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/adaptiveSizePolicy.inline.hpp"
29 #include "memory/allocation.hpp"
30 #include "memory/allocation.inline.hpp"
31 #include "runtime/atomic.inline.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. Still, return a boolean so that
40 // a future implementation can possibly do a partial
41 // initialization of the workers and report such to the
42 // caller.
43 bool AbstractWorkGang::initialize_workers() {
44 log_develop_trace(gc, workgang)("Constructing work gang %s with %u threads", name(), total_workers());
45 _workers = NEW_C_HEAP_ARRAY(AbstractGangWorker*, total_workers(), mtInternal);
46 if (_workers == NULL) {
47 vm_exit_out_of_memory(0, OOM_MALLOC_ERROR, "Cannot create GangWorker array.");
48 return false;
49 }
50
51 _active_workers = AdaptiveSizePolicy::initial_number_of_workers();
52 return add_workers(true);
53 }
54
55
56 AbstractGangWorker* AbstractWorkGang::install_worker(uint worker_id) {
57 AbstractGangWorker* new_worker = allocate_worker(worker_id);
58 set_thread(worker_id, new_worker);
59 return new_worker;
60 }
61
62 bool AbstractWorkGang::add_workers(bool initializing) {
63
64 os::ThreadType worker_type;
65 if (are_ConcurrentGC_threads()) {
66 worker_type = os::cgc_thread;
67 } else {
68 worker_type = os::pgc_thread;
69 }
70
71 return AdaptiveSizePolicy::add_workers(this,
72 _active_workers,
73 _total_workers,
74 _created_workers,
75 worker_type,
76 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 = (uint) Atomic::add(1, (volatile jint*)&_started);
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 = (uint) Atomic::add(-1, (volatile jint*)&_not_finished);
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 _monitor(new Monitor(Monitor::leaf, "WorkGang dispatcher lock", false, Monitor::_safepoint_check_never)),
190 _started(0),
191 _finished(0),
192 _num_workers(0) {}
193
194 ~MutexGangTaskDispatcher() {
195 delete _monitor;
196 }
197
198 void coordinator_execute_on_workers(AbstractGangTask* task, uint num_workers) {
199 MutexLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag);
200
201 _task = task;
202 _num_workers = num_workers;
203
204 // Tell the workers to get to work.
205 _monitor->notify_all();
206
207 // Wait for them to finish.
208 while (_finished < _num_workers) {
209 _monitor->wait(/* no_safepoint_check */ true);
210 }
211
212 _task = NULL;
213 _num_workers = 0;
214 _started = 0;
215 _finished = 0;
216 }
217
218 WorkData worker_wait_for_task() {
219 MonitorLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag);
220
221 while (_num_workers == 0 || _started == _num_workers) {
222 _monitor->wait(/* no_safepoint_check */ true);
223 }
224
225 _started++;
226
227 // Subtract one to get a zero-indexed worker id.
228 uint worker_id = _started - 1;
229
230 return WorkData(_task, worker_id);
231 }
232
233 void worker_done_with_task() {
234 MonitorLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag);
235
236 _finished++;
237
238 if (_finished == _num_workers) {
239 // This will wake up all workers and not only the coordinator.
