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 _active_workers = ParallelGCThreads; 48 if (UseDynamicNumberOfGCThreads && !FLAG_IS_CMDLINE(ParallelGCThreads)) { 49 _active_workers = 1U; 50 } 51 52 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 void AbstractWorkGang::add_workers(bool initializing) { 63 add_workers(_active_workers, initializing); 64 } 65 66 void AbstractWorkGang::add_workers(uint active_workers, bool initializing) { 67 68 os::ThreadType worker_type; 69 if (are_ConcurrentGC_threads()) { 70 worker_type = os::cgc_thread; 71 } else { 72 worker_type = os::pgc_thread; 73 } 74 75 _created_workers = WorkerManager::add_workers(this, 76 active_workers, 77 _total_workers, 78 _created_workers, 79 worker_type, 80 initializing); 81 _active_workers = MIN2(_created_workers, _active_workers); 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 run_task(task, active_workers()); 272 } 273 274 void WorkGang::run_task(AbstractGangTask* task, uint num_workers) { 275 guarantee(num_workers <= total_workers(), 276 "Trying to execute task %s with %u workers which is more than the amount of total workers %u.", 277 task->name(), num_workers, total_workers()); 278 guarantee(num_workers > 0, "Trying to execute task %s with zero workers", task->name()); 279 _dispatcher->coordinator_execute_on_workers(task, num_workers); 280 } 281 282 AbstractGangWorker::AbstractGangWorker(AbstractWorkGang* gang, uint id) { 283 _gang = gang; 284 set_id(id); 285 set_name("%s#%d", gang->name(), id); 286 } 287 288 void AbstractGangWorker::run() { 289 initialize(); 290 loop(); 291 } 292 293 void AbstractGangWorker::initialize() { 294 this->record_stack_base_and_size(); 295 this->initialize_named_thread(); 296 assert(_gang != NULL, "No gang to run in"); 297 os::set_priority(this, NearMaxPriority); 298 log_develop_trace(gc, workgang)("Running gang worker for gang %s id %u", gang()->name(), id()); 299 // The VM thread should not execute here because MutexLocker's are used 300 // as (opposed to MutexLockerEx's). 301 assert(!Thread::current()->is_VM_thread(), "VM thread should not be part" 302 " of a work gang"); 303 } 304 305 bool AbstractGangWorker::is_GC_task_thread() const { 306 return gang()->are_GC_task_threads(); 307 } 308 309 bool AbstractGangWorker::is_ConcurrentGC_thread() const { 310 return gang()->are_ConcurrentGC_threads(); 311 } 312 313 void AbstractGangWorker::print_on(outputStream* st) const { 314 st->print("\"%s\" ", name()); 315 Thread::print_on(st); 316 st->cr(); 317 } 318 319 WorkData GangWorker::wait_for_task() { 320 return gang()->dispatcher()->worker_wait_for_task(); 321 } 322 323 void GangWorker::signal_task_done() { 324 gang()->dispatcher()->worker_done_with_task(); 325 } 326 327 void GangWorker::run_task(WorkData data) { 328 GCIdMark gc_id_mark(data._task->gc_id()); 329 log_develop_trace(gc, workgang)("Running work gang: %s task: %s worker: %u", name(), data._task->name(), data._worker_id); 330 331 data._task->work(data._worker_id); 332 333 log_develop_trace(gc, workgang)("Finished work gang: %s task: %s worker: %u thread: " PTR_FORMAT, 334 name(), data._task->name(), data._worker_id, p2i(Thread::current())); 335 } 336 337 void GangWorker::loop() { 338 while (true) { 339 WorkData data = wait_for_task(); 340 341 run_task(data); 342 343 signal_task_done(); 344 } 345 } 346 347 // *** WorkGangBarrierSync 348 349 WorkGangBarrierSync::WorkGangBarrierSync() 350 : _monitor(Mutex::safepoint, "work gang barrier sync", true, 351 Monitor::_safepoint_check_never), 352 _n_workers(0), _n_completed(0), _should_reset(false), _aborted(false) { 353 } 354 355 WorkGangBarrierSync::WorkGangBarrierSync(uint n_workers, const char* name) 356 : _monitor(Mutex::safepoint, name, true, Monitor::_safepoint_check_never), 357 _n_workers(n_workers), _n_completed(0), _should_reset(false), _aborted(false) { 358 } 359 360 void WorkGangBarrierSync::set_n_workers(uint n_workers) { 361 _n_workers = n_workers; 362 _n_completed = 0; 363 _should_reset = false; 364 _aborted = false; 365 } 366 367 bool WorkGangBarrierSync::enter() { 368 MutexLockerEx x(monitor(), Mutex::_no_safepoint_check_flag); 369 if (should_reset()) { 370 // The should_reset() was set and we are the first worker to enter 371 // the sync barrier. We will zero the n_completed() count which 372 // effectively resets the barrier. 