1 /*
   2  * Copyright (c) 2001, 2014, 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 "memory/allocation.hpp"
  27 #include "memory/allocation.inline.hpp"
  28 #include "runtime/atomic.inline.hpp"
  29 #include "runtime/os.hpp"
  30 #include "utilities/workgroup.hpp"
  31 
  32 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
  33 
  34 // Definitions of WorkGang methods.
  35 
  36 AbstractWorkGang::AbstractWorkGang(const char* name,
  37                                    bool  are_GC_task_threads,
  38                                    bool  are_ConcurrentGC_threads) :
  39   _name(name),
  40   _are_GC_task_threads(are_GC_task_threads),
  41   _are_ConcurrentGC_threads(are_ConcurrentGC_threads) {
  42 
  43   assert(!(are_GC_task_threads && are_ConcurrentGC_threads),
  44          "They cannot both be STW GC and Concurrent threads" );
  45 
  46   // Other initialization.
  47   _monitor = new Monitor(/* priority */       Mutex::leaf,
  48                          /* name */           "WorkGroup monitor",
  49                          /* allow_vm_block */ are_GC_task_threads,
  50                                               Monitor::_safepoint_check_sometimes);
  51   assert(monitor() != NULL, "Failed to allocate monitor");
  52   _terminate = false;
  53   _task = NULL;
  54   _sequence_number = 0;
  55   _started_workers = 0;
  56   _finished_workers = 0;
  57 }
  58 
  59 WorkGang::WorkGang(const char* name,
  60                    uint        workers,
  61                    bool        are_GC_task_threads,
  62                    bool        are_ConcurrentGC_threads) :
  63   AbstractWorkGang(name, are_GC_task_threads, are_ConcurrentGC_threads) {
  64   _total_workers = workers;
  65 }
  66 
  67 GangWorker* WorkGang::allocate_worker(uint which) {
  68   GangWorker* new_worker = new GangWorker(this, which);
  69   return new_worker;
  70 }
  71 
  72 // The current implementation will exit if the allocation
  73 // of any worker fails.  Still, return a boolean so that
  74 // a future implementation can possibly do a partial
  75 // initialization of the workers and report such to the
  76 // caller.
  77 bool WorkGang::initialize_workers() {
  78 
  79   if (TraceWorkGang) {
  80     tty->print_cr("Constructing work gang %s with %d threads",
  81                   name(),
  82                   total_workers());
  83   }
  84   _gang_workers = NEW_C_HEAP_ARRAY(GangWorker*, total_workers(), mtInternal);
  85   if (gang_workers() == NULL) {
  86     vm_exit_out_of_memory(0, OOM_MALLOC_ERROR, "Cannot create GangWorker array.");
  87     return false;
  88   }
  89   os::ThreadType worker_type;
  90   if (are_ConcurrentGC_threads()) {
  91     worker_type = os::cgc_thread;
  92   } else {
  93     worker_type = os::pgc_thread;
  94   }
  95   for (uint worker = 0; worker < total_workers(); worker += 1) {
  96     GangWorker* new_worker = allocate_worker(worker);
  97     assert(new_worker != NULL, "Failed to allocate GangWorker");
  98     _gang_workers[worker] = new_worker;
  99     if (new_worker == NULL || !os::create_thread(new_worker, worker_type)) {
 100       vm_exit_out_of_memory(0, OOM_MALLOC_ERROR,
 101               "Cannot create worker GC thread. Out of system resources.");
 102       return false;
 103     }
 104     if (!DisableStartThread) {
 105       os::start_thread(new_worker);
 106     }
 107   }
 108   return true;
 109 }
 110 
 111 AbstractWorkGang::~AbstractWorkGang() {
 112   if (TraceWorkGang) {
 113     tty->print_cr("Destructing work gang %s", name());
 114   }
 115   stop();   // stop all the workers
 116   for (uint worker = 0; worker < total_workers(); worker += 1) {
 117     delete gang_worker(worker);
 118   }
 119   delete gang_workers();
 120   delete monitor();
 121 }
 122 
 123 GangWorker* AbstractWorkGang::gang_worker(uint i) const {
 124   // Array index bounds checking.
