1 /*
   2  * Copyright (c) 1998, 2018, 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 "classfile/vmSymbols.hpp"
  27 #include "jfr/jfrEvents.hpp"
  28 #include "jfr/support/jfrThreadId.hpp"
  29 #include "memory/allocation.inline.hpp"
  30 #include "memory/resourceArea.hpp"
  31 #include "oops/markOop.hpp"
  32 #include "oops/oop.inline.hpp"
  33 #include "runtime/atomic.hpp"
  34 #include "runtime/handles.inline.hpp"
  35 #include "runtime/interfaceSupport.inline.hpp"
  36 #include "runtime/mutexLocker.hpp"
  37 #include "runtime/objectMonitor.hpp"
  38 #include "runtime/objectMonitor.inline.hpp"
  39 #include "runtime/orderAccess.hpp"
  40 #include "runtime/osThread.hpp"
  41 #include "runtime/safepointMechanism.inline.hpp"
  42 #include "runtime/sharedRuntime.hpp"
  43 #include "runtime/stubRoutines.hpp"
  44 #include "runtime/thread.inline.hpp"
  45 #include "services/threadService.hpp"
  46 #include "utilities/dtrace.hpp"
  47 #include "utilities/macros.hpp"
  48 #include "utilities/preserveException.hpp"
  49 #if INCLUDE_JFR
  50 #include "jfr/support/jfrFlush.hpp"
  51 #endif
  52 
  53 #ifdef DTRACE_ENABLED
  54 
  55 // Only bother with this argument setup if dtrace is available
  56 // TODO-FIXME: probes should not fire when caller is _blocked.  assert() accordingly.
  57 
  58 
  59 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread)                           \
  60   char* bytes = NULL;                                                      \
  61   int len = 0;                                                             \
  62   jlong jtid = SharedRuntime::get_java_tid(thread);                        \
  63   Symbol* klassname = ((oop)obj)->klass()->name();                         \
  64   if (klassname != NULL) {                                                 \
  65     bytes = (char*)klassname->bytes();                                     \
  66     len = klassname->utf8_length();                                        \
  67   }
  68 
  69 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis)            \
  70   {                                                                        \
  71     if (DTraceMonitorProbes) {                                             \
  72       DTRACE_MONITOR_PROBE_COMMON(obj, thread);                            \
  73       HOTSPOT_MONITOR_WAIT(jtid,                                           \
  74                            (monitor), bytes, len, (millis));               \
  75     }                                                                      \
  76   }
  77 
  78 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER
  79 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED
  80 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT
  81 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY
  82 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL
  83 
  84 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread)                  \
  85   {                                                                        \
  86     if (DTraceMonitorProbes) {                                             \
  87       DTRACE_MONITOR_PROBE_COMMON(obj, thread);                            \
  88       HOTSPOT_MONITOR_##probe(jtid,                                        \
  89                               (uintptr_t)(monitor), bytes, len);           \
  90     }                                                                      \
  91   }
  92 
  93 #else //  ndef DTRACE_ENABLED
  94 
  95 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon)    {;}
  96 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon)          {;}
  97 
  98 #endif // ndef DTRACE_ENABLED
  99 
 100 // Tunables ...
 101 // The knob* variables are effectively final.  Once set they should
 102 // never be modified hence.  Consider using __read_mostly with GCC.
 103 
 104 int ObjectMonitor::Knob_SpinLimit    = 5000;    // derived by an external tool -
 105 
 106 static int Knob_SuccRestrict        = 0;       // Limit successors + spinners to at-most-one
 107 static int Knob_MaxSpinners         = -1;      // Should be a function of # CPUs
 108 static int Knob_Bonus               = 100;     // spin success bonus
 109 static int Knob_BonusB              = 100;     // spin success bonus
 110 static int Knob_Penalty             = 200;     // spin failure penalty
 111 static int Knob_Poverty             = 1000;
 112 static int Knob_SpinAfterFutile     = 1;       // Spin after returning from park()
 113 static int Knob_FixedSpin           = 0;
 114 static int Knob_OState              = 3;       // Spinner checks thread state of _owner
 115 static int Knob_UsePause            = 1;
 116 static int Knob_ExitPolicy          = 0;
 117 static int Knob_PreSpin             = 10;      // 20-100 likely better
 118 static int Knob_ResetEvent          = 0;
 119 
 120 static int Knob_FastHSSEC           = 0;
 121 static int Knob_MoveNotifyee        = 2;       // notify() - disposition of notifyee
 122 static int Knob_QMode               = 0;       // EntryList-cxq policy - queue discipline
 123 static volatile int InitDone        = 0;
 124 
 125 // -----------------------------------------------------------------------------
 126 // Theory of operations -- Monitors lists, thread residency, etc:
 127 //
 128 // * A thread acquires ownership of a monitor by successfully
 129 //   CAS()ing the _owner field from null to non-null.
 130 //
 131 // * Invariant: A thread appears on at most one monitor list --
 132 //   cxq, EntryList or WaitSet -- at any one time.
 133 //
 134 // * Contending threads "push" themselves onto the cxq with CAS
 135 //   and then spin/park.
 136 //
 137 // * After a contending thread eventually acquires the lock it must
 138 //   dequeue itself from either the EntryList or the cxq.
 139 //
 140 // * The exiting thread identifies and unparks an "heir presumptive"
 141 //   tentative successor thread on the EntryList.  Critically, the
 142 //   exiting thread doesn't unlink the successor thread from the EntryList.
 143 //   After having been unparked, the wakee will recontend for ownership of
 144 //   the monitor.   The successor (wakee) will either acquire the lock or
 145 //   re-park itself.
 146 //
 147 //   Succession is provided for by a policy of competitive handoff.
 148 //   The exiting thread does _not_ grant or pass ownership to the
 149 //   successor thread.  (This is also referred to as "handoff" succession").
 150 //   Instead the exiting thread releases ownership and possibly wakes
 151 //   a successor, so the successor can (re)compete for ownership of the lock.
 152 //   If the EntryList is empty but the cxq is populated the exiting
 153 //   thread will drain the cxq into the EntryList.  It does so by
 154 //   by detaching the cxq (installing null with CAS) and folding
 155 //   the threads from the cxq into the EntryList.  The EntryList is
 156 //   doubly linked, while the cxq is singly linked because of the
 157 //   CAS-based "push" used to enqueue recently arrived threads (RATs).
 158 //
 159 // * Concurrency invariants:
 160 //
 161 //   -- only the monitor owner may access or mutate the EntryList.
 162 //      The mutex property of the monitor itself protects the EntryList
 163 //      from concurrent interference.
 164 //   -- Only the monitor owner may detach the cxq.
 165 //
 166 // * The monitor entry list operations avoid locks, but strictly speaking
 167 //   they're not lock-free.  Enter is lock-free, exit is not.
 168 //   For a description of 'Methods and apparatus providing non-blocking access
 169 //   to a resource,' see U.S. Pat. No. 7844973.
 170 //
 171 // * The cxq can have multiple concurrent "pushers" but only one concurrent
 172 //   detaching thread.  This mechanism is immune from the ABA corruption.
 173 //   More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
 174 //
 175 // * Taken together, the cxq and the EntryList constitute or form a
 176 //   single logical queue of threads stalled trying to acquire the lock.
 177 //   We use two distinct lists to improve the odds of a constant-time
 178 //   dequeue operation after acquisition (in the ::enter() epilogue) and
 179 //   to reduce heat on the list ends.  (c.f. Michael Scott's "2Q" algorithm).
 180 //   A key desideratum is to minimize queue & monitor metadata manipulation
 181 //   that occurs while holding the monitor lock -- that is, we want to
 182 //   minimize monitor lock holds times.  Note that even a small amount of
 183 //   fixed spinning will greatly reduce the # of enqueue-dequeue operations
 184 //   on EntryList|cxq.  That is, spinning relieves contention on the "inner"
 185 //   locks and monitor metadata.
 186 //
 187 //   Cxq points to the set of Recently Arrived Threads attempting entry.
 188 //   Because we push threads onto _cxq with CAS, the RATs must take the form of
 189 //   a singly-linked LIFO.  We drain _cxq into EntryList  at unlock-time when
 190 //   the unlocking thread notices that EntryList is null but _cxq is != null.
 191 //
 192 //   The EntryList is ordered by the prevailing queue discipline and
 193 //   can be organized in any convenient fashion, such as a doubly-linked list or
 194 //   a circular doubly-linked list.  Critically, we want insert and delete operations
 195 //   to operate in constant-time.  If we need a priority queue then something akin
 196 //   to Solaris' sleepq would work nicely.  Viz.,
 197 //   http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
 198 //   Queue discipline is enforced at ::exit() time, when the unlocking thread
 199 //   drains the cxq into the EntryList, and orders or reorders the threads on the
 200 //   EntryList accordingly.
 201 //
 202 //   Barring "lock barging", this mechanism provides fair cyclic ordering,
 203 //   somewhat similar to an elevator-scan.
 204 //
 205 // * The monitor synchronization subsystem avoids the use of native
 206 //   synchronization primitives except for the narrow platform-specific
 207 //   park-unpark abstraction.  See the comments in os_solaris.cpp regarding
 208 //   the semantics of park-unpark.  Put another way, this monitor implementation
 209 //   depends only on atomic operations and park-unpark.  The monitor subsystem
 210 //   manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
 211 //   underlying OS manages the READY<->RUN transitions.
 212 //
 213 // * Waiting threads reside on the WaitSet list -- wait() puts
 214 //   the caller onto the WaitSet.
 215 //
 216 // * notify() or notifyAll() simply transfers threads from the WaitSet to
 217 //   either the EntryList or cxq.  Subsequent exit() operations will
 218 //   unpark the notifyee.  Unparking a notifee in notify() is inefficient -
 219 //   it's likely the notifyee would simply impale itself on the lock held
 220 //   by the notifier.
 221 //
 222 // * An interesting alternative is to encode cxq as (List,LockByte) where
 223 //   the LockByte is 0 iff the monitor is owned.  _owner is simply an auxiliary
 224 //   variable, like _recursions, in the scheme.  The threads or Events that form
 225 //   the list would have to be aligned in 256-byte addresses.  A thread would
 226 //   try to acquire the lock or enqueue itself with CAS, but exiting threads
 227 //   could use a 1-0 protocol and simply STB to set the LockByte to 0.
 228 //   Note that is is *not* word-tearing, but it does presume that full-word
 229 //   CAS operations are coherent with intermix with STB operations.  That's true
 230 //   on most common processors.
 231 //
 232 // * See also http://blogs.sun.com/dave
 233 
 234 
 235 void* ObjectMonitor::operator new (size_t size) throw() {
 236   return AllocateHeap(size, mtInternal);
 237 }
 238 void* ObjectMonitor::operator new[] (size_t size) throw() {
 239   return operator new (size);
 240 }
 241 void ObjectMonitor::operator delete(void* p) {
 242   FreeHeap(p);
 243 }
 244 void ObjectMonitor::operator delete[] (void *p) {
 245   operator delete(p);
 246 }
 247 
 248 // -----------------------------------------------------------------------------
 249 // Enter support
 250 
 251 void ObjectMonitor::enter(TRAPS) {
 252   // The following code is ordered to check the most common cases first
 253   // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
 254   Thread * const Self = THREAD;
 255 
 256   void * cur = Atomic::cmpxchg(Self, &_owner, (void*)NULL);
 257   if (cur == NULL) {
 258     // Either ASSERT _recursions == 0 or explicitly set _recursions = 0.
