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