1 /*
2 * Copyright (c) 2001, 2018, 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.
initialize_workers()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
install_worker(uint worker_id)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
add_workers(bool initializing)57 void AbstractWorkGang::add_workers(bool initializing) {
58 add_workers(_active_workers, initializing);
59 }
60
add_workers(uint active_workers,bool initializing)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
worker(uint i) const82 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
print_worker_threads_on(outputStream * st) const92 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
threads_do(ThreadClosure * tc) const100 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:
SemaphoreGangTaskDispatcher()125 SemaphoreGangTaskDispatcher() :
126 _task(NULL),
127 _started(0),
128 _not_finished(0),
129 _start_semaphore(new Semaphore()),
130 _end_semaphore(new Semaphore())
131 { }
132
~SemaphoreGangTaskDispatcher()133 ~SemaphoreGangTaskDispatcher() {
134 delete _start_semaphore;
135 delete _end_semaphore;
136 }
137
coordinator_execute_on_workers(AbstractGangTask * task,uint num_workers)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
worker_wait_for_task()156 WorkData worker_wait_for_task() {
157 // Wait for the coordinator to dispatch a task.
158 _start_semaphore->wait();
159
160 uint num_started = Atomic::add(1u, &_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
worker_done_with_task()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 = Atomic::sub(1u, &_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:
MutexGangTaskDispatcher()190 MutexGangTaskDispatcher() :
191 _task(NULL),
192 _started(0),
193 _finished(0),
194 _num_workers(0),
195 _monitor(new Monitor(Monitor::leaf, "WorkGang dispatcher lock", false, Monitor::_safepoint_check_never)) {
196 }
197
~MutexGangTaskDispatcher()198 ~MutexGangTaskDispatcher() {
199 delete _monitor;
200 }
201
coordinator_execute_on_workers(AbstractGangTask * task,uint num_workers)202 void coordinator_execute_on_workers(AbstractGangTask* task, uint num_workers) {
203 MutexLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag);
204
205 _task = task;
206 _num_workers = num_workers;
207
208 // Tell the workers to get to work.
209 _monitor->notify_all();
210
211 // Wait for them to finish.
212 while (_finished < _num_workers) {
213 _monitor->wait(/* no_safepoint_check */ true);
214 }
215
216 _task = NULL;
217 _num_workers = 0;
218 _started = 0;
219 _finished = 0;
220 }
221
worker_wait_for_task()222 WorkData worker_wait_for_task() {
223 MonitorLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag);
224
225 while (_num_workers == 0 || _started == _num_workers) {
226 _monitor->wait(/* no_safepoint_check */ true);
227 }
228
229 _started++;
230
231 // Subtract one to get a zero-indexed worker id.
232 uint worker_id = _started - 1;
233
234 return WorkData(_task, worker_id);
235 }
236
worker_done_with_task()237 void worker_done_with_task() {
238 MonitorLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag);
239
240 _finished++;
241
242 if (_finished == _num_workers) {
243 // This will wake up all workers and not only the coordinator.
244 _monitor->notify_all();
245 }
246 }
247 };
248
create_dispatcher()249 static GangTaskDispatcher* create_dispatcher() {
250 if (UseSemaphoreGCThreadsSynchronization) {
251 return new SemaphoreGangTaskDispatcher();
252 }
253
254 return new MutexGangTaskDispatcher();
255 }
256
WorkGang(const char * name,uint workers,bool are_GC_task_threads,bool are_ConcurrentGC_threads)257 WorkGang::WorkGang(const char* name,
258 uint workers,
259 bool are_GC_task_threads,
260 bool are_ConcurrentGC_threads) :
261 AbstractWorkGang(name, workers, are_GC_task_threads, are_ConcurrentGC_threads),
262 _dispatcher(create_dispatcher())
263 { }
264
~WorkGang()265 WorkGang::~WorkGang() {
266 delete _dispatcher;
267 }
268
allocate_worker(uint worker_id)269 AbstractGangWorker* WorkGang::allocate_worker(uint worker_id) {
270 return new GangWorker(this, worker_id);