 125   GangWorker* result = NULL;
 126   assert(gang_workers() != NULL, "No workers for indexing");
 127   assert(i < total_workers(), "Worker index out of bounds");
 128   result = _gang_workers[i];
 129   assert(result != NULL, "Indexing to null worker");
 130   return result;
 131 }
 132 
 133 void WorkGang::run_task(AbstractGangTask* task) {
 134   run_task(task, total_workers());
 135 }
 136 
 137 void WorkGang::run_task(AbstractGangTask* task, uint no_of_parallel_workers) {
 138   task->set_for_termination(no_of_parallel_workers);
 139 
 140   // This thread is executed by the VM thread which does not block
 141   // on ordinary MutexLocker's.
 142   MutexLockerEx ml(monitor(), Mutex::_no_safepoint_check_flag);
 143   if (TraceWorkGang) {
 144     tty->print_cr("Running work gang %s task %s", name(), task->name());
 145   }
 146   // Tell all the workers to run a task.
 147   assert(task != NULL, "Running a null task");
 148   // Initialize.
 149   _task = task;
 150   _sequence_number += 1;
 151   _started_workers = 0;
 152   _finished_workers = 0;
 153   // Tell the workers to get to work.
 154   monitor()->notify_all();
 155   // Wait for them to be finished
 156   while (finished_workers() < no_of_parallel_workers) {
 157     if (TraceWorkGang) {
 158       tty->print_cr("Waiting in work gang %s: %d/%d finished sequence %d",
 159                     name(), finished_workers(), no_of_parallel_workers,
 160                     _sequence_number);
 161     }
 162     monitor()->wait(/* no_safepoint_check */ true);
 163   }
 164   _task = NULL;
 165   if (TraceWorkGang) {
 166     tty->print_cr("\nFinished work gang %s: %d/%d sequence %d",
 167                   name(), finished_workers(), no_of_parallel_workers,
 168                   _sequence_number);
 169     Thread* me = Thread::current();
 170     tty->print_cr("  T: 0x%x  VM_thread: %d", me, me->is_VM_thread());
 171   }
 172 }
 173 
 174 void FlexibleWorkGang::run_task(AbstractGangTask* task) {
 175   // If active_workers() is passed, _finished_workers
 176   // must only be incremented for workers that find non_null
 177   // work (as opposed to all those that just check that the
 178   // task is not null).
 179   WorkGang::run_task(task, (uint) active_workers());
 180 }
 181 
 182 void AbstractWorkGang::stop() {
 183   // Tell all workers to terminate, then wait for them to become inactive.
 184   MutexLockerEx ml(monitor(), Mutex::_no_safepoint_check_flag);
 185   if (TraceWorkGang) {
 186     tty->print_cr("Stopping work gang %s task %s", name(), task()->name());
 187   }
 188   _task = NULL;
 189   _terminate = true;
 190   monitor()->notify_all();
 191   while (finished_workers() < active_workers()) {
 192     if (TraceWorkGang) {
 193       tty->print_cr("Waiting in work gang %s: %d/%d finished",
 194                     name(), finished_workers(), active_workers());
 195     }
 196     monitor()->wait(/* no_safepoint_check */ true);
 197   }
 198 }
 199 
 200 void AbstractWorkGang::internal_worker_poll(WorkData* data) const {
 201   assert(monitor()->owned_by_self(), "worker_poll is an internal method");
 202   assert(data != NULL, "worker data is null");
 203   data->set_terminate(terminate());
 204   data->set_task(task());
 205   data->set_sequence_number(sequence_number());
 206 }
 207 
 208 void AbstractWorkGang::internal_note_start() {
 209   assert(monitor()->owned_by_self(), "note_finish is an internal method");
 210   _started_workers += 1;
 211 }
 212 
 213 void AbstractWorkGang::internal_note_finish() {
 214   assert(monitor()->owned_by_self(), "note_finish is an internal method");
 215   _finished_workers += 1;
 216 }
 217 
 218 void AbstractWorkGang::print_worker_threads_on(outputStream* st) const {
 219   uint    num_thr = total_workers();
 220   for (uint i = 0; i < num_thr; i++) {
 221     gang_worker(i)->print_on(st);
 222     st->cr();
 223   }
 224 }
 225 
 226 void AbstractWorkGang::threads_do(ThreadClosure* tc) const {
 227   assert(tc != NULL, "Null ThreadClosure");
 228   uint num_thr = total_workers();
 229   for (uint i = 0; i < num_thr; i++) {
 230     tc->do_thread(gang_worker(i));
 231   }
 232 }
 233 
 234 // GangWorker methods.
