1 /* 2 * Copyright (c) 2001, 2020, 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 "memory/iterator.hpp" 32 #include "runtime/atomic.hpp" 33 #include "runtime/os.hpp" 34 #include "runtime/semaphore.hpp" 35 #include "runtime/thread.inline.hpp" 36 37 // Definitions of WorkGang methods. 38 39 // The current implementation will exit if the allocation 40 // of any worker fails. 41 void AbstractWorkGang::initialize_workers() { 42 log_develop_trace(gc, workgang)("Constructing work gang %s with %u threads", name(), total_workers()); 43 _workers = NEW_C_HEAP_ARRAY(AbstractGangWorker*, total_workers(), mtInternal); 44 add_workers(true); 45 } 46 47 48 AbstractGangWorker* AbstractWorkGang::install_worker(uint worker_id) { 49 AbstractGangWorker* new_worker = allocate_worker(worker_id); 50 set_thread(worker_id, new_worker); 51 return new_worker; 52 } 53 54 void AbstractWorkGang::add_workers(bool initializing) { 55 add_workers(_active_workers, initializing); 56 } 57 58 void AbstractWorkGang::add_workers(uint active_workers, bool initializing) { 59 60 os::ThreadType worker_type; 61 if (are_ConcurrentGC_threads()) { 62 worker_type = os::cgc_thread; 63 } else { 64 worker_type = os::pgc_thread; 65 } 66 uint previous_created_workers = _created_workers; 67 68 _created_workers = WorkerManager::add_workers(this, 69 active_workers, 70 _total_workers, 71 _created_workers, 72 worker_type, 73 initializing); 74 _active_workers = MIN2(_created_workers, _active_workers); 75 76 WorkerManager::log_worker_creation(this, previous_created_workers, _active_workers, _created_workers, initializing); 77 } 78 79 AbstractGangWorker* AbstractWorkGang::worker(uint i) const { 80 // Array index bounds checking. 81 AbstractGangWorker* result = NULL; 82 assert(_workers != NULL, "No workers for indexing"); 83 assert(i < total_workers(), "Worker index out of bounds"); 84 result = _workers[i]; 85 assert(result != NULL, "Indexing to null worker"); 86 return result; 87 } 88 89 void AbstractWorkGang::threads_do(ThreadClosure* tc) const { 90 assert(tc != NULL, "Null ThreadClosure"); 91 uint workers = created_workers(); 92 for (uint i = 0; i < workers; i++) { 93 tc->do_thread(worker(i)); 94 } 95 } 96 97 static void run_foreground_task_if_needed(AbstractGangTask* task, uint num_workers, 98 bool add_foreground_work) { 99 if (add_foreground_work) { 100 log_develop_trace(gc, workgang)("Running work gang: %s task: %s worker: foreground", 101 Thread::current()->name(), task->name()); 102 task->work(num_workers); 103 log_develop_trace(gc, workgang)("Finished work gang: %s task: %s worker: foreground " 104 "thread: " PTR_FORMAT, Thread::current()->name(), task->name(), p2i(Thread::current())); 105 } 106 } 107 108 // WorkGang dispatcher implemented with semaphores. 109 // 110 // Semaphores don't require the worker threads to re-claim the lock when they wake up. 111 // This helps lowering the latency when starting and stopping the worker threads. 112 class SemaphoreGangTaskDispatcher : public GangTaskDispatcher { 113 // The task currently being dispatched to the GangWorkers. 114 AbstractGangTask* _task; 115 116 volatile uint _started; 117 volatile uint _not_finished; 118 119 // Semaphore used to start the GangWorkers. 120 Semaphore* _start_semaphore; 121 // Semaphore used to notify the coordinator that all workers are done. 122 Semaphore* _end_semaphore; 123 124 public: 125 SemaphoreGangTaskDispatcher() : 126 _task(NULL), 127 _started(0), 128 _not_finished(0), 129 _start_semaphore(new Semaphore()), 130 _end_semaphore(new Semaphore()) 131 { } 132 133 ~SemaphoreGangTaskDispatcher() { 134 delete _start_semaphore; 135 delete _end_semaphore; 136 } 137 138 void coordinator_execute_on_workers(AbstractGangTask* task, uint num_workers, bool add_foreground_work) { 139 // No workers are allowed to read the state variables until they have been signaled. 