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 "memory/allocation.hpp" 29 #include "memory/allocation.inline.hpp" 30 #include "runtime/atomic.inline.hpp" 31 #include "runtime/os.hpp" 32 #include "runtime/semaphore.hpp" 33 #include "runtime/thread.inline.hpp" 34 35 // Definitions of WorkGang methods. 36 37 // The current implementation will exit if the allocation 38 // of any worker fails. Still, return a boolean so that 39 // a future implementation can possibly do a partial 40 // initialization of the workers and report such to the 41 // caller. 42 bool AbstractWorkGang::initialize_workers() { 43 log_develop_trace(gc, workgang)("Constructing work gang %s with %u threads", name(), total_workers()); 44 _workers = NEW_C_HEAP_ARRAY(AbstractGangWorker*, total_workers(), mtInternal); 45 if (_workers == NULL) { 46 vm_exit_out_of_memory(0, OOM_MALLOC_ERROR, "Cannot create GangWorker array."); 47 return false; 48 } 49 os::ThreadType worker_type; 50 if (are_ConcurrentGC_threads()) { 51 worker_type = os::cgc_thread; 52 } else { 53 worker_type = os::pgc_thread; 54 } 55 for (uint worker = 0; worker < total_workers(); worker += 1) { 56 AbstractGangWorker* new_worker = allocate_worker(worker); 57 assert(new_worker != NULL, "Failed to allocate GangWorker"); 58 _workers[worker] = new_worker; 59 if (new_worker == NULL || !os::create_thread(new_worker, worker_type)) { 60 vm_exit_out_of_memory(0, OOM_MALLOC_ERROR, 61 "Cannot create worker GC thread. Out of system resources."); 62 return false; 63 } 64 if (!DisableStartThread) { 65 os::start_thread(new_worker); 66 } 67 } 68 return true; 69 } 70 71 AbstractGangWorker* AbstractWorkGang::worker(uint i) const { 72 // Array index bounds checking. 73 AbstractGangWorker* result = NULL; 74 assert(_workers != NULL, "No workers for indexing"); 75 assert(i < total_workers(), "Worker index out of bounds"); 76 result = _workers[i]; 77 assert(result != NULL, "Indexing to null worker"); 78 return result; 79 } 80 81 void AbstractWorkGang::print_worker_threads_on(outputStream* st) const { 82 uint workers = total_workers(); 83 for (uint i = 0; i < workers; i++) { 84 worker(i)->print_on(st); 85 st->cr(); 86 } 87 } 88 89 void AbstractWorkGang::threads_do(ThreadClosure* tc) const { 90 assert(tc != NULL, "Null ThreadClosure"); 91 uint workers = total_workers(); 92 for (uint i = 0; i < workers; i++) { 93 tc->do_thread(worker(i)); 94 } 95 } 96 97 // WorkGang dispatcher implemented with semaphores. 98 // 99 // Semaphores don't require the worker threads to re-claim the lock when they wake up. 100 // This helps lowering the latency when starting and stopping the worker threads. 101 class SemaphoreGangTaskDispatcher : public GangTaskDispatcher { 102 // The task currently being dispatched to the GangWorkers. 103 AbstractGangTask* _task; 104 105 volatile uint _started; 106 volatile uint _not_finished; 107 108 // Semaphore used to start the GangWorkers. 109 Semaphore* _start_semaphore; 110 // Semaphore used to notify the coordinator that all workers are done. 111 Semaphore* _end_semaphore; 112 113 public: 114 SemaphoreGangTaskDispatcher() : 115 _task(NULL), 116 _started(0), 117 _not_finished(0), 118 _start_semaphore(new Semaphore()), 119 _end_semaphore(new Semaphore()) 120 { } 121 122 ~SemaphoreGangTaskDispatcher() { 123 delete _start_semaphore; 124 delete _end_semaphore; 125 } 126 127 void coordinator_execute_on_workers(AbstractGangTask* task, uint num_workers) { 128 // No workers are allowed to read the state variables until they have been signaled. 