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 WorkerContext worker_contexts [GENERATION_MAX];
26 * Allowed transitions:
28 * | from \ to | NOT WORKING | WORKING | WORK ENQUEUED |
29 * |--------------------+-------------+---------+---------------+
30 * | NOT WORKING | - | - | main / worker |
31 * | WORKING | worker | - | main / worker |
32 * | WORK ENQUEUED | - | worker | - |
34 * The WORK ENQUEUED state guarantees that the worker thread will inspect the queue again at
35 * least once. Only after looking at the queue will it go back to WORKING, and then,
36 * eventually, to NOT WORKING. After enqueuing work the main thread transitions the state
37 * to WORK ENQUEUED. Signalling the worker thread to wake up is only necessary if the old
38 * state was NOT WORKING.
47 #define SGEN_WORKER_MIN_SECTIONS_SIGNAL 4
49 static guint64 stat_workers_num_finished;
52 set_state (WorkerData *data, State old_state, State new_state)
54 SGEN_ASSERT (0, old_state != new_state, "Why are we transitioning to the same state?");
55 if (new_state == STATE_NOT_WORKING)
56 SGEN_ASSERT (0, old_state == STATE_WORKING, "We can only transition to NOT WORKING from WORKING");
57 else if (new_state == STATE_WORKING)
58 SGEN_ASSERT (0, old_state == STATE_WORK_ENQUEUED, "We can only transition to WORKING from WORK ENQUEUED");
60 return InterlockedCompareExchange (&data->state, new_state, old_state) == old_state;
64 state_is_working_or_enqueued (State state)
66 return state == STATE_WORKING || state == STATE_WORK_ENQUEUED;
70 sgen_workers_ensure_awake (WorkerContext *context)
73 gboolean need_signal = FALSE;
76 * All workers are awaken, make sure we reset the parallel context.
77 * We call this function only when starting the workers so nobody is running,
78 * or when the last worker is enqueuing preclean work. In both cases we can't
79 * have a worker working using a nopar context, which means it is safe.
81 context->idle_func_object_ops = (context->active_workers_num > 1) ? context->idle_func_object_ops_par : context->idle_func_object_ops_nopar;
82 context->workers_finished = FALSE;
84 for (i = 0; i < context->active_workers_num; i++) {
86 gboolean did_set_state;
89 old_state = context->workers_data [i].state;
91 if (old_state == STATE_WORK_ENQUEUED)
94 did_set_state = set_state (&context->workers_data [i], old_state, STATE_WORK_ENQUEUED);
96 if (did_set_state && old_state == STATE_NOT_WORKING)
97 context->workers_data [i].last_start = sgen_timestamp ();
98 } while (!did_set_state);
100 if (!state_is_working_or_enqueued (old_state))
105 sgen_thread_pool_idle_signal (context->thread_pool_context);
109 worker_try_finish (WorkerData *data)
113 WorkerContext *context = data->context;
114 gint64 last_start = data->last_start;
116 ++stat_workers_num_finished;
118 mono_os_mutex_lock (&context->finished_lock);
120 for (i = 0; i < context->active_workers_num; i++) {
121 if (state_is_working_or_enqueued (context->workers_data [i].state))
126 SgenWorkersFinishCallback callback = context->finish_callback;
127 SGEN_ASSERT (0, context->idle_func_object_ops == context->idle_func_object_ops_nopar, "Why are we finishing with parallel context");
128 /* We are the last one left. Enqueue preclean job if we have one and awake everybody */
129 SGEN_ASSERT (0, data->state != STATE_NOT_WORKING, "How did we get from doing idle work to NOT WORKING without setting it ourselves?");
131 context->finish_callback = NULL;
133 context->worker_awakenings = 0;
134 /* Make sure each worker has a chance of seeing the enqueued jobs */
135 sgen_workers_ensure_awake (context);
136 SGEN_ASSERT (0, data->state == STATE_WORK_ENQUEUED, "Why did we fail to set our own state to ENQUEUED");
139 * Log to be able to get the duration of normal concurrent M&S phase.
140 * Worker indexes are 1 based, since 0 is logically considered gc thread.
