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 forced_stop;
26 static WorkerData *workers_data;
27 static SgenWorkerCallback worker_init_cb;
30 * When using multiple workers, we need to have the last worker
31 * enqueue the preclean jobs (if there are any). This lock ensures
32 * that when the last worker takes it, all the other workers have
33 * gracefully finished, so it can restart them.
35 static mono_mutex_t finished_lock;
36 static volatile gboolean workers_finished;
37 static int worker_awakenings;
39 static SgenSectionGrayQueue workers_distribute_gray_queue;
40 static gboolean workers_distribute_gray_queue_inited;
43 * Allowed transitions:
45 * | from \ to | NOT WORKING | WORKING | WORK ENQUEUED |
46 * |--------------------+-------------+---------+---------------+
47 * | NOT WORKING | - | - | main / worker |
48 * | WORKING | worker | - | main / worker |
49 * | WORK ENQUEUED | - | worker | - |
51 * The WORK ENQUEUED state guarantees that the worker thread will inspect the queue again at
52 * least once. Only after looking at the queue will it go back to WORKING, and then,
53 * eventually, to NOT WORKING. After enqueuing work the main thread transitions the state
54 * to WORK ENQUEUED. Signalling the worker thread to wake up is only necessary if the old
55 * state was NOT WORKING.
64 #define SGEN_WORKER_MIN_SECTIONS_SIGNAL 4
68 static SgenObjectOperations * volatile idle_func_object_ops;
69 static SgenObjectOperations *idle_func_object_ops_par, *idle_func_object_ops_nopar;
71 * finished_callback is called only when the workers finish work normally (when they
72 * are not forced to finish). The callback is used to enqueue preclean jobs.
74 static volatile SgenWorkersFinishCallback finish_callback;
76 static guint64 stat_workers_num_finished;
79 set_state (WorkerData *data, State old_state, State new_state)
81 SGEN_ASSERT (0, old_state != new_state, "Why are we transitioning to the same state?");
82 if (new_state == STATE_NOT_WORKING)
83 SGEN_ASSERT (0, old_state == STATE_WORKING, "We can only transition to NOT WORKING from WORKING");
84 else if (new_state == STATE_WORKING)
85 SGEN_ASSERT (0, old_state == STATE_WORK_ENQUEUED, "We can only transition to WORKING from WORK ENQUEUED");
86 if (new_state == STATE_NOT_WORKING || new_state == STATE_WORKING)
87 SGEN_ASSERT (6, sgen_thread_pool_is_thread_pool_thread (mono_native_thread_id_get ()), "Only the worker thread is allowed to transition to NOT_WORKING or WORKING");
89 return InterlockedCompareExchange (&data->state, new_state, old_state) == old_state;
93 state_is_working_or_enqueued (State state)
95 return state == STATE_WORKING || state == STATE_WORK_ENQUEUED;
99 sgen_workers_ensure_awake (void)
102 gboolean need_signal = FALSE;
105 * All workers are awaken, make sure we reset the parallel context.
106 * We call this function only when starting the workers so nobody is running,
107 * or when the last worker is enqueuing preclean work. In both cases we can't
108 * have a worker working using a nopar context, which means it is safe.
