1 /*
2 * Copyright (c) 2001, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/shared/gcId.hpp"
27 #include "gc/shared/workgroup.hpp"
28 #include "gc/shared/workerManager.hpp"
29 #include "memory/allocation.hpp"
30 #include "memory/allocation.inline.hpp"
31 #include "memory/iterator.hpp"
32 #include "runtime/atomic.hpp"
33 #include "runtime/os.hpp"
34 #include "runtime/semaphore.hpp"
35 #include "runtime/thread.inline.hpp"
36
37 // Definitions of WorkGang methods.
38
39 // The current implementation will exit if the allocation
40 // of any worker fails.
41 void AbstractWorkGang::initialize_workers() {
42 log_develop_trace(gc, workgang)("Constructing work gang %s with %u threads", name(), total_workers());
43 _workers = NEW_C_HEAP_ARRAY(AbstractGangWorker*, total_workers(), mtInternal);
44 add_workers(true);
45 }
46
47
48 AbstractGangWorker* AbstractWorkGang::install_worker(uint worker_id) {
49 AbstractGangWorker* new_worker = allocate_worker(worker_id);
50 set_thread(worker_id, new_worker);
51 return new_worker;
52 }
53
54 void AbstractWorkGang::add_workers(bool initializing) {
55 add_workers(_active_workers, initializing);
56 }
57
58 void AbstractWorkGang::add_workers(uint active_workers, bool initializing) {
59
60 os::ThreadType worker_type;
61 if (are_ConcurrentGC_threads()) {
62 worker_type = os::cgc_thread;
63 } else {
64 worker_type = os::pgc_thread;
65 }
66 uint previous_created_workers = _created_workers;
67
68 _created_workers = WorkerManager::add_workers(this,
69 active_workers,
70 _total_workers,
71 _created_workers,
72 worker_type,
73 initializing);
74 _active_workers = MIN2(_created_workers, _active_workers);
75
76 WorkerManager::log_worker_creation(this, previous_created_workers, _active_workers, _created_workers, initializing);
77 }
78
79 AbstractGangWorker* AbstractWorkGang::worker(uint i) const {
80 // Array index bounds checking.
81 AbstractGangWorker* result = NULL;
82 assert(_workers != NULL, "No workers for indexing");
83 assert(i < total_workers(), "Worker index out of bounds");
84 result = _workers[i];
85 assert(result != NULL, "Indexing to null worker");
86 return result;
87 }
88
89 void AbstractWorkGang::print_worker_threads_on(outputStream* st) const {
90 uint workers = created_workers();
91 for (uint i = 0; i < workers; i++) {
92 worker(i)->print_on(st);
93 st->cr();
94 }
95 }
96
97 void AbstractWorkGang::threads_do(ThreadClosure* tc) const {
98 assert(tc != NULL, "Null ThreadClosure");
99 uint workers = created_workers();
100 for (uint i = 0; i < workers; i++) {
101 tc->do_thread(worker(i));
102 }
103 }
104
105 // WorkGang dispatcher implemented with semaphores.
106 //
107 // Semaphores don't require the worker threads to re-claim the lock when they wake up.
108 // This helps lowering the latency when starting and stopping the worker threads.
109 class SemaphoreGangTaskDispatcher : public GangTaskDispatcher {
110 // The task currently being dispatched to the GangWorkers.
111 AbstractGangTask* _task;
112
113 volatile uint _started;
114 volatile uint _not_finished;
115
116 // Semaphore used to start the GangWorkers.
117 Semaphore* _start_semaphore;
118 // Semaphore used to notify the coordinator that all workers are done.
119 Semaphore* _end_semaphore;
120
121 public:
122 SemaphoreGangTaskDispatcher() :
123 _task(NULL),
124 _started(0),
125 _not_finished(0),
126 _start_semaphore(new Semaphore()),
127 _end_semaphore(new Semaphore())
128 { }
129
130 ~SemaphoreGangTaskDispatcher() {
131 delete _start_semaphore;
132 delete _end_semaphore;
133 }
134
135 void coordinator_execute_on_workers(AbstractGangTask* task, uint num_workers) {
136 // No workers are allowed to read the state variables until they have been signaled.