 259     assert(_recursions == 0, "invariant");
 260     assert(_owner == Self, "invariant");
 261     return;
 262   }
 263 
 264   if (cur == Self) {
 265     // TODO-FIXME: check for integer overflow!  BUGID 6557169.
 266     _recursions++;
 267     return;
 268   }
 269 
 270   if (Self->is_lock_owned ((address)cur)) {
 271     assert(_recursions == 0, "internal state error");
 272     _recursions = 1;
 273     // Commute owner from a thread-specific on-stack BasicLockObject address to
 274     // a full-fledged "Thread *".
 275     _owner = Self;
 276     return;
 277   }
 278 
 279   // We've encountered genuine contention.
 280   assert(Self->_Stalled == 0, "invariant");
 281   Self->_Stalled = intptr_t(this);
 282 
 283   // Try one round of spinning *before* enqueueing Self
 284   // and before going through the awkward and expensive state
 285   // transitions.  The following spin is strictly optional ...
 286   // Note that if we acquire the monitor from an initial spin
 287   // we forgo posting JVMTI events and firing DTRACE probes.
 288   if (TrySpin(Self) > 0) {
 289     assert(_owner == Self, "invariant");
 290     assert(_recursions == 0, "invariant");
 291     assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 292     Self->_Stalled = 0;
 293     return;
 294   }
 295 
 296   assert(_owner != Self, "invariant");
 297   assert(_succ != Self, "invariant");
 298   assert(Self->is_Java_thread(), "invariant");
 299   JavaThread * jt = (JavaThread *) Self;
 300   assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
 301   assert(jt->thread_state() != _thread_blocked, "invariant");
 302   assert(this->object() != NULL, "invariant");
 303   assert(_count >= 0, "invariant");
 304 
 305   // Prevent deflation at STW-time.  See deflate_idle_monitors() and is_busy().
 306   // Ensure the object-monitor relationship remains stable while there's contention.
 307   Atomic::inc(&_count);
 308 
 309   JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);)
 310   EventJavaMonitorEnter event;
 311   if (event.should_commit()) {
 312     event.set_monitorClass(((oop)this->object())->klass());
 313     event.set_address((uintptr_t)(this->object_addr()));
 314   }
 315 
 316   { // Change java thread status to indicate blocked on monitor enter.
 317     JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this);
 318 
 319     Self->set_current_pending_monitor(this);
 320 
 321     DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt);
 322     if (JvmtiExport::should_post_monitor_contended_enter()) {
 323       JvmtiExport::post_monitor_contended_enter(jt, this);
 324 
 325       // The current thread does not yet own the monitor and does not
 326       // yet appear on any queues that would get it made the successor.
 327       // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
 328       // handler cannot accidentally consume an unpark() meant for the
 329       // ParkEvent associated with this ObjectMonitor.
 330     }
 331 
 332     OSThreadContendState osts(Self->osthread());
 333     ThreadBlockInVM tbivm(jt);
 334 
 335     // TODO-FIXME: change the following for(;;) loop to straight-line code.
 336     for (;;) {
 337       jt->set_suspend_equivalent();
 338       // cleared by handle_special_suspend_equivalent_condition()
 339       // or java_suspend_self()
 340 
 341       EnterI(THREAD);
 342 
 343       if (!ExitSuspendEquivalent(jt)) break;
 344 
 345       // We have acquired the contended monitor, but while we were
 346       // waiting another thread suspended us. We don't want to enter
 347       // the monitor while suspended because that would surprise the
 348       // thread that suspended us.
 349       //
 350       _recursions = 0;
 351       _succ = NULL;
 352       exit(false, Self);
 353 
 354       jt->java_suspend_self();
 355     }
 356     Self->set_current_pending_monitor(NULL);
 357 
 358     // We cleared the pending monitor info since we've just gotten past
 359     // the enter-check-for-suspend dance and we now own the monitor free
 360     // and clear, i.e., it is no longer pending. The ThreadBlockInVM
 361     // destructor can go to a safepoint at the end of this block. If we
 362     // do a thread dump during that safepoint, then this thread will show
 363     // as having "-locked" the monitor, but the OS and java.lang.Thread
 364     // states will still report that the thread is blocked trying to
 365     // acquire it.
 366   }
 367 
 368   Atomic::dec(&_count);
 369   assert(_count >= 0, "invariant");
 370   Self->_Stalled = 0;
 371 
 372   // Must either set _recursions = 0 or ASSERT _recursions == 0.
 373   assert(_recursions == 0, "invariant");
 374   assert(_owner == Self, "invariant");
 375   assert(_succ != Self, "invariant");
 376   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 377 
 378   // The thread -- now the owner -- is back in vm mode.
 379   // Report the glorious news via TI,DTrace and jvmstat.
 380   // The probe effect is non-trivial.  All the reportage occurs
 381   // while we hold the monitor, increasing the length of the critical
 382   // section.  Amdahl's parallel speedup law comes vividly into play.
 383   //
 384   // Another option might be to aggregate the events (thread local or
 385   // per-monitor aggregation) and defer reporting until a more opportune
 386   // time -- such as next time some thread encounters contention but has
 387   // yet to acquire the lock.  While spinning that thread could
 388   // spinning we could increment JVMStat counters, etc.
 389 
 390   DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt);
 391   if (JvmtiExport::should_post_monitor_contended_entered()) {
 392     JvmtiExport::post_monitor_contended_entered(jt, this);
 393 
 394     // The current thread already owns the monitor and is not going to
 395     // call park() for the remainder of the monitor enter protocol. So
 396     // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
 397     // event handler consumed an unpark() issued by the thread that
 398     // just exited the monitor.
 399   }
 400   if (event.should_commit()) {
 401     event.set_previousOwner((uintptr_t)_previous_owner_tid);
 402     event.commit();
 403   }
 404   OM_PERFDATA_OP(ContendedLockAttempts, inc());
 405 }
 406 
 407 // Caveat: TryLock() is not necessarily serializing if it returns failure.
 408 // Callers must compensate as needed.
 409 
 410 int ObjectMonitor::TryLock(Thread * Self) {
 411   void * own = _owner;
 412   if (own != NULL) return 0;
 413   if (Atomic::replace_if_null(Self, &_owner)) {
 414     // Either guarantee _recursions == 0 or set _recursions = 0.
 415     assert(_recursions == 0, "invariant");
 416     assert(_owner == Self, "invariant");
 417     return 1;
 418   }
 419   // The lock had been free momentarily, but we lost the race to the lock.
 420   // Interference -- the CAS failed.
 421   // We can either return -1 or retry.
 422   // Retry doesn't make as much sense because the lock was just acquired.
 423   return -1;
 424 }
 425 
 426 #define MAX_RECHECK_INTERVAL 1000
 427 
 428 void ObjectMonitor::EnterI(TRAPS) {
 429   Thread * const Self = THREAD;
 430   assert(Self->is_Java_thread(), "invariant");
 431   assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant");
 432 
 433   // Try the lock - TATAS
 434   if (TryLock (Self) > 0) {
 435     assert(_succ != Self, "invariant");
 436     assert(_owner == Self, "invariant");
 437     assert(_Responsible != Self, "invariant");
 438     return;
 439   }
 440 
 441   DeferredInitialize();
 442 
 443   // We try one round of spinning *before* enqueueing Self.
 444   //
 445   // If the _owner is ready but OFFPROC we could use a YieldTo()
 446   // operation to donate the remainder of this thread's quantum
 447   // to the owner.  This has subtle but beneficial affinity
 448   // effects.
 449 
 450   if (TrySpin(Self) > 0) {
 451     assert(_owner == Self, "invariant");
 452     assert(_succ != Self, "invariant");
 453     assert(_Responsible != Self, "invariant");
 454     return;
 455   }
 456 
 457   // The Spin failed -- Enqueue and park the thread ...
 458   assert(_succ != Self, "invariant");
 459   assert(_owner != Self, "invariant");
 460   assert(_Responsible != Self, "invariant");
 461 
 462   // Enqueue "Self" on ObjectMonitor's _cxq.
 463   //
 464   // Node acts as a proxy for Self.
 465   // As an aside, if were to ever rewrite the synchronization code mostly
 466   // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
 467   // Java objects.  This would avoid awkward lifecycle and liveness issues,
 468   // as well as eliminate a subset of ABA issues.
 469   // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
 470 
 471   ObjectWaiter node(Self);
 472   Self->_ParkEvent->reset();
 473   node._prev   = (ObjectWaiter *) 0xBAD;
 474   node.TState  = ObjectWaiter::TS_CXQ;
 475 
 476   // Push "Self" onto the front of the _cxq.
 477   // Once on cxq/EntryList, Self stays on-queue until it acquires the lock.
 478   // Note that spinning tends to reduce the rate at which threads
 479   // enqueue and dequeue on EntryList|cxq.
 480   ObjectWaiter * nxt;
 481   for (;;) {
 482     node._next = nxt = _cxq;
 483     if (Atomic::cmpxchg(&node, &_cxq, nxt) == nxt) break;
 484 
 485     // Interference - the CAS failed because _cxq changed.  Just retry.
 486     // As an optional optimization we retry the lock.
 487     if (TryLock (Self) > 0) {
 488       assert(_succ != Self, "invariant");
 489       assert(_owner == Self, "invariant");
 490       assert(_Responsible != Self, "invariant");
 491       return;
 492     }
 493   }
 494 
 495   // Check for cxq|EntryList edge transition to non-null.  This indicates
 496   // the onset of contention.  While contention persists exiting threads
 497   // will use a ST:MEMBAR:LD 1-1 exit protocol.  When contention abates exit
 498   // operations revert to the faster 1-0 mode.  This enter operation may interleave
 499   // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
 500   // arrange for one of the contending thread to use a timed park() operations
 501   // to detect and recover from the race.  (Stranding is form of progress failure
 502   // where the monitor is unlocked but all the contending threads remain parked).
 503   // That is, at least one of the contended threads will periodically poll _owner.
 504   // One of the contending threads will become the designated "Responsible" thread.
 505   // The Responsible thread uses a timed park instead of a normal indefinite park
 506   // operation -- it periodically wakes and checks for and recovers from potential
 507   // strandings admitted by 1-0 exit operations.   We need at most one Responsible
 508   // thread per-monitor at any given moment.  Only threads on cxq|EntryList may
 509   // be responsible for a monitor.
 510   //
 511   // Currently, one of the contended threads takes on the added role of "Responsible".
 512   // A viable alternative would be to use a dedicated "stranding checker" thread
 513   // that periodically iterated over all the threads (or active monitors) and unparked
 514   // successors where there was risk of stranding.  This would help eliminate the
 515   // timer scalability issues we see on some platforms as we'd only have one thread
 516   // -- the checker -- parked on a timer.
 517 
 518   if (nxt == NULL && _EntryList == NULL) {
 519     // Try to assume the role of responsible thread for the monitor.
 520     // CONSIDER:  ST vs CAS vs { if (Responsible==null) Responsible=Self }
 521     Atomic::replace_if_null(Self, &_Responsible);
 522   }
 523 
 524   // The lock might have been released while this thread was occupied queueing
 525   // itself onto _cxq.  To close the race and avoid "stranding" and
 526   // progress-liveness failure we must resample-retry _owner before parking.