142 binary_protocol_worker_finish_stats (data - &context->workers_data [0] + 1, context->generation, context->forced_stop, data->major_scan_time, data->los_scan_time, data->total_time + sgen_timestamp () - last_start);
148 old_state = data->state;
150 SGEN_ASSERT (0, old_state != STATE_NOT_WORKING, "How did we get from doing idle work to NOT WORKING without setting it ourselves?");
151 if (old_state == STATE_WORK_ENQUEUED)
153 SGEN_ASSERT (0, old_state == STATE_WORKING, "What other possibility is there?");
154 } while (!set_state (data, old_state, STATE_NOT_WORKING));
157 * If we are second to last to finish, we set the scan context to the non-parallel
158 * version so we can speed up the last worker. This helps us maintain same level
159 * of performance as non-parallel mode even if we fail to distribute work properly.
162 context->idle_func_object_ops = context->idle_func_object_ops_nopar;
164 context->workers_finished = TRUE;
165 mono_os_mutex_unlock (&context->finished_lock);
167 data->total_time += (sgen_timestamp () - last_start);
168 binary_protocol_worker_finish_stats (data - &context->workers_data [0] + 1, context->generation, context->forced_stop, data->major_scan_time, data->los_scan_time, data->total_time);
170 binary_protocol_worker_finish (sgen_timestamp (), context->forced_stop);
172 sgen_gray_object_queue_trim_free_list (&data->private_gray_queue);
176 mono_os_mutex_unlock (&context->finished_lock);
180 sgen_workers_enqueue_job (int generation, SgenThreadPoolJob *job, gboolean enqueue)
183 job->func (NULL, job);
184 sgen_thread_pool_job_free (job);
188 sgen_thread_pool_job_enqueue (worker_contexts [generation].thread_pool_context, job);
192 workers_get_work (WorkerData *data)
194 SgenMajorCollector *major = sgen_get_major_collector ();
195 SgenMinorCollector *minor = sgen_get_minor_collector ();
196 GrayQueueSection *section;
198 g_assert (sgen_gray_object_queue_is_empty (&data->private_gray_queue));
199 g_assert (major->is_concurrent || minor->is_parallel);
201 section = sgen_section_gray_queue_dequeue (&data->context->workers_distribute_gray_queue);
203 sgen_gray_object_enqueue_section (&data->private_gray_queue, section, major->is_parallel);
207 /* Nobody to steal from */
208 g_assert (sgen_gray_object_queue_is_empty (&data->private_gray_queue));
213 workers_steal_work (WorkerData *data)
215 SgenMajorCollector *major = sgen_get_major_collector ();
216 SgenMinorCollector *minor = sgen_get_minor_collector ();
217 int generation = sgen_get_current_collection_generation ();
218 GrayQueueSection *section = NULL;
219 WorkerContext *context = data->context;
220 int i, current_worker;
222 if ((generation == GENERATION_OLD && !major->is_parallel) ||
223 (generation == GENERATION_NURSERY && !minor->is_parallel))
226 /* If we're parallel, steal from other workers' private gray queues */
227 g_assert (sgen_gray_object_queue_is_empty (&data->private_gray_queue));
229 current_worker = (int) (data - context->workers_data);
231 for (i = 1; i < context->active_workers_num && !section; i++) {
232 int steal_worker = (current_worker + i) % context->active_workers_num;
233 if (state_is_working_or_enqueued (context->workers_data [steal_worker].state))
234 section = sgen_gray_object_steal_section (&context->workers_data [steal_worker].private_gray_queue);
238 sgen_gray_object_enqueue_section (&data->private_gray_queue, section, TRUE);
242 /* Nobody to steal from */
243 g_assert (sgen_gray_object_queue_is_empty (&data->private_gray_queue));
248 concurrent_enqueue_check (GCObject *obj)
250 g_assert (sgen_concurrent_collection_in_progress ());
251 g_assert (!sgen_ptr_in_nursery (obj));
252 g_assert (SGEN_LOAD_VTABLE (obj));
256 init_private_gray_queue (WorkerData *data)
258 sgen_gray_object_queue_init (&data->private_gray_queue,
259 sgen_get_major_collector ()->is_concurrent ? concurrent_enqueue_check : NULL,
264 thread_pool_init_func (void *data_untyped)