 235 
 236 GangWorker::GangWorker(AbstractWorkGang* gang, uint id) {
 237   _gang = gang;
 238   set_id(id);
 239   set_name("Gang worker#%d (%s)", id, gang->name());
 240 }
 241 
 242 void GangWorker::run() {
 243   initialize();
 244   loop();
 245 }
 246 
 247 void GangWorker::initialize() {
 248   this->initialize_thread_local_storage();
 249   this->record_stack_base_and_size();
 250   assert(_gang != NULL, "No gang to run in");
 251   os::set_priority(this, NearMaxPriority);
 252   if (TraceWorkGang) {
 253     tty->print_cr("Running gang worker for gang %s id %d",
 254                   gang()->name(), id());
 255   }
 256   // The VM thread should not execute here because MutexLocker's are used
 257   // as (opposed to MutexLockerEx's).
 258   assert(!Thread::current()->is_VM_thread(), "VM thread should not be part"
 259          " of a work gang");
 260 }
 261 
 262 void GangWorker::loop() {
 263   int previous_sequence_number = 0;
 264   Monitor* gang_monitor = gang()->monitor();
 265   for ( ; /* !terminate() */; ) {
 266     WorkData data;
 267     int part;  // Initialized below.
 268     {
 269       // Grab the gang mutex.
 270       MutexLocker ml(gang_monitor);
 271       // Wait for something to do.
 272       // Polling outside the while { wait } avoids missed notifies
 273       // in the outer loop.
 274       gang()->internal_worker_poll(&data);
 275       if (TraceWorkGang) {
 276         tty->print("Polled outside for work in gang %s worker %d",
 277                    gang()->name(), id());
 278         tty->print("  terminate: %s",
 279                    data.terminate() ? "true" : "false");
 280         tty->print("  sequence: %d (prev: %d)",
 281                    data.sequence_number(), previous_sequence_number);
 282         if (data.task() != NULL) {
 283           tty->print("  task: %s", data.task()->name());
 284         } else {
 285           tty->print("  task: NULL");
 286         }
 287         tty->cr();
 288       }
 289       for ( ; /* break or return */; ) {
 290         // Terminate if requested.
 291         if (data.terminate()) {
 292           gang()->internal_note_finish();
 293           gang_monitor->notify_all();
 294           return;
 295         }
 296         // Check for new work.
 297         if ((data.task() != NULL) &&
 298             (data.sequence_number() != previous_sequence_number)) {
 299           if (gang()->needs_more_workers()) {
 300             gang()->internal_note_start();
 301             gang_monitor->notify_all();
 302             part = gang()->started_workers() - 1;
 303             break;
 304           }
 305         }
 306         // Nothing to do.
 307         gang_monitor->wait(/* no_safepoint_check */ true);
 308         gang()->internal_worker_poll(&data);
 309         if (TraceWorkGang) {
 310           tty->print("Polled inside for work in gang %s worker %d",
 311                      gang()->name(), id());
 312           tty->print("  terminate: %s",
 313                      data.terminate() ? "true" : "false");
 314           tty->print("  sequence: %d (prev: %d)",
 315                      data.sequence_number(), previous_sequence_number);
 316           if (data.task() != NULL) {
 317             tty->print("  task: %s", data.task()->name());
 318           } else {
 319             tty->print("  task: NULL");
 320           }
 321           tty->cr();
 322         }
 323       }
 324       // Drop gang mutex.
