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