266 WorkerData *data = (WorkerData *)data_untyped;
267 SgenMajorCollector *major = sgen_get_major_collector ();
268 SgenMinorCollector *minor = sgen_get_minor_collector ();
270 if (!major->is_concurrent && !minor->is_parallel)
273 init_private_gray_queue (data);
275 /* Separate WorkerData for same thread share free_block_lists */
276 if (major->is_parallel || minor->is_parallel)
277 major->init_block_free_lists (&data->free_block_lists);
281 sgen_workers_are_working (WorkerContext *context)
285 for (i = 0; i < context->active_workers_num; i++) {
286 if (state_is_working_or_enqueued (context->workers_data [i].state))
293 continue_idle_func (void *data_untyped, int thread_pool_context)
296 return state_is_working_or_enqueued (((WorkerData*)data_untyped)->state);
298 /* Return if any of the threads is working in the context */
299 if (worker_contexts [GENERATION_NURSERY].workers_num && worker_contexts [GENERATION_NURSERY].thread_pool_context == thread_pool_context)
300 return sgen_workers_are_working (&worker_contexts [GENERATION_NURSERY]);
301 if (worker_contexts [GENERATION_OLD].workers_num && worker_contexts [GENERATION_OLD].thread_pool_context == thread_pool_context)
302 return sgen_workers_are_working (&worker_contexts [GENERATION_OLD]);
304 g_assert_not_reached ();
309 should_work_func (void *data_untyped)
311 WorkerData *data = (WorkerData*)data_untyped;
312 WorkerContext *context = data->context;
313 int current_worker = (int) (data - context->workers_data);
315 return context->started && current_worker < context->active_workers_num && state_is_working_or_enqueued (data->state);
319 marker_idle_func (void *data_untyped)
321 WorkerData *data = (WorkerData *)data_untyped;
322 WorkerContext *context = data->context;
324 SGEN_ASSERT (0, continue_idle_func (data_untyped, context->thread_pool_context), "Why are we called when we're not supposed to work?");
326 if (data->state == STATE_WORK_ENQUEUED) {
327 set_state (data, STATE_WORK_ENQUEUED, STATE_WORKING);
328 SGEN_ASSERT (0, data->state != STATE_NOT_WORKING, "How did we get from WORK ENQUEUED to NOT WORKING?");
331 if (!context->forced_stop && (!sgen_gray_object_queue_is_empty (&data->private_gray_queue) || workers_get_work (data) || workers_steal_work (data))) {
332 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (context->idle_func_object_ops, &data->private_gray_queue);
334 SGEN_ASSERT (0, !sgen_gray_object_queue_is_empty (&data->private_gray_queue), "How is our gray queue empty if we just got work?");
336 sgen_drain_gray_stack (ctx);
338 if (data->private_gray_queue.num_sections >= SGEN_WORKER_MIN_SECTIONS_SIGNAL
339 && context->workers_finished && context->worker_awakenings < context->active_workers_num) {
340 /* We bound the number of worker awakenings just to be sure */
341 context->worker_awakenings++;
342 mono_os_mutex_lock (&context->finished_lock);
343 sgen_workers_ensure_awake (context);
344 mono_os_mutex_unlock (&context->finished_lock);
347 worker_try_finish (data);
352 init_distribute_gray_queue (WorkerContext *context)
354 sgen_section_gray_queue_init (&context->workers_distribute_gray_queue, TRUE,
355 sgen_get_major_collector ()->is_concurrent ? concurrent_enqueue_check : NULL);
359 sgen_workers_create_context (int generation, int num_workers)
361 static gboolean stat_inited = FALSE;
363 WorkerData **workers_data_ptrs = (WorkerData**)sgen_alloc_internal_dynamic (num_workers * sizeof(WorkerData*), INTERNAL_MEM_WORKER_DATA, TRUE);
364 WorkerContext *context = &worker_contexts [generation];
366 SGEN_ASSERT (0, !context->workers_num, "We can't init the worker context for a generation twice");
368 mono_os_mutex_init (&context->finished_lock);
370 context->generation = generation;
371 context->workers_num = num_workers;
372 context->active_workers_num = num_workers;
374 context->workers_data = (WorkerData *)sgen_alloc_internal_dynamic (sizeof (WorkerData) * num_workers, INTERNAL_MEM_WORKER_DATA, TRUE);