 325     }
 326     if (TraceWorkGang) {
 327       tty->print("Work for work gang %s id %d task %s part %d",
 328                  gang()->name(), id(), data.task()->name(), part);
 329     }
 330     assert(data.task() != NULL, "Got null task");
 331     data.task()->work(part);
 332     {
 333       if (TraceWorkGang) {
 334         tty->print("Finish for work gang %s id %d task %s part %d",
 335                    gang()->name(), id(), data.task()->name(), part);
 336       }
 337       // Grab the gang mutex.
 338       MutexLocker ml(gang_monitor);
 339       gang()->internal_note_finish();
 340       // Tell the gang you are done.
 341       gang_monitor->notify_all();
 342       // Drop the gang mutex.
 343     }
 344     previous_sequence_number = data.sequence_number();
 345   }
 346 }
 347 
 348 bool GangWorker::is_GC_task_thread() const {
 349   return gang()->are_GC_task_threads();
 350 }
 351 
 352 bool GangWorker::is_ConcurrentGC_thread() const {
 353   return gang()->are_ConcurrentGC_threads();
 354 }
 355 
 356 void GangWorker::print_on(outputStream* st) const {
 357   st->print("\"%s\" ", name());
 358   Thread::print_on(st);
 359   st->cr();
 360 }
 361 
 362 // Printing methods
 363 
 364 const char* AbstractWorkGang::name() const {
 365   return _name;
 366 }
 367 
 368 #ifndef PRODUCT
 369 
 370 const char* AbstractGangTask::name() const {
 371   return _name;
 372 }
 373 
 374 #endif /* PRODUCT */
 375 
 376 // FlexibleWorkGang
 377 
 378 
 379 // *** WorkGangBarrierSync
 380 
 381 WorkGangBarrierSync::WorkGangBarrierSync()
 382   : _monitor(Mutex::safepoint, "work gang barrier sync", true,
 383              Monitor::_safepoint_check_never),
 384     _n_workers(0), _n_completed(0), _should_reset(false), _aborted(false) {
 385 }
 386 
 387 WorkGangBarrierSync::WorkGangBarrierSync(uint n_workers, const char* name)
 388   : _monitor(Mutex::safepoint, name, true, Monitor::_safepoint_check_never),
 389     _n_workers(n_workers), _n_completed(0), _should_reset(false), _aborted(false) {
 390 }
 391 
 392 void WorkGangBarrierSync::set_n_workers(uint n_workers) {
 393   _n_workers    = n_workers;
 394   _n_completed  = 0;
 395   _should_reset = false;
 396   _aborted      = false;
 397 }
 398 
 399 bool WorkGangBarrierSync::enter() {
 400   MutexLockerEx x(monitor(), Mutex::_no_safepoint_check_flag);
 401   if (should_reset()) {
 402     // The should_reset() was set and we are the first worker to enter
 403     // the sync barrier. We will zero the n_completed() count which
 404     // effectively resets the barrier.
 405     zero_completed();
 406     set_should_reset(false);
 407   }
 408   inc_completed();
 409   if (n_completed() == n_workers()) {
 410     // At this point we would like to reset the barrier to be ready in
 411     // case it is used again. However, we cannot set n_completed() to
 412     // 0, even after the notify_all(), given that some other workers
 413     // might still be waiting for n_completed() to become ==
 414     // n_workers(). So, if we set n_completed() to 0, those workers
 415     // will get stuck (as they will wake up, see that n_completed() !=
 416     // n_workers() and go back to sleep). Instead, we raise the
 417     // should_reset() flag and the barrier will be reset the first
 418     // time a worker enters it again.
 419     set_should_reset(true);
 420     monitor()->notify_all();
 421   } else {
 422     while (n_completed() != n_workers() && !aborted()) {
 423       monitor()->wait(/* no_safepoint_check */ true);
 424     }
 425   }
 426   return !aborted();
 427 }
 428 
 429 void WorkGangBarrierSync::abort() {
 430   MutexLockerEx x(monitor(), Mutex::_no_safepoint_check_flag);
 431   set_aborted();
 432   monitor()->notify_all();
 433 }
 434 
 435 // SubTasksDone functions.