137 _task = task;
138 _not_finished = num_workers;
139
140 // Dispatch 'num_workers' number of tasks.
141 _start_semaphore->signal(num_workers);
142
143 // Wait for the last worker to signal the coordinator.
144 _end_semaphore->wait();
145
146 // No workers are allowed to read the state variables after the coordinator has been signaled.
147 assert(_not_finished == 0, "%d not finished workers?", _not_finished);
148 _task = NULL;
149 _started = 0;
150
151 }
152
153 WorkData worker_wait_for_task() {
154 // Wait for the coordinator to dispatch a task.
155 _start_semaphore->wait();
156
157 uint num_started = Atomic::add(&_started, 1u);
158
159 // Subtract one to get a zero-indexed worker id.
160 uint worker_id = num_started - 1;
161
162 return WorkData(_task, worker_id);
163 }
164
165 void worker_done_with_task() {
166 // Mark that the worker is done with the task.
167 // The worker is not allowed to read the state variables after this line.
168 uint not_finished = Atomic::sub(&_not_finished, 1u);
169
170 // The last worker signals to the coordinator that all work is completed.
171 if (not_finished == 0) {
172 _end_semaphore->signal();
173 }
174 }
175 };
176
177 class MutexGangTaskDispatcher : public GangTaskDispatcher {
178 AbstractGangTask* _task;
179
180 volatile uint _started;
181 volatile uint _finished;
182 volatile uint _num_workers;
183
184 Monitor* _monitor;
185
186 public:
187 MutexGangTaskDispatcher() :
188 _task(NULL),
189 _started(0),
190 _finished(0),
191 _num_workers(0),
192 _monitor(new Monitor(Monitor::leaf, "WorkGang dispatcher lock", false, Monitor::_safepoint_check_never)) {
193 }
194
195 ~MutexGangTaskDispatcher() {
196 delete _monitor;
197 }
198
199 void coordinator_execute_on_workers(AbstractGangTask* task, uint num_workers) {
200 MonitorLocker ml(_monitor, Mutex::_no_safepoint_check_flag);
201
202 _task = task;
203 _num_workers = num_workers;
204
205 // Tell the workers to get to work.
206 _monitor->notify_all();
207
208 // Wait for them to finish.
209 while (_finished < _num_workers) {
210 ml.wait();
211 }
212
213 _task = NULL;
214 _num_workers = 0;
215 _started = 0;
216 _finished = 0;
217 }
218
219 WorkData worker_wait_for_task() {
220 MonitorLocker ml(_monitor, Mutex::_no_safepoint_check_flag);
221
222 while (_num_workers == 0 || _started == _num_workers) {
223 _monitor->wait();
224 }
225
226 _started++;
227
228 // Subtract one to get a zero-indexed worker id.
229 uint worker_id = _started - 1;
230
231 return WorkData(_task, worker_id);
232 }
233
234 void worker_done_with_task() {
235 MonitorLocker ml(_monitor, Mutex::_no_safepoint_check_flag);
236
237 _finished++;
238
239 if (_finished == _num_workers) {
240 // This will wake up all workers and not only the coordinator.