 527   // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
 528   // In this case the ST-MEMBAR is accomplished with CAS().
 529   //
 530   // TODO: Defer all thread state transitions until park-time.
 531   // Since state transitions are heavy and inefficient we'd like
 532   // to defer the state transitions until absolutely necessary,
 533   // and in doing so avoid some transitions ...
 534 
 535   int nWakeups = 0;
 536   int recheckInterval = 1;
 537 
 538   for (;;) {
 539 
 540     if (TryLock(Self) > 0) break;
 541     assert(_owner != Self, "invariant");
 542 
 543     // park self
 544     if (_Responsible == Self) {
 545       Self->_ParkEvent->park((jlong) recheckInterval);
 546       // Increase the recheckInterval, but clamp the value.
 547       recheckInterval *= 8;
 548       if (recheckInterval > MAX_RECHECK_INTERVAL) {
 549         recheckInterval = MAX_RECHECK_INTERVAL;
 550       }
 551     } else {
 552       Self->_ParkEvent->park();
 553     }
 554 
 555     if (TryLock(Self) > 0) break;
 556 
 557     // The lock is still contested.
 558     // Keep a tally of the # of futile wakeups.
 559     // Note that the counter is not protected by a lock or updated by atomics.
 560     // That is by design - we trade "lossy" counters which are exposed to
 561     // races during updates for a lower probe effect.
 562 
 563     // This PerfData object can be used in parallel with a safepoint.
 564     // See the work around in PerfDataManager::destroy().
 565     OM_PERFDATA_OP(FutileWakeups, inc());
 566     ++nWakeups;
 567 
 568     // Assuming this is not a spurious wakeup we'll normally find _succ == Self.
 569     // We can defer clearing _succ until after the spin completes
 570     // TrySpin() must tolerate being called with _succ == Self.
 571     // Try yet another round of adaptive spinning.
 572     if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break;
 573 
 574     // We can find that we were unpark()ed and redesignated _succ while
 575     // we were spinning.  That's harmless.  If we iterate and call park(),
 576     // park() will consume the event and return immediately and we'll
 577     // just spin again.  This pattern can repeat, leaving _succ to simply
 578     // spin on a CPU.  Enable Knob_ResetEvent to clear pending unparks().
 579     // Alternately, we can sample fired() here, and if set, forgo spinning
 580     // in the next iteration.
 581 
 582     if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) {
 583       Self->_ParkEvent->reset();
 584       OrderAccess::fence();
 585     }
 586     if (_succ == Self) _succ = NULL;
 587 
 588     // Invariant: after clearing _succ a thread *must* retry _owner before parking.
 589     OrderAccess::fence();
 590   }
 591 
 592   // Egress :
 593   // Self has acquired the lock -- Unlink Self from the cxq or EntryList.
 594   // Normally we'll find Self on the EntryList .
 595   // From the perspective of the lock owner (this thread), the
 596   // EntryList is stable and cxq is prepend-only.
 597   // The head of cxq is volatile but the interior is stable.
 598   // In addition, Self.TState is stable.
 599 
 600   assert(_owner == Self, "invariant");
 601   assert(object() != NULL, "invariant");
 602   // I'd like to write:
 603   //   guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
 604   // but as we're at a safepoint that's not safe.
 605 
 606   UnlinkAfterAcquire(Self, &node);
 607   if (_succ == Self) _succ = NULL;
 608 
 609   assert(_succ != Self, "invariant");
 610   if (_Responsible == Self) {
 611     _Responsible = NULL;
 612     OrderAccess::fence(); // Dekker pivot-point
 613 
 614     // We may leave threads on cxq|EntryList without a designated
 615     // "Responsible" thread.  This is benign.  When this thread subsequently
 616     // exits the monitor it can "see" such preexisting "old" threads --
 617     // threads that arrived on the cxq|EntryList before the fence, above --
 618     // by LDing cxq|EntryList.  Newly arrived threads -- that is, threads
 619     // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
 620     // non-null and elect a new "Responsible" timer thread.
 621     //
 622     // This thread executes:
 623     //    ST Responsible=null; MEMBAR    (in enter epilogue - here)
 624     //    LD cxq|EntryList               (in subsequent exit)
 625     //
 626     // Entering threads in the slow/contended path execute:
 627     //    ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
 628     //    The (ST cxq; MEMBAR) is accomplished with CAS().
 629     //
 630     // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
 631     // exit operation from floating above the ST Responsible=null.
 632   }
 633 
 634   // We've acquired ownership with CAS().
 635   // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
 636   // But since the CAS() this thread may have also stored into _succ,
 637   // EntryList, cxq or Responsible.  These meta-data updates must be
 638   // visible __before this thread subsequently drops the lock.
 639   // Consider what could occur if we didn't enforce this constraint --
 640   // STs to monitor meta-data and user-data could reorder with (become
 641   // visible after) the ST in exit that drops ownership of the lock.
 642   // Some other thread could then acquire the lock, but observe inconsistent
 643   // or old monitor meta-data and heap data.  That violates the JMM.
 644   // To that end, the 1-0 exit() operation must have at least STST|LDST
 645   // "release" barrier semantics.  Specifically, there must be at least a
 646   // STST|LDST barrier in exit() before the ST of null into _owner that drops
 647   // the lock.   The barrier ensures that changes to monitor meta-data and data
 648   // protected by the lock will be visible before we release the lock, and
 649   // therefore before some other thread (CPU) has a chance to acquire the lock.
 650   // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
 651   //
 652   // Critically, any prior STs to _succ or EntryList must be visible before
 653   // the ST of null into _owner in the *subsequent* (following) corresponding
 654   // monitorexit.  Recall too, that in 1-0 mode monitorexit does not necessarily
 655   // execute a serializing instruction.
 656 
 657   return;
 658 }
 659 
 660 // ReenterI() is a specialized inline form of the latter half of the
 661 // contended slow-path from EnterI().  We use ReenterI() only for
 662 // monitor reentry in wait().
 663 //
 664 // In the future we should reconcile EnterI() and ReenterI(), adding
 665 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the
 666 // loop accordingly.
 667 
 668 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) {
 669   assert(Self != NULL, "invariant");
 670   assert(SelfNode != NULL, "invariant");
 671   assert(SelfNode->_thread == Self, "invariant");
 672   assert(_waiters > 0, "invariant");
 673   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 674   assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant");
 675   JavaThread * jt = (JavaThread *) Self;
 676 
 677   int nWakeups = 0;
 678   for (;;) {
 679     ObjectWaiter::TStates v = SelfNode->TState;
 680     guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
 681     assert(_owner != Self, "invariant");
 682 
 683     if (TryLock(Self) > 0) break;
 684     if (TrySpin(Self) > 0) break;
 685 
 686     // State transition wrappers around park() ...
 687     // ReenterI() wisely defers state transitions until
 688     // it's clear we must park the thread.
 689     {
 690       OSThreadContendState osts(Self->osthread());
 691       ThreadBlockInVM tbivm(jt);
 692 
 693       // cleared by handle_special_suspend_equivalent_condition()
 694       // or java_suspend_self()
 695       jt->set_suspend_equivalent();
 696       Self->_ParkEvent->park();
 697 
 698       // were we externally suspended while we were waiting?
 699       for (;;) {
 700         if (!ExitSuspendEquivalent(jt)) break;
 701         if (_succ == Self) { _succ = NULL; OrderAccess::fence(); }
 702         jt->java_suspend_self();
 703         jt->set_suspend_equivalent();
 704       }
 705     }
 706 
 707     // Try again, but just so we distinguish between futile wakeups and
 708     // successful wakeups.  The following test isn't algorithmically
 709     // necessary, but it helps us maintain sensible statistics.
 710     if (TryLock(Self) > 0) break;
 711 
 712     // The lock is still contested.
 713     // Keep a tally of the # of futile wakeups.
 714     // Note that the counter is not protected by a lock or updated by atomics.
 715     // That is by design - we trade "lossy" counters which are exposed to
 716     // races during updates for a lower probe effect.
 717     ++nWakeups;
 718 
 719     // Assuming this is not a spurious wakeup we'll normally
 720     // find that _succ == Self.
 721     if (_succ == Self) _succ = NULL;
 722 
 723     // Invariant: after clearing _succ a contending thread
 724     // *must* retry  _owner before parking.
 725     OrderAccess::fence();
 726 
 727     // This PerfData object can be used in parallel with a safepoint.
 728     // See the work around in PerfDataManager::destroy().
 729     OM_PERFDATA_OP(FutileWakeups, inc());
 730   }
 731 
 732   // Self has acquired the lock -- Unlink Self from the cxq or EntryList .
 733   // Normally we'll find Self on the EntryList.
 734   // Unlinking from the EntryList is constant-time and atomic-free.
 735   // From the perspective of the lock owner (this thread), the
 736   // EntryList is stable and cxq is prepend-only.
 737   // The head of cxq is volatile but the interior is stable.
 738   // In addition, Self.TState is stable.
 739 
 740   assert(_owner == Self, "invariant");
 741   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
 742   UnlinkAfterAcquire(Self, SelfNode);
 743   if (_succ == Self) _succ = NULL;
 744   assert(_succ != Self, "invariant");
 745   SelfNode->TState = ObjectWaiter::TS_RUN;
 746   OrderAccess::fence();      // see comments at the end of EnterI()
 747 }
 748 
 749 // By convention we unlink a contending thread from EntryList|cxq immediately
 750 // after the thread acquires the lock in ::enter().  Equally, we could defer
 751 // unlinking the thread until ::exit()-time.
 752 
 753 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) {
 754   assert(_owner == Self, "invariant");
 755   assert(SelfNode->_thread == Self, "invariant");
 756 
 757   if (SelfNode->TState == ObjectWaiter::TS_ENTER) {
 758     // Normal case: remove Self from the DLL EntryList .
 759     // This is a constant-time operation.
 760     ObjectWaiter * nxt = SelfNode->_next;
 761     ObjectWaiter * prv = SelfNode->_prev;
 762     if (nxt != NULL) nxt->_prev = prv;
 763     if (prv != NULL) prv->_next = nxt;
 764     if (SelfNode == _EntryList) _EntryList = nxt;
 765     assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant");
 766     assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant");
 767   } else {
 768     assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant");
 769     // Inopportune interleaving -- Self is still on the cxq.
 770     // This usually means the enqueue of self raced an exiting thread.
 771     // Normally we'll find Self near the front of the cxq, so
 772     // dequeueing is typically fast.  If needbe we can accelerate
 773     // this with some MCS/CHL-like bidirectional list hints and advisory
 774     // back-links so dequeueing from the interior will normally operate
 775     // in constant-time.
 776     // Dequeue Self from either the head (with CAS) or from the interior
 777     // with a linear-time scan and normal non-atomic memory operations.
 778     // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList
 779     // and then unlink Self from EntryList.  We have to drain eventually,
 780     // so it might as well be now.
 781 
 782     ObjectWaiter * v = _cxq;
 783     assert(v != NULL, "invariant");
 784     if (v != SelfNode || Atomic::cmpxchg(SelfNode->_next, &_cxq, v) != v) {
 785       // The CAS above can fail from interference IFF a "RAT" arrived.