373 zero_completed(); 374 set_should_reset(false); 375 } 376 inc_completed(); 377 if (n_completed() == n_workers()) { 378 // At this point we would like to reset the barrier to be ready in 379 // case it is used again. However, we cannot set n_completed() to 380 // 0, even after the notify_all(), given that some other workers 381 // might still be waiting for n_completed() to become == 382 // n_workers(). So, if we set n_completed() to 0, those workers 383 // will get stuck (as they will wake up, see that n_completed() != 384 // n_workers() and go back to sleep). Instead, we raise the 385 // should_reset() flag and the barrier will be reset the first 386 // time a worker enters it again. 387 set_should_reset(true); 388 monitor()->notify_all(); 389 } else { 390 while (n_completed() != n_workers() && !aborted()) { 391 monitor()->wait(/* no_safepoint_check */ true); 392 } 393 } 394 return !aborted(); 395 } 396 397 void WorkGangBarrierSync::abort() { 398 MutexLockerEx x(monitor(), Mutex::_no_safepoint_check_flag); 399 set_aborted(); 400 monitor()->notify_all(); 401 } 402 403 // SubTasksDone functions. 404 405 SubTasksDone::SubTasksDone(uint n) : 406 _n_tasks(n), _tasks(NULL) { 407 _tasks = NEW_C_HEAP_ARRAY(uint, n, mtInternal); 408 guarantee(_tasks != NULL, "alloc failure"); 409 clear(); 410 } 411 412 bool SubTasksDone::valid() { 413 return _tasks != NULL; 414 } 415 416 void SubTasksDone::clear() { 417 for (uint i = 0; i < _n_tasks; i++) { 418 _tasks[i] = 0; 419 } 420 _threads_completed = 0; 421 #ifdef ASSERT 422 _claimed = 0; 423 #endif 424 } 425 426 bool SubTasksDone::is_task_claimed(uint t) { 427 assert(t < _n_tasks, "bad task id."); 428 uint old = _tasks[t]; 429 if (old == 0) { 430 old = Atomic::cmpxchg(1, &_tasks[t], 0); 431 } 432 assert(_tasks[t] == 1, "What else?"); 433 bool res = old != 0; 434 #ifdef ASSERT 435 if (!res) { 436 assert(_claimed < _n_tasks, "Too many tasks claimed; missing clear?"); 437 Atomic::inc((volatile jint*) &_claimed); 438 } 439 #endif 440 return res; 441 } 442 443 void SubTasksDone::all_tasks_completed(uint n_threads) { 444 jint observed = _threads_completed; 445 jint old; 446 do { 447 old = observed; 448 observed = Atomic::cmpxchg(old+1, &_threads_completed, old); 449 } while (observed != old); 450 // If this was the last thread checking in, clear the tasks. 451 uint adjusted_thread_count = (n_threads == 0 ? 1 : n_threads); 452 if (observed + 1 == (jint)adjusted_thread_count) { 453 clear(); 454 } 455 } 456 457 458 SubTasksDone::~SubTasksDone() { 459 if (_tasks != NULL) FREE_C_HEAP_ARRAY(jint, _tasks); 460 } 461 462 // *** SequentialSubTasksDone 463 464 void SequentialSubTasksDone::clear() { 465 _n_tasks = _n_claimed = 0; 466 _n_threads = _n_completed = 0; 467 } 468 469 bool SequentialSubTasksDone::valid() { 470 return _n_threads > 0; 471 } 472 473 bool SequentialSubTasksDone::is_task_claimed(uint& t) { 474 uint* n_claimed_ptr = &_n_claimed; 475 t = *n_claimed_ptr; 476 while (t < _n_tasks) { 477 jint res = Atomic::cmpxchg(t+1, n_claimed_ptr, t); 478 if (res == (jint)t) { 479 return false; 480 } 481 t = *n_claimed_ptr; 482 } 483 return true; 484 } 485 486 bool SequentialSubTasksDone::all_tasks_completed() { 487 uint* n_completed_ptr = &_n_completed; 488 uint complete = *n_completed_ptr; 489 while (true) { 490 uint res = Atomic::cmpxchg(complete+1, n_completed_ptr, complete); 491 if (res == complete) { 492 break; 493 } 494 complete = res; 495 } 496 if (complete+1 == _n_threads) { 497 clear(); 498 return true; 499 } 500 return false; 501 }