110 idle_func_object_ops = (active_workers_num > 1) ? idle_func_object_ops_par : idle_func_object_ops_nopar;
111 workers_finished = FALSE;
113 for (i = 0; i < active_workers_num; i++) {
115 gboolean did_set_state;
118 old_state = workers_data [i].state;
120 if (old_state == STATE_WORK_ENQUEUED)
123 did_set_state = set_state (&workers_data [i], old_state, STATE_WORK_ENQUEUED);
124 } while (!did_set_state);
126 if (!state_is_working_or_enqueued (old_state))
131 sgen_thread_pool_idle_signal ();
135 worker_try_finish (WorkerData *data)
140 ++stat_workers_num_finished;
142 mono_os_mutex_lock (&finished_lock);
144 for (i = 0; i < active_workers_num; i++) {
145 if (state_is_working_or_enqueued (workers_data [i].state))
150 SgenWorkersFinishCallback callback = finish_callback;
151 SGEN_ASSERT (0, idle_func_object_ops == idle_func_object_ops_nopar, "Why are we finishing with parallel context");
152 /* We are the last one left. Enqueue preclean job if we have one and awake everybody */
153 SGEN_ASSERT (0, data->state != STATE_NOT_WORKING, "How did we get from doing idle work to NOT WORKING without setting it ourselves?");
155 finish_callback = NULL;
157 worker_awakenings = 0;
158 /* Make sure each worker has a chance of seeing the enqueued jobs */
159 sgen_workers_ensure_awake ();
160 SGEN_ASSERT (0, data->state == STATE_WORK_ENQUEUED, "Why did we fail to set our own state to ENQUEUED");
166 old_state = data->state;
168 SGEN_ASSERT (0, old_state != STATE_NOT_WORKING, "How did we get from doing idle work to NOT WORKING without setting it ourselves?");
169 if (old_state == STATE_WORK_ENQUEUED)
171 SGEN_ASSERT (0, old_state == STATE_WORKING, "What other possibility is there?");
172 } while (!set_state (data, old_state, STATE_NOT_WORKING));
175 * If we are second to last to finish, we set the scan context to the non-parallel
176 * version so we can speed up the last worker. This helps us maintain same level
177 * of performance as non-parallel mode even if we fail to distribute work properly.
180 idle_func_object_ops = idle_func_object_ops_nopar;
182 workers_finished = TRUE;
183 mono_os_mutex_unlock (&finished_lock);
185 binary_protocol_worker_finish (sgen_timestamp (), forced_stop);
187 sgen_gray_object_queue_trim_free_list (&data->private_gray_queue);
191 mono_os_mutex_unlock (&finished_lock);
195 sgen_workers_enqueue_job (SgenThreadPoolJob *job, gboolean enqueue)
198 job->func (NULL, job);
199 sgen_thread_pool_job_free (job);
203 sgen_thread_pool_job_enqueue (job);
207 workers_get_work (WorkerData *data)
209 SgenMajorCollector *major = sgen_get_major_collector ();
210 SgenMinorCollector *minor = sgen_get_minor_collector ();
211 GrayQueueSection *section;
213 g_assert (sgen_gray_object_queue_is_empty (&data->private_gray_queue));
214 g_assert (major->is_concurrent || minor->is_parallel);
216 section = sgen_section_gray_queue_dequeue (&workers_distribute_gray_queue);
218 sgen_gray_object_enqueue_section (&data->private_gray_queue, section, major->is_parallel);
222 /* Nobody to steal from */
223 g_assert (sgen_gray_object_queue_is_empty (&data->private_gray_queue));
228 workers_steal_work (WorkerData *data)
230 SgenMajorCollector *major = sgen_get_major_collector ();
231 SgenMinorCollector *minor = sgen_get_minor_collector ();
232 int generation = sgen_get_current_collection_generation ();
233 GrayQueueSection *section = NULL;
234 int i, current_worker;
236 if ((generation == GENERATION_OLD && !major->is_parallel) ||
237 (generation == GENERATION_NURSERY && !minor->is_parallel))
240 /* If we're parallel, steal from other workers' private gray queues */
241 g_assert (sgen_gray_object_queue_is_empty (&data->private_gray_queue));
243 current_worker = (int) (data - workers_data);
245 for (i = 1; i < active_workers_num && !section; i++) {
246 int steal_worker = (current_worker + i) % active_workers_num;
247 if (state_is_working_or_enqueued (workers_data [steal_worker].state))
248 section = sgen_gray_object_steal_section (&workers_data [steal_worker].private_gray_queue);
252 sgen_gray_object_enqueue_section (&data->private_gray_queue, section, TRUE);
256 /* Nobody to steal from */
257 g_assert (sgen_gray_object_queue_is_empty (&data->private_gray_queue));
262 concurrent_enqueue_check (GCObject *obj)
264 g_assert (sgen_concurrent_collection_in_progress ());
265 g_assert (!sgen_ptr_in_nursery (obj));
266 g_assert (SGEN_LOAD_VTABLE (obj));
270 init_private_gray_queue (WorkerData *data)
272 sgen_gray_object_queue_init (&data->private_gray_queue,
273 sgen_get_major_collector ()->is_concurrent ? concurrent_enqueue_check : NULL,
278 thread_pool_init_func (void *data_untyped)
280 WorkerData *data = (WorkerData *)data_untyped;