 786       // In that case Self must be in the interior and can no longer be
 787       // at the head of cxq.
 788       if (v == SelfNode) {
 789         assert(_cxq != v, "invariant");
 790         v = _cxq;          // CAS above failed - start scan at head of list
 791       }
 792       ObjectWaiter * p;
 793       ObjectWaiter * q = NULL;
 794       for (p = v; p != NULL && p != SelfNode; p = p->_next) {
 795         q = p;
 796         assert(p->TState == ObjectWaiter::TS_CXQ, "invariant");
 797       }
 798       assert(v != SelfNode, "invariant");
 799       assert(p == SelfNode, "Node not found on cxq");
 800       assert(p != _cxq, "invariant");
 801       assert(q != NULL, "invariant");
 802       assert(q->_next == p, "invariant");
 803       q->_next = p->_next;
 804     }
 805   }
 806 
 807 #ifdef ASSERT
 808   // Diagnostic hygiene ...
 809   SelfNode->_prev  = (ObjectWaiter *) 0xBAD;
 810   SelfNode->_next  = (ObjectWaiter *) 0xBAD;
 811   SelfNode->TState = ObjectWaiter::TS_RUN;
 812 #endif
 813 }
 814 
 815 // -----------------------------------------------------------------------------
 816 // Exit support
 817 //
 818 // exit()
 819 // ~~~~~~
 820 // Note that the collector can't reclaim the objectMonitor or deflate
 821 // the object out from underneath the thread calling ::exit() as the
 822 // thread calling ::exit() never transitions to a stable state.
 823 // This inhibits GC, which in turn inhibits asynchronous (and
 824 // inopportune) reclamation of "this".
 825 //
 826 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
 827 // There's one exception to the claim above, however.  EnterI() can call
 828 // exit() to drop a lock if the acquirer has been externally suspended.
 829 // In that case exit() is called with _thread_state as _thread_blocked,
 830 // but the monitor's _count field is > 0, which inhibits reclamation.
 831 //
 832 // 1-0 exit
 833 // ~~~~~~~~
 834 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
 835 // the fast-path operators have been optimized so the common ::exit()
 836 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock().
 837 // The code emitted by fast_unlock() elides the usual MEMBAR.  This
 838 // greatly improves latency -- MEMBAR and CAS having considerable local
 839 // latency on modern processors -- but at the cost of "stranding".  Absent the
 840 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
 841 // ::enter() path, resulting in the entering thread being stranding
 842 // and a progress-liveness failure.   Stranding is extremely rare.
 843 // We use timers (timed park operations) & periodic polling to detect
 844 // and recover from stranding.  Potentially stranded threads periodically
 845 // wake up and poll the lock.  See the usage of the _Responsible variable.
 846 //
 847 // The CAS() in enter provides for safety and exclusion, while the CAS or
 848 // MEMBAR in exit provides for progress and avoids stranding.  1-0 locking
 849 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding.
 850 // We detect and recover from stranding with timers.
 851 //
 852 // If a thread transiently strands it'll park until (a) another
 853 // thread acquires the lock and then drops the lock, at which time the
 854 // exiting thread will notice and unpark the stranded thread, or, (b)
 855 // the timer expires.  If the lock is high traffic then the stranding latency
 856 // will be low due to (a).  If the lock is low traffic then the odds of
 857 // stranding are lower, although the worst-case stranding latency
 858 // is longer.  Critically, we don't want to put excessive load in the
 859 // platform's timer subsystem.  We want to minimize both the timer injection
 860 // rate (timers created/sec) as well as the number of timers active at
 861 // any one time.  (more precisely, we want to minimize timer-seconds, which is
 862 // the integral of the # of active timers at any instant over time).
 863 // Both impinge on OS scalability.  Given that, at most one thread parked on
 864 // a monitor will use a timer.
 865 //
 866 // There is also the risk of a futile wake-up. If we drop the lock
 867 // another thread can reacquire the lock immediately, and we can
 868 // then wake a thread unnecessarily. This is benign, and we've
 869 // structured the code so the windows are short and the frequency
 870 // of such futile wakups is low.
 871 
 872 void ObjectMonitor::exit(bool not_suspended, TRAPS) {
 873   Thread * const Self = THREAD;
 874   if (THREAD != _owner) {
 875     if (THREAD->is_lock_owned((address) _owner)) {
 876       // Transmute _owner from a BasicLock pointer to a Thread address.
 877       // We don't need to hold _mutex for this transition.
 878       // Non-null to Non-null is safe as long as all readers can
 879       // tolerate either flavor.
 880       assert(_recursions == 0, "invariant");
 881       _owner = THREAD;
 882       _recursions = 0;
 883     } else {
 884       // Apparent unbalanced locking ...
 885       // Naively we'd like to throw IllegalMonitorStateException.
 886       // As a practical matter we can neither allocate nor throw an
 887       // exception as ::exit() can be called from leaf routines.
 888       // see x86_32.ad Fast_Unlock() and the I1 and I2 properties.
 889       // Upon deeper reflection, however, in a properly run JVM the only
 890       // way we should encounter this situation is in the presence of
 891       // unbalanced JNI locking. TODO: CheckJNICalls.
 892       // See also: CR4414101
 893       assert(false, "Non-balanced monitor enter/exit! Likely JNI locking");
 894       return;
 895     }
 896   }
 897 
 898   if (_recursions != 0) {
 899     _recursions--;        // this is simple recursive enter
 900     return;
 901   }
 902 
 903   // Invariant: after setting Responsible=null an thread must execute
 904   // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
 905   _Responsible = NULL;
 906 
 907 #if INCLUDE_JFR
 908   // get the owner's thread id for the MonitorEnter event
 909   // if it is enabled and the thread isn't suspended
 910   if (not_suspended && EventJavaMonitorEnter::is_enabled()) {
 911     _previous_owner_tid = JFR_THREAD_ID(Self);
 912   }
 913 #endif
 914 
 915   for (;;) {
 916     assert(THREAD == _owner, "invariant");
 917 
 918     if (Knob_ExitPolicy == 0) {
 919       // release semantics: prior loads and stores from within the critical section
 920       // must not float (reorder) past the following store that drops the lock.
 921       // On SPARC that requires MEMBAR #loadstore|#storestore.
 922       // But of course in TSO #loadstore|#storestore is not required.
 923       // I'd like to write one of the following:
 924       // A.  OrderAccess::release() ; _owner = NULL
 925       // B.  OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL;
 926       // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both
 927       // store into a _dummy variable.  That store is not needed, but can result
 928       // in massive wasteful coherency traffic on classic SMP systems.
 929       // Instead, I use release_store(), which is implemented as just a simple
 930       // ST on x64, x86 and SPARC.
 931       OrderAccess::release_store(&_owner, (void*)NULL);   // drop the lock
 932       OrderAccess::storeload();                        // See if we need to wake a successor
 933       if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
 934         return;
 935       }
 936       // Other threads are blocked trying to acquire the lock.
 937 
 938       // Normally the exiting thread is responsible for ensuring succession,
 939       // but if other successors are ready or other entering threads are spinning
 940       // then this thread can simply store NULL into _owner and exit without
 941       // waking a successor.  The existence of spinners or ready successors
 942       // guarantees proper succession (liveness).  Responsibility passes to the
 943       // ready or running successors.  The exiting thread delegates the duty.
 944       // More precisely, if a successor already exists this thread is absolved
 945       // of the responsibility of waking (unparking) one.
 946       //
 947       // The _succ variable is critical to reducing futile wakeup frequency.
 948       // _succ identifies the "heir presumptive" thread that has been made
 949       // ready (unparked) but that has not yet run.  We need only one such
 950       // successor thread to guarantee progress.
 951       // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
 952       // section 3.3 "Futile Wakeup Throttling" for details.
 953       //
 954       // Note that spinners in Enter() also set _succ non-null.
 955       // In the current implementation spinners opportunistically set
 956       // _succ so that exiting threads might avoid waking a successor.
 957       // Another less appealing alternative would be for the exiting thread
 958       // to drop the lock and then spin briefly to see if a spinner managed
 959       // to acquire the lock.  If so, the exiting thread could exit
 960       // immediately without waking a successor, otherwise the exiting
 961       // thread would need to dequeue and wake a successor.
 962       // (Note that we'd need to make the post-drop spin short, but no
 963       // shorter than the worst-case round-trip cache-line migration time.
 964       // The dropped lock needs to become visible to the spinner, and then
 965       // the acquisition of the lock by the spinner must become visible to
 966       // the exiting thread).
 967 
 968       // It appears that an heir-presumptive (successor) must be made ready.
 969       // Only the current lock owner can manipulate the EntryList or
 970       // drain _cxq, so we need to reacquire the lock.  If we fail
 971       // to reacquire the lock the responsibility for ensuring succession
 972       // falls to the new owner.
 973       //
 974       if (!Atomic::replace_if_null(THREAD, &_owner)) {
 975         return;
 976       }
 977     } else {
 978       if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
 979         OrderAccess::release_store(&_owner, (void*)NULL);   // drop the lock
 980         OrderAccess::storeload();
 981         // Ratify the previously observed values.
 982         if (_cxq == NULL || _succ != NULL) {
 983           return;
 984         }
 985 
 986         // inopportune interleaving -- the exiting thread (this thread)
 987         // in the fast-exit path raced an entering thread in the slow-enter
 988         // path.
 989         // We have two choices:
 990         // A.  Try to reacquire the lock.
 991         //     If the CAS() fails return immediately, otherwise
 992         //     we either restart/rerun the exit operation, or simply
 993         //     fall-through into the code below which wakes a successor.
 994         // B.  If the elements forming the EntryList|cxq are TSM
 995         //     we could simply unpark() the lead thread and return
 996         //     without having set _succ.
 997         if (!Atomic::replace_if_null(THREAD, &_owner)) {
 998           return;
 999         }
1000       }
1001     }
1002 
1003     guarantee(_owner == THREAD, "invariant");
1004 
1005     ObjectWaiter * w = NULL;
1006     int QMode = Knob_QMode;
1007 
1008     if (QMode == 2 && _cxq != NULL) {
1009       // QMode == 2 : cxq has precedence over EntryList.
1010       // Try to directly wake a successor from the cxq.
1011       // If successful, the successor will need to unlink itself from cxq.
1012       w = _cxq;
1013       assert(w != NULL, "invariant");
1014       assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant");
1015       ExitEpilog(Self, w);
1016       return;
1017     }
1018 
1019     if (QMode == 3 && _cxq != NULL) {
1020       // Aggressively drain cxq into EntryList at the first opportunity.
1021       // This policy ensure that recently-run threads live at the head of EntryList.
1022       // Drain _cxq into EntryList - bulk transfer.
1023       // First, detach _cxq.
1024       // The following loop is tantamount to: w = swap(&cxq, NULL)
1025       w = _cxq;
1026       for (;;) {
1027         assert(w != NULL, "Invariant");
1028         ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1029         if (u == w) break;
1030         w = u;
1031       }
1032       assert(w != NULL, "invariant");
1033 
1034       ObjectWaiter * q = NULL;
1035       ObjectWaiter * p;
1036       for (p = w; p != NULL; p = p->_next) {
1037         guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1038         p->TState = ObjectWaiter::TS_ENTER;
1039         p->_prev = q;
1040         q = p;
1041       }
1042 
1043       // Append the RATs to the EntryList
1044       // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time.