129 _task = task; 130 _not_finished = num_workers; 131 132 // Dispatch 'num_workers' number of tasks. 133 _start_semaphore->signal(num_workers); 134 135 // Wait for the last worker to signal the coordinator. 136 _end_semaphore->wait(); 137 138 // No workers are allowed to read the state variables after the coordinator has been signaled. 139 assert(_not_finished == 0, "%d not finished workers?", _not_finished); 140 _task = NULL; 141 _started = 0; 142 143 } 144 145 WorkData worker_wait_for_task() { 146 // Wait for the coordinator to dispatch a task. 147 _start_semaphore->wait(); 148 149 uint num_started = (uint) Atomic::add(1, (volatile jint*)&_started); 150 151 // Subtract one to get a zero-indexed worker id. 152 uint worker_id = num_started - 1; 153 154 return WorkData(_task, worker_id); 155 } 156 157 void worker_done_with_task() { 158 // Mark that the worker is done with the task. 159 // The worker is not allowed to read the state variables after this line. 160 uint not_finished = (uint) Atomic::add(-1, (volatile jint*)&_not_finished); 161 162 // The last worker signals to the coordinator that all work is completed. 163 if (not_finished == 0) { 164 _end_semaphore->signal(); 165 } 166 } 167 }; 168 169 class MutexGangTaskDispatcher : public GangTaskDispatcher { 170 AbstractGangTask* _task; 171 172 volatile uint _started; 173 volatile uint _finished; 174 volatile uint _num_workers; 175 176 Monitor* _monitor; 177 178 public: 179 MutexGangTaskDispatcher() 180 : _task(NULL), 181 _monitor(new Monitor(Monitor::leaf, "WorkGang dispatcher lock", false, Monitor::_safepoint_check_never)), 182 _started(0), 183 _finished(0), 184 _num_workers(0) {} 185 186 ~MutexGangTaskDispatcher() { 187 delete _monitor; 188 } 189 190 void coordinator_execute_on_workers(AbstractGangTask* task, uint num_workers) { 191 MutexLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag); 192 193 _task = task; 194 _num_workers = num_workers; 195 196 // Tell the workers to get to work. 197 _monitor->notify_all(); 198 199 // Wait for them to finish. 200 while (_finished < _num_workers) { 201 _monitor->wait(/* no_safepoint_check */ true); 202 } 203 204 _task = NULL; 205 _num_workers = 0; 206 _started = 0; 207 _finished = 0; 208 } 209 210 WorkData worker_wait_for_task() { 211 MonitorLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag); 212 213 while (_num_workers == 0 || _started == _num_workers) { 214 _monitor->wait(/* no_safepoint_check */ true); 215 } 216 217 _started++; 218 219 // Subtract one to get a zero-indexed worker id. 220 uint worker_id = _started - 1; 221 222 return WorkData(_task, worker_id); 223 } 224 225 void worker_done_with_task() { 226 MonitorLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag); 227 228 _finished++; 229 230 if (_finished == _num_workers) { 231 // This will wake up all workers and not only the coordinator. 