1 /*
2 * Copyright (c) 2001, 2015, 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/workgroup.hpp"
27 #include "memory/allocation.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "runtime/atomic.inline.hpp"
30 #include "runtime/os.hpp"
31 #include "runtime/thread.inline.hpp"
32 #include "utilities/semaphore.hpp"
33
34 // Definitions of WorkGang methods.
35
36 // The current implementation will exit if the allocation
37 // of any worker fails. Still, return a boolean so that
38 // a future implementation can possibly do a partial
39 // initialization of the workers and report such to the
40 // caller.
41 bool AbstractWorkGang::initialize_workers() {
42
43 if (TraceWorkGang) {
44 tty->print_cr("Constructing work gang %s with %d threads",
45 name(),
46 total_workers());
47 }
48 _workers = NEW_C_HEAP_ARRAY(AbstractGangWorker*, total_workers(), mtInternal);
49 if (_workers == NULL) {
50 vm_exit_out_of_memory(0, OOM_MALLOC_ERROR, "Cannot create GangWorker array.");
51 return false;
52 }
53 os::ThreadType worker_type;
54 if (are_ConcurrentGC_threads()) {
55 worker_type = os::cgc_thread;
56 } else {
57 worker_type = os::pgc_thread;
58 }
59 for (uint worker = 0; worker < total_workers(); worker += 1) {
60 AbstractGangWorker* new_worker = allocate_worker(worker);
61 assert(new_worker != NULL, "Failed to allocate GangWorker");
62 _workers[worker] = new_worker;
63 if (new_worker == NULL || !os::create_thread(new_worker, worker_type)) {
64 vm_exit_out_of_memory(0, OOM_MALLOC_ERROR,
65 "Cannot create worker GC thread. Out of system resources.");
66 return false;
67 }
68 if (!DisableStartThread) {
69 os::start_thread(new_worker);
70 }
71 }
72 return true;
73 }
74
75 AbstractGangWorker* AbstractWorkGang::worker(uint i) const {
76 // Array index bounds checking.
77 AbstractGangWorker* result = NULL;
78 assert(_workers != NULL, "No workers for indexing");
79 assert(i < total_workers(), "Worker index out of bounds");
80 result = _workers[i];
81 assert(result != NULL, "Indexing to null worker");
82 return result;
83 }
84
85 void AbstractWorkGang::print_worker_threads_on(outputStream* st) const {
86 uint workers = total_workers();
87 for (uint i = 0; i < workers; i++) {
88 worker(i)->print_on(st);
89 st->cr();
90 }
91 }
92
93 void AbstractWorkGang::threads_do(ThreadClosure* tc) const {
94 assert(tc != NULL, "Null ThreadClosure");
95 uint workers = total_workers();
96 for (uint i = 0; i < workers; i++) {
97 tc->do_thread(worker(i));
98 }
99 }
100
101 #if IMPLEMENTS_SEMAPHORE_CLASS
102
103 // WorkGang dispatcher implemented with semaphores.
104 //
105 // Semaphores don't require the worker threads to re-claim the lock when they wake up.
106 // This helps lowering the latency when starting and stopping the worker threads.
107 class SemaphoreGangTaskDispatcher : public GangTaskDispatcher {
108 // The task currently being dispatched to the GangWorkers.
109 AbstractGangTask* _task;
110
111 volatile uint _started;
112 volatile uint _not_finished;
113
114 // Semaphore used to start the GangWorkers.
115 Semaphore* _start_semaphore;
116 // Semaphore used to notify the coordinator that all workers are done.
117 Semaphore* _end_semaphore;
118
119 public:
120 SemaphoreGangTaskDispatcher(uint workers) :
121 _task(NULL),
122 _started(0),
123 _not_finished(0),
124 // Limit the semaphore value to the number of workers.
