1 /*
2 * Copyright (c) 2001, 2016, 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 "runtime/atomic.hpp"
32 #include "runtime/os.hpp"
33 #include "runtime/semaphore.hpp"
34 #include "runtime/thread.inline.hpp"
35
36 // Definitions of WorkGang methods.
37
38 // The current implementation will exit if the allocation
39 // of any worker fails.
40 void AbstractWorkGang::initialize_workers() {
41 log_develop_trace(gc, workgang)("Constructing work gang %s with %u threads", name(), total_workers());
42 _workers = NEW_C_HEAP_ARRAY(AbstractGangWorker*, total_workers(), mtInternal);
43 if (_workers == NULL) {
44 vm_exit_out_of_memory(0, OOM_MALLOC_ERROR, "Cannot create GangWorker array.");
45 }
46
47 add_workers(true);
48 }
49
50
51 AbstractGangWorker* AbstractWorkGang::install_worker(uint worker_id) {
52 AbstractGangWorker* new_worker = allocate_worker(worker_id);
53 set_thread(worker_id, new_worker);
54 return new_worker;
55 }
56
57 void AbstractWorkGang::add_workers(bool initializing) {
58 add_workers(_active_workers, initializing);
59 }
60
61 void AbstractWorkGang::add_workers(uint active_workers, bool initializing) {
62
63 os::ThreadType worker_type;
64 if (are_ConcurrentGC_threads()) {
65 worker_type = os::cgc_thread;
66 } else {
67 worker_type = os::pgc_thread;
68 }
69 uint previous_created_workers = _created_workers;
70
71 _created_workers = WorkerManager::add_workers(this,
72 active_workers,
73 _total_workers,
74 _created_workers,
75 worker_type,
76 initializing);
77 _active_workers = MIN2(_created_workers, _active_workers);
78
79 WorkerManager::log_worker_creation(this, previous_created_workers, _active_workers, _created_workers, initializing);
80 }
81
82 AbstractGangWorker* AbstractWorkGang::worker(uint i) const {
83 // Array index bounds checking.
84 AbstractGangWorker* result = NULL;
85 assert(_workers != NULL, "No workers for indexing");
86 assert(i < total_workers(), "Worker index out of bounds");
87 result = _workers[i];
88 assert(result != NULL, "Indexing to null worker");
89 return result;
90 }
91
92 void AbstractWorkGang::print_worker_threads_on(outputStream* st) const {
93 uint workers = created_workers();
94 for (uint i = 0; i < workers; i++) {
95 worker(i)->print_on(st);
96 st->cr();
97 }
98 }
99
100 void AbstractWorkGang::threads_do(ThreadClosure* tc) const {
101 assert(tc != NULL, "Null ThreadClosure");
102 uint workers = created_workers();
103 for (uint i = 0; i < workers; i++) {
104 tc->do_thread(worker(i));
105 }
106 }
107
108 // WorkGang dispatcher implemented with semaphores.
109 //
110 // Semaphores don't require the worker threads to re-claim the lock when they wake up.
111 // This helps lowering the latency when starting and stopping the worker threads.
112 class SemaphoreGangTaskDispatcher : public GangTaskDispatcher {
113 // The task currently being dispatched to the GangWorkers.
114 AbstractGangTask* _task;
115
116 volatile uint _started;
117 volatile uint _not_finished;
118
119 // Semaphore used to start the GangWorkers.
120 Semaphore* _start_semaphore;
121 // Semaphore used to notify the coordinator that all workers are done.
122 Semaphore* _end_semaphore;
123
124 public:
125 SemaphoreGangTaskDispatcher() :
126 _task(NULL),
127 _started(0),
128 _not_finished(0),
129 _start_semaphore(new Semaphore()),
130 _end_semaphore(new Semaphore())
131 { }
132
133 ~SemaphoreGangTaskDispatcher() {
134 delete _start_semaphore;
135 delete _end_semaphore;
136 }
137
138 void coordinator_execute_on_workers(AbstractGangTask* task, uint num_workers) {
139 // No workers are allowed to read the state variables until they have been signaled.
140 _task = task;
141 _not_finished = num_workers;
142
143 // Dispatch 'num_workers' number of tasks.