240 _monitor->notify_all();
241 }
242 }
243 };
244
245 static GangTaskDispatcher* create_dispatcher() {
246 if (UseSemaphoreGCThreadsSynchronization) {
247 return new SemaphoreGangTaskDispatcher();
248 }
249
250 return new MutexGangTaskDispatcher();
251 }
252
253 WorkGang::WorkGang(const char* name,
254 uint workers,
255 bool are_GC_task_threads,
256 bool are_ConcurrentGC_threads) :
257 AbstractWorkGang(name, workers, are_GC_task_threads, are_ConcurrentGC_threads),
258 _dispatcher(create_dispatcher())
259 { }
260
261 AbstractGangWorker* WorkGang::allocate_worker(uint worker_id) {
262 return new GangWorker(this, worker_id);
263 }
264
265 void WorkGang::run_task(AbstractGangTask* task) {
266 _dispatcher->coordinator_execute_on_workers(task, active_workers());
267 }
268
269 AbstractGangWorker::AbstractGangWorker(AbstractWorkGang* gang, uint id) {
270 _gang = gang;
271 set_id(id);
272 set_name("%s#%d", gang->name(), id);
273 }
274
275 void AbstractGangWorker::run() {
276 initialize();
277 loop();
278 }
279
280 void AbstractGangWorker::initialize() {
281 this->record_stack_base_and_size();
282 this->initialize_named_thread();
283 assert(_gang != NULL, "No gang to run in");
284 os::set_priority(this, NearMaxPriority);
285 log_develop_trace(gc, workgang)("Running gang worker for gang %s id %u", gang()->name(), id());
286 // The VM thread should not execute here because MutexLocker's are used
287 // as (opposed to MutexLockerEx's).
288 assert(!Thread::current()->is_VM_thread(), "VM thread should not be part"
289 " of a work gang");
290 }
291
292 bool AbstractGangWorker::is_GC_task_thread() const {
293 return gang()->are_GC_task_threads();
294 }
295
296 bool AbstractGangWorker::is_ConcurrentGC_thread() const {
297 return gang()->are_ConcurrentGC_threads();
298 }
299
300 void AbstractGangWorker::print_on(outputStream* st) const {
301 st->print("\"%s\" ", name());
302 Thread::print_on(st);
303 st->cr();
304 }
305
306 WorkData GangWorker::wait_for_task() {
307 return gang()->dispatcher()->worker_wait_for_task();
308 }
309
310 void GangWorker::signal_task_done() {
311 gang()->dispatcher()->worker_done_with_task();
312 }
313
314 void GangWorker::run_task(WorkData data) {
315 GCIdMark gc_id_mark(data._task->gc_id());
316 log_develop_trace(gc, workgang)("Running work gang: %s task: %s worker: %u", name(), data._task->name(), data._worker_id);
317
318 data._task->work(data._worker_id);
319
320 log_develop_trace(gc, workgang)("Finished work gang: %s task: %s worker: %u thread: " PTR_FORMAT,
321 name(), data._task->name(), data._worker_id, p2i(Thread::current()));
322 }
323
324 void GangWorker::loop() {
325 while (true) {
326 WorkData data = wait_for_task();
327
328 run_task(data);
329
330 signal_task_done();
331 }
332 }
333
334 // *** WorkGangBarrierSync
335
336 WorkGangBarrierSync::WorkGangBarrierSync()
337 : _monitor(Mutex::safepoint, "work gang barrier sync", true,
338 Monitor::_safepoint_check_never),
339 _n_workers(0), _n_completed(0), _should_reset(false), _aborted(false) {
340 }
341
342 WorkGangBarrierSync::WorkGangBarrierSync(uint n_workers, const char* name)
343 : _monitor(Mutex::safepoint, name, true, Monitor::_safepoint_check_never),
344 _n_workers(n_workers), _n_completed(0), _should_reset(false), _aborted(false) {
345 }
346
347 void WorkGangBarrierSync::set_n_workers(uint n_workers) {
348 _n_workers = n_workers;
349 _n_completed = 0;
350 _should_reset = false;
351 _aborted = false;
352 }
353
354 bool WorkGangBarrierSync::enter() {
355 MutexLockerEx x(monitor(), Mutex::_no_safepoint_check_flag);
356 if (should_reset()) {
357 // The should_reset() was set and we are the first worker to enter
358 // the sync barrier. We will zero the n_completed() count which
359 // effectively resets the barrier.
360 zero_completed();
361 set_should_reset(false);
362 }
363 inc_completed();
364 if (n_completed() == n_workers()) {
365 // At this point we would like to reset the barrier to be ready in
366 // case it is used again. However, we cannot set n_completed() to
367 // 0, even after the notify_all(), given that some other workers
368 // might still be waiting for n_completed() to become ==
369 // n_workers(). So, if we set n_completed() to 0, those workers
370 // will get stuck (as they will wake up, see that n_completed() !=
371 // n_workers() and go back to sleep). Instead, we raise the
372 // should_reset() flag and the barrier will be reset the first
373 // time a worker enters it again.