140 _task = task; 141 _not_finished = num_workers; 142 143 // Dispatch 'num_workers' number of tasks. 144 _start_semaphore->signal(num_workers); 145 146 run_foreground_task_if_needed(task, num_workers, add_foreground_work); 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 = Atomic::add(&_started, 1u); 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 = Atomic::sub(&_not_finished, 1u); 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 WorkGang::WorkGang(const char* name, 183 uint workers, 184 bool are_GC_task_threads, 185 bool are_ConcurrentGC_threads) : 186 AbstractWorkGang(name, workers, are_GC_task_threads, are_ConcurrentGC_threads), 187 _dispatcher(new SemaphoreGangTaskDispatcher()) 188 { } 189 190 WorkGang::~WorkGang() { 191 delete _dispatcher; 192 } 193 194 AbstractGangWorker* WorkGang::allocate_worker(uint worker_id) { 195 return new GangWorker(this, worker_id); 196 } 197 198 void WorkGang::run_task(AbstractGangTask* task) { 199 run_task(task, active_workers()); 200 } 201 202 void WorkGang::run_task(AbstractGangTask* task, uint num_workers, bool add_foreground_work) { 203 guarantee(num_workers <= total_workers(), 204 "Trying to execute task %s with %u workers which is more than the amount of total workers %u.", 205 task->name(), num_workers, total_workers()); 206 guarantee(num_workers > 0, "Trying to execute task %s with zero workers", task->name()); 207 uint old_num_workers = _active_workers; 208 update_active_workers(num_workers); 209 _dispatcher->coordinator_execute_on_workers(task, num_workers, add_foreground_work); 210 update_active_workers(old_num_workers); 211 } 212 213 AbstractGangWorker::AbstractGangWorker(AbstractWorkGang* gang, uint id) { 214 _gang = gang; 215 set_id(id); 216 set_name("%s#%d", gang->name(), id); 217 } 218 219 void AbstractGangWorker::run() { 220 initialize(); 221 loop(); 222 } 223 224 void AbstractGangWorker::initialize() { 225 assert(_gang != NULL, "No gang to run in"); 226 os::set_priority(this, NearMaxPriority); 227 log_develop_trace(gc, workgang)("Running gang worker for gang %s id %u", gang()->name(), id()); 228 assert(!Thread::current()->is_VM_thread(), "VM thread should not be part" 229 " of a work gang"); 230 } 231 232 bool AbstractGangWorker::is_GC_task_thread() const { 233 return gang()->are_GC_task_threads(); 234 } 235 236 bool AbstractGangWorker::is_ConcurrentGC_thread() const { 237 return gang()->are_ConcurrentGC_threads(); 238 } 239 240 void AbstractGangWorker::print_on(outputStream* st) const { 241 st->print("\"%s\" ", name()); 242 Thread::print_on(st); 243 st->cr(); 244 } 245 246 void AbstractGangWorker::print() const { print_on(tty); } 247 248 WorkData GangWorker::wait_for_task() { 249 return gang()->dispatcher()->worker_wait_for_task(); 250 } 251 252 void GangWorker::signal_task_done() { 253 gang()->dispatcher()->worker_done_with_task(); 254 } 255 256 void GangWorker::run_task(WorkData data) { 257 GCIdMark gc_id_mark(data._task->gc_id()); 258 log_develop_trace(gc, workgang)("Running work gang: %s task: %s worker: %u", name(), data._task->name(), data._worker_id); 259 260 data._task->work(data._worker_id); 261 262 log_develop_trace(gc, workgang)("Finished work gang: %s task: %s worker: %u thread: " PTR_FORMAT, 263 name(), data._task->name(), data._worker_id, p2i(Thread::current())); 264 } 265 266 void GangWorker::loop() { 267 while (true) { 268 WorkData data = wait_for_task(); 269 270 run_task(data); 271 272 signal_task_done(); 273 } 274 } 275 276 // *** WorkGangBarrierSync 277 278 WorkGangBarrierSync::WorkGangBarrierSync() 279 : _monitor(Mutex::safepoint, "work gang barrier sync", true, 280 Monitor::_safepoint_check_never), 281 _n_workers(0), _n_completed(0), _should_reset(false), _aborted(false) { 282 } 283 284 WorkGangBarrierSync::WorkGangBarrierSync(uint n_workers, const char* name) 285 : _monitor(Mutex::safepoint, name, true, Monitor::_safepoint_check_never), 