1045       ObjectWaiter * Tail;
1046       for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL;
1047            Tail = Tail->_next)
1048         /* empty */;
1049       if (Tail == NULL) {
1050         _EntryList = w;
1051       } else {
1052         Tail->_next = w;
1053         w->_prev = Tail;
1054       }
1055 
1056       // Fall thru into code that tries to wake a successor from EntryList
1057     }
1058 
1059     if (QMode == 4 && _cxq != NULL) {
1060       // Aggressively drain cxq into EntryList at the first opportunity.
1061       // This policy ensure that recently-run threads live at the head of EntryList.
1062 
1063       // Drain _cxq into EntryList - bulk transfer.
1064       // First, detach _cxq.
1065       // The following loop is tantamount to: w = swap(&cxq, NULL)
1066       w = _cxq;
1067       for (;;) {
1068         assert(w != NULL, "Invariant");
1069         ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1070         if (u == w) break;
1071         w = u;
1072       }
1073       assert(w != NULL, "invariant");
1074 
1075       ObjectWaiter * q = NULL;
1076       ObjectWaiter * p;
1077       for (p = w; p != NULL; p = p->_next) {
1078         guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1079         p->TState = ObjectWaiter::TS_ENTER;
1080         p->_prev = q;
1081         q = p;
1082       }
1083 
1084       // Prepend the RATs to the EntryList
1085       if (_EntryList != NULL) {
1086         q->_next = _EntryList;
1087         _EntryList->_prev = q;
1088       }
1089       _EntryList = w;
1090 
1091       // Fall thru into code that tries to wake a successor from EntryList
1092     }
1093 
1094     w = _EntryList;
1095     if (w != NULL) {
1096       // I'd like to write: guarantee (w->_thread != Self).
1097       // But in practice an exiting thread may find itself on the EntryList.
1098       // Let's say thread T1 calls O.wait().  Wait() enqueues T1 on O's waitset and
1099       // then calls exit().  Exit release the lock by setting O._owner to NULL.
1100       // Let's say T1 then stalls.  T2 acquires O and calls O.notify().  The
1101       // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1102       // release the lock "O".  T2 resumes immediately after the ST of null into
1103       // _owner, above.  T2 notices that the EntryList is populated, so it
1104       // reacquires the lock and then finds itself on the EntryList.
1105       // Given all that, we have to tolerate the circumstance where "w" is
1106       // associated with Self.
1107       assert(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1108       ExitEpilog(Self, w);
1109       return;
1110     }
1111 
1112     // If we find that both _cxq and EntryList are null then just
1113     // re-run the exit protocol from the top.
1114     w = _cxq;
1115     if (w == NULL) continue;
1116 
1117     // Drain _cxq into EntryList - bulk transfer.
1118     // First, detach _cxq.
1119     // The following loop is tantamount to: w = swap(&cxq, NULL)
1120     for (;;) {
1121       assert(w != NULL, "Invariant");
1122       ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1123       if (u == w) break;
1124       w = u;
1125     }
1126 
1127     assert(w != NULL, "invariant");
1128     assert(_EntryList == NULL, "invariant");
1129 
1130     // Convert the LIFO SLL anchored by _cxq into a DLL.
1131     // The list reorganization step operates in O(LENGTH(w)) time.
1132     // It's critical that this step operate quickly as
1133     // "Self" still holds the outer-lock, restricting parallelism
1134     // and effectively lengthening the critical section.
1135     // Invariant: s chases t chases u.
1136     // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1137     // we have faster access to the tail.
1138 
1139     if (QMode == 1) {
1140       // QMode == 1 : drain cxq to EntryList, reversing order
1141       // We also reverse the order of the list.
1142       ObjectWaiter * s = NULL;
1143       ObjectWaiter * t = w;
1144       ObjectWaiter * u = NULL;
1145       while (t != NULL) {
1146         guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant");
1147         t->TState = ObjectWaiter::TS_ENTER;
1148         u = t->_next;
1149         t->_prev = u;
1150         t->_next = s;
1151         s = t;
1152         t = u;
1153       }
1154       _EntryList  = s;
1155       assert(s != NULL, "invariant");
1156     } else {
1157       // QMode == 0 or QMode == 2
1158       _EntryList = w;
1159       ObjectWaiter * q = NULL;
1160       ObjectWaiter * p;
1161       for (p = w; p != NULL; p = p->_next) {
1162         guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1163         p->TState = ObjectWaiter::TS_ENTER;
1164         p->_prev = q;
1165         q = p;
1166       }
1167     }
1168 
1169     // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL
1170     // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1171 
1172     // See if we can abdicate to a spinner instead of waking a thread.
1173     // A primary goal of the implementation is to reduce the
1174     // context-switch rate.
1175     if (_succ != NULL) continue;
1176 
1177     w = _EntryList;
1178     if (w != NULL) {
1179       guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1180       ExitEpilog(Self, w);
1181       return;
1182     }
1183   }
1184 }
1185 
1186 // ExitSuspendEquivalent:
1187 // A faster alternate to handle_special_suspend_equivalent_condition()
1188 //
1189 // handle_special_suspend_equivalent_condition() unconditionally
1190 // acquires the SR_lock.  On some platforms uncontended MutexLocker()
1191 // operations have high latency.  Note that in ::enter() we call HSSEC
1192 // while holding the monitor, so we effectively lengthen the critical sections.
1193 //
1194 // There are a number of possible solutions:
1195 //
1196 // A.  To ameliorate the problem we might also defer state transitions
1197 //     to as late as possible -- just prior to parking.
1198 //     Given that, we'd call HSSEC after having returned from park(),
1199 //     but before attempting to acquire the monitor.  This is only a
1200 //     partial solution.  It avoids calling HSSEC while holding the
1201 //     monitor (good), but it still increases successor reacquisition latency --
1202 //     the interval between unparking a successor and the time the successor
1203 //     resumes and retries the lock.  See ReenterI(), which defers state transitions.
1204 //     If we use this technique we can also avoid EnterI()-exit() loop
1205 //     in ::enter() where we iteratively drop the lock and then attempt
1206 //     to reacquire it after suspending.
1207 //
1208 // B.  In the future we might fold all the suspend bits into a
1209 //     composite per-thread suspend flag and then update it with CAS().
1210 //     Alternately, a Dekker-like mechanism with multiple variables
1211 //     would suffice:
1212 //       ST Self->_suspend_equivalent = false
1213 //       MEMBAR
1214 //       LD Self_>_suspend_flags
1215 //
1216 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem
1217 // with a more efficient implementation, the need to use "FastHSSEC" has
1218 // decreased. - Dave
1219 
1220 
1221 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) {
1222   const int Mode = Knob_FastHSSEC;
1223   if (Mode && !jSelf->is_external_suspend()) {
1224     assert(jSelf->is_suspend_equivalent(), "invariant");
1225     jSelf->clear_suspend_equivalent();
1226     if (2 == Mode) OrderAccess::storeload();
1227     if (!jSelf->is_external_suspend()) return false;
1228     // We raced a suspension -- fall thru into the slow path
1229     jSelf->set_suspend_equivalent();
1230   }
1231   return jSelf->handle_special_suspend_equivalent_condition();
1232 }
1233 
1234 
1235 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) {
1236   assert(_owner == Self, "invariant");
1237 
1238   // Exit protocol:
1239   // 1. ST _succ = wakee
1240   // 2. membar #loadstore|#storestore;
1241   // 2. ST _owner = NULL
1242   // 3. unpark(wakee)
1243 
1244   _succ = Wakee->_thread;
1245   ParkEvent * Trigger = Wakee->_event;
1246 
1247   // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again.
1248   // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1249   // out-of-scope (non-extant).
1250   Wakee  = NULL;
1251 
1252   // Drop the lock
1253   OrderAccess::release_store(&_owner, (void*)NULL);
1254   OrderAccess::fence();                               // ST _owner vs LD in unpark()
1255 
1256   DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self);
1257   Trigger->unpark();
1258 
1259   // Maintain stats and report events to JVMTI
1260   OM_PERFDATA_OP(Parks, inc());
1261 }
1262 
1263 
1264 // -----------------------------------------------------------------------------
1265 // Class Loader deadlock handling.
1266 //
1267 // complete_exit exits a lock returning recursion count
1268 // complete_exit/reenter operate as a wait without waiting
1269 // complete_exit requires an inflated monitor
1270 // The _owner field is not always the Thread addr even with an
1271 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1272 // thread due to contention.
1273 intptr_t ObjectMonitor::complete_exit(TRAPS) {
1274   Thread * const Self = THREAD;
1275   assert(Self->is_Java_thread(), "Must be Java thread!");
1276   JavaThread *jt = (JavaThread *)THREAD;
1277 
1278   DeferredInitialize();
1279 
1280   if (THREAD != _owner) {
1281     if (THREAD->is_lock_owned ((address)_owner)) {
1282       assert(_recursions == 0, "internal state error");
1283       _owner = THREAD;   // Convert from basiclock addr to Thread addr
1284       _recursions = 0;
1285     }
1286   }
1287 
1288   guarantee(Self == _owner, "complete_exit not owner");
1289   intptr_t save = _recursions; // record the old recursion count
1290   _recursions = 0;        // set the recursion level to be 0
1291   exit(true, Self);           // exit the monitor
1292   guarantee(_owner != Self, "invariant");
1293   return save;
1294 }
1295 
1296 // reenter() enters a lock and sets recursion count
1297 // complete_exit/reenter operate as a wait without waiting
1298 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) {
1299   Thread * const Self = THREAD;
1300   assert(Self->is_Java_thread(), "Must be Java thread!");
1301   JavaThread *jt = (JavaThread *)THREAD;
1302 
1303   guarantee(_owner != Self, "reenter already owner");
1304   enter(THREAD);       // enter the monitor
1305   guarantee(_recursions == 0, "reenter recursion");
1306   _recursions = recursions;
1307   return;
1308 }
1309 
1310 
1311 // -----------------------------------------------------------------------------
1312 // A macro is used below because there may already be a pending
1313 // exception which should not abort the execution of the routines
1314 // which use this (which is why we don't put this into check_slow and
1315 // call it with a CHECK argument).
1316 
1317 #define CHECK_OWNER()                                                       \
1318   do {                                                                      \
1319     if (THREAD != _owner) {                                                 \
1320       if (THREAD->is_lock_owned((address) _owner)) {                        \
1321         _owner = THREAD;  /* Convert from basiclock addr to Thread addr */  \
1322         _recursions = 0;                                                    \
1323       } else {                                                              \
1324         THROW(vmSymbols::java_lang_IllegalMonitorStateException());         \
1325       }                                                                     \
1326     }                                                                       \
1327   } while (false)
1328 
1329 // check_slow() is a misnomer.  It's called to simply to throw an IMSX exception.
1330 // TODO-FIXME: remove check_slow() -- it's likely dead.
