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/semaphore.hpp" 32 #include "runtime/thread.inline.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 // WorkGang dispatcher implemented with semaphores. 102 // 103 // Semaphores don't require the worker threads to re-claim the lock when they wake up. 104 // This helps lowering the latency when starting and stopping the worker threads. 105 class SemaphoreGangTaskDispatcher : public GangTaskDispatcher { 106 // The task currently being dispatched to the GangWorkers. 107 AbstractGangTask* _task; 108 109 volatile uint _started; 110 volatile uint _not_finished; 111 112 // Semaphore used to start the GangWorkers. 113 Semaphore* _start_semaphore; 114 // Semaphore used to notify the coordinator that all workers are done. 115 Semaphore* _end_semaphore; 116 117 public: 118 SemaphoreGangTaskDispatcher() : 119 _task(NULL), 120 _started(0), 121 _not_finished(0), 122 _start_semaphore(new Semaphore()), 123 _end_semaphore(new Semaphore()) 124 { } 125 126 ~SemaphoreGangTaskDispatcher() { 127 delete _start_semaphore; 128 delete _end_semaphore; 129 } 130 131 void coordinator_execute_on_workers(AbstractGangTask* task, uint num_workers) { 132 // No workers are allowed to read the state variables until they have been signaled. 133 _task = task; 134 _not_finished = num_workers; 135 136 // Dispatch 'num_workers' number of tasks. 137 _start_semaphore->signal(num_workers); 138 139 // Wait for the last worker to signal the coordinator. 140 _end_semaphore->wait(); 141 142 // No workers are allowed to read the state variables after the coordinator has been signaled. 143 assert(_not_finished == 0, err_msg("%d not finished workers?", _not_finished)); 144 _task = NULL; 145 _started = 0; 146 147 } 148 149 WorkData worker_wait_for_task() { 150 // Wait for the coordinator to dispatch a task. 151 _start_semaphore->wait(); 152 153 uint num_started = (uint) Atomic::add(1, (volatile jint*)&_started); 154 155 // Subtract one to get a zero-indexed worker id. 156 uint worker_id = num_started - 1; 157 158 return WorkData(_task, worker_id); 159 } 160 161 void worker_done_with_task() { 162 // Mark that the worker is done with the task. 163 // The worker is not allowed to read the state variables after this line. 164 uint not_finished = (uint) Atomic::add(-1, (volatile jint*)&_not_finished); 165 166 // The last worker signals to the coordinator that all work is completed. 167 if (not_finished == 0) { 168 _end_semaphore->signal(); 169 } 170 } 171 }; 172 173 class MutexGangTaskDispatcher : public GangTaskDispatcher { 174 AbstractGangTask* _task; 175 176 volatile uint _started; 177 volatile uint _finished; 178 volatile uint _num_workers; 179 180 Monitor* _monitor; 181 182 public: 183 MutexGangTaskDispatcher() 184 : _task(NULL), 185 _monitor(new Monitor(Monitor::leaf, "WorkGang dispatcher lock", false, Monitor::_safepoint_check_never)), 186 _started(0), 187 _finished(0), 188 _num_workers(0) {} 189 190 ~MutexGangTaskDispatcher() { 191 delete _monitor; 192 } 193 194 void coordinator_execute_on_workers(AbstractGangTask* task, uint num_workers) { 195 MutexLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag); 196 197 _task = task; 198 _num_workers = num_workers; 199 200 // Tell the workers to get to work. 201 _monitor->notify_all(); 202 203 // Wait for them to finish. 