125 _start_semaphore(new Semaphore(0, workers)),
126 _end_semaphore(new Semaphore(0, workers))
127 { }
128
129 ~SemaphoreGangTaskDispatcher() {
130 delete _start_semaphore;
131 delete _end_semaphore;
132 }
133
134 void coordinator_execute_on_workers(AbstractGangTask* task, uint num_workers) {
135 // No workers are allowed to read the state variables until they have been signaled.
136 _task = task;
137 _not_finished = num_workers;
138
139 // Dispatch 'num_workers' number of tasks.
140 _start_semaphore->signal(num_workers);
141
142 // Wait for the last worker to signal the coordinator.
143 _end_semaphore->wait();
144
145 // No workers are allowed to read the state variables after the coordinator has been signaled.
146 _task = NULL;
147 _started = 0;
148 _not_finished = 0;
149 }
150
151 WorkData worker_wait_for_task() {
152 // Wait for the coordinator to dispatch a task.
153 _start_semaphore->wait();
154
155 uint num_started = (uint) Atomic::add(1, (volatile jint*)&_started);
156
157 // Subtract one to get a zero-indexed worker id.
158 uint worker_id = num_started - 1;
159
160 return WorkData(_task, worker_id);
161 }
162
163 void worker_done_with_task() {
164 // Mark that the worker is done with the task.
165 // The worker is not allowed to read the state variables after this line.
166 uint not_finished = (uint) Atomic::add(-1, (volatile jint*)&_not_finished);
167
168 // The last worker signals to the coordinator that all work is completed.
169 if (not_finished == 0) {
170 _end_semaphore->signal();
171 }
172 }
173 };
174 #endif // IMPLEMENTS_SEMAPHORE_CLASS
175
176 class MutexGangTaskDispatcher : public GangTaskDispatcher {
177 AbstractGangTask* _task;
178
179 volatile uint _started;
180 volatile uint _finished;
181 volatile uint _num_workers;
182
183 Monitor* _monitor;
184
185 public:
186 MutexGangTaskDispatcher()
187 : _task(NULL),
188 _monitor(new Monitor(Monitor::leaf, "WorkGang dispatcher lock", false, Monitor::_safepoint_check_never)),
189 _started(0),
190 _finished(0),
191 _num_workers(0) {}
192
193 ~MutexGangTaskDispatcher() {
194 delete _monitor;
195 }
196
197 void coordinator_execute_on_workers(AbstractGangTask* task, uint num_workers) {
198 MutexLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag);
199
200 _task = task;
201 _num_workers = num_workers;
202
203 // Tell the workers to get to work.
204 _monitor->notify_all();
205
206 // Wait for them to finish.
207 while (_finished < _num_workers) {
208 _monitor->wait(/* no_safepoint_check */ true);
209 }
210
211 _task = NULL;
212 _num_workers = 0;
213 _started = 0;
214 _finished = 0;
215 }
216
217 WorkData worker_wait_for_task() {
218 MonitorLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag);
219
220 while (_num_workers == 0 || _started == _num_workers) {
221 _monitor->wait(/* no_safepoint_check */ true);
222 }
223
224 _started++;
225
226 // Subtract one to get a zero-indexed worker id.
227 uint worker_id = _started - 1;
228
229 return WorkData(_task, worker_id);
230 }
231
232 void worker_done_with_task() {
233 MonitorLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag);
234
235 _finished++;
236
237 if (_finished == _num_workers) {
238 // This will wake up all workers and not only the coordinator.
