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 32 // Definitions of WorkGang methods. 33 34 AbstractWorkGang::AbstractWorkGang(const char* name, 35 bool are_GC_task_threads, 36 bool are_ConcurrentGC_threads) : 37 _name(name), 38 _are_GC_task_threads(are_GC_task_threads), 39 _are_ConcurrentGC_threads(are_ConcurrentGC_threads) { 40 41 assert(!(are_GC_task_threads && are_ConcurrentGC_threads), 42 "They cannot both be STW GC and Concurrent threads" ); 43 44 // Other initialization. 45 _monitor = new Monitor(/* priority */ Mutex::leaf, 46 /* name */ "WorkGroup monitor", 47 /* allow_vm_block */ are_GC_task_threads, 48 Monitor::_safepoint_check_sometimes); 49 assert(monitor() != NULL, "Failed to allocate monitor"); 50 _task = NULL; 51 _sequence_number = 0; 52 _started_workers = 0; 53 _finished_workers = 0; 54 } 55 56 WorkGang::WorkGang(const char* name, 57 uint workers, 58 bool are_GC_task_threads, 59 bool are_ConcurrentGC_threads) : 60 AbstractWorkGang(name, are_GC_task_threads, are_ConcurrentGC_threads) { 61 _total_workers = workers; 62 } 63 64 GangWorker* WorkGang::allocate_worker(uint which) { 65 GangWorker* new_worker = new GangWorker(this, which); 66 return new_worker; 67 } 68 69 // The current implementation will exit if the allocation 70 // of any worker fails. Still, return a boolean so that 71 // a future implementation can possibly do a partial 72 // initialization of the workers and report such to the 73 // caller. 74 bool WorkGang::initialize_workers() { 75 76 if (TraceWorkGang) { 77 tty->print_cr("Constructing work gang %s with %d threads", 78 name(), 79 total_workers()); 80 } 81 _gang_workers = NEW_C_HEAP_ARRAY(GangWorker*, total_workers(), mtInternal); 82 if (gang_workers() == NULL) { 83 vm_exit_out_of_memory(0, OOM_MALLOC_ERROR, "Cannot create GangWorker array."); 84 return false; 85 } 86 os::ThreadType worker_type; 87 if (are_ConcurrentGC_threads()) { 88 worker_type = os::cgc_thread; 89 } else { 90 worker_type = os::pgc_thread; 91 } 92 for (uint worker = 0; worker < total_workers(); worker += 1) { 93 GangWorker* new_worker = allocate_worker(worker); 94 assert(new_worker != NULL, "Failed to allocate GangWorker"); 95 _gang_workers[worker] = new_worker; 96 if (new_worker == NULL || !os::create_thread(new_worker, worker_type)) { 97 vm_exit_out_of_memory(0, OOM_MALLOC_ERROR, 98 "Cannot create worker GC thread. Out of system resources."); 99 return false; 100 } 101 if (!DisableStartThread) { 102 os::start_thread(new_worker); 103 } 104 } 105 return true; 106 } 107 108 GangWorker* AbstractWorkGang::gang_worker(uint i) const { 109 // Array index bounds checking. 110 GangWorker* result = NULL; 111 assert(gang_workers() != NULL, "No workers for indexing"); 112 assert(i < total_workers(), "Worker index out of bounds"); 113 result = _gang_workers[i]; 114 assert(result != NULL, "Indexing to null worker"); 115 return result; 116 } 117 118 void WorkGang::run_task(AbstractGangTask* task) { 119 run_task(task, total_workers()); 120 } 121 122 void WorkGang::run_task(AbstractGangTask* task, uint no_of_parallel_workers) { 123 // This thread is executed by the VM thread which does not block 124 // on ordinary MutexLocker's. 125 MutexLockerEx ml(monitor(), Mutex::_no_safepoint_check_flag); 126 if (TraceWorkGang) { 127 tty->print_cr("Running work gang %s task %s", name(), task->name()); 128 } 129 // Tell all the workers to run a task. 130 assert(task != NULL, "Running a null task"); 131 // Initialize. 