375 memset (context->workers_data, 0, sizeof (WorkerData) * num_workers);
377 init_distribute_gray_queue (context);
379 for (i = 0; i < num_workers; ++i) {
380 workers_data_ptrs [i] = &context->workers_data [i];
381 context->workers_data [i].context = context;
384 context->thread_pool_context = sgen_thread_pool_create_context (num_workers, thread_pool_init_func, marker_idle_func, continue_idle_func, should_work_func, (void**)workers_data_ptrs);
387 mono_counters_register ("# workers finished", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_workers_num_finished);
392 /* This is called with thread pool lock so no context switch can happen */
394 continue_idle_wait (int calling_context, int *threads_context)
396 WorkerContext *context;
399 if (worker_contexts [GENERATION_OLD].workers_num && calling_context == worker_contexts [GENERATION_OLD].thread_pool_context)
400 context = &worker_contexts [GENERATION_OLD];
401 else if (worker_contexts [GENERATION_NURSERY].workers_num && calling_context == worker_contexts [GENERATION_NURSERY].thread_pool_context)
402 context = &worker_contexts [GENERATION_NURSERY];
404 g_assert_not_reached ();
407 * We assume there are no pending jobs, since this is called only after
408 * we waited for all the jobs.
410 for (i = 0; i < context->active_workers_num; i++) {
411 if (threads_context [i] == calling_context)
415 if (sgen_workers_have_idle_work (context->generation) && !context->forced_stop)
419 * At this point there are no jobs to be done, and no objects to be scanned
420 * in the gray queues. We can simply asynchronously finish all the workers
421 * from the context that were not finished already (due to being stuck working
422 * in another context)
425 for (i = 0; i < context->active_workers_num; i++) {
426 if (context->workers_data [i].state == STATE_WORK_ENQUEUED)
427 set_state (&context->workers_data [i], STATE_WORK_ENQUEUED, STATE_WORKING);
428 if (context->workers_data [i].state == STATE_WORKING)
429 worker_try_finish (&context->workers_data [i]);
437 sgen_workers_stop_all_workers (int generation)
439 WorkerContext *context = &worker_contexts [generation];
441 mono_os_mutex_lock (&context->finished_lock);
442 context->finish_callback = NULL;
443 mono_os_mutex_unlock (&context->finished_lock);
445 context->forced_stop = TRUE;
447 sgen_thread_pool_wait_for_all_jobs (context->thread_pool_context);
448 sgen_thread_pool_idle_wait (context->thread_pool_context, continue_idle_wait);
449 SGEN_ASSERT (0, !sgen_workers_are_working (context), "Can only signal enqueue work when in no work state");
451 context->started = FALSE;
455 sgen_workers_set_num_active_workers (int generation, int num_workers)
457 WorkerContext *context = &worker_contexts [generation];
459 SGEN_ASSERT (0, num_workers <= context->workers_num, "We can't start more workers than we initialized");
460 context->active_workers_num = num_workers;
462 context->active_workers_num = context->workers_num;
467 sgen_workers_start_all_workers (int generation, SgenObjectOperations *object_ops_nopar, SgenObjectOperations *object_ops_par, SgenWorkersFinishCallback callback)
469 WorkerContext *context = &worker_contexts [generation];
471 SGEN_ASSERT (0, !context->started, "Why are we starting to work without finishing previous cycle");
473 context->idle_func_object_ops_par = object_ops_par;
474 context->idle_func_object_ops_nopar = object_ops_nopar;
475 context->forced_stop = FALSE;
476 context->finish_callback = callback;
477 context->worker_awakenings = 0;
478 context->started = TRUE;
480 for (i = 0; i < context->active_workers_num; i++) {
481 context->workers_data [i].major_scan_time = 0;
482 context->workers_data [i].los_scan_time = 0;
483 context->workers_data [i].total_time = 0;
484 context->workers_data [i].last_start = 0;
486 mono_memory_write_barrier ();
489 * We expect workers to start finishing only after all of them were awaken.