286 _n_workers(n_workers), _n_completed(0), _should_reset(false), _aborted(false) { 287 } 288 289 void WorkGangBarrierSync::set_n_workers(uint n_workers) { 290 _n_workers = n_workers; 291 _n_completed = 0; 292 _should_reset = false; 293 _aborted = false; 294 } 295 296 bool WorkGangBarrierSync::enter() { 297 MonitorLocker ml(monitor(), Mutex::_no_safepoint_check_flag); 298 if (should_reset()) { 299 // The should_reset() was set and we are the first worker to enter 300 // the sync barrier. We will zero the n_completed() count which 301 // effectively resets the barrier. 302 zero_completed(); 303 set_should_reset(false); 304 } 305 inc_completed(); 306 if (n_completed() == n_workers()) { 307 // At this point we would like to reset the barrier to be ready in 308 // case it is used again. However, we cannot set n_completed() to 309 // 0, even after the notify_all(), given that some other workers 310 // might still be waiting for n_completed() to become == 311 // n_workers(). So, if we set n_completed() to 0, those workers 312 // will get stuck (as they will wake up, see that n_completed() != 313 // n_workers() and go back to sleep). Instead, we raise the 314 // should_reset() flag and the barrier will be reset the first 315 // time a worker enters it again. 316 set_should_reset(true); 317 ml.notify_all(); 318 } else { 319 while (n_completed() != n_workers() && !aborted()) { 320 ml.wait(); 321 } 322 } 323 return !aborted(); 324 } 325 326 void WorkGangBarrierSync::abort() { 327 MutexLocker x(monitor(), Mutex::_no_safepoint_check_flag); 328 set_aborted(); 329 monitor()->notify_all(); 330 } 331 332 // SubTasksDone functions. 333 334 SubTasksDone::SubTasksDone(uint n) : 335 _tasks(NULL), _n_tasks(n), _threads_completed(0) { 336 _tasks = NEW_C_HEAP_ARRAY(uint, n, mtInternal); 337 clear(); 338 } 339 340 bool SubTasksDone::valid() { 341 return _tasks != NULL; 342 } 343 344 void SubTasksDone::clear() { 345 for (uint i = 0; i < _n_tasks; i++) { 346 _tasks[i] = 0; 347 } 348 _threads_completed = 0; 349 #ifdef ASSERT 350 _claimed = 0; 351 #endif 352 } 353 354 bool SubTasksDone::try_claim_task(uint t) { 355 assert(t < _n_tasks, "bad task id."); 356 uint old = _tasks[t]; 357 if (old == 0) { 358 old = Atomic::cmpxchg(&_tasks[t], 0u, 1u); 359 } 360 bool res = old == 0; 361 #ifdef ASSERT 362 if (res) { 363 assert(_claimed < _n_tasks, "Too many tasks claimed; missing clear?"); 364 Atomic::inc(&_claimed); 365 } 366 #endif 367 return res; 368 } 369 370 void SubTasksDone::all_tasks_completed(uint n_threads) { 371 uint observed = _threads_completed; 372 uint old; 373 do { 374 old = observed; 375 observed = Atomic::cmpxchg(&_threads_completed, old, old+1); 376 } while (observed != old); 377 // If this was the last thread checking in, clear the tasks. 378 uint adjusted_thread_count = (n_threads == 0 ? 1 : n_threads); 379 if (observed + 1 == adjusted_thread_count) { 380 clear(); 381 } 382 } 383 384 385 SubTasksDone::~SubTasksDone() { 386 FREE_C_HEAP_ARRAY(uint, _tasks); 387 } 388 389 // *** SequentialSubTasksDone 390 391 void SequentialSubTasksDone::clear() { 392 _n_tasks = _n_claimed = 0; 393 _n_threads = _n_completed = 0; 394 } 395 396 bool SequentialSubTasksDone::valid() { 397 return _n_threads > 0; 398 } 399 400 bool SequentialSubTasksDone::try_claim_task(uint& t) { 401 t = _n_claimed; 402 while (t < _n_tasks) { 403 uint res = Atomic::cmpxchg(&_n_claimed, t, t+1); 404 if (res == t) { 405 return true; 406 } 407 t = res; 408 } 409 return false; 410 } 411 412 bool SequentialSubTasksDone::all_tasks_completed() { 413 uint complete = _n_completed; 414 while (true) { 415 uint res = Atomic::cmpxchg(&_n_completed, complete, complete+1); 416 if (res == complete) { 417 break; 418 } 419 complete = res; 420 } 421 if (complete+1 == _n_threads) { 422 clear(); 423 return true; 424 } 425 return false; 426 }