239 _monitor->notify_all();
240 }
241 }
242 };
243
244 static GangTaskDispatcher* create_dispatcher(uint workers) {
245 #if IMPLEMENTS_SEMAPHORE_CLASS
246 if (UseSemaphoreGCThreadsSynchronization) {
247 return new SemaphoreGangTaskDispatcher(workers);
248 }
249 #endif
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 GangTaskDispatcher* dispatcher) :
259 AbstractWorkGang(name, workers, are_GC_task_threads, are_ConcurrentGC_threads),
260 _dispatcher(dispatcher != NULL ? dispatcher: create_dispatcher(workers))
261 { }
262
263 AbstractGangWorker* WorkGang::allocate_worker(uint worker_id) {
264 return new GangWorker(this, worker_id);
265 }
266
267 void WorkGang::run_task(AbstractGangTask* task) {
268 _dispatcher->coordinator_execute_on_workers(task, active_workers());
269 }
270
271 AbstractGangWorker::AbstractGangWorker(AbstractWorkGang* gang, uint id) {
272 _gang = gang;
273 set_id(id);
274 set_name("%s#%d", gang->name(), id);
275 }
276
277 void AbstractGangWorker::run() {
278 initialize();
279 loop();
280 }
281
282 void AbstractGangWorker::initialize() {
283 this->initialize_thread_local_storage();
284 this->record_stack_base_and_size();
285 this->initialize_named_thread();
286 assert(_gang != NULL, "No gang to run in");
287 os::set_priority(this, NearMaxPriority);
288 if (TraceWorkGang) {
289 tty->print_cr("Running gang worker for gang %s id %u",
290 gang()->name(), id());
291 }
292 // The VM thread should not execute here because MutexLocker's are used
293 // as (opposed to MutexLockerEx's).
294 assert(!Thread::current()->is_VM_thread(), "VM thread should not be part"
295 " of a work gang");
296 }
297
298 bool AbstractGangWorker::is_GC_task_thread() const {
299 return gang()->are_GC_task_threads();
300 }
301
302 bool AbstractGangWorker::is_ConcurrentGC_thread() const {
303 return gang()->are_ConcurrentGC_threads();
304 }
305
306 void AbstractGangWorker::print_on(outputStream* st) const {
307 st->print("\"%s\" ", name());
308 Thread::print_on(st);
309 st->cr();
310 }
311
312 WorkData GangWorker::wait_for_task() {
313 return gang()->dispatcher()->worker_wait_for_task();
314 }
315
316 void GangWorker::signal_task_done() {
317 gang()->dispatcher()->worker_done_with_task();
318 }
319
320 void GangWorker::print_task_started(WorkData data) {
321 if (TraceWorkGang) {
322 tty->print_cr("Running work gang %s task %s worker %u", name(), data._task->name(), data._worker_id);
323 }
324 }
325
326 void GangWorker::print_task_done(WorkData data) {
327 if (TraceWorkGang) {
328 tty->print_cr("\nFinished work gang %s task %s worker %u", name(), data._task->name(), data._worker_id);
329 Thread* me = Thread::current();
330 tty->print_cr(" T: " PTR_FORMAT " VM_thread: %d", p2i(me), me->is_VM_thread());
331 }
332 }
333
334 void GangWorker::run_task(WorkData data) {
335 print_task_started(data);
336
337 data._task->work(data._worker_id);
338
339 print_task_done(data);
340 }
341
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
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
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
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
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
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
410 SubTasksDone::SubTasksDone(uint n) :
411 _n_tasks(n), _tasks(NULL) {
412 _tasks = NEW_C_HEAP_ARRAY(uint, n, mtInternal);
413 guarantee(_tasks != NULL, "alloc failure");
414 clear();
415 }
416
417 bool SubTasksDone::valid() {
418 return _tasks != NULL;
419 }
420
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