 436 
 437 SubTasksDone::SubTasksDone(uint n) :
 438   _n_tasks(n), _n_threads(1), _tasks(NULL) {
 439   _tasks = NEW_C_HEAP_ARRAY(uint, n, mtInternal);
 440   guarantee(_tasks != NULL, "alloc failure");
 441   clear();
 442 }
 443 
 444 bool SubTasksDone::valid() {
 445   return _tasks != NULL;
 446 }
 447 
 448 void SubTasksDone::set_n_threads(uint t) {
 449   assert(_claimed == 0 || _threads_completed == _n_threads,
 450          "should not be called while tasks are being processed!");
 451   _n_threads = (t == 0 ? 1 : t);
 452 }
 453 
 454 void SubTasksDone::clear() {
 455   for (uint i = 0; i < _n_tasks; i++) {
 456     _tasks[i] = 0;
 457   }
 458   _threads_completed = 0;
 459 #ifdef ASSERT
 460   _claimed = 0;
 461 #endif
 462 }
 463 
 464 bool SubTasksDone::is_task_claimed(uint t) {
 465   assert(t < _n_tasks, "bad task id.");
 466   uint old = _tasks[t];
 467   if (old == 0) {
 468     old = Atomic::cmpxchg(1, &_tasks[t], 0);
 469   }
 470   assert(_tasks[t] == 1, "What else?");
 471   bool res = old != 0;
 472 #ifdef ASSERT
 473   if (!res) {
 474     assert(_claimed < _n_tasks, "Too many tasks claimed; missing clear?");
 475     Atomic::inc((volatile jint*) &_claimed);
 476   }
 477 #endif
 478   return res;
 479 }
 480 
 481 void SubTasksDone::all_tasks_completed() {
 482   jint observed = _threads_completed;
 483   jint old;
 484   do {
 485     old = observed;
 486     observed = Atomic::cmpxchg(old+1, &_threads_completed, old);
 487   } while (observed != old);
 488   // If this was the last thread checking in, clear the tasks.
 489   if (observed+1 == (jint)_n_threads) clear();
 490 }
 491 
 492 
 493 SubTasksDone::~SubTasksDone() {
 494   if (_tasks != NULL) FREE_C_HEAP_ARRAY(jint, _tasks);
 495 }
 496 
 497 // *** SequentialSubTasksDone
 498 
 499 void SequentialSubTasksDone::clear() {
 500   _n_tasks   = _n_claimed   = 0;
 501   _n_threads = _n_completed = 0;
 502 }
 503 
 504 bool SequentialSubTasksDone::valid() {
 505   return _n_threads > 0;
 506 }
 507 
 508 bool SequentialSubTasksDone::is_task_claimed(uint& t) {
 509   uint* n_claimed_ptr = &_n_claimed;
 510   t = *n_claimed_ptr;
 511   while (t < _n_tasks) {
 512     jint res = Atomic::cmpxchg(t+1, n_claimed_ptr, t);
 513     if (res == (jint)t) {
 514       return false;
 515     }
 516     t = *n_claimed_ptr;
 517   }
 518   return true;
 519 }
 520 
 521 bool SequentialSubTasksDone::all_tasks_completed() {
 522   uint* n_completed_ptr = &_n_completed;
 523   uint  complete        = *n_completed_ptr;
 524   while (true) {
 525     uint res = Atomic::cmpxchg(complete+1, n_completed_ptr, complete);
 526     if (res == complete) {
 527       break;
 528     }
 529     complete = res;
 530   }
 531   if (complete+1 == _n_threads) {
 532     clear();
 533     return true;
 534   }
 535   return false;
 536 }
 537 
 538 bool FreeIdSet::_stat_init = false;
 539 FreeIdSet* FreeIdSet::_sets[NSets];
 540 bool FreeIdSet::_safepoint;
 541 
 542 FreeIdSet::FreeIdSet(int sz, Monitor* mon) :
 543   _sz(sz), _mon(mon), _hd(0), _waiters(0), _index(-1), _claimed(0)
 544 {
 545   _ids = NEW_C_HEAP_ARRAY(int, sz, mtInternal);
 546   for (int i = 0; i < sz; i++) _ids[i] = i+1;
 547   _ids[sz-1] = end_of_list; // end of list.