281 SgenMajorCollector *major = sgen_get_major_collector ();
282 SgenMinorCollector *minor = sgen_get_minor_collector ();
284 sgen_client_thread_register_worker ();
286 if (!major->is_concurrent && !minor->is_parallel)
289 init_private_gray_queue (data);
292 worker_init_cb (data);
296 continue_idle_func (void *data_untyped)
299 WorkerData *data = (WorkerData *)data_untyped;
300 return state_is_working_or_enqueued (data->state);
302 /* Return if any of the threads is working */
303 return !sgen_workers_all_done ();
308 should_work_func (void *data_untyped)
310 WorkerData *data = (WorkerData*)data_untyped;
311 int current_worker = (int) (data - workers_data);
313 return current_worker < active_workers_num;
317 marker_idle_func (void *data_untyped)
319 WorkerData *data = (WorkerData *)data_untyped;
321 SGEN_ASSERT (0, continue_idle_func (data_untyped), "Why are we called when we're not supposed to work?");
323 if (data->state == STATE_WORK_ENQUEUED) {
324 set_state (data, STATE_WORK_ENQUEUED, STATE_WORKING);
325 SGEN_ASSERT (0, data->state != STATE_NOT_WORKING, "How did we get from WORK ENQUEUED to NOT WORKING?");
328 if (!forced_stop && (!sgen_gray_object_queue_is_empty (&data->private_gray_queue) || workers_get_work (data) || workers_steal_work (data))) {
329 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (idle_func_object_ops, &data->private_gray_queue);
331 SGEN_ASSERT (0, !sgen_gray_object_queue_is_empty (&data->private_gray_queue), "How is our gray queue empty if we just got work?");
333 sgen_drain_gray_stack (ctx);
335 if (data->private_gray_queue.num_sections >= SGEN_WORKER_MIN_SECTIONS_SIGNAL
336 && workers_finished && worker_awakenings < active_workers_num) {
337 /* We bound the number of worker awakenings just to be sure */
339 mono_os_mutex_lock (&finished_lock);
340 sgen_workers_ensure_awake ();
341 mono_os_mutex_unlock (&finished_lock);
344 worker_try_finish (data);
349 init_distribute_gray_queue (void)
351 if (workers_distribute_gray_queue_inited) {
352 g_assert (sgen_section_gray_queue_is_empty (&workers_distribute_gray_queue));
353 g_assert (workers_distribute_gray_queue.locked);
357 sgen_section_gray_queue_init (&workers_distribute_gray_queue, TRUE,
358 sgen_get_major_collector ()->is_concurrent ? concurrent_enqueue_check : NULL);
359 workers_distribute_gray_queue_inited = TRUE;
363 sgen_workers_init_distribute_gray_queue (void)
365 SGEN_ASSERT (0, sgen_get_major_collector ()->is_concurrent || sgen_get_minor_collector ()->is_parallel,
366 "Why should we init the distribute gray queue if we don't need it?");
367 init_distribute_gray_queue ();
371 sgen_workers_init (int num_workers, SgenWorkerCallback callback)
374 void **workers_data_ptrs = (void **)alloca(num_workers * sizeof(void *));
376 if (!sgen_get_major_collector ()->is_concurrent && !sgen_get_minor_collector ()->is_parallel) {
377 sgen_thread_pool_init (num_workers, thread_pool_init_func, NULL, NULL, NULL, NULL);
381 mono_os_mutex_init (&finished_lock);
382 //g_print ("initing %d workers\n", num_workers);
384 workers_num = num_workers;
385 active_workers_num = num_workers;
387 workers_data = (WorkerData *)sgen_alloc_internal_dynamic (sizeof (WorkerData) * num_workers, INTERNAL_MEM_WORKER_DATA, TRUE);
388 memset (workers_data, 0, sizeof (WorkerData) * num_workers);
390 init_distribute_gray_queue ();
392 for (i = 0; i < num_workers; ++i)
393 workers_data_ptrs [i] = (void *) &workers_data [i];
395 worker_init_cb = callback;
397 sgen_thread_pool_init (num_workers, thread_pool_init_func, marker_idle_func, continue_idle_func, should_work_func, workers_data_ptrs);
399 mono_counters_register ("# workers finished", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_workers_num_finished);
403 sgen_workers_stop_all_workers (void)
405 finish_callback = NULL;
406 mono_memory_write_barrier ();
409 sgen_thread_pool_wait_for_all_jobs ();
410 sgen_thread_pool_idle_wait ();
411 SGEN_ASSERT (0, sgen_workers_all_done (), "Can only signal enqueue work when in no work state");
415 sgen_workers_set_num_active_workers (int num_workers)
418 SGEN_ASSERT (0, active_workers_num <= workers_num, "We can't start more workers than we initialized");
419 active_workers_num = num_workers;
421 active_workers_num = workers_num;
426 sgen_workers_start_all_workers (SgenObjectOperations *object_ops_nopar, SgenObjectOperations *object_ops_par, SgenWorkersFinishCallback callback)
428 idle_func_object_ops_par = object_ops_par;
429 idle_func_object_ops_nopar = object_ops_nopar;
431 finish_callback = callback;
432 worker_awakenings = 0;
433 mono_memory_write_barrier ();
436 * We expect workers to start finishing only after all of them were awaken.