132 _task = task; 133 _sequence_number += 1; 134 _started_workers = 0; 135 _finished_workers = 0; 136 // Tell the workers to get to work. 137 monitor()->notify_all(); 138 // Wait for them to be finished 139 while (finished_workers() < no_of_parallel_workers) { 140 if (TraceWorkGang) { 141 tty->print_cr("Waiting in work gang %s: %u/%u finished sequence %d", 142 name(), finished_workers(), no_of_parallel_workers, 143 _sequence_number); 144 } 145 monitor()->wait(/* no_safepoint_check */ true); 146 } 147 _task = NULL; 148 if (TraceWorkGang) { 149 tty->print_cr("\nFinished work gang %s: %u/%u sequence %d", 150 name(), finished_workers(), no_of_parallel_workers, 151 _sequence_number); 152 Thread* me = Thread::current(); 153 tty->print_cr(" T: " PTR_FORMAT " VM_thread: %d", p2i(me), me->is_VM_thread()); 154 } 155 } 156 157 void FlexibleWorkGang::run_task(AbstractGangTask* task) { 158 // If active_workers() is passed, _finished_workers 159 // must only be incremented for workers that find non_null 160 // work (as opposed to all those that just check that the 161 // task is not null). 162 WorkGang::run_task(task, (uint) active_workers()); 163 } 164 165 void AbstractWorkGang::internal_worker_poll(WorkData* data) const { 166 assert(monitor()->owned_by_self(), "worker_poll is an internal method"); 167 assert(data != NULL, "worker data is null"); 168 data->set_task(task()); 169 data->set_sequence_number(sequence_number()); 170 } 171 172 void AbstractWorkGang::internal_note_start() { 173 assert(monitor()->owned_by_self(), "note_finish is an internal method"); 174 _started_workers += 1; 175 } 176 177 void AbstractWorkGang::internal_note_finish() { 178 assert(monitor()->owned_by_self(), "note_finish is an internal method"); 179 _finished_workers += 1; 180 } 181 182 void AbstractWorkGang::print_worker_threads_on(outputStream* st) const { 183 uint num_thr = total_workers(); 184 for (uint i = 0; i < num_thr; i++) { 185 gang_worker(i)->print_on(st); 186 st->cr(); 187 } 188 } 189 190 void AbstractWorkGang::threads_do(ThreadClosure* tc) const { 191 assert(tc != NULL, "Null ThreadClosure"); 192 uint num_thr = total_workers(); 193 for (uint i = 0; i < num_thr; i++) { 194 tc->do_thread(gang_worker(i)); 195 } 196 } 197 198 // GangWorker methods. 199 200 GangWorker::GangWorker(AbstractWorkGang* gang, uint id) { 201 _gang = gang; 202 set_id(id); 203 set_name("%s#%d", gang->name(), id); 204 } 205 206 void GangWorker::run() { 207 initialize(); 208 loop(); 209 } 210 211 void GangWorker::initialize() { 212 this->initialize_thread_local_storage(); 213 this->record_stack_base_and_size(); 214 this->initialize_named_thread(); 215 assert(_gang != NULL, "No gang to run in"); 216 os::set_priority(this, NearMaxPriority); 217 if (TraceWorkGang) { 218 tty->print_cr("Running gang worker for gang %s id %u", 219 gang()->name(), id()); 220 } 221 // The VM thread should not execute here because MutexLocker's are used 222 // as (opposed to MutexLockerEx's). 223 assert(!Thread::current()->is_VM_thread(), "VM thread should not be part" 224 " of a work gang"); 225 } 226 227 void GangWorker::loop() { 228 int previous_sequence_number = 0; 229 Monitor* gang_monitor = gang()->monitor(); 230 for ( ; ; ) { 231 WorkData data; 232 int part; // Initialized below. 233 { 234 // Grab the gang mutex. 