431 bool SubTasksDone::is_task_claimed(uint t) {
432 assert(t < _n_tasks, "bad task id.");
433 uint old = _tasks[t];
434 if (old == 0) {
435 old = Atomic::cmpxchg(1, &_tasks[t], 0);
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((volatile jint*) &_claimed);
443 }
444 #endif
445 return res;
446 }
447
448 void SubTasksDone::all_tasks_completed(uint n_threads) {
449 jint observed = _threads_completed;
450 jint 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 == (jint)adjusted_thread_count) {
458 clear();
459 }
460 }
461
462
463 SubTasksDone::~SubTasksDone() {
464 if (_tasks != NULL) FREE_C_HEAP_ARRAY(jint, _tasks);
465 }
466
467 // *** SequentialSubTasksDone
468
469 void SequentialSubTasksDone::clear() {
470 _n_tasks = _n_claimed = 0;
471 _n_threads = _n_completed = 0;
472 }
473
474 bool SequentialSubTasksDone::valid() {
475 return _n_threads > 0;
476 }
477
478 bool SequentialSubTasksDone::is_task_claimed(uint& t) {
479 uint* n_claimed_ptr = &_n_claimed;
480 t = *n_claimed_ptr;
481 while (t < _n_tasks) {
482 jint res = Atomic::cmpxchg(t+1, n_claimed_ptr, t);
483 if (res == (jint)t) {
484 return false;
485 }
486 t = *n_claimed_ptr;
487 }
488 return true;
489 }
490
491 bool SequentialSubTasksDone::all_tasks_completed() {
492 uint* n_completed_ptr = &_n_completed;
493 uint complete = *n_completed_ptr;
494 while (true) {
495 uint res = Atomic::cmpxchg(complete+1, n_completed_ptr, complete);
496 if (res == complete) {
497 break;
498 }
499 complete = res;
500 }
501 if (complete+1 == _n_threads) {
502 clear();
503 return true;
504 }
505 return false;
506 }
507
508 bool FreeIdSet::_stat_init = false;
509 FreeIdSet* FreeIdSet::_sets[NSets];
510 bool FreeIdSet::_safepoint;
511
512 FreeIdSet::FreeIdSet(int sz, Monitor* mon) :
513 _sz(sz), _mon(mon), _hd(0), _waiters(0), _index(-1), _claimed(0)
514 {
515 _ids = NEW_C_HEAP_ARRAY(int, sz, mtInternal);
516 for (int i = 0; i < sz; i++) _ids[i] = i+1;
517 _ids[sz-1] = end_of_list; // end of list.
518 if (_stat_init) {
519 for (int j = 0; j < NSets; j++) _sets[j] = NULL;
520 _stat_init = true;
521 }
522 // Add to sets. (This should happen while the system is still single-threaded.)
523 for (int j = 0; j < NSets; j++) {
524 if (_sets[j] == NULL) {
525 _sets[j] = this;
526 _index = j;
527 break;
528 }
529 }
530 guarantee(_index != -1, "Too many FreeIdSets in use!");
531 }
532
533 FreeIdSet::~FreeIdSet() {
534 _sets[_index] = NULL;
535 FREE_C_HEAP_ARRAY(int, _ids);
536 }
537
538 void FreeIdSet::set_safepoint(bool b) {
539 _safepoint = b;
540 if (b) {
541 for (int j = 0; j < NSets; j++) {
542 if (_sets[j] != NULL && _sets[j]->_waiters > 0) {
543 Monitor* mon = _sets[j]->_mon;
544 mon->lock_without_safepoint_check();
545 mon->notify_all();
546 mon->unlock();
547 }
548 }
549 }
550 }
551
552 #define FID_STATS 0
553
554 int FreeIdSet::claim_par_id() {
555 #if FID_STATS
556 thread_t tslf = thr_self();
557 tty->print("claim_par_id[%d]: sz = %d, claimed = %d\n", tslf, _sz, _claimed);
558 #endif
559 MutexLockerEx x(_mon, Mutex::_no_safepoint_check_flag);
560 while (!_safepoint && _hd == end_of_list) {
561 _waiters++;
562 #if FID_STATS
563 if (_waiters > 5) {
564 tty->print("claim_par_id waiting[%d]: %d waiters, %d claimed.\n",
565 tslf, _waiters, _claimed);
566 }
567 #endif
568 _mon->wait(Mutex::_no_safepoint_check_flag);
569 _waiters--;
570 }
571 if (_hd == end_of_list) {
572 #if FID_STATS
573 tty->print("claim_par_id[%d]: returning EOL.\n", tslf);
574 #endif
575 return -1;
576 } else {
577 int res = _hd;
578 _hd = _ids[res];
579 _ids[res] = claimed; // For debugging.
