3 * Worker threads for parallel and concurrent GC.
5 * Copyright 2001-2003 Ximian, Inc
6 * Copyright 2003-2010 Novell, Inc.
7 * Copyright (C) 2012 Xamarin Inc
9 * Licensed under the MIT license. See LICENSE file in the project root for full license information.
17 #include "mono/sgen/sgen-gc.h"
18 #include "mono/sgen/sgen-workers.h"
19 #include "mono/sgen/sgen-thread-pool.h"
20 #include "mono/utils/mono-membar.h"
21 #include "mono/sgen/sgen-client.h"
23 static int workers_num;
24 static int active_workers_num;
25 static volatile gboolean started;
26 static volatile gboolean forced_stop;
27 static WorkerData *workers_data;
28 static SgenWorkerCallback worker_init_cb;
30 static SgenThreadPool pool_inst;
31 static SgenThreadPool *pool; /* null if we're not using workers */
34 * When using multiple workers, we need to have the last worker
35 * enqueue the preclean jobs (if there are any). This lock ensures
36 * that when the last worker takes it, all the other workers have
37 * gracefully finished, so it can restart them.
39 static mono_mutex_t finished_lock;
40 static volatile gboolean workers_finished;
41 static int worker_awakenings;
43 static SgenSectionGrayQueue workers_distribute_gray_queue;
44 static gboolean workers_distribute_gray_queue_inited;
47 * Allowed transitions:
49 * | from \ to | NOT WORKING | WORKING | WORK ENQUEUED |
50 * |--------------------+-------------+---------+---------------+
51 * | NOT WORKING | - | - | main / worker |
52 * | WORKING | worker | - | main / worker |
53 * | WORK ENQUEUED | - | worker | - |
55 * The WORK ENQUEUED state guarantees that the worker thread will inspect the queue again at
56 * least once. Only after looking at the queue will it go back to WORKING, and then,
57 * eventually, to NOT WORKING. After enqueuing work the main thread transitions the state
58 * to WORK ENQUEUED. Signalling the worker thread to wake up is only necessary if the old
59 * state was NOT WORKING.
68 #define SGEN_WORKER_MIN_SECTIONS_SIGNAL 4
72 static SgenObjectOperations * volatile idle_func_object_ops;
73 static SgenObjectOperations *idle_func_object_ops_par, *idle_func_object_ops_nopar;
75 * finished_callback is called only when the workers finish work normally (when they
76 * are not forced to finish). The callback is used to enqueue preclean jobs.
78 static volatile SgenWorkersFinishCallback finish_callback;
80 static guint64 stat_workers_num_finished;
83 set_state (WorkerData *data, State old_state, State new_state)
85 SGEN_ASSERT (0, old_state != new_state, "Why are we transitioning to the same state?");
86 if (new_state == STATE_NOT_WORKING)
87 SGEN_ASSERT (0, old_state == STATE_WORKING, "We can only transition to NOT WORKING from WORKING");
88 else if (new_state == STATE_WORKING)
89 SGEN_ASSERT (0, old_state == STATE_WORK_ENQUEUED, "We can only transition to WORKING from WORK ENQUEUED");
90 if (new_state == STATE_NOT_WORKING || new_state == STATE_WORKING)
91 SGEN_ASSERT (6, sgen_thread_pool_is_thread_pool_thread (pool, mono_native_thread_id_get ()), "Only the worker thread is allowed to transition to NOT_WORKING or WORKING");
93 return InterlockedCompareExchange (&data->state, new_state, old_state) == old_state;
97 state_is_working_or_enqueued (State state)
99 return state == STATE_WORKING || state == STATE_WORK_ENQUEUED;
103 sgen_workers_ensure_awake (void)
106 gboolean need_signal = FALSE;
109 * All workers are awaken, make sure we reset the parallel context.
110 * We call this function only when starting the workers so nobody is running,
111 * or when the last worker is enqueuing preclean work. In both cases we can't
112 * have a worker working using a nopar context, which means it is safe.