232 _monitor->notify_all(); 233 } 234 } 235 }; 236 237 static GangTaskDispatcher* create_dispatcher() { 238 if (UseSemaphoreGCThreadsSynchronization) { 239 return new SemaphoreGangTaskDispatcher(); 240 } 241 242 return new MutexGangTaskDispatcher(); 243 } 244 245 WorkGang::WorkGang(const char* name, 246 uint workers, 247 bool are_GC_task_threads, 248 bool are_ConcurrentGC_threads) : 249 AbstractWorkGang(name, workers, are_GC_task_threads, are_ConcurrentGC_threads), 250 _dispatcher(create_dispatcher()) 251 { } 252 253 AbstractGangWorker* WorkGang::allocate_worker(uint worker_id) { 254 return new GangWorker(this, worker_id); 255 } 256 257 void WorkGang::run_task(AbstractGangTask* task) { 258 _dispatcher->coordinator_execute_on_workers(task, active_workers()); 259 } 260 261 AbstractGangWorker::AbstractGangWorker(AbstractWorkGang* gang, uint id) { 262 _gang = gang; 263 set_id(id); 264 set_name("%s#%d", gang->name(), id); 265 } 266 267 void AbstractGangWorker::run() { 268 initialize(); 269 loop(); 270 } 271 272 void AbstractGangWorker::initialize() { 273 this->record_stack_base_and_size(); 274 this->initialize_named_thread(); 275 assert(_gang != NULL, "No gang to run in"); 276 os::set_priority(this, NearMaxPriority); 277 log_develop_trace(gc, workgang)("Running gang worker for gang %s id %u", gang()->name(), id()); 278 // The VM thread should not execute here because MutexLocker's are used 279 // as (opposed to MutexLockerEx's). 280 assert(!Thread::current()->is_VM_thread(), "VM thread should not be part" 281 " of a work gang"); 282 } 283 284 bool AbstractGangWorker::is_GC_task_thread() const { 285 return gang()->are_GC_task_threads(); 286 } 287 288 bool AbstractGangWorker::is_ConcurrentGC_thread() const { 289 return gang()->are_ConcurrentGC_threads(); 290 } 291 292 void AbstractGangWorker::print_on(outputStream* st) const { 293 st->print("\"%s\" ", name()); 294 Thread::print_on(st); 295 st->cr(); 296 } 297 298 WorkData GangWorker::wait_for_task() { 299 return gang()->dispatcher()->worker_wait_for_task(); 300 } 301 302 void GangWorker::signal_task_done() { 303 gang()->dispatcher()->worker_done_with_task(); 304 } 305 306 void GangWorker::run_task(WorkData data) { 307 GCIdMark gc_id_mark(data._task->gc_id()); 308 log_develop_trace(gc, workgang)("Running work gang: %s task: %s worker: %u", name(), data._task->name(), data._worker_id); 309 310 data._task->work(data._worker_id); 311 312 log_develop_trace(gc, workgang)("Finished work gang: %s task: %s worker: %u thread: " PTR_FORMAT, 313 name(), data._task->name(), data._worker_id, p2i(Thread::current())); 314 } 315 316 void GangWorker::loop() { 317 while (true) { 318 WorkData data = wait_for_task(); 319 320 run_task(data); 321 322 signal_task_done(); 323 } 324 } 325 326 // *** WorkGangBarrierSync 327 328 WorkGangBarrierSync::WorkGangBarrierSync() 329 : _monitor(Mutex::safepoint, "work gang barrier sync", true, 330 Monitor::_safepoint_check_never), 331 _n_workers(0), _n_completed(0), _should_reset(false), _aborted(false) { 332 } 333 334 WorkGangBarrierSync::WorkGangBarrierSync(uint n_workers, const char* name) 335 : _monitor(Mutex::safepoint, name, true, Monitor::_safepoint_check_never), 336 _n_workers(n_workers), _n_completed(0), _should_reset(false), _aborted(false) { 337 } 338 339 void WorkGangBarrierSync::set_n_workers(uint n_workers) { 340 _n_workers = n_workers; 341 _n_completed = 0; 342 _should_reset = false; 343 _aborted = false; 344 } 345 346 bool WorkGangBarrierSync::enter() { 347 MutexLockerEx x(monitor(), Mutex::_no_safepoint_check_flag); 348 if (should_reset()) { 349 // The should_reset() was set and we are the first worker to enter 350 // the sync barrier. We will zero the n_completed() count which 351 // effectively resets the barrier. 