580 _claimed++;
581 #if FID_STATS
582 tty->print("claim_par_id[%d]: returning %d, claimed = %d.\n",
583 tslf, res, _claimed);
584 #endif
585 return res;
586 }
587 }
588
589 bool FreeIdSet::claim_perm_id(int i) {
590 assert(0 <= i && i < _sz, "Out of range.");
591 MutexLockerEx x(_mon, Mutex::_no_safepoint_check_flag);
592 int prev = end_of_list;
593 int cur = _hd;
594 while (cur != end_of_list) {
595 if (cur == i) {
596 if (prev == end_of_list) {
597 _hd = _ids[cur];
598 } else {
599 _ids[prev] = _ids[cur];
600 }
601 _ids[cur] = claimed;
602 _claimed++;
603 return true;
604 } else {
605 prev = cur;
606 cur = _ids[cur];
607 }
608 }
609 return false;
610
611 }
612
613 void FreeIdSet::release_par_id(int id) {
614 MutexLockerEx x(_mon, Mutex::_no_safepoint_check_flag);
615 assert(_ids[id] == claimed, "Precondition.");
616 _ids[id] = _hd;
617 _hd = id;
618 _claimed--;
619 #if FID_STATS
620 tty->print("[%d] release_par_id(%d), waiters =%d, claimed = %d.\n",
621 thr_self(), id, _waiters, _claimed);
622 #endif
623 if (_waiters > 0)
624 // Notify all would be safer, but this is OK, right?
625 _mon->notify_all();
626 }
627
628 /////////////// Unit tests ///////////////
629
630 #ifndef PRODUCT
631
632 class CountTask : public AbstractGangTask {
633 volatile jint _count;
634 bool _sleep;
635 public:
636 CountTask(bool sleep) : AbstractGangTask("CountTask"), _sleep(sleep), _count(0) {}
637 virtual void work(uint worker_id) {
638 if (_sleep) {
639 // Sleep a while, to delay the task to allow the coordinator to run.
640 os::sleep(Thread::current(), 100, false);
641 }
642
643 // Count number of times executed.
644 Atomic::inc(&_count);
645 }
646 uint count() { return _count; }
647 };
648
649
650 static void test_workgang_dispatch(bool use_semaphore,
651 uint total_workers,
652 uint active_workers,
653 bool sleep) {
654
655 GangTaskDispatcher* dispatcher = use_semaphore
656 ? (GangTaskDispatcher*) new SemGangTaskDispatcher(total_workers)
657 : (GangTaskDispatcher*) new MutexGangTaskDispatcher();
658
659 // Intentionally leaking WorkGang, since there's no support to delete WorkGangs.
660 WorkGang* gang = new WorkGang("Test WorkGang", total_workers, false, false, dispatcher);
661 gang->initialize_workers();
662
663 gang->set_active_workers(active_workers);
664
665 CountTask task(sleep);
666
667 gang->run_task(&task);
668
669 uint task_count = task.count();
670
671 assert(task_count == active_workers, err_msg("Expected count: %u got: %u", active_workers, task_count));
672 }
673
674 static void test_workgang_dispatch(bool use_semaphore) {
675 const bool sleep = true;
676 const uint total_workers = 8;
677 for (uint active_workers = 1; active_workers <= 4; active_workers++) {
678 test_workgang_dispatch(use_semaphore, total_workers, active_workers, sleep);
679 test_workgang_dispatch(use_semaphore, total_workers, active_workers, !sleep);
680 }
681 }
682
683 static void test_workgang_dispatch_mutex() {
684 test_workgang_dispatch(false);
685 }
686
687 static void test_workgang_dispatch_semaphore() {
688 test_workgang_dispatch(true);
689 }
690
691 void test_workgang() {
692 // Needed to use dynamic number of active workers.
693 FlagSetting fs(UseDynamicNumberOfGCThreads, true);
694
695 test_workgang_dispatch_semaphore();
696 test_workgang_dispatch_mutex();
697 }
698
699 #endif // PRODUCT