490 * Otherwise we might think that we have fewer workers and use wrong context.
492 mono_os_mutex_lock (&context->finished_lock);
493 sgen_workers_ensure_awake (context);
494 mono_os_mutex_unlock (&context->finished_lock);
498 sgen_workers_join (int generation)
500 WorkerContext *context = &worker_contexts [generation];
503 SGEN_ASSERT (0, !context->finish_callback, "Why are we joining concurrent mark early");
505 sgen_thread_pool_wait_for_all_jobs (context->thread_pool_context);
506 sgen_thread_pool_idle_wait (context->thread_pool_context, continue_idle_wait);
507 SGEN_ASSERT (0, !sgen_workers_are_working (context), "Can only signal enqueue work when in no work state");
509 /* At this point all the workers have stopped. */
511 SGEN_ASSERT (0, sgen_section_gray_queue_is_empty (&context->workers_distribute_gray_queue), "Why is there still work left to do?");
512 for (i = 0; i < context->active_workers_num; ++i)
513 SGEN_ASSERT (0, sgen_gray_object_queue_is_empty (&context->workers_data [i].private_gray_queue), "Why is there still work left to do?");
515 context->started = FALSE;
519 * Can only be called if the workers are not working in the
520 * context and there are no pending jobs.
523 sgen_workers_have_idle_work (int generation)
525 WorkerContext *context = &worker_contexts [generation];
528 if (!sgen_section_gray_queue_is_empty (&context->workers_distribute_gray_queue))
531 for (i = 0; i < context->active_workers_num; ++i) {
532 if (!sgen_gray_object_queue_is_empty (&context->workers_data [i].private_gray_queue))
540 sgen_workers_all_done (void)
542 if (worker_contexts [GENERATION_NURSERY].workers_num && sgen_workers_are_working (&worker_contexts [GENERATION_NURSERY]))
544 if (worker_contexts [GENERATION_OLD].workers_num && sgen_workers_are_working (&worker_contexts [GENERATION_OLD]))
551 sgen_workers_assert_gray_queue_is_empty (int generation)
553 SGEN_ASSERT (0, sgen_section_gray_queue_is_empty (&worker_contexts [generation].workers_distribute_gray_queue), "Why is the workers gray queue not empty?");
557 sgen_workers_take_from_queue (int generation, SgenGrayQueue *queue)
559 WorkerContext *context = &worker_contexts [generation];
561 sgen_gray_object_spread (queue, sgen_workers_get_job_split_count (generation));
564 GrayQueueSection *section = sgen_gray_object_dequeue_section (queue);
567 sgen_section_gray_queue_enqueue (&context->workers_distribute_gray_queue, section);
570 SGEN_ASSERT (0, !sgen_workers_are_working (context), "We should fully populate the distribute gray queue before we start the workers");
573 SgenObjectOperations*
574 sgen_workers_get_idle_func_object_ops (WorkerData *worker)
576 g_assert (worker->context->idle_func_object_ops);
577 return worker->context->idle_func_object_ops;
581 * If we have a single worker, splitting into multiple jobs makes no sense. With
582 * more than one worker, we split into a larger number of jobs so that, in case
583 * the work load is uneven, a worker that finished quickly can take up more jobs
586 * We also return 1 if there is no worker context for that generation.
589 sgen_workers_get_job_split_count (int generation)
591 return (worker_contexts [generation].active_workers_num > 1) ? worker_contexts [generation].active_workers_num * 4 : 1;
595 sgen_workers_foreach (int generation, SgenWorkerCallback callback)
597 WorkerContext *context = &worker_contexts [generation];
600 for (i = 0; i < context->workers_num; i++)
601 callback (&context->workers_data [i]);
605 sgen_workers_is_worker_thread (MonoNativeThreadId id)
607 return sgen_thread_pool_is_thread_pool_thread (id);