241 _monitor->notify_all();
242 }
243 }
244 };
245
246 static GangTaskDispatcher* create_dispatcher() {
247 if (UseSemaphoreGCThreadsSynchronization) {
248 return new SemaphoreGangTaskDispatcher();
249 }
250
251 return new MutexGangTaskDispatcher();
252 }
253
254 WorkGang::WorkGang(const char* name,
255 uint workers,
256 bool are_GC_task_threads,
257 bool are_ConcurrentGC_threads) :
258 AbstractWorkGang(name, workers, are_GC_task_threads, are_ConcurrentGC_threads),
259 _dispatcher(create_dispatcher())
260 { }
261
262 WorkGang::~WorkGang() {
263 delete _dispatcher;
264 }
265
266 AbstractGangWorker* WorkGang::allocate_worker(uint worker_id) {
267 return new GangWorker(this, worker_id);
268 }
269
270 void WorkGang::run_task(AbstractGangTask* task) {
271 run_task(task, active_workers());
272 }
273
274 void WorkGang::run_task(AbstractGangTask* task, uint num_workers) {
275 guarantee(num_workers <= total_workers(),
276 "Trying to execute task %s with %u workers which is more than the amount of total workers %u.",
277 task->name(), num_workers, total_workers());
278 guarantee(num_workers > 0, "Trying to execute task %s with zero workers", task->name());
279 uint old_num_workers = _active_workers;
280 update_active_workers(num_workers);
281 _dispatcher->coordinator_execute_on_workers(task, num_workers);
282 update_active_workers(old_num_workers);
283 }
284
285 AbstractGangWorker::AbstractGangWorker(AbstractWorkGang* gang, uint id) {
286 _gang = gang;
287 set_id(id);
288 set_name("%s#%d", gang->name(), id);
289 }
290
291 void AbstractGangWorker::run() {
292 initialize();
293 loop();
294 }
295
296 void AbstractGangWorker::initialize() {
297 assert(_gang != NULL, "No gang to run in");
298 os::set_priority(this, NearMaxPriority);
299 log_develop_trace(gc, workgang)("Running gang worker for gang %s id %u", gang()->name(), id());
300 assert(!Thread::current()->is_VM_thread(), "VM thread should not be part"
301 " of a work gang");
302 }
303
304 bool AbstractGangWorker::is_GC_task_thread() const {
305 return gang()->are_GC_task_threads();
306 }
307
308 bool AbstractGangWorker::is_ConcurrentGC_thread() const {
309 return gang()->are_ConcurrentGC_threads();
310 }
311
312 void AbstractGangWorker::print_on(outputStream* st) const {
313 st->print("\"%s\" ", name());
314 Thread::print_on(st);
315 st->cr();
316 }
317
318 void AbstractGangWorker::print() const { print_on(tty); }
319
320 WorkData GangWorker::wait_for_task() {
321 return gang()->dispatcher()->worker_wait_for_task();
322 }
323
324 void GangWorker::signal_task_done() {
325 gang()->dispatcher()->worker_done_with_task();
326 }
327
328 void GangWorker::run_task(WorkData data) {
329 GCIdMark gc_id_mark(data._task->gc_id());
330 log_develop_trace(gc, workgang)("Running work gang: %s task: %s worker: %u", name(), data._task->name(), data._worker_id);
331
332 data._task->work(data._worker_id);
333
334 log_develop_trace(gc, workgang)("Finished work gang: %s task: %s worker: %u thread: " PTR_FORMAT,
335 name(), data._task->name(), data._worker_id, p2i(Thread::current()));
336 }
337
338 void GangWorker::loop() {
339 while (true) {
340 WorkData data = wait_for_task();
341
342 run_task(data);
343
344 signal_task_done();
345 }
346 }
347
348 // *** WorkGangBarrierSync
349
350 WorkGangBarrierSync::WorkGangBarrierSync()
351 : _monitor(Mutex::safepoint, "work gang barrier sync", true,
352 Monitor::_safepoint_check_never),
353 _n_workers(0), _n_completed(0), _should_reset(false), _aborted(false) {
354 }
355
356 WorkGangBarrierSync::WorkGangBarrierSync(uint n_workers, const char* name)
357 : _monitor(Mutex::safepoint, name, true, Monitor::_safepoint_check_never),
358 _n_workers(n_workers), _n_completed(0), _should_reset(false), _aborted(false) {
359 }
360
361 void WorkGangBarrierSync::set_n_workers(uint n_workers) {
362 _n_workers = n_workers;
363 _n_completed = 0;
364 _should_reset = false;
365 _aborted = false;
366 }
367
368 bool WorkGangBarrierSync::enter() {
369 MonitorLocker ml(monitor(), Mutex::_no_safepoint_check_flag);
370 if (should_reset()) {
371 // The should_reset() was set and we are the first worker to enter
372 // the sync barrier. We will zero the n_completed() count which
373 // effectively resets the barrier.