 548   if (_stat_init) {
 549     for (int j = 0; j < NSets; j++) _sets[j] = NULL;
 550     _stat_init = true;
 551   }
 552   // Add to sets.  (This should happen while the system is still single-threaded.)
 553   for (int j = 0; j < NSets; j++) {
 554     if (_sets[j] == NULL) {
 555       _sets[j] = this;
 556       _index = j;
 557       break;
 558     }
 559   }
 560   guarantee(_index != -1, "Too many FreeIdSets in use!");
 561 }
 562 
 563 FreeIdSet::~FreeIdSet() {
 564   _sets[_index] = NULL;
 565   FREE_C_HEAP_ARRAY(int, _ids);
 566 }
 567 
 568 void FreeIdSet::set_safepoint(bool b) {
 569   _safepoint = b;
 570   if (b) {
 571     for (int j = 0; j < NSets; j++) {
 572       if (_sets[j] != NULL && _sets[j]->_waiters > 0) {
 573         Monitor* mon = _sets[j]->_mon;
 574         mon->lock_without_safepoint_check();
 575         mon->notify_all();
 576         mon->unlock();
 577       }
 578     }
 579   }
 580 }
 581 
 582 #define FID_STATS 0
 583 
 584 int FreeIdSet::claim_par_id() {
 585 #if FID_STATS
 586   thread_t tslf = thr_self();
 587   tty->print("claim_par_id[%d]: sz = %d, claimed = %d\n", tslf, _sz, _claimed);
 588 #endif
 589   MutexLockerEx x(_mon, Mutex::_no_safepoint_check_flag);
 590   while (!_safepoint && _hd == end_of_list) {
 591     _waiters++;
 592 #if FID_STATS
 593     if (_waiters > 5) {
 594       tty->print("claim_par_id waiting[%d]: %d waiters, %d claimed.\n",
 595                  tslf, _waiters, _claimed);
 596     }
 597 #endif
 598     _mon->wait(Mutex::_no_safepoint_check_flag);
 599     _waiters--;
 600   }
 601   if (_hd == end_of_list) {
 602 #if FID_STATS
 603     tty->print("claim_par_id[%d]: returning EOL.\n", tslf);
 604 #endif
 605     return -1;
 606   } else {
 607     int res = _hd;
 608     _hd = _ids[res];
 609     _ids[res] = claimed;  // For debugging.
 610     _claimed++;
 611 #if FID_STATS
 612     tty->print("claim_par_id[%d]: returning %d, claimed = %d.\n",
 613                tslf, res, _claimed);
 614 #endif
 615     return res;
 616   }
 617 }
 618 
 619 bool FreeIdSet::claim_perm_id(int i) {
 620   assert(0 <= i && i < _sz, "Out of range.");
 621   MutexLockerEx x(_mon, Mutex::_no_safepoint_check_flag);
 622   int prev = end_of_list;
 623   int cur = _hd;
 624   while (cur != end_of_list) {
 625     if (cur == i) {
 626       if (prev == end_of_list) {
 627         _hd = _ids[cur];
 628       } else {
 629         _ids[prev] = _ids[cur];
 630       }
 631       _ids[cur] = claimed;
 632       _claimed++;
 633       return true;
 634     } else {
 635       prev = cur;
 636       cur = _ids[cur];
 637     }
 638   }
 639   return false;
 640 
 641 }
 642 
 643 void FreeIdSet::release_par_id(int id) {
 644   MutexLockerEx x(_mon, Mutex::_no_safepoint_check_flag);
 645   assert(_ids[id] == claimed, "Precondition.");
 646   _ids[id] = _hd;
 647   _hd = id;
 648   _claimed--;
 649 #if FID_STATS
 650   tty->print("[%d] release_par_id(%d), waiters =%d,  claimed = %d.\n",
 651              thr_self(), id, _waiters, _claimed);
 652 #endif
 653   if (_waiters > 0)
 654     // Notify all would be safer, but this is OK, right?
 655     _mon->notify_all();
 656 }