271 }
272
run_task(AbstractGangTask * task)273 void WorkGang::run_task(AbstractGangTask* task) {
274 run_task(task, active_workers());
275 }
276
run_task(AbstractGangTask * task,uint num_workers)277 void WorkGang::run_task(AbstractGangTask* task, uint num_workers) {
278 guarantee(num_workers <= total_workers(),
279 "Trying to execute task %s with %u workers which is more than the amount of total workers %u.",
280 task->name(), num_workers, total_workers());
281 guarantee(num_workers > 0, "Trying to execute task %s with zero workers", task->name());
282 uint old_num_workers = _active_workers;
283 update_active_workers(num_workers);
284 _dispatcher->coordinator_execute_on_workers(task, num_workers);
285 update_active_workers(old_num_workers);
286 }
287
AbstractGangWorker(AbstractWorkGang * gang,uint id)288 AbstractGangWorker::AbstractGangWorker(AbstractWorkGang* gang, uint id) {
289 _gang = gang;
290 set_id(id);
291 set_name("%s#%d", gang->name(), id);
292 }
293
run()294 void AbstractGangWorker::run() {
295 initialize();
296 loop();
297 }
298
initialize()299 void AbstractGangWorker::initialize() {
300 this->initialize_named_thread();
301 assert(_gang != NULL, "No gang to run in");
302 os::set_priority(this, NearMaxPriority);
303 log_develop_trace(gc, workgang)("Running gang worker for gang %s id %u", gang()->name(), id());
304 // The VM thread should not execute here because MutexLocker's are used
305 // as (opposed to MutexLockerEx's).
306 assert(!Thread::current()->is_VM_thread(), "VM thread should not be part"
307 " of a work gang");
308 }
309
is_GC_task_thread() const310 bool AbstractGangWorker::is_GC_task_thread() const {
311 return gang()->are_GC_task_threads();
312 }
313
is_ConcurrentGC_thread() const314 bool AbstractGangWorker::is_ConcurrentGC_thread() const {
315 return gang()->are_ConcurrentGC_threads();
316 }
317
print_on(outputStream * st) const318 void AbstractGangWorker::print_on(outputStream* st) const {
319 st->print("\"%s\" ", name());
320 Thread::print_on(st);
321 st->cr();
322 }
323
wait_for_task()324 WorkData GangWorker::wait_for_task() {
325 return gang()->dispatcher()->worker_wait_for_task();
326 }
327
signal_task_done()328 void GangWorker::signal_task_done() {
329 gang()->dispatcher()->worker_done_with_task();
330 }
331
run_task(WorkData data)332 void GangWorker::run_task(WorkData data) {
333 GCIdMark gc_id_mark(data._task->gc_id());
334 log_develop_trace(gc, workgang)("Running work gang: %s task: %s worker: %u", name(), data._task->name(), data._worker_id);
335
336 data._task->work(data._worker_id);
337
338 log_develop_trace(gc, workgang)("Finished work gang: %s task: %s worker: %u thread: " PTR_FORMAT,
339 name(), data._task->name(), data._worker_id, p2i(Thread::current()));
340 }
341
loop()342 void GangWorker::loop() {
343 while (true) {
344 WorkData data = wait_for_task();
345
346 run_task(data);
347
348 signal_task_done();
349 }
350 }
351
352 // *** WorkGangBarrierSync
353
WorkGangBarrierSync()354 WorkGangBarrierSync::WorkGangBarrierSync()
355 : _monitor(Mutex::safepoint, "work gang barrier sync", true,
356 Monitor::_safepoint_check_never),
357 _n_workers(0), _n_completed(0), _should_reset(false), _aborted(false) {
358 }
359
WorkGangBarrierSync(uint n_workers,const char * name)360 WorkGangBarrierSync::WorkGangBarrierSync(uint n_workers, const char* name)
361 : _monitor(Mutex::safepoint, name, true, Monitor::_safepoint_check_never),
362 _n_workers(n_workers), _n_completed(0), _should_reset(false), _aborted(false) {
363 }
364
set_n_workers(uint n_workers)365 void WorkGangBarrierSync::set_n_workers(uint n_workers) {
366 _n_workers = n_workers;
367 _n_completed = 0;
368 _should_reset = false;
369 _aborted = false;
370 }
371
enter()372 bool WorkGangBarrierSync::enter() {
373 MutexLockerEx x(monitor(), Mutex::_no_safepoint_check_flag);
374 if (should_reset()) {
375 // The should_reset() was set and we are the first worker to enter
376 // the sync barrier. We will zero the n_completed() count which
377 // effectively resets the barrier.