235 MutexLocker ml(gang_monitor); 236 // Wait for something to do. 237 // Polling outside the while { wait } avoids missed notifies 238 // in the outer loop. 239 gang()->internal_worker_poll(&data); 240 if (TraceWorkGang) { 241 tty->print("Polled outside for work in gang %s worker %u", 242 gang()->name(), id()); 243 tty->print(" sequence: %d (prev: %d)", 244 data.sequence_number(), previous_sequence_number); 245 if (data.task() != NULL) { 246 tty->print(" task: %s", data.task()->name()); 247 } else { 248 tty->print(" task: NULL"); 249 } 250 tty->cr(); 251 } 252 for ( ; /* break */; ) { 253 // Check for new work. 254 if ((data.task() != NULL) && 255 (data.sequence_number() != previous_sequence_number)) { 256 if (gang()->needs_more_workers()) { 257 gang()->internal_note_start(); 258 gang_monitor->notify_all(); 259 part = gang()->started_workers() - 1; 260 break; 261 } 262 } 263 // Nothing to do. 264 gang_monitor->wait(/* no_safepoint_check */ true); 265 gang()->internal_worker_poll(&data); 266 if (TraceWorkGang) { 267 tty->print("Polled inside for work in gang %s worker %u", 268 gang()->name(), id()); 269 tty->print(" sequence: %d (prev: %d)", 270 data.sequence_number(), previous_sequence_number); 271 if (data.task() != NULL) { 272 tty->print(" task: %s", data.task()->name()); 273 } else { 274 tty->print(" task: NULL"); 275 } 276 tty->cr(); 277 } 278 } 279 // Drop gang mutex. 280 } 281 if (TraceWorkGang) { 282 tty->print("Work for work gang %s id %u task %s part %d", 283 gang()->name(), id(), data.task()->name(), part); 284 } 285 assert(data.task() != NULL, "Got null task"); 286 data.task()->work(part); 287 { 288 if (TraceWorkGang) { 289 tty->print("Finish for work gang %s id %u task %s part %d", 290 gang()->name(), id(), data.task()->name(), part); 291 } 292 // Grab the gang mutex. 293 MutexLocker ml(gang_monitor); 294 gang()->internal_note_finish(); 295 // Tell the gang you are done. 296 gang_monitor->notify_all(); 297 // Drop the gang mutex. 298 } 299 previous_sequence_number = data.sequence_number(); 300 } 301 } 302 303 bool GangWorker::is_GC_task_thread() const { 304 return gang()->are_GC_task_threads(); 305 } 306 307 bool GangWorker::is_ConcurrentGC_thread() const { 308 return gang()->are_ConcurrentGC_threads(); 309 } 310 311 void GangWorker::print_on(outputStream* st) const { 312 st->print("\"%s\" ", name()); 313 Thread::print_on(st); 314 st->cr(); 315 } 316 317 // Printing methods 318 319 const char* AbstractWorkGang::name() const { 320 return _name; 321 } 322 323 #ifndef PRODUCT 324 325 const char* AbstractGangTask::name() const { 326 return _name; 327 } 328 329 #endif /* PRODUCT */ 330 331 // FlexibleWorkGang 332 333 334 // *** WorkGangBarrierSync 335 336 WorkGangBarrierSync::WorkGangBarrierSync() 337 : _monitor(Mutex::safepoint, "work gang barrier sync", true, 338 Monitor::_safepoint_check_never), 339 _n_workers(0), _n_completed(0), _should_reset(false), _aborted(false) { 340 } 341 342 WorkGangBarrierSync::WorkGangBarrierSync(uint n_workers, const char* name) 343 : _monitor(Mutex::safepoint, name, true, Monitor::_safepoint_check_never), 344 _n_workers(n_workers), _n_completed(0), _should_reset(false), _aborted(false) { 345 } 346 347 void WorkGangBarrierSync::set_n_workers(uint n_workers) { 348 _n_workers = n_workers; 349 _n_completed = 0; 350 _should_reset = false; 351 _aborted = false; 352 } 353 354 bool WorkGangBarrierSync::enter() { 355 MutexLockerEx x(monitor(), Mutex::_no_safepoint_check_flag); 356 if (should_reset()) { 357 // The should_reset() was set and we are the first worker to enter 358 // the sync barrier. We will zero the n_completed() count which 359 // effectively resets the barrier. 