352 zero_completed(); 353 set_should_reset(false); 354 } 355 inc_completed(); 356 if (n_completed() == n_workers()) { 357 // At this point we would like to reset the barrier to be ready in 358 // case it is used again. However, we cannot set n_completed() to 359 // 0, even after the notify_all(), given that some other workers 360 // might still be waiting for n_completed() to become == 361 // n_workers(). So, if we set n_completed() to 0, those workers 362 // will get stuck (as they will wake up, see that n_completed() != 363 // n_workers() and go back to sleep). Instead, we raise the 364 // should_reset() flag and the barrier will be reset the first 365 // time a worker enters it again. 366 set_should_reset(true); 367 monitor()->notify_all(); 368 } else { 369 while (n_completed() != n_workers() && !aborted()) { 370 monitor()->wait(/* no_safepoint_check */ true); 371 } 372 } 373 return !aborted(); 374 } 375 376 void WorkGangBarrierSync::abort() { 377 MutexLockerEx x(monitor(), Mutex::_no_safepoint_check_flag); 378 set_aborted(); 379 monitor()->notify_all(); 380 } 381 382 // SubTasksDone functions. 383 384 SubTasksDone::SubTasksDone(uint n) : 385 _n_tasks(n), _tasks(NULL) { 386 _tasks = NEW_C_HEAP_ARRAY(uint, n, mtInternal); 387 guarantee(_tasks != NULL, "alloc failure"); 388 clear(); 389 } 390 391 bool SubTasksDone::valid() { 392 return _tasks != NULL; 393 } 394 395 void SubTasksDone::clear() { 396 for (uint i = 0; i < _n_tasks; i++) { 397 _tasks[i] = 0; 398 } 399 _threads_completed = 0; 400 #ifdef ASSERT 401 _claimed = 0; 402 #endif 403 } 404 405 bool SubTasksDone::is_task_claimed(uint t) { 406 assert(t < _n_tasks, "bad task id."); 407 uint old = _tasks[t]; 408 if (old == 0) { 409 old = Atomic::cmpxchg(1, &_tasks[t], 0); 410 } 411 assert(_tasks[t] == 1, "What else?"); 412 bool res = old != 0; 413 #ifdef ASSERT 414 if (!res) { 415 assert(_claimed < _n_tasks, "Too many tasks claimed; missing clear?"); 416 Atomic::inc((volatile jint*) &_claimed); 417 } 418 #endif 419 return res; 420 } 421 422 void SubTasksDone::all_tasks_completed(uint n_threads) { 423 jint observed = _threads_completed; 424 jint old; 425 do { 426 old = observed; 427 observed = Atomic::cmpxchg(old+1, &_threads_completed, old); 428 } while (observed != old); 429 // If this was the last thread checking in, clear the tasks. 430 uint adjusted_thread_count = (n_threads == 0 ? 1 : n_threads); 431 if (observed + 1 == (jint)adjusted_thread_count) { 432 clear(); 433 } 434 } 435 436 437 SubTasksDone::~SubTasksDone() { 438 if (_tasks != NULL) FREE_C_HEAP_ARRAY(jint, _tasks); 439 } 440 441 // *** SequentialSubTasksDone 442 443 void SequentialSubTasksDone::clear() { 444 _n_tasks = _n_claimed = 0; 445 _n_threads = _n_completed = 0; 446 } 447 448 bool SequentialSubTasksDone::valid() { 449 return _n_threads > 0; 450 } 451 452 bool SequentialSubTasksDone::is_task_claimed(uint& t) { 453 uint* n_claimed_ptr = &_n_claimed; 454 t = *n_claimed_ptr; 455 while (t < _n_tasks) { 456 jint res = Atomic::cmpxchg(t+1, n_claimed_ptr, t); 457 if (res == (jint)t) { 458 return false; 459 } 460 t = *n_claimed_ptr; 461 } 462 return true; 463 } 464 465 bool SequentialSubTasksDone::all_tasks_completed() { 466 uint* n_completed_ptr = &_n_completed; 467 uint complete = *n_completed_ptr; 468 while (true) { 469 uint res = Atomic::cmpxchg(complete+1, n_completed_ptr, complete); 470 if (res == complete) { 471 break; 472 } 473 complete = res; 474 } 475 if (complete+1 == _n_threads) { 476 clear(); 477 return true; 478 } 479 return false; 480 }