144 _start_semaphore->signal(num_workers);
145
146 // Wait for the last worker to signal the coordinator.
147 _end_semaphore->wait();
148
149 // No workers are allowed to read the state variables after the coordinator has been signaled.
150 assert(_not_finished == 0, "%d not finished workers?", _not_finished);
151 _task = NULL;
152 _started = 0;
153
154 }
155
156 WorkData worker_wait_for_task() {
157 // Wait for the coordinator to dispatch a task.
158 _start_semaphore->wait();
159
160 uint num_started = (uint) Atomic::add(1, (volatile jint*)&_started);
161
162 // Subtract one to get a zero-indexed worker id.
163 uint worker_id = num_started - 1;
164
165 return WorkData(_task, worker_id);
166 }
167
168 void worker_done_with_task() {
169 // Mark that the worker is done with the task.
170 // The worker is not allowed to read the state variables after this line.
171 uint not_finished = (uint) Atomic::add(-1, (volatile jint*)&_not_finished);
172
173 // The last worker signals to the coordinator that all work is completed.
174 if (not_finished == 0) {
175 _end_semaphore->signal();
176 }
177 }
178 };
179
180 class MutexGangTaskDispatcher : public GangTaskDispatcher {
181 AbstractGangTask* _task;
182
183 volatile uint _started;
184 volatile uint _finished;
185 volatile uint _num_workers;
186
187 Monitor* _monitor;
188
189 public:
190 MutexGangTaskDispatcher()
191 : _task(NULL),
192 _monitor(new Monitor(Monitor::leaf, "WorkGang dispatcher lock", false, Monitor::_safepoint_check_never)),
193 _started(0),
194 _finished(0),
195 _num_workers(0) {}
196
197 ~MutexGangTaskDispatcher() {
198 delete _monitor;
199 }
200
201 void coordinator_execute_on_workers(AbstractGangTask* task, uint num_workers) {
202 MutexLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag);
203
204 _task = task;
205 _num_workers = num_workers;
206
207 // Tell the workers to get to work.
208 _monitor->notify_all();
209
210 // Wait for them to finish.
211 while (_finished < _num_workers) {
212 _monitor->wait(/* no_safepoint_check */ true);
213 }
214
215 _task = NULL;
216 _num_workers = 0;
217 _started = 0;
218 _finished = 0;
219 }
220
221 WorkData worker_wait_for_task() {
222 MonitorLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag);
223
224 while (_num_workers == 0 || _started == _num_workers) {
225 _monitor->wait(/* no_safepoint_check */ true);
226 }
227
228 _started++;
229
230 // Subtract one to get a zero-indexed worker id.
231 uint worker_id = _started - 1;
232
233 return WorkData(_task, worker_id);
234 }
235
236 void worker_done_with_task() {
237 MonitorLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag);
238
239 _finished++;
240
241 if (_finished == _num_workers) {
242 // This will wake up all workers and not only the coordinator.
