1585 if (heuristic_should_adjust ())
1586 heuristic_adjust ();
1590 ves_icall_System_Threading_ThreadPool_ReportThreadStatus (MonoBoolean is_working)
1594 mono_error_set_not_implemented (&error, "");
1595 mono_error_set_pending_exception (&error);
1599 ves_icall_System_Threading_ThreadPool_RequestWorkerThread (void)
1601 return worker_request (mono_domain_get ());
1604 MonoBoolean G_GNUC_UNUSED
1605 ves_icall_System_Threading_ThreadPool_PostQueuedCompletionStatus (MonoNativeOverlapped *native_overlapped)
1607 /* This copy the behavior of the current Mono implementation */
1609 mono_error_set_not_implemented (&error, "");
1610 mono_error_set_pending_exception (&error);
1614 MonoBoolean G_GNUC_UNUSED
1615 ves_icall_System_Threading_ThreadPool_BindIOCompletionCallbackNative (gpointer file_handle)
1617 /* This copy the behavior of the current Mono implementation */
1621 MonoBoolean G_GNUC_UNUSED
1622 ves_icall_System_Threading_ThreadPool_IsThreadPoolHosted (void)