1331 
1332 void ObjectMonitor::check_slow(TRAPS) {
1333   assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner");
1334   THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner");
1335 }
1336 
1337 static int Adjust(volatile int * adr, int dx) {
1338   int v;
1339   for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */;
1340   return v;
1341 }
1342 
1343 static void post_monitor_wait_event(EventJavaMonitorWait* event,
1344                                     ObjectMonitor* monitor,
1345                                     jlong notifier_tid,
1346                                     jlong timeout,
1347                                     bool timedout) {
1348   assert(event != NULL, "invariant");
1349   assert(monitor != NULL, "invariant");
1350   event->set_monitorClass(((oop)monitor->object())->klass());
1351   event->set_timeout(timeout);
1352   event->set_address((uintptr_t)monitor->object_addr());
1353   event->set_notifier(notifier_tid);
1354   event->set_timedOut(timedout);
1355   event->commit();
1356 }
1357 
1358 // -----------------------------------------------------------------------------
1359 // Wait/Notify/NotifyAll
1360 //
1361 // Note: a subset of changes to ObjectMonitor::wait()
1362 // will need to be replicated in complete_exit
1363 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1364   Thread * const Self = THREAD;
1365   assert(Self->is_Java_thread(), "Must be Java thread!");
1366   JavaThread *jt = (JavaThread *)THREAD;
1367 
1368   DeferredInitialize();
1369 
1370   // Throw IMSX or IEX.
1371   CHECK_OWNER();
1372 
1373   EventJavaMonitorWait event;
1374 
1375   // check for a pending interrupt
1376   if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1377     // post monitor waited event.  Note that this is past-tense, we are done waiting.
1378     if (JvmtiExport::should_post_monitor_waited()) {
1379       // Note: 'false' parameter is passed here because the
1380       // wait was not timed out due to thread interrupt.
1381       JvmtiExport::post_monitor_waited(jt, this, false);
1382 
1383       // In this short circuit of the monitor wait protocol, the
1384       // current thread never drops ownership of the monitor and
1385       // never gets added to the wait queue so the current thread
1386       // cannot be made the successor. This means that the
1387       // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1388       // consume an unpark() meant for the ParkEvent associated with
1389       // this ObjectMonitor.
1390     }
1391     if (event.should_commit()) {
1392       post_monitor_wait_event(&event, this, 0, millis, false);
1393     }
1394     THROW(vmSymbols::java_lang_InterruptedException());
1395     return;
1396   }
1397 
1398   assert(Self->_Stalled == 0, "invariant");
1399   Self->_Stalled = intptr_t(this);
1400   jt->set_current_waiting_monitor(this);
1401 
1402   // create a node to be put into the queue
1403   // Critically, after we reset() the event but prior to park(), we must check
1404   // for a pending interrupt.
1405   ObjectWaiter node(Self);
1406   node.TState = ObjectWaiter::TS_WAIT;
1407   Self->_ParkEvent->reset();
1408   OrderAccess::fence();          // ST into Event; membar ; LD interrupted-flag
1409 
1410   // Enter the waiting queue, which is a circular doubly linked list in this case
1411   // but it could be a priority queue or any data structure.
1412   // _WaitSetLock protects the wait queue.  Normally the wait queue is accessed only
1413   // by the the owner of the monitor *except* in the case where park()
1414   // returns because of a timeout of interrupt.  Contention is exceptionally rare
1415   // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1416 
1417   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add");
1418   AddWaiter(&node);
1419   Thread::SpinRelease(&_WaitSetLock);
1420 
1421   _Responsible = NULL;
1422 
1423   intptr_t save = _recursions; // record the old recursion count
1424   _waiters++;                  // increment the number of waiters
1425   _recursions = 0;             // set the recursion level to be 1
1426   exit(true, Self);                    // exit the monitor
1427   guarantee(_owner != Self, "invariant");
1428 
1429   // The thread is on the WaitSet list - now park() it.
1430   // On MP systems it's conceivable that a brief spin before we park
1431   // could be profitable.
1432   //
1433   // TODO-FIXME: change the following logic to a loop of the form
1434   //   while (!timeout && !interrupted && _notified == 0) park()
1435 
1436   int ret = OS_OK;
1437   int WasNotified = 0;
1438   { // State transition wrappers
1439     OSThread* osthread = Self->osthread();
1440     OSThreadWaitState osts(osthread, true);
1441     {
1442       ThreadBlockInVM tbivm(jt);
1443       // Thread is in thread_blocked state and oop access is unsafe.
1444       jt->set_suspend_equivalent();
1445 
1446       if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) {
1447         // Intentionally empty
1448       } else if (node._notified == 0) {
1449         if (millis <= 0) {
1450           Self->_ParkEvent->park();
1451         } else {
1452           ret = Self->_ParkEvent->park(millis);
1453         }
1454       }
1455 
1456       // were we externally suspended while we were waiting?
1457       if (ExitSuspendEquivalent (jt)) {
1458         // TODO-FIXME: add -- if succ == Self then succ = null.
1459         jt->java_suspend_self();
1460       }
1461 
1462     } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm
1463 
1464     // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1465     // from the WaitSet to the EntryList.
1466     // See if we need to remove Node from the WaitSet.
1467     // We use double-checked locking to avoid grabbing _WaitSetLock
1468     // if the thread is not on the wait queue.
1469     //
1470     // Note that we don't need a fence before the fetch of TState.
1471     // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1472     // written by the is thread. (perhaps the fetch might even be satisfied
1473     // by a look-aside into the processor's own store buffer, although given
1474     // the length of the code path between the prior ST and this load that's
1475     // highly unlikely).  If the following LD fetches a stale TS_WAIT value
1476     // then we'll acquire the lock and then re-fetch a fresh TState value.
1477     // That is, we fail toward safety.
1478 
1479     if (node.TState == ObjectWaiter::TS_WAIT) {
1480       Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink");
1481       if (node.TState == ObjectWaiter::TS_WAIT) {
1482         DequeueSpecificWaiter(&node);       // unlink from WaitSet
1483         assert(node._notified == 0, "invariant");
1484         node.TState = ObjectWaiter::TS_RUN;
1485       }
1486       Thread::SpinRelease(&_WaitSetLock);
1487     }
1488 
1489     // The thread is now either on off-list (TS_RUN),
1490     // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1491     // The Node's TState variable is stable from the perspective of this thread.
1492     // No other threads will asynchronously modify TState.
1493     guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant");
1494     OrderAccess::loadload();
1495     if (_succ == Self) _succ = NULL;
1496     WasNotified = node._notified;
1497 
1498     // Reentry phase -- reacquire the monitor.
1499     // re-enter contended monitor after object.wait().
1500     // retain OBJECT_WAIT state until re-enter successfully completes
1501     // Thread state is thread_in_vm and oop access is again safe,
1502     // although the raw address of the object may have changed.
1503     // (Don't cache naked oops over safepoints, of course).
1504 
1505     // post monitor waited event. Note that this is past-tense, we are done waiting.
1506     if (JvmtiExport::should_post_monitor_waited()) {
1507       JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT);
1508 
1509       if (node._notified != 0 && _succ == Self) {
1510         // In this part of the monitor wait-notify-reenter protocol it
1511         // is possible (and normal) for another thread to do a fastpath
1512         // monitor enter-exit while this thread is still trying to get
1513         // to the reenter portion of the protocol.
1514         //
1515         // The ObjectMonitor was notified and the current thread is
1516         // the successor which also means that an unpark() has already
1517         // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1518         // consume the unpark() that was done when the successor was
1519         // set because the same ParkEvent is shared between Java
1520         // monitors and JVM/TI RawMonitors (for now).
1521         //
1522         // We redo the unpark() to ensure forward progress, i.e., we
1523         // don't want all pending threads hanging (parked) with none
1524         // entering the unlocked monitor.
1525         node._event->unpark();
1526       }
1527     }
1528 
1529     if (event.should_commit()) {
1530       post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT);
1531     }
1532 
1533     OrderAccess::fence();
1534 
1535     assert(Self->_Stalled != 0, "invariant");
1536     Self->_Stalled = 0;
1537 
1538     assert(_owner != Self, "invariant");
1539     ObjectWaiter::TStates v = node.TState;
1540     if (v == ObjectWaiter::TS_RUN) {
1541       enter(Self);
1542     } else {
1543       guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1544       ReenterI(Self, &node);
1545       node.wait_reenter_end(this);
1546     }
1547 
1548     // Self has reacquired the lock.
1549     // Lifecycle - the node representing Self must not appear on any queues.
1550     // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1551     // want residual elements associated with this thread left on any lists.
1552     guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant");
1553     assert(_owner == Self, "invariant");
1554     assert(_succ != Self, "invariant");
1555   } // OSThreadWaitState()
1556 
1557   jt->set_current_waiting_monitor(NULL);
1558 
1559   guarantee(_recursions == 0, "invariant");
1560   _recursions = save;     // restore the old recursion count
1561   _waiters--;             // decrement the number of waiters
1562 
1563   // Verify a few postconditions
1564   assert(_owner == Self, "invariant");
1565   assert(_succ != Self, "invariant");
1566   assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
1567 
1568   // check if the notification happened
1569   if (!WasNotified) {
1570     // no, it could be timeout or Thread.interrupt() or both
1571     // check for interrupt event, otherwise it is timeout
1572     if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1573       THROW(vmSymbols::java_lang_InterruptedException());
1574     }
1575   }
1576 
1577   // NOTE: Spurious wake up will be consider as timeout.
1578   // Monitor notify has precedence over thread interrupt.
1579 }
1580 
1581 
1582 // Consider:
1583 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1584 // then instead of transferring a thread from the WaitSet to the EntryList
1585 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1586 
1587 void ObjectMonitor::INotify(Thread * Self) {
1588   const int policy = Knob_MoveNotifyee;
1589 
1590   Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify");
1591   ObjectWaiter * iterator = DequeueWaiter();
1592   if (iterator != NULL) {
1593     guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant");
1594     guarantee(iterator->_notified == 0, "invariant");
1595     // Disposition - what might we do with iterator ?
1596     // a.  add it directly to the EntryList - either tail (policy == 1)
1597     //     or head (policy == 0).
1598     // b.  push it onto the front of the _cxq (policy == 2).
1599     // For now we use (b).
1600     if (policy != 4) {
1601       iterator->TState = ObjectWaiter::TS_ENTER;
1602     }
1603     iterator->_notified = 1;
1604     iterator->_notifier_tid = JFR_THREAD_ID(Self);
1605 
1606     ObjectWaiter * list = _EntryList;
1607     if (list != NULL) {
1608       assert(list->_prev == NULL, "invariant");
1609       assert(list->TState == ObjectWaiter::TS_ENTER, "invariant");
1610       assert(list != iterator, "invariant");
1611     }
1612 
1613     if (policy == 0) {       // prepend to EntryList
1614       if (list == NULL) {
1615         iterator->_next = iterator->_prev = NULL;
1616         _EntryList = iterator;
1617       } else {
1618         list->_prev = iterator;
1619         iterator->_next = list;
1620         iterator->_prev = NULL;
1621         _EntryList = iterator;
1622       }
1623     } else if (policy == 1) {      // append to EntryList
1624       if (list == NULL) {
1625         iterator->_next = iterator->_prev = NULL;
1626         _EntryList = iterator;
1627       } else {
1628         // CONSIDER:  finding the tail currently requires a linear-time walk of
1629         // the EntryList.  We can make tail access constant-time by converting to
1630         // a CDLL instead of using our current DLL.