243 _monitor->notify_all();
244 }
245 }
246 };
247
248 static GangTaskDispatcher* create_dispatcher() {
249 if (UseSemaphoreGCThreadsSynchronization) {
250 return new SemaphoreGangTaskDispatcher();
251 }
252
253 return new MutexGangTaskDispatcher();
254 }
255
256 WorkGang::WorkGang(const char* name,
257 uint workers,
258 bool are_GC_task_threads,
259 bool are_ConcurrentGC_threads) :
260 AbstractWorkGang(name, workers, are_GC_task_threads, are_ConcurrentGC_threads),
261 _dispatcher(create_dispatcher())
262 { }
263
264
265 WorkGang::WorkGang(const char* name,
266 uint workers,
267 bool are_GC_task_threads,
268 bool are_ConcurrentGC_threads,
269 GangTaskDispatcher* dispatcher) :
270 AbstractWorkGang(name, workers, are_GC_task_threads, are_ConcurrentGC_threads),
271 _dispatcher(dispatcher)
272 { }
273
274 AbstractGangWorker* WorkGang::allocate_worker(uint worker_id) {
275 return new GangWorker(this, worker_id);
276 }
277
278 void WorkGang::run_task(AbstractGangTask* task) {
279 run_task(task, active_workers());
280 }
281
282 void WorkGang::run_task(AbstractGangTask* task, uint num_workers) {
283 guarantee(num_workers <= total_workers(),
284 "Trying to execute task %s with %u workers which is more than the amount of total workers %u.",
285 task->name(), num_workers, total_workers());
286 guarantee(num_workers > 0, "Trying to execute task %s with zero workers", task->name());
287 uint old_num_workers = _active_workers;
288 update_active_workers(num_workers);
289 _dispatcher->coordinator_execute_on_workers(task, num_workers);
290 update_active_workers(old_num_workers);
291 }
292
293 AbstractGangWorker::AbstractGangWorker(AbstractWorkGang* gang, uint id) {
294 _gang = gang;
295 set_id(id);
296 set_name("%s#%d", gang->name(), id);
297 }
298
299 void AbstractGangWorker::run() {
300 initialize();
301 loop();
302 }
303
304 void AbstractGangWorker::initialize() {
305 this->record_stack_base_and_size();
306 this->initialize_named_thread();
307 assert(_gang != NULL, "No gang to run in");
308 os::set_priority(this, NearMaxPriority);
309 log_develop_trace(gc, workgang)("Running gang worker for gang %s id %u", gang()->name(), id());
310 // The VM thread should not execute here because MutexLocker's are used
311 // as (opposed to MutexLockerEx's).
312 assert(!Thread::current()->is_VM_thread(), "VM thread should not be part"
313 " of a work gang");
314 }
315
316 bool AbstractGangWorker::is_GC_task_thread() const {
317 return gang()->are_GC_task_threads();
318 }
319
320 bool AbstractGangWorker::is_ConcurrentGC_thread() const {
321 return gang()->are_ConcurrentGC_threads();
322 }
323
324 void AbstractGangWorker::print_on(outputStream* st) const {
325 st->print("\"%s\" ", name());
326 Thread::print_on(st);
327 st->cr();
328 }
329
330 WorkData GangWorker::wait_for_task() {
331 return gang()->dispatcher()->worker_wait_for_task();
332 }
333
334 void GangWorker::signal_task_done() {
335 gang()->dispatcher()->worker_done_with_task();
336 }
337
338 void GangWorker::run_task(WorkData data) {
339 GCIdMark gc_id_mark(data._task->gc_id());
340 log_develop_trace(gc, workgang)("Running work gang: %s task: %s worker: %u", name(), data._task->name(), data._worker_id);
341
342 data._task->work(data._worker_id);
343
344 log_develop_trace(gc, workgang)("Finished work gang: %s task: %s worker: %u thread: " PTR_FORMAT,
345 name(), data._task->name(), data._worker_id, p2i(Thread::current()));
346 }
347
348 void GangWorker::loop() {
349 while (true) {
350 WorkData data = wait_for_task();
351
352 run_task(data);
353
354 signal_task_done();
355 }
356 }
357
358 // *** WorkGangBarrierSync
359
360 WorkGangBarrierSync::WorkGangBarrierSync()
361 : _monitor(Mutex::safepoint, "work gang barrier sync", true,
362 Monitor::_safepoint_check_never),
363 _n_workers(0), _n_completed(0), _should_reset(false), _aborted(false) {
364 }
365
366 WorkGangBarrierSync::WorkGangBarrierSync(uint n_workers, const char* name)
367 : _monitor(Mutex::safepoint, name, true, Monitor::_safepoint_check_never),
368 _n_workers(n_workers), _n_completed(0), _should_reset(false), _aborted(false) {
369 }
370
371 void WorkGangBarrierSync::set_n_workers(uint n_workers) {
372 _n_workers = n_workers;
373 _n_completed = 0;
374 _should_reset = false;
375 _aborted = false;
376 }
377
378 bool WorkGangBarrierSync::enter() {
379 MutexLockerEx x(monitor(), Mutex::_no_safepoint_check_flag);
380 if (should_reset()) {
381 // The should_reset() was set and we are the first worker to enter
382 // the sync barrier. We will zero the n_completed() count which
383 // effectively resets the barrier.