360 zero_completed(); 361 set_should_reset(false); 362 } 363 inc_completed(); 364 if (n_completed() == n_workers()) { 365 // At this point we would like to reset the barrier to be ready in 366 // case it is used again. However, we cannot set n_completed() to 367 // 0, even after the notify_all(), given that some other workers 368 // might still be waiting for n_completed() to become == 369 // n_workers(). So, if we set n_completed() to 0, those workers 370 // will get stuck (as they will wake up, see that n_completed() != 371 // n_workers() and go back to sleep). Instead, we raise the 372 // should_reset() flag and the barrier will be reset the first 373 // time a worker enters it again. 374 set_should_reset(true); 375 monitor()->notify_all(); 376 } else { 377 while (n_completed() != n_workers() && !aborted()) { 378 monitor()->wait(/* no_safepoint_check */ true); 379 } 380 } 381 return !aborted(); 382 } 383 384 void WorkGangBarrierSync::abort() { 385 MutexLockerEx x(monitor(), Mutex::_no_safepoint_check_flag); 386 set_aborted(); 387 monitor()->notify_all(); 388 } 389 390 // SubTasksDone functions. 391 392 SubTasksDone::SubTasksDone(uint n) : 393 _n_tasks(n), _tasks(NULL) { 394 _tasks = NEW_C_HEAP_ARRAY(uint, n, mtInternal); 395 guarantee(_tasks != NULL, "alloc failure"); 396 clear(); 397 } 398 399 bool SubTasksDone::valid() { 400 return _tasks != NULL; 401 } 402 403 void SubTasksDone::clear() { 404 for (uint i = 0; i < _n_tasks; i++) { 405 _tasks[i] = 0; 406 } 407 _threads_completed = 0; 408 #ifdef ASSERT 409 _claimed = 0; 410 #endif 411 } 412 413 bool SubTasksDone::is_task_claimed(uint t) { 414 assert(t < _n_tasks, "bad task id."); 415 uint old = _tasks[t]; 416 if (old == 0) { 417 old = Atomic::cmpxchg(1, &_tasks[t], 0); 418 } 419 assert(_tasks[t] == 1, "What else?"); 420 bool res = old != 0; 421 #ifdef ASSERT 422 if (!res) { 423 assert(_claimed < _n_tasks, "Too many tasks claimed; missing clear?"); 424 Atomic::inc((volatile jint*) &_claimed); 425 } 426 #endif 427 return res; 428 } 429 430 void SubTasksDone::all_tasks_completed(uint n_threads) { 431 jint observed = _threads_completed; 432 jint old; 433 do { 434 old = observed; 435 observed = Atomic::cmpxchg(old+1, &_threads_completed, old); 436 } while (observed != old); 437 // If this was the last thread checking in, clear the tasks. 438 uint adjusted_thread_count = (n_threads == 0 ? 1 : n_threads); 439 if (observed + 1 == (jint)adjusted_thread_count) { 440 clear(); 441 } 442 } 443 444 445 SubTasksDone::~SubTasksDone() { 446 if (_tasks != NULL) FREE_C_HEAP_ARRAY(jint, _tasks); 447 } 448 449 // *** SequentialSubTasksDone 450 451 void SequentialSubTasksDone::clear() { 452 _n_tasks = _n_claimed = 0; 453 _n_threads = _n_completed = 0; 454 } 455 456 bool SequentialSubTasksDone::valid() { 457 return _n_threads > 0; 458 } 459 460 bool SequentialSubTasksDone::is_task_claimed(uint& t) { 461 uint* n_claimed_ptr = &_n_claimed; 462 t = *n_claimed_ptr; 463 while (t < _n_tasks) { 464 jint res = Atomic::cmpxchg(t+1, n_claimed_ptr, t); 465 if (res == (jint)t) { 466 return false; 467 } 468 t = *n_claimed_ptr; 469 } 470 return true; 471 } 472 473 bool SequentialSubTasksDone::all_tasks_completed() { 474 uint* n_completed_ptr = &_n_completed; 475 uint complete = *n_completed_ptr; 476 while (true) { 477 uint res = Atomic::cmpxchg(complete+1, n_completed_ptr, complete); 478 if (res == complete) { 479 break; 480 } 481 complete = res; 482 } 483 if (complete+1 == _n_threads) { 484 clear(); 485 return true; 486 } 487 return false; 488 } 489 490 bool FreeIdSet::_stat_init = false; 491 FreeIdSet* FreeIdSet::_sets[NSets]; 492 bool FreeIdSet::_safepoint; 493 494 FreeIdSet::FreeIdSet(int sz, Monitor* mon) : 495 _sz(sz), _mon(mon), _hd(0), _waiters(0), _index(-1), _claimed(0) 496 { 497 _ids = NEW_C_HEAP_ARRAY(int, sz, mtInternal); 498 for (int i = 0; i < sz; i++) _ids[i] = i+1; 499 _ids[sz-1] = end_of_list; // end of list. 500 if (_stat_init) { 501 for (int j = 0; j < NSets; j++) _sets[j] = NULL; 502 _stat_init = true; 503 } 504 // Add to sets. (This should happen while the system is still single-threaded.) 505 for (int j = 0; j < NSets; j++) { 506 if (_sets[j] == NULL) { 507 _sets[j] = this; 508 _index = j; 509 break; 510 } 511 } 512 guarantee(_index != -1, "Too many FreeIdSets in use!"); 513 } 514 515 FreeIdSet::~FreeIdSet() { 516 _sets[_index] = NULL; 517 FREE_C_HEAP_ARRAY(int, _ids); 518 } 519 520 void FreeIdSet::set_safepoint(bool b) { 521 _safepoint = b; 522 if (b) { 523 for (int j = 0; j < NSets; j++) { 524 if (_sets[j] != NULL && _sets[j]->_waiters > 0) { 525 Monitor* mon = _sets[j]->_mon; 526 mon->lock_without_safepoint_check(); 527 mon->notify_all(); 528 mon->unlock(); 529 } 530 } 531 } 532 } 533 534 #define FID_STATS 0 535 536 int FreeIdSet::claim_par_id() { 537 #if FID_STATS 538 thread_t tslf = thr_self(); 539 tty->print("claim_par_id[%d]: sz = %d, claimed = %d\n", tslf, _sz, _claimed); 540 #endif 541 MutexLockerEx x(_mon, Mutex::_no_safepoint_check_flag); 542 while (!_safepoint && _hd == end_of_list) { 543 _waiters++; 544 #if FID_STATS 545 if (_waiters > 5) { 546 tty->print("claim_par_id waiting[%d]: %d waiters, %d claimed.\n", 547 tslf, _waiters, _claimed); 548 } 549 #endif 550 _mon->wait(Mutex::_no_safepoint_check_flag); 551 _waiters--; 552 } 553 if (_hd == end_of_list) { 554 #if FID_STATS 555 tty->print("claim_par_id[%d]: returning EOL.\n", tslf); 556 #endif 557 return -1; 558 } else { 559 int res = _hd; 560 _hd = _ids[res]; 561 _ids[res] = claimed; // For debugging. 562 _claimed++; 563 #if FID_STATS 564 tty->print("claim_par_id[%d]: returning %d, claimed = %d.\n", 565 tslf, res, _claimed); 566 #endif 567 return res; 568 } 569 } 570 571 bool FreeIdSet::claim_perm_id(int i) { 572 assert(0 <= i && i < _sz, "Out of range."); 573 MutexLockerEx x(_mon, Mutex::_no_safepoint_check_flag); 574 int prev = end_of_list; 575 int cur = _hd; 576 while (cur != end_of_list) { 577 if (cur == i) { 578 if (prev == end_of_list) { 579 _hd = _ids[cur]; 580 } else { 581 _ids[prev] = _ids[cur]; 582 } 583 _ids[cur] = claimed; 584 _claimed++; 585 return true; 586 } else { 587 prev = cur; 588 cur = _ids[cur]; 589 } 590 } 591 return false; 592 593 } 594 595 void FreeIdSet::release_par_id(int id) { 596 MutexLockerEx x(_mon, Mutex::_no_safepoint_check_flag); 597 assert(_ids[id] == claimed, "Precondition."); 598 _ids[id] = _hd; 599 _hd = id; 600 _claimed--; 601 #if FID_STATS 602 tty->print("[%d] release_par_id(%d), waiters =%d, claimed = %d.\n", 603 thr_self(), id, _waiters, _claimed); 604 #endif 605 if (_waiters > 0) 606 // Notify all would be safer, but this is OK, right? 607 _mon->notify_all(); 608 }