114 idle_func_object_ops = (active_workers_num > 1) ? idle_func_object_ops_par : idle_func_object_ops_nopar;
115 workers_finished = FALSE;
117 for (i = 0; i < active_workers_num; i++) {
119 gboolean did_set_state;
122 old_state = workers_data [i].state;
124 if (old_state == STATE_WORK_ENQUEUED)
127 did_set_state = set_state (&workers_data [i], old_state, STATE_WORK_ENQUEUED);
128 } while (!did_set_state);
130 if (!state_is_working_or_enqueued (old_state))
135 sgen_thread_pool_idle_signal (pool);
139 worker_try_finish (WorkerData *data)
144 ++stat_workers_num_finished;
146 mono_os_mutex_lock (&finished_lock);
148 for (i = 0; i < active_workers_num; i++) {
149 if (state_is_working_or_enqueued (workers_data [i].state))
154 SgenWorkersFinishCallback callback = finish_callback;
155 SGEN_ASSERT (0, idle_func_object_ops == idle_func_object_ops_nopar, "Why are we finishing with parallel context");
156 /* We are the last one left. Enqueue preclean job if we have one and awake everybody */
157 SGEN_ASSERT (0, data->state != STATE_NOT_WORKING, "How did we get from doing idle work to NOT WORKING without setting it ourselves?");
159 finish_callback = NULL;
161 worker_awakenings = 0;
162 /* Make sure each worker has a chance of seeing the enqueued jobs */
163 sgen_workers_ensure_awake ();
164 SGEN_ASSERT (0, data->state == STATE_WORK_ENQUEUED, "Why did we fail to set our own state to ENQUEUED");
170 old_state = data->state;
172 SGEN_ASSERT (0, old_state != STATE_NOT_WORKING, "How did we get from doing idle work to NOT WORKING without setting it ourselves?");
173 if (old_state == STATE_WORK_ENQUEUED)
175 SGEN_ASSERT (0, old_state == STATE_WORKING, "What other possibility is there?");
176 } while (!set_state (data, old_state, STATE_NOT_WORKING));
179 * If we are second to last to finish, we set the scan context to the non-parallel
180 * version so we can speed up the last worker. This helps us maintain same level
181 * of performance as non-parallel mode even if we fail to distribute work properly.
184 idle_func_object_ops = idle_func_object_ops_nopar;
186 workers_finished = TRUE;
187 mono_os_mutex_unlock (&finished_lock);
189 binary_protocol_worker_finish (sgen_timestamp (), forced_stop);
191 sgen_gray_object_queue_trim_free_list (&data->private_gray_queue);
195 mono_os_mutex_unlock (&finished_lock);
199 sgen_workers_enqueue_job (SgenThreadPoolJob *job, gboolean enqueue)
202 job->func (NULL, job);
203 sgen_thread_pool_job_free (job);
207 sgen_thread_pool_job_enqueue (pool, job);
211 workers_get_work (WorkerData *data)
213 SgenMajorCollector *major = sgen_get_major_collector ();
214 SgenMinorCollector *minor = sgen_get_minor_collector ();
215 GrayQueueSection *section;
217 g_assert (sgen_gray_object_queue_is_empty (&data->private_gray_queue));
218 g_assert (major->is_concurrent || minor->is_parallel);
220 section = sgen_section_gray_queue_dequeue (&workers_distribute_gray_queue);
222 sgen_gray_object_enqueue_section (&data->private_gray_queue, section, major->is_parallel);
226 /* Nobody to steal from */
227 g_assert (sgen_gray_object_queue_is_empty (&data->private_gray_queue));
232 workers_steal_work (WorkerData *data)
234 SgenMajorCollector *major = sgen_get_major_collector ();
235 SgenMinorCollector *minor = sgen_get_minor_collector ();
236 int generation = sgen_get_current_collection_generation ();
237 GrayQueueSection *section = NULL;
238 int i, current_worker;
240 if ((generation == GENERATION_OLD && !major->is_parallel) ||
241 (generation == GENERATION_NURSERY && !minor->is_parallel))
244 /* If we're parallel, steal from other workers' private gray queues */
245 g_assert (sgen_gray_object_queue_is_empty (&data->private_gray_queue));
247 current_worker = (int) (data - workers_data);
249 for (i = 1; i < active_workers_num && !section; i++) {
250 int steal_worker = (current_worker + i) % active_workers_num;
251 if (state_is_working_or_enqueued (workers_data [steal_worker].state))
252 section = sgen_gray_object_steal_section (&workers_data [steal_worker].private_gray_queue);
256 sgen_gray_object_enqueue_section (&data->private_gray_queue, section, TRUE);
260 /* Nobody to steal from */
261 g_assert (sgen_gray_object_queue_is_empty (&data->private_gray_queue));
266 concurrent_enqueue_check (GCObject *obj)
268 g_assert (sgen_concurrent_collection_in_progress ());
269 g_assert (!sgen_ptr_in_nursery (obj));
270 g_assert (SGEN_LOAD_VTABLE (obj));
274 init_private_gray_queue (WorkerData *data)
276 sgen_gray_object_queue_init (&data->private_gray_queue,
277 sgen_get_major_collector ()->is_concurrent ? concurrent_enqueue_check : NULL,
282 thread_pool_init_func (void *data_untyped)
284 WorkerData *data = (WorkerData *)data_untyped;