1631         ObjectWaiter * tail;
1632         for (tail = list; tail->_next != NULL; tail = tail->_next) {}
1633         assert(tail != NULL && tail->_next == NULL, "invariant");
1634         tail->_next = iterator;
1635         iterator->_prev = tail;
1636         iterator->_next = NULL;
1637       }
1638     } else if (policy == 2) {      // prepend to cxq
1639       if (list == NULL) {
1640         iterator->_next = iterator->_prev = NULL;
1641         _EntryList = iterator;
1642       } else {
1643         iterator->TState = ObjectWaiter::TS_CXQ;
1644         for (;;) {
1645           ObjectWaiter * front = _cxq;
1646           iterator->_next = front;
1647           if (Atomic::cmpxchg(iterator, &_cxq, front) == front) {
1648             break;
1649           }
1650         }
1651       }
1652     } else if (policy == 3) {      // append to cxq
1653       iterator->TState = ObjectWaiter::TS_CXQ;
1654       for (;;) {
1655         ObjectWaiter * tail = _cxq;
1656         if (tail == NULL) {
1657           iterator->_next = NULL;
1658           if (Atomic::replace_if_null(iterator, &_cxq)) {
1659             break;
1660           }
1661         } else {
1662           while (tail->_next != NULL) tail = tail->_next;
1663           tail->_next = iterator;
1664           iterator->_prev = tail;
1665           iterator->_next = NULL;
1666           break;
1667         }
1668       }
1669     } else {
1670       ParkEvent * ev = iterator->_event;
1671       iterator->TState = ObjectWaiter::TS_RUN;
1672       OrderAccess::fence();
1673       ev->unpark();
1674     }
1675 
1676     // _WaitSetLock protects the wait queue, not the EntryList.  We could
1677     // move the add-to-EntryList operation, above, outside the critical section
1678     // protected by _WaitSetLock.  In practice that's not useful.  With the
1679     // exception of  wait() timeouts and interrupts the monitor owner
1680     // is the only thread that grabs _WaitSetLock.  There's almost no contention
1681     // on _WaitSetLock so it's not profitable to reduce the length of the
1682     // critical section.
1683 
1684     if (policy < 4) {
1685       iterator->wait_reenter_begin(this);
1686     }
1687   }
1688   Thread::SpinRelease(&_WaitSetLock);
1689 }
1690 
1691 // Consider: a not-uncommon synchronization bug is to use notify() when
1692 // notifyAll() is more appropriate, potentially resulting in stranded
1693 // threads; this is one example of a lost wakeup. A useful diagnostic
1694 // option is to force all notify() operations to behave as notifyAll().
1695 //
1696 // Note: We can also detect many such problems with a "minimum wait".
1697 // When the "minimum wait" is set to a small non-zero timeout value
1698 // and the program does not hang whereas it did absent "minimum wait",
1699 // that suggests a lost wakeup bug.
1700 
1701 void ObjectMonitor::notify(TRAPS) {
1702   CHECK_OWNER();
1703   if (_WaitSet == NULL) {
1704     return;
1705   }
1706   DTRACE_MONITOR_PROBE(notify, this, object(), THREAD);
1707   INotify(THREAD);
1708   OM_PERFDATA_OP(Notifications, inc(1));
1709 }
1710 
1711 
1712 // The current implementation of notifyAll() transfers the waiters one-at-a-time
1713 // from the waitset to the EntryList. This could be done more efficiently with a
1714 // single bulk transfer but in practice it's not time-critical. Beware too,
1715 // that in prepend-mode we invert the order of the waiters. Let's say that the
1716 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend
1717 // mode the waitset will be empty and the EntryList will be "DCBAXYZ".
1718 
1719 void ObjectMonitor::notifyAll(TRAPS) {
1720   CHECK_OWNER();
1721   if (_WaitSet == NULL) {
1722     return;
1723   }
1724 
1725   DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD);
1726   int tally = 0;
1727   while (_WaitSet != NULL) {
1728     tally++;
1729     INotify(THREAD);
1730   }
1731 
1732   OM_PERFDATA_OP(Notifications, inc(tally));
1733 }
1734 
1735 // -----------------------------------------------------------------------------
1736 // Adaptive Spinning Support
1737 //
1738 // Adaptive spin-then-block - rational spinning
1739 //
1740 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1741 // algorithm.  On high order SMP systems it would be better to start with
1742 // a brief global spin and then revert to spinning locally.  In the spirit of MCS/CLH,
1743 // a contending thread could enqueue itself on the cxq and then spin locally
1744 // on a thread-specific variable such as its ParkEvent._Event flag.
1745 // That's left as an exercise for the reader.  Note that global spinning is
1746 // not problematic on Niagara, as the L2 cache serves the interconnect and
1747 // has both low latency and massive bandwidth.
1748 //
1749 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1750 // acquisition attempts where we opt to spin --  at 100% and vary the spin count
1751 // (duration) or we can fix the count at approximately the duration of
1752 // a context switch and vary the frequency.   Of course we could also
1753 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1754 // For a description of 'Adaptive spin-then-block mutual exclusion in
1755 // multi-threaded processing,' see U.S. Pat. No. 8046758.
1756 //
1757 // This implementation varies the duration "D", where D varies with
1758 // the success rate of recent spin attempts. (D is capped at approximately
1759 // length of a round-trip context switch).  The success rate for recent
1760 // spin attempts is a good predictor of the success rate of future spin
1761 // attempts.  The mechanism adapts automatically to varying critical
1762 // section length (lock modality), system load and degree of parallelism.
1763 // D is maintained per-monitor in _SpinDuration and is initialized
1764 // optimistically.  Spin frequency is fixed at 100%.
1765 //
1766 // Note that _SpinDuration is volatile, but we update it without locks
1767 // or atomics.  The code is designed so that _SpinDuration stays within
1768 // a reasonable range even in the presence of races.  The arithmetic
1769 // operations on _SpinDuration are closed over the domain of legal values,
1770 // so at worst a race will install and older but still legal value.
1771 // At the very worst this introduces some apparent non-determinism.
1772 // We might spin when we shouldn't or vice-versa, but since the spin
1773 // count are relatively short, even in the worst case, the effect is harmless.
1774 //
1775 // Care must be taken that a low "D" value does not become an
1776 // an absorbing state.  Transient spinning failures -- when spinning
1777 // is overall profitable -- should not cause the system to converge
1778 // on low "D" values.  We want spinning to be stable and predictable
1779 // and fairly responsive to change and at the same time we don't want
1780 // it to oscillate, become metastable, be "too" non-deterministic,
1781 // or converge on or enter undesirable stable absorbing states.
1782 //
1783 // We implement a feedback-based control system -- using past behavior
1784 // to predict future behavior.  We face two issues: (a) if the
1785 // input signal is random then the spin predictor won't provide optimal
1786 // results, and (b) if the signal frequency is too high then the control
1787 // system, which has some natural response lag, will "chase" the signal.
1788 // (b) can arise from multimodal lock hold times.  Transient preemption
1789 // can also result in apparent bimodal lock hold times.
1790 // Although sub-optimal, neither condition is particularly harmful, as
1791 // in the worst-case we'll spin when we shouldn't or vice-versa.
1792 // The maximum spin duration is rather short so the failure modes aren't bad.
1793 // To be conservative, I've tuned the gain in system to bias toward
1794 // _not spinning.  Relatedly, the system can sometimes enter a mode where it
1795 // "rings" or oscillates between spinning and not spinning.  This happens
1796 // when spinning is just on the cusp of profitability, however, so the
1797 // situation is not dire.  The state is benign -- there's no need to add
1798 // hysteresis control to damp the transition rate between spinning and
1799 // not spinning.
1800 
1801 // Spinning: Fixed frequency (100%), vary duration
1802 int ObjectMonitor::TrySpin(Thread * Self) {
1803   // Dumb, brutal spin.  Good for comparative measurements against adaptive spinning.
1804   int ctr = Knob_FixedSpin;
1805   if (ctr != 0) {
1806     while (--ctr >= 0) {
1807       if (TryLock(Self) > 0) return 1;
1808       SpinPause();
1809     }
1810     return 0;
1811   }
1812 
1813   for (ctr = Knob_PreSpin + 1; --ctr >= 0;) {
1814     if (TryLock(Self) > 0) {
1815       // Increase _SpinDuration ...
1816       // Note that we don't clamp SpinDuration precisely at SpinLimit.
1817       // Raising _SpurDuration to the poverty line is key.
1818       int x = _SpinDuration;
1819       if (x < Knob_SpinLimit) {
1820         if (x < Knob_Poverty) x = Knob_Poverty;
1821         _SpinDuration = x + Knob_BonusB;
1822       }
1823       return 1;
1824     }
1825     SpinPause();
1826   }
1827 
1828   // Admission control - verify preconditions for spinning
1829   //
1830   // We always spin a little bit, just to prevent _SpinDuration == 0 from
1831   // becoming an absorbing state.  Put another way, we spin briefly to
1832   // sample, just in case the system load, parallelism, contention, or lock
1833   // modality changed.
1834   //
1835   // Consider the following alternative:
1836   // Periodically set _SpinDuration = _SpinLimit and try a long/full
1837   // spin attempt.  "Periodically" might mean after a tally of
1838   // the # of failed spin attempts (or iterations) reaches some threshold.
1839   // This takes us into the realm of 1-out-of-N spinning, where we
1840   // hold the duration constant but vary the frequency.
1841 
1842   ctr = _SpinDuration;
1843   if (ctr <= 0) return 0;
1844 
1845   if (Knob_SuccRestrict && _succ != NULL) return 0;
1846   if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) {
1847     return 0;
1848   }
1849 
1850   int MaxSpin = Knob_MaxSpinners;
1851   if (MaxSpin >= 0) {
1852     if (_Spinner > MaxSpin) {
1853       return 0;
1854     }
1855     // Slightly racy, but benign ...
1856     Adjust(&_Spinner, 1);
1857   }
1858 
1859   // We're good to spin ... spin ingress.
1860   // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
1861   // when preparing to LD...CAS _owner, etc and the CAS is likely
1862   // to succeed.
1863   if (_succ == NULL) {
1864     _succ = Self;
1865   }
1866   Thread * prv = NULL;
1867 
1868   // There are three ways to exit the following loop:
1869   // 1.  A successful spin where this thread has acquired the lock.
1870   // 2.  Spin failure with prejudice
1871   // 3.  Spin failure without prejudice
1872 
1873   while (--ctr >= 0) {
1874 
1875     // Periodic polling -- Check for pending GC
1876     // Threads may spin while they're unsafe.
1877     // We don't want spinning threads to delay the JVM from reaching
1878     // a stop-the-world safepoint or to steal cycles from GC.
1879     // If we detect a pending safepoint we abort in order that
1880     // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
1881     // this thread, if safe, doesn't steal cycles from GC.
1882     // This is in keeping with the "no loitering in runtime" rule.
1883     // We periodically check to see if there's a safepoint pending.