437 * Otherwise we might think that we have fewer workers and use wrong context.
439 mono_os_mutex_lock (&finished_lock);
440 sgen_workers_ensure_awake ();
441 mono_os_mutex_unlock (&finished_lock);
445 sgen_workers_join (void)
449 sgen_thread_pool_wait_for_all_jobs ();
450 sgen_thread_pool_idle_wait ();
451 SGEN_ASSERT (0, sgen_workers_all_done (), "Can only signal enqueue work when in no work state");
453 /* At this point all the workers have stopped. */
455 SGEN_ASSERT (0, sgen_section_gray_queue_is_empty (&workers_distribute_gray_queue), "Why is there still work left to do?");
456 for (i = 0; i < active_workers_num; ++i)
457 SGEN_ASSERT (0, sgen_gray_object_queue_is_empty (&workers_data [i].private_gray_queue), "Why is there still work left to do?");
461 * Can only be called if the workers are stopped.
462 * If we're stopped, there are also no pending jobs.
465 sgen_workers_have_idle_work (void)
469 SGEN_ASSERT (0, forced_stop && sgen_workers_all_done (), "Checking for idle work should only happen if the workers are stopped.");
471 if (!sgen_section_gray_queue_is_empty (&workers_distribute_gray_queue))
474 for (i = 0; i < active_workers_num; ++i) {
475 if (!sgen_gray_object_queue_is_empty (&workers_data [i].private_gray_queue))
483 sgen_workers_all_done (void)
487 for (i = 0; i < active_workers_num; i++) {
488 if (state_is_working_or_enqueued (workers_data [i].state))
494 /* Must only be used for debugging */
496 sgen_workers_are_working (void)
498 return !sgen_workers_all_done ();
502 sgen_workers_assert_gray_queue_is_empty (void)
504 SGEN_ASSERT (0, sgen_section_gray_queue_is_empty (&workers_distribute_gray_queue), "Why is the workers gray queue not empty?");
508 sgen_workers_take_from_queue (SgenGrayQueue *queue)
510 sgen_gray_object_spread (queue, sgen_workers_get_job_split_count ());
513 GrayQueueSection *section = sgen_gray_object_dequeue_section (queue);
516 sgen_section_gray_queue_enqueue (&workers_distribute_gray_queue, section);
519 SGEN_ASSERT (0, !sgen_workers_are_working (), "We should fully populate the distribute gray queue before we start the workers");
522 SgenObjectOperations*
523 sgen_workers_get_idle_func_object_ops (void)
525 return (idle_func_object_ops_par) ? idle_func_object_ops_par : idle_func_object_ops_nopar;
529 * If we have a single worker, splitting into multiple jobs makes no sense. With
530 * more than one worker, we split into a larger number of jobs so that, in case
531 * the work load is uneven, a worker that finished quickly can take up more jobs
535 sgen_workers_get_job_split_count (void)
537 return (active_workers_num > 1) ? active_workers_num * 4 : 1;
541 sgen_workers_foreach (SgenWorkerCallback callback)
545 for (i = 0; i < workers_num; i++)
546 callback (&workers_data [i]);