285 SgenMajorCollector *major = sgen_get_major_collector ();
286 SgenMinorCollector *minor = sgen_get_minor_collector ();
288 sgen_client_thread_register_worker ();
290 if (!major->is_concurrent && !minor->is_parallel)
293 init_private_gray_queue (data);
296 worker_init_cb (data);
300 continue_idle_func (void *data_untyped)
303 WorkerData *data = (WorkerData *)data_untyped;
304 return state_is_working_or_enqueued (data->state);
306 /* Return if any of the threads is working */
307 return !sgen_workers_all_done ();
312 should_work_func (void *data_untyped)
314 WorkerData *data = (WorkerData*)data_untyped;
315 int current_worker = (int) (data - workers_data);
317 return started && current_worker < active_workers_num && state_is_working_or_enqueued (data->state);
321 marker_idle_func (void *data_untyped)
323 WorkerData *data = (WorkerData *)data_untyped;
325 SGEN_ASSERT (0, continue_idle_func (data_untyped), "Why are we called when we're not supposed to work?");
327 if (data->state == STATE_WORK_ENQUEUED) {
328 set_state (data, STATE_WORK_ENQUEUED, STATE_WORKING);
329 SGEN_ASSERT (0, data->state != STATE_NOT_WORKING, "How did we get from WORK ENQUEUED to NOT WORKING?");
332 if (!forced_stop && (!sgen_gray_object_queue_is_empty (&data->private_gray_queue) || workers_get_work (data) || workers_steal_work (data))) {
333 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (idle_func_object_ops, &data->private_gray_queue);
335 SGEN_ASSERT (0, !sgen_gray_object_queue_is_empty (&data->private_gray_queue), "How is our gray queue empty if we just got work?");
337 sgen_drain_gray_stack (ctx);
339 if (data->private_gray_queue.num_sections >= SGEN_WORKER_MIN_SECTIONS_SIGNAL
340 && workers_finished && worker_awakenings < active_workers_num) {
341 /* We bound the number of worker awakenings just to be sure */
343 mono_os_mutex_lock (&finished_lock);
344 sgen_workers_ensure_awake ();
345 mono_os_mutex_unlock (&finished_lock);
348 worker_try_finish (data);
353 init_distribute_gray_queue (void)
355 if (workers_distribute_gray_queue_inited) {
356 g_assert (sgen_section_gray_queue_is_empty (&workers_distribute_gray_queue));
357 g_assert (workers_distribute_gray_queue.locked);
361 sgen_section_gray_queue_init (&workers_distribute_gray_queue, TRUE,
362 sgen_get_major_collector ()->is_concurrent ? concurrent_enqueue_check : NULL);
363 workers_distribute_gray_queue_inited = TRUE;
367 sgen_workers_init_distribute_gray_queue (void)
369 SGEN_ASSERT (0, sgen_get_major_collector ()->is_concurrent || sgen_get_minor_collector ()->is_parallel,
370 "Why should we init the distribute gray queue if we don't need it?");
371 init_distribute_gray_queue ();
375 sgen_workers_init (int num_workers, SgenWorkerCallback callback)
378 WorkerData **workers_data_ptrs = (WorkerData**)alloca(num_workers * sizeof(WorkerData*));
380 mono_os_mutex_init (&finished_lock);
381 //g_print ("initing %d workers\n", num_workers);
383 workers_num = num_workers;
384 active_workers_num = num_workers;
386 workers_data = (WorkerData *)sgen_alloc_internal_dynamic (sizeof (WorkerData) * num_workers, INTERNAL_MEM_WORKER_DATA, TRUE);
387 memset (workers_data, 0, sizeof (WorkerData) * num_workers);
389 init_distribute_gray_queue ();
391 for (i = 0; i < num_workers; ++i)
392 workers_data_ptrs [i] = &workers_data [i];
394 worker_init_cb = callback;
397 sgen_thread_pool_init (pool, num_workers, thread_pool_init_func, marker_idle_func, continue_idle_func, should_work_func, (SgenThreadPoolData**)workers_data_ptrs);
399 mono_counters_register ("# workers finished", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_workers_num_finished);
403 sgen_workers_shutdown (void)
406 sgen_thread_pool_shutdown (pool);
410 sgen_workers_stop_all_workers (void)
412 finish_callback = NULL;
413 mono_memory_write_barrier ();
416 sgen_thread_pool_wait_for_all_jobs (pool);
417 sgen_thread_pool_idle_wait (pool);
418 SGEN_ASSERT (0, sgen_workers_all_done (), "Can only signal enqueue work when in no work state");
424 sgen_workers_set_num_active_workers (int num_workers)
427 SGEN_ASSERT (0, active_workers_num <= workers_num, "We can't start more workers than we initialized");
428 active_workers_num = num_workers;
430 active_workers_num = workers_num;
435 sgen_workers_start_all_workers (SgenObjectOperations *object_ops_nopar, SgenObjectOperations *object_ops_par, SgenWorkersFinishCallback callback)
437 SGEN_ASSERT (0, !started, "Why are we starting to work without finishing previous cycle");
439 idle_func_object_ops_par = object_ops_par;
440 idle_func_object_ops_nopar = object_ops_nopar;
442 finish_callback = callback;
443 worker_awakenings = 0;
445 mono_memory_write_barrier ();
448 * We expect workers to start finishing only after all of them were awaken.