204 while (_finished < _num_workers) { 205 _monitor->wait(/* no_safepoint_check */ true); 206 } 207 208 _task = NULL; 209 _num_workers = 0; 210 _started = 0; 211 _finished = 0; 212 } 213 214 WorkData worker_wait_for_task() { 215 MonitorLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag); 216 217 while (_num_workers == 0 || _started == _num_workers) { 218 _monitor->wait(/* no_safepoint_check */ true); 219 } 220 221 _started++; 222 223 // Subtract one to get a zero-indexed worker id. 224 uint worker_id = _started - 1; 225 226 return WorkData(_task, worker_id); 227 } 228 229 void worker_done_with_task() { 230 MonitorLockerEx ml(_monitor, Mutex::_no_safepoint_check_flag); 231 232 _finished++; 233 234 if (_finished == _num_workers) { 235 // This will wake up all workers and not only the coordinator. 236 _monitor->notify_all(); 237 } 238 } 239 }; 240 241 static GangTaskDispatcher* create_dispatcher() { 242 if (UseSemaphoreGCThreadsSynchronization) { 243 return new SemaphoreGangTaskDispatcher(); 244 } 245 246 return new MutexGangTaskDispatcher(); 247 } 248 249 WorkGang::WorkGang(const char* name, 250 uint workers, 251 bool are_GC_task_threads, 252 bool are_ConcurrentGC_threads) : 253 AbstractWorkGang(name, workers, are_GC_task_threads, are_ConcurrentGC_threads), 254 _dispatcher(create_dispatcher()) 255 { } 256 257 AbstractGangWorker* WorkGang::allocate_worker(uint worker_id) { 258 return new GangWorker(this, worker_id); 259 } 260 261 void WorkGang::run_task(AbstractGangTask* task) { 262 _dispatcher->coordinator_execute_on_workers(task, active_workers()); 263 } 264 265 AbstractGangWorker::AbstractGangWorker(AbstractWorkGang* gang, uint id) { 266 _gang = gang; 267 set_id(id); 268 set_name("%s#%d", gang->name(), id); 269 } 270 271 void AbstractGangWorker::run() { 272 initialize(); 273 loop(); 274 } 275 276 void AbstractGangWorker::initialize() { 277 this->initialize_thread_local_storage(); 278 this->record_stack_base_and_size(); 279 this->initialize_named_thread(); 280 assert(_gang != NULL, "No gang to run in"); 281 os::set_priority(this, NearMaxPriority); 282 if (TraceWorkGang) { 283 tty->print_cr("Running gang worker for gang %s id %u", 284 gang()->name(), id()); 285 } 286 // The VM thread should not execute here because MutexLocker's are used 287 // as (opposed to MutexLockerEx's). 288 assert(!Thread::current()->is_VM_thread(), "VM thread should not be part" 289 " of a work gang"); 290 } 291 292 bool AbstractGangWorker::is_GC_task_thread() const { 293 return gang()->are_GC_task_threads(); 294 } 295 296 bool AbstractGangWorker::is_ConcurrentGC_thread() const { 297 return gang()->are_ConcurrentGC_threads(); 298 } 299 300 void AbstractGangWorker::print_on(outputStream* st) const { 301 st->print("\"%s\" ", name()); 302 Thread::print_on(st); 303 st->cr(); 304 } 305 306 WorkData GangWorker::wait_for_task() { 307 return gang()->dispatcher()->worker_wait_for_task(); 308 } 309 310 void GangWorker::signal_task_done() { 311 gang()->dispatcher()->worker_done_with_task(); 312 } 313 314 void GangWorker::print_task_started(WorkData data) { 315 if (TraceWorkGang) { 316 tty->print_cr("Running work gang %s task %s worker %u", name(), data._task->name(), data._worker_id); 317 } 318 } 319 320 void GangWorker::print_task_done(WorkData data) { 321 if (TraceWorkGang) { 322 tty->print_cr("\nFinished work gang %s task %s worker %u", name(), data._task->name(), data._worker_id); 323 Thread* me = Thread::current(); 324 tty->print_cr(" T: " PTR_FORMAT " VM_thread: %d", p2i(me), me->is_VM_thread()); 325 } 326 } 327 328 void GangWorker::run_task(WorkData data) { 329 print_task_started(data); 330 331 data._task->work(data._worker_id); 332 333 print_task_done(data); 334 } 335 336 void GangWorker::loop() { 337 while (true) { 338 WorkData data = wait_for_task(); 339 340 run_task(data); 341 342 signal_task_done(); 343 } 344 } 345 346 // *** WorkGangBarrierSync 347 348 WorkGangBarrierSync::WorkGangBarrierSync() 349 : _monitor(Mutex::safepoint, "work gang barrier sync", true, 350 Monitor::_safepoint_check_never), 351 _n_workers(0), _n_completed(0), _should_reset(false), _aborted(false) { 352 } 353 354 WorkGangBarrierSync::WorkGangBarrierSync(uint n_workers, const char* name) 355 : _monitor(Mutex::safepoint, name, true, Monitor::_safepoint_check_never), 356 _n_workers(n_workers), _n_completed(0), _should_reset(false), _aborted(false) { 357 } 358 359 void WorkGangBarrierSync::set_n_workers(uint n_workers) { 360 _n_workers = n_workers; 361 _n_completed = 0; 362 _should_reset = false; 363 _aborted = false; 364 } 365 366 bool WorkGangBarrierSync::enter() { 367 MutexLockerEx x(monitor(), Mutex::_no_safepoint_check_flag); 368 if (should_reset()) { 369 // The should_reset() was set and we are the first worker to enter 370 // the sync barrier. We will zero the n_completed() count which 371 // effectively resets the barrier. 372 zero_completed(); 373 set_should_reset(false); 374 } 375 inc_completed(); 376 if (n_completed() == n_workers()) { 377 // At this point we would like to reset the barrier to be ready in 378 // case it is used again. However, we cannot set n_completed() to 379 // 0, even after the notify_all(), given that some other workers 380 // might still be waiting for n_completed() to become == 381 // n_workers(). So, if we set n_completed() to 0, those workers 382 // will get stuck (as they will wake up, see that n_completed() != 383 // n_workers() and go back to sleep). Instead, we raise the 384 // should_reset() flag and the barrier will be reset the first 385 // time a worker enters it again. 386 set_should_reset(true); 387 monitor()->notify_all(); 388 } else { 389 while (n_completed() != n_workers() && !aborted()) { 390 monitor()->wait(/* no_safepoint_check */ true); 391 } 392 } 393 return !aborted(); 394 } 395 396 void WorkGangBarrierSync::abort() { 397 MutexLockerEx x(monitor(), Mutex::_no_safepoint_check_flag); 398 set_aborted(); 399 monitor()->notify_all(); 400 } 401 402 // SubTasksDone functions. 403 404 SubTasksDone::SubTasksDone(uint n) : 405 _n_tasks(n), _tasks(NULL) { 406 _tasks = NEW_C_HEAP_ARRAY(uint, n, mtInternal); 407 guarantee(_tasks != NULL, "alloc failure"); 408 clear(); 409 } 410 411 bool SubTasksDone::valid() { 412 return _tasks != NULL; 413 } 414 415 void SubTasksDone::clear() { 416 for (uint i = 0; i < _n_tasks; i++) { 417 _tasks[i] = 0; 418 } 419 _threads_completed = 0; 420 #ifdef ASSERT 421 _claimed = 0; 422 #endif 423 } 424 425 bool SubTasksDone::is_task_claimed(uint t) { 426 assert(t < _n_tasks, "bad task id."); 427 uint old = _tasks[t]; 428 if (old == 0) { 429 old = Atomic::cmpxchg(1, &_tasks[t], 0); 430 } 431 assert(_tasks[t] == 1, "What else?"); 432 bool res = old != 0; 433 #ifdef ASSERT 434 if (!res) { 435 assert(_claimed < _n_tasks, "Too many tasks claimed; missing clear?"); 436 Atomic::inc((volatile jint*) &_claimed); 437 } 438 #endif 439 return res; 440 } 441 442 void SubTasksDone::all_tasks_completed(uint n_threads) { 443 jint observed = _threads_completed; 444 jint old; 445 do { 446 old = observed; 447 observed = Atomic::cmpxchg(old+1, &_threads_completed, old); 448 } while (observed != old); 449 // If this was the last thread checking in, clear the tasks. 450 uint adjusted_thread_count = (n_threads == 0 ? 