1884     if ((ctr & 0xFF) == 0) {
1885       if (SafepointMechanism::poll(Self)) {
1886         goto Abort;           // abrupt spin egress
1887       }
1888       if (Knob_UsePause & 1) SpinPause();
1889     }
1890 
1891     if (Knob_UsePause & 2) SpinPause();
1892 
1893     // Probe _owner with TATAS
1894     // If this thread observes the monitor transition or flicker
1895     // from locked to unlocked to locked, then the odds that this
1896     // thread will acquire the lock in this spin attempt go down
1897     // considerably.  The same argument applies if the CAS fails
1898     // or if we observe _owner change from one non-null value to
1899     // another non-null value.   In such cases we might abort
1900     // the spin without prejudice or apply a "penalty" to the
1901     // spin count-down variable "ctr", reducing it by 100, say.
1902 
1903     Thread * ox = (Thread *) _owner;
1904     if (ox == NULL) {
1905       ox = (Thread*)Atomic::cmpxchg(Self, &_owner, (void*)NULL);
1906       if (ox == NULL) {
1907         // The CAS succeeded -- this thread acquired ownership
1908         // Take care of some bookkeeping to exit spin state.
1909         if (_succ == Self) {
1910           _succ = NULL;
1911         }
1912         if (MaxSpin > 0) Adjust(&_Spinner, -1);
1913 
1914         // Increase _SpinDuration :
1915         // The spin was successful (profitable) so we tend toward
1916         // longer spin attempts in the future.
1917         // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
1918         // If we acquired the lock early in the spin cycle it
1919         // makes sense to increase _SpinDuration proportionally.
1920         // Note that we don't clamp SpinDuration precisely at SpinLimit.
1921         int x = _SpinDuration;
1922         if (x < Knob_SpinLimit) {
1923           if (x < Knob_Poverty) x = Knob_Poverty;
1924           _SpinDuration = x + Knob_Bonus;
1925         }
1926         return 1;
1927       }
1928 
1929       // The CAS failed ... we can take any of the following actions:
1930       // * penalize: ctr -= CASPenalty
1931       // * exit spin with prejudice -- goto Abort;
1932       // * exit spin without prejudice.
1933       // * Since CAS is high-latency, retry again immediately.
1934       prv = ox;
1935       goto Abort;
1936     }
1937 
1938     // Did lock ownership change hands ?
1939     if (ox != prv && prv != NULL) {
1940       goto Abort;
1941     }
1942     prv = ox;
1943 
1944     // Abort the spin if the owner is not executing.
1945     // The owner must be executing in order to drop the lock.
1946     // Spinning while the owner is OFFPROC is idiocy.
1947     // Consider: ctr -= RunnablePenalty ;
1948     if (Knob_OState && NotRunnable (Self, ox)) {
1949       goto Abort;
1950     }
1951     if (_succ == NULL) {
1952       _succ = Self;
1953     }
1954   }
1955 
1956   // Spin failed with prejudice -- reduce _SpinDuration.
1957   // TODO: Use an AIMD-like policy to adjust _SpinDuration.
1958   // AIMD is globally stable.
1959   {
1960     int x = _SpinDuration;
1961     if (x > 0) {
1962       // Consider an AIMD scheme like: x -= (x >> 3) + 100
1963       // This is globally sample and tends to damp the response.
1964       x -= Knob_Penalty;
1965       if (x < 0) x = 0;
1966       _SpinDuration = x;
1967     }
1968   }
1969 
1970  Abort:
1971   if (MaxSpin >= 0) Adjust(&_Spinner, -1);
1972   if (_succ == Self) {
1973     _succ = NULL;
1974     // Invariant: after setting succ=null a contending thread
1975     // must recheck-retry _owner before parking.  This usually happens
1976     // in the normal usage of TrySpin(), but it's safest
1977     // to make TrySpin() as foolproof as possible.
1978     OrderAccess::fence();
1979     if (TryLock(Self) > 0) return 1;
1980   }
1981   return 0;
1982 }
1983 
1984 // NotRunnable() -- informed spinning
1985 //
1986 // Don't bother spinning if the owner is not eligible to drop the lock.
1987 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm.
1988 // The thread must be runnable in order to drop the lock in timely fashion.
1989 // If the _owner is not runnable then spinning will not likely be
1990 // successful (profitable).
1991 //
1992 // Beware -- the thread referenced by _owner could have died
1993 // so a simply fetch from _owner->_thread_state might trap.
1994 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state.
1995 // Because of the lifecycle issues, the _thread_state values
1996 // observed by NotRunnable() might be garbage.  NotRunnable must
1997 // tolerate this and consider the observed _thread_state value
1998 // as advisory.
1999 //
2000 // Beware too, that _owner is sometimes a BasicLock address and sometimes
2001 // a thread pointer.
2002 // Alternately, we might tag the type (thread pointer vs basiclock pointer)
2003 // with the LSB of _owner.  Another option would be to probabilistically probe
2004 // the putative _owner->TypeTag value.
2005 //
2006 // Checking _thread_state isn't perfect.  Even if the thread is
2007 // in_java it might be blocked on a page-fault or have been preempted
2008 // and sitting on a ready/dispatch queue.
2009 //
2010 // The return value from NotRunnable() is *advisory* -- the
2011 // result is based on sampling and is not necessarily coherent.
2012 // The caller must tolerate false-negative and false-positive errors.
2013 // Spinning, in general, is probabilistic anyway.
2014 
2015 
2016 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) {
2017   // Check ox->TypeTag == 2BAD.
2018   if (ox == NULL) return 0;
2019 
2020   // Avoid transitive spinning ...
2021   // Say T1 spins or blocks trying to acquire L.  T1._Stalled is set to L.
2022   // Immediately after T1 acquires L it's possible that T2, also
2023   // spinning on L, will see L.Owner=T1 and T1._Stalled=L.
2024   // This occurs transiently after T1 acquired L but before
2025   // T1 managed to clear T1.Stalled.  T2 does not need to abort
2026   // its spin in this circumstance.
2027   intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1));
2028 
2029   if (BlockedOn == 1) return 1;
2030   if (BlockedOn != 0) {
2031     return BlockedOn != intptr_t(this) && _owner == ox;
2032   }
2033 
2034   assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant");
2035   int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);;
2036   // consider also: jst != _thread_in_Java -- but that's overspecific.
2037   return jst == _thread_blocked || jst == _thread_in_native;
2038 }
2039 
2040 
2041 // -----------------------------------------------------------------------------
2042 // WaitSet management ...
2043 
2044 ObjectWaiter::ObjectWaiter(Thread* thread) {
2045   _next     = NULL;
2046   _prev     = NULL;
2047   _notified = 0;
2048   _notifier_tid = 0;
2049   TState    = TS_RUN;
2050   _thread   = thread;
2051   _event    = thread->_ParkEvent;
2052   _active   = false;
2053   assert(_event != NULL, "invariant");
2054 }
2055 
2056 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) {
2057   JavaThread *jt = (JavaThread *)this->_thread;
2058   _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon);
2059 }
2060 
2061 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) {
2062   JavaThread *jt = (JavaThread *)this->_thread;
2063   JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active);
2064 }
2065 
2066 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2067   assert(node != NULL, "should not add NULL node");
2068   assert(node->_prev == NULL, "node already in list");
2069   assert(node->_next == NULL, "node already in list");
2070   // put node at end of queue (circular doubly linked list)
2071   if (_WaitSet == NULL) {
2072     _WaitSet = node;
2073     node->_prev = node;
2074     node->_next = node;
2075   } else {
2076     ObjectWaiter* head = _WaitSet;
2077     ObjectWaiter* tail = head->_prev;
2078     assert(tail->_next == head, "invariant check");
2079     tail->_next = node;
2080     head->_prev = node;
2081     node->_next = head;
2082     node->_prev = tail;
2083   }
2084 }
2085 
2086 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2087   // dequeue the very first waiter
2088   ObjectWaiter* waiter = _WaitSet;
2089   if (waiter) {
2090     DequeueSpecificWaiter(waiter);
2091   }
2092   return waiter;
2093 }
2094 
2095 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2096   assert(node != NULL, "should not dequeue NULL node");
2097   assert(node->_prev != NULL, "node already removed from list");
2098   assert(node->_next != NULL, "node already removed from list");
2099   // when the waiter has woken up because of interrupt,
2100   // timeout or other spurious wake-up, dequeue the
2101   // waiter from waiting list
2102   ObjectWaiter* next = node->_next;
2103   if (next == node) {
2104     assert(node->_prev == node, "invariant check");
2105     _WaitSet = NULL;
2106   } else {
2107     ObjectWaiter* prev = node->_prev;
2108     assert(prev->_next == node, "invariant check");
2109     assert(next->_prev == node, "invariant check");
2110     next->_prev = prev;
2111     prev->_next = next;
2112     if (_WaitSet == node) {
2113       _WaitSet = next;
2114     }
2115   }
2116   node->_next = NULL;
2117   node->_prev = NULL;
2118 }
2119 
2120 // -----------------------------------------------------------------------------
2121 // PerfData support
2122 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts       = NULL;
2123 PerfCounter * ObjectMonitor::_sync_FutileWakeups               = NULL;
2124 PerfCounter * ObjectMonitor::_sync_Parks                       = NULL;
2125 PerfCounter * ObjectMonitor::_sync_Notifications               = NULL;
2126 PerfCounter * ObjectMonitor::_sync_Inflations                  = NULL;
2127 PerfCounter * ObjectMonitor::_sync_Deflations                  = NULL;
2128 PerfLongVariable * ObjectMonitor::_sync_MonExtant              = NULL;
2129 
2130 // One-shot global initialization for the sync subsystem.
2131 // We could also defer initialization and initialize on-demand
2132 // the first time we call inflate().  Initialization would
2133 // be protected - like so many things - by the MonitorCache_lock.
2134 
2135 void ObjectMonitor::Initialize() {
2136   static int InitializationCompleted = 0;
2137   assert(InitializationCompleted == 0, "invariant");
2138   InitializationCompleted = 1;
2139   if (UsePerfData) {
2140     EXCEPTION_MARK;
2141 #define NEWPERFCOUNTER(n)                                                \
2142   {                                                                      \
2143     n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,  \
2144                                         CHECK);                          \
2145   }
2146 #define NEWPERFVARIABLE(n)                                                \
2147   {                                                                       \
2148     n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,  \
2149                                          CHECK);                          \
2150   }
2151     NEWPERFCOUNTER(_sync_Inflations);
2152     NEWPERFCOUNTER(_sync_Deflations);
2153     NEWPERFCOUNTER(_sync_ContendedLockAttempts);
2154     NEWPERFCOUNTER(_sync_FutileWakeups);
2155     NEWPERFCOUNTER(_sync_Parks);
2156     NEWPERFCOUNTER(_sync_Notifications);
2157     NEWPERFVARIABLE(_sync_MonExtant);
2158 #undef NEWPERFCOUNTER
2159 #undef NEWPERFVARIABLE
2160   }
2161 }
2162 
2163 void ObjectMonitor::DeferredInitialize() {
2164   if (InitDone > 0) return;
2165   if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) {
2166     while (InitDone != 1) /* empty */;
2167     return;
2168   }
2169 
2170   // One-shot global initialization ...
2171   // The initialization is idempotent, so we don't need locks.
2172   // In the future consider doing this via os::init_2().
2173 
2174   if (!os::is_MP()) {
2175     Knob_SpinLimit = 0;
2176     Knob_PreSpin   = 0;
2177     Knob_FixedSpin = -1;
2178   }
2179 
2180   OrderAccess::fence();
2181   InitDone = 1;
2182 }
2183