449 * Otherwise we might think that we have fewer workers and use wrong context.
451 mono_os_mutex_lock (&finished_lock);
452 sgen_workers_ensure_awake ();
453 mono_os_mutex_unlock (&finished_lock);
457 sgen_workers_join (void)
461 sgen_thread_pool_wait_for_all_jobs (pool);
462 sgen_thread_pool_idle_wait (pool);
463 SGEN_ASSERT (0, sgen_workers_all_done (), "Can only signal enqueue work when in no work state");
465 /* At this point all the workers have stopped. */
467 SGEN_ASSERT (0, sgen_section_gray_queue_is_empty (&workers_distribute_gray_queue), "Why is there still work left to do?");
468 for (i = 0; i < active_workers_num; ++i)
469 SGEN_ASSERT (0, sgen_gray_object_queue_is_empty (&workers_data [i].private_gray_queue), "Why is there still work left to do?");
475 * Can only be called if the workers are stopped.
476 * If we're stopped, there are also no pending jobs.
479 sgen_workers_have_idle_work (void)
483 SGEN_ASSERT (0, forced_stop && sgen_workers_all_done (), "Checking for idle work should only happen if the workers are stopped.");
485 if (!sgen_section_gray_queue_is_empty (&workers_distribute_gray_queue))
488 for (i = 0; i < active_workers_num; ++i) {
489 if (!sgen_gray_object_queue_is_empty (&workers_data [i].private_gray_queue))
497 sgen_workers_all_done (void)
501 for (i = 0; i < active_workers_num; i++) {
502 if (state_is_working_or_enqueued (workers_data [i].state))
508 /* Must only be used for debugging */
510 sgen_workers_are_working (void)
512 return !sgen_workers_all_done ();
516 sgen_workers_assert_gray_queue_is_empty (void)
518 SGEN_ASSERT (0, sgen_section_gray_queue_is_empty (&workers_distribute_gray_queue), "Why is the workers gray queue not empty?");
522 sgen_workers_take_from_queue (SgenGrayQueue *queue)
524 sgen_gray_object_spread (queue, sgen_workers_get_job_split_count ());
527 GrayQueueSection *section = sgen_gray_object_dequeue_section (queue);
530 sgen_section_gray_queue_enqueue (&workers_distribute_gray_queue, section);
533 SGEN_ASSERT (0, !sgen_workers_are_working (), "We should fully populate the distribute gray queue before we start the workers");
536 SgenObjectOperations*
537 sgen_workers_get_idle_func_object_ops (void)
539 g_assert (idle_func_object_ops);
540 return idle_func_object_ops;
544 * If we have a single worker, splitting into multiple jobs makes no sense. With
545 * more than one worker, we split into a larger number of jobs so that, in case
546 * the work load is uneven, a worker that finished quickly can take up more jobs
550 sgen_workers_get_job_split_count (void)
552 return (active_workers_num > 1) ? active_workers_num * 4 : 1;
556 sgen_workers_foreach (SgenWorkerCallback callback)
560 for (i = 0; i < workers_num; i++)
561 callback (&workers_data [i]);
565 sgen_workers_is_worker_thread (MonoNativeThreadId id)
569 return sgen_thread_pool_is_thread_pool_thread (pool, id);