374 zero_completed();
375 set_should_reset(false);
376 }
377 inc_completed();
378 if (n_completed() == n_workers()) {
379 // At this point we would like to reset the barrier to be ready in
380 // case it is used again. However, we cannot set n_completed() to
381 // 0, even after the notify_all(), given that some other workers
382 // might still be waiting for n_completed() to become ==
383 // n_workers(). So, if we set n_completed() to 0, those workers
384 // will get stuck (as they will wake up, see that n_completed() !=
385 // n_workers() and go back to sleep). Instead, we raise the
386 // should_reset() flag and the barrier will be reset the first
387 // time a worker enters it again.
388 set_should_reset(true);
389 ml.notify_all();
390 } else {
391 while (n_completed() != n_workers() && !aborted()) {
392 ml.wait();
393 }
394 }
395 return !aborted();
396 }
397
398 void WorkGangBarrierSync::abort() {
399 MutexLocker x(monitor(), Mutex::_no_safepoint_check_flag);
400 set_aborted();
401 monitor()->notify_all();
402 }
403
404 // SubTasksDone functions.
405
406 SubTasksDone::SubTasksDone(uint n) :
407 _tasks(NULL), _n_tasks(n), _threads_completed(0) {
408 _tasks = NEW_C_HEAP_ARRAY(uint, n, mtInternal);
409 clear();
410 }
411
412 bool SubTasksDone::valid() {
413 return _tasks != NULL;
414 }
415
416 void SubTasksDone::clear() {
417 for (uint i = 0; i < _n_tasks; i++) {
418 _tasks[i] = 0;
419 }
420 _threads_completed = 0;
421 #ifdef ASSERT
422 _claimed = 0;
423 #endif
424 }
425
426 bool SubTasksDone::try_claim_task(uint t) {
427 assert(t < _n_tasks, "bad task id.");
428 uint old = _tasks[t];
429 if (old == 0) {
430 old = Atomic::cmpxchg(&_tasks[t], 0u, 1u);
431 }
432 bool res = old == 0;
433 #ifdef ASSERT
434 if (res) {
435 assert(_claimed < _n_tasks, "Too many tasks claimed; missing clear?");
436 Atomic::inc(&_claimed);
437 }
438 #endif
439 return res;
440 }
441
442 void SubTasksDone::all_tasks_completed(uint n_threads) {
443 uint observed = _threads_completed;
444 uint old;
445 do {
446 old = observed;
447 observed = Atomic::cmpxchg(&_threads_completed, old, old+1);
448 } while (observed != old);
449 // If this was the last thread checking in, clear the tasks.
450 uint adjusted_thread_count = (n_threads == 0 ? 1 : n_threads);
451 if (observed + 1 == adjusted_thread_count) {
452 clear();
453 }
454 }
455
456
457 SubTasksDone::~SubTasksDone() {
458 FREE_C_HEAP_ARRAY(uint, _tasks);
459 }
460
461 // *** SequentialSubTasksDone
462
463 void SequentialSubTasksDone::clear() {
464 _n_tasks = _n_claimed = 0;
465 _n_threads = _n_completed = 0;
466 }
467
468 bool SequentialSubTasksDone::valid() {
469 return _n_threads > 0;
470 }
471
472 bool SequentialSubTasksDone::try_claim_task(uint& t) {
473 t = _n_claimed;
474 while (t < _n_tasks) {
475 uint res = Atomic::cmpxchg(&_n_claimed, t, t+1);
476 if (res == t) {
477 return true;
478 }
479 t = res;
480 }
481 return false;
482 }
483
484 bool SequentialSubTasksDone::all_tasks_completed() {
485 uint complete = _n_completed;
486 while (true) {
487 uint res = Atomic::cmpxchg(&_n_completed, complete, complete+1);
488 if (res == complete) {
489 break;
490 }
491 complete = res;
492 }
493 if (complete+1 == _n_threads) {
494 clear();
495 return true;
496 }
497 return false;
498 }