374 set_should_reset(true);
375 monitor()->notify_all();
376 } else {
377 while (n_completed() != n_workers() && !aborted()) {
378 monitor()->wait(/* no_safepoint_check */ true);
379 }
380 }
381 return !aborted();
382 }
383
384 void WorkGangBarrierSync::abort() {
385 MutexLockerEx x(monitor(), Mutex::_no_safepoint_check_flag);
386 set_aborted();
387 monitor()->notify_all();
388 }
389
390 // SubTasksDone functions.
391
392 SubTasksDone::SubTasksDone(uint n) :
393 _n_tasks(n), _tasks(NULL) {
394 _tasks = NEW_C_HEAP_ARRAY(uint, n, mtInternal);
395 guarantee(_tasks != NULL, "alloc failure");
396 clear();
397 }
398
399 bool SubTasksDone::valid() {
400 return _tasks != NULL;
401 }
402
403 void SubTasksDone::clear() {
404 for (uint i = 0; i < _n_tasks; i++) {
405 _tasks[i] = 0;
406 }
407 _threads_completed = 0;
408 #ifdef ASSERT
409 _claimed = 0;
410 #endif
411 }
412
413 bool SubTasksDone::is_task_claimed(uint t) {
414 assert(t < _n_tasks, "bad task id.");
415 uint old = _tasks[t];
416 if (old == 0) {
417 old = Atomic::cmpxchg(1, &_tasks[t], 0);
418 }
419 assert(_tasks[t] == 1, "What else?");
420 bool res = old != 0;
421 #ifdef ASSERT
422 if (!res) {
423 assert(_claimed < _n_tasks, "Too many tasks claimed; missing clear?");
424 Atomic::inc((volatile jint*) &_claimed);
425 }
426 #endif
427 return res;
428 }
429
430 void SubTasksDone::all_tasks_completed(uint n_threads) {
431 jint observed = _threads_completed;
432 jint old;
433 do {
434 old = observed;
435 observed = Atomic::cmpxchg(old+1, &_threads_completed, old);
436 } while (observed != old);
437 // If this was the last thread checking in, clear the tasks.
438 uint adjusted_thread_count = (n_threads == 0 ? 1 : n_threads);
439 if (observed + 1 == (jint)adjusted_thread_count) {
440 clear();
441 }
442 }
443
444
445 SubTasksDone::~SubTasksDone() {
446 if (_tasks != NULL) FREE_C_HEAP_ARRAY(jint, _tasks);
447 }
448
449 // *** SequentialSubTasksDone
450
451 void SequentialSubTasksDone::clear() {
452 _n_tasks = _n_claimed = 0;
453 _n_threads = _n_completed = 0;
454 }
455
456 bool SequentialSubTasksDone::valid() {
457 return _n_threads > 0;
458 }
459
460 bool SequentialSubTasksDone::is_task_claimed(uint& t) {
461 uint* n_claimed_ptr = &_n_claimed;
462 t = *n_claimed_ptr;
463 while (t < _n_tasks) {
464 jint res = Atomic::cmpxchg(t+1, n_claimed_ptr, t);
465 if (res == (jint)t) {
466 return false;
467 }
468 t = *n_claimed_ptr;
469 }
470 return true;
471 }
472
473 bool SequentialSubTasksDone::all_tasks_completed() {
474 uint* n_completed_ptr = &_n_completed;
475 uint complete = *n_completed_ptr;
476 while (true) {
477 uint res = Atomic::cmpxchg(complete+1, n_completed_ptr, complete);
478 if (res == complete) {
479 break;
480 }
481 complete = res;
482 }
483 if (complete+1 == _n_threads) {
484 clear();
485 return true;
486 }
487 return false;
488 }