378 zero_completed();
379 set_should_reset(false);
380 }
381 inc_completed();
382 if (n_completed() == n_workers()) {
383 // At this point we would like to reset the barrier to be ready in
384 // case it is used again. However, we cannot set n_completed() to
385 // 0, even after the notify_all(), given that some other workers
386 // might still be waiting for n_completed() to become ==
387 // n_workers(). So, if we set n_completed() to 0, those workers
388 // will get stuck (as they will wake up, see that n_completed() !=
389 // n_workers() and go back to sleep). Instead, we raise the
390 // should_reset() flag and the barrier will be reset the first
391 // time a worker enters it again.
392 set_should_reset(true);
393 monitor()->notify_all();
394 } else {
395 while (n_completed() != n_workers() && !aborted()) {
396 monitor()->wait(/* no_safepoint_check */ true);
397 }
398 }
399 return !aborted();
400 }
401
abort()402 void WorkGangBarrierSync::abort() {
403 MutexLockerEx x(monitor(), Mutex::_no_safepoint_check_flag);
404 set_aborted();
405 monitor()->notify_all();
406 }
407
408 // SubTasksDone functions.
409
SubTasksDone(uint n)410 SubTasksDone::SubTasksDone(uint n) :
411 _tasks(NULL), _n_tasks(n), _threads_completed(0) {
412 _tasks = NEW_C_HEAP_ARRAY(uint, n, mtInternal);
413 guarantee(_tasks != NULL, "alloc failure");
414 clear();
415 }
416
valid()417 bool SubTasksDone::valid() {
418 return _tasks != NULL;
419 }
420
clear()421 void SubTasksDone::clear() {
422 for (uint i = 0; i < _n_tasks; i++) {
423 _tasks[i] = 0;
424 }
425 _threads_completed = 0;
426 #ifdef ASSERT
427 _claimed = 0;
428 #endif
429 }
430
try_claim_task(uint t)431 bool SubTasksDone::try_claim_task(uint t) {
432 assert(t < _n_tasks, "bad task id.");
433 uint old = _tasks[t];
434 if (old == 0) {
435 old = Atomic::cmpxchg(1u, &_tasks[t], 0u);
436 }
437 assert(_tasks[t] == 1, "What else?");
438 bool res = old == 0;
439 #ifdef ASSERT
440 if (res) {
441 assert(_claimed < _n_tasks, "Too many tasks claimed; missing clear?");
442 Atomic::inc(&_claimed);
443 }
444 #endif
445 return res;
446 }
447
all_tasks_completed(uint n_threads)448 void SubTasksDone::all_tasks_completed(uint n_threads) {
449 uint observed = _threads_completed;
450 uint old;
451 do {
452 old = observed;
453 observed = Atomic::cmpxchg(old+1, &_threads_completed, old);
454 } while (observed != old);
455 // If this was the last thread checking in, clear the tasks.
456 uint adjusted_thread_count = (n_threads == 0 ? 1 : n_threads);
457 if (observed + 1 == adjusted_thread_count) {
458 clear();
459 }
460 }
461
462
~SubTasksDone()463 SubTasksDone::~SubTasksDone() {
464 if (_tasks != NULL) FREE_C_HEAP_ARRAY(uint, _tasks);
465 }
466
467 // *** SequentialSubTasksDone
468
clear()469 void SequentialSubTasksDone::clear() {
470 _n_tasks = _n_claimed = 0;
471 _n_threads = _n_completed = 0;
472 }
473
valid()474 bool SequentialSubTasksDone::valid() {
475 return _n_threads > 0;
476 }
477
try_claim_task(uint & t)478 bool SequentialSubTasksDone::try_claim_task(uint& t) {
479 t = _n_claimed;
480 while (t < _n_tasks) {
481 uint res = Atomic::cmpxchg(t+1, &_n_claimed, t);
482 if (res == t) {
483 return true;
484 }
485 t = res;
486 }
487 return false;
488 }
489
all_tasks_completed()490 bool SequentialSubTasksDone::all_tasks_completed() {
491 uint complete = _n_completed;
492 while (true) {
493 uint res = Atomic::cmpxchg(complete+1, &_n_completed, complete);
494 if (res == complete) {
495 break;
496 }
497 complete = res;
498 }
499 if (complete+1 == _n_threads) {
500 clear();
501 return true;
502 }
503 return false;
504 }
505