384 zero_completed();
385 set_should_reset(false);
386 }
387 inc_completed();
388 if (n_completed() == n_workers()) {
389 // At this point we would like to reset the barrier to be ready in
390 // case it is used again. However, we cannot set n_completed() to
391 // 0, even after the notify_all(), given that some other workers
392 // might still be waiting for n_completed() to become ==
393 // n_workers(). So, if we set n_completed() to 0, those workers
394 // will get stuck (as they will wake up, see that n_completed() !=
395 // n_workers() and go back to sleep). Instead, we raise the
396 // should_reset() flag and the barrier will be reset the first
397 // time a worker enters it again.
398 set_should_reset(true);
399 monitor()->notify_all();
400 } else {
401 while (n_completed() != n_workers() && !aborted()) {
402 monitor()->wait(/* no_safepoint_check */ true);
403 }
404 }
405 return !aborted();
406 }
407
408 void WorkGangBarrierSync::abort() {
409 MutexLockerEx x(monitor(), Mutex::_no_safepoint_check_flag);
410 set_aborted();
411 monitor()->notify_all();
412 }
413
414 // SubTasksDone functions.
415
416 SubTasksDone::SubTasksDone(uint n) :
417 _n_tasks(n), _tasks(NULL) {
418 _tasks = NEW_C_HEAP_ARRAY(uint, n, mtInternal);
419 guarantee(_tasks != NULL, "alloc failure");
420 clear();
421 }
422
423 bool SubTasksDone::valid() {
424 return _tasks != NULL;
425 }
426
427 void SubTasksDone::clear() {
428 for (uint i = 0; i < _n_tasks; i++) {
429 _tasks[i] = 0;
430 }
431 _threads_completed = 0;
432 #ifdef ASSERT
433 _claimed = 0;
434 #endif
435 }
436
437 bool SubTasksDone::is_task_claimed(uint t) {
438 assert(t < _n_tasks, "bad task id.");
439 uint old = _tasks[t];
440 if (old == 0) {
441 old = Atomic::cmpxchg(1, &_tasks[t], 0);
442 }
443 assert(_tasks[t] == 1, "What else?");
444 bool res = old != 0;
445 #ifdef ASSERT
446 if (!res) {
447 assert(_claimed < _n_tasks, "Too many tasks claimed; missing clear?");
448 Atomic::inc((volatile jint*) &_claimed);
449 }
450 #endif
451 return res;
452 }
453
454 void SubTasksDone::all_tasks_completed(uint n_threads) {
455 jint observed = _threads_completed;
456 jint old;
457 do {
458 old = observed;
459 observed = Atomic::cmpxchg(old+1, &_threads_completed, old);
460 } while (observed != old);
461 // If this was the last thread checking in, clear the tasks.
462 uint adjusted_thread_count = (n_threads == 0 ? 1 : n_threads);
463 if (observed + 1 == (jint)adjusted_thread_count) {
464 clear();
465 }
466 }
467
468
469 SubTasksDone::~SubTasksDone() {
470 if (_tasks != NULL) FREE_C_HEAP_ARRAY(jint, _tasks);
471 }
472
473 // *** SequentialSubTasksDone
474
475 void SequentialSubTasksDone::clear() {
476 _n_tasks = _n_claimed = 0;
477 _n_threads = _n_completed = 0;
478 }
479
480 bool SequentialSubTasksDone::valid() {
481 return _n_threads > 0;
482 }
483
484 bool SequentialSubTasksDone::is_task_claimed(uint& t) {
485 t = _n_claimed;
486 while (t < _n_tasks) {
487 jint res = Atomic::cmpxchg(t+1, &_n_claimed, t);
488 if (res == (jint)t) {
489 return false;
490 }
491 t = res;
492 }
493 return true;
494 }
495
496 bool SequentialSubTasksDone::all_tasks_completed() {
497 uint complete = _n_completed;
498 while (true) {
499 uint res = Atomic::cmpxchg(complete+1, &_n_completed, complete);
500 if (res == complete) {
501 break;
502 }
503 complete = res;
504 }
505 if (complete+1 == _n_threads) {
506 clear();
507 return true;
508 }
509 return false;
510 }