1 : n_threads); 451 if (observed + 1 == (jint)adjusted_thread_count) { 452 clear(); 453 } 454 } 455 456 457 SubTasksDone::~SubTasksDone() { 458 if (_tasks != NULL) FREE_C_HEAP_ARRAY(jint, _tasks); 459 } 460 461 // *** SequentialSubTasksDone 462 463 void SequentialSubTasksDone::clear() { 464 _n_tasks = _n_claimed = 0; 465 _n_threads = _n_completed = 0; 466 } 467 468 bool SequentialSubTasksDone::valid() { 469 return _n_threads > 0; 470 } 471 472 bool SequentialSubTasksDone::is_task_claimed(uint& t) { 473 uint* n_claimed_ptr = &_n_claimed; 474 t = *n_claimed_ptr; 475 while (t < _n_tasks) { 476 jint res = Atomic::cmpxchg(t+1, n_claimed_ptr, t); 477 if (res == (jint)t) { 478 return false; 479 } 480 t = *n_claimed_ptr; 481 } 482 return true; 483 } 484 485 bool SequentialSubTasksDone::all_tasks_completed() { 486 uint* n_completed_ptr = &_n_completed; 487 uint complete = *n_completed_ptr; 488 while (true) { 489 uint res = Atomic::cmpxchg(complete+1, n_completed_ptr, complete); 490 if (res == complete) { 491 break; 492 } 493 complete = res; 494 } 495 if (complete+1 == _n_threads) { 496 clear(); 497 return true; 498 } 499 return false; 500 } 501 502 bool FreeIdSet::_stat_init = false; 503 FreeIdSet* FreeIdSet::_sets[NSets]; 504 bool FreeIdSet::_safepoint; 505 506 FreeIdSet::FreeIdSet(int sz, Monitor* mon) : 507 _sz(sz), _mon(mon), _hd(0), _waiters(0), _index(-1), _claimed(0) 508 { 509 _ids = NEW_C_HEAP_ARRAY(int, sz, mtInternal); 510 for (int i = 0; i < sz; i++) _ids[i] = i+1; 511 _ids[sz-1] = end_of_list; // end of list. 512 if (_stat_init) { 513 for (int j = 0; j < NSets; j++) _sets[j] = NULL; 514 _stat_init = true; 515 } 516 // Add to sets. (This should happen while the system is still single-threaded.) 517 for (int j = 0; j < NSets; j++) { 518 if (_sets[j] == NULL) { 519 _sets[j] = this; 520 _index = j; 521 break; 522 } 523 } 524 guarantee(_index != -1, "Too many FreeIdSets in use!"); 525 } 526 527 FreeIdSet::~FreeIdSet() { 528 _sets[_index] = NULL; 529 FREE_C_HEAP_ARRAY(int, _ids); 530 } 531 532 void FreeIdSet::set_safepoint(bool b) { 533 _safepoint = b; 534 if (b) { 535 for (int j = 0; j < NSets; j++) { 536 if (_sets[j] != NULL && _sets[j]->_waiters > 0) { 537 Monitor* mon = _sets[j]->_mon; 538 mon->lock_without_safepoint_check(); 539 mon->notify_all(); 540 mon->unlock(); 541 } 542 } 543 } 544 } 545 546 #define FID_STATS 0 547 548 int FreeIdSet::claim_par_id() { 549 #if FID_STATS 550 thread_t tslf = thr_self(); 551 tty->print("claim_par_id[%d]: sz = %d, claimed = %d\n", tslf, _sz, _claimed); 552 #endif 553 MutexLockerEx x(_mon, Mutex::_no_safepoint_check_flag); 554 while (!_safepoint && _hd == end_of_list) { 555 _waiters++; 556 #if FID_STATS 557 if (_waiters > 5) { 558 tty->print("claim_par_id waiting[%d]: %d waiters, %d claimed.\n", 559 tslf, _waiters, _claimed); 560 } 561 #endif 562 _mon->wait(Mutex::_no_safepoint_check_flag); 563 _waiters--; 564 } 565 if (_hd == end_of_list) { 566 #if FID_STATS 567 tty->print("claim_par_id[%d]: returning EOL.\n", tslf); 568 #endif 569 return -1; 570 } else { 571 int res = _hd; 572 _hd = _ids[res]; 573 _ids[res] = claimed; // For debugging. 574 _claimed++; 575 #if FID_STATS 576 tty->print("claim_par_id[%d]: returning %d, claimed = %d.\n", 577 tslf, res, _claimed); 578 #endif 579 return res; 580 } 581 } 582 583 bool FreeIdSet::claim_perm_id(int i) { 584 assert(0 <= i && i < _sz, "Out of range."); 585 MutexLockerEx x(_mon, Mutex::_no_safepoint_check_flag); 586 int prev = end_of_list; 587 int cur = _hd; 588 while (cur != end_of_list) { 589 if (cur == i) { 590 if (prev == end_of_list) { 591 _hd = _ids[cur]; 592 } else { 593 _ids[prev] = _ids[cur]; 594 } 595 _ids[cur] = claimed; 596 _claimed++; 597 return true; 598 } else { 599 prev = cur; 600 cur = _ids[cur]; 601 } 602 } 603 return false; 604 605 } 606 607 void FreeIdSet::release_par_id(int id) { 608 MutexLockerEx x(_mon, Mutex::_no_safepoint_check_flag); 609 assert(_ids[id] == claimed, "Precondition."); 610 _ids[id] = _hd; 611 _hd = id; 612 _claimed--; 613 #if FID_STATS 614 tty->print("[%d] release_par_id(%d), waiters =%d, claimed = %d.\n", 615 thr_self(), id, _waiters, _claimed); 616 #endif 617 if (_waiters > 0) 618 // Notify all would be safer, but this